code_generator_x86.cc revision b2bd1c5f9171f35fa5b71ada42d1a9e11189428d 
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_x86.h"
18
19#include "entrypoints/quick/quick_entrypoints.h"
20#include "entrypoints/quick/quick_entrypoints_enum.h"
21#include "gc/accounting/card_table.h"
22#include "mirror/array-inl.h"
23#include "mirror/art_method.h"
24#include "mirror/class.h"
25#include "thread.h"
26#include "utils/assembler.h"
27#include "utils/stack_checks.h"
28#include "utils/x86/assembler_x86.h"
29#include "utils/x86/managed_register_x86.h"
30
31namespace art {
32
33namespace x86 {
34
35static constexpr int kCurrentMethodStackOffset = 0;
36
37static constexpr Register kRuntimeParameterCoreRegisters[] = { EAX, ECX, EDX, EBX };
38static constexpr size_t kRuntimeParameterCoreRegistersLength =
39    arraysize(kRuntimeParameterCoreRegisters);
40static constexpr Register kCoreCalleeSaves[] = { EBP, ESI, EDI };
41static constexpr XmmRegister kRuntimeParameterFpuRegisters[] = { XMM0, XMM1, XMM2, XMM3 };
42static constexpr size_t kRuntimeParameterFpuRegistersLength =
43    arraysize(kRuntimeParameterFpuRegisters);
44
45static constexpr int kC2ConditionMask = 0x400;
46
47static constexpr int kFakeReturnRegister = Register(8);
48
49class InvokeRuntimeCallingConvention : public CallingConvention<Register, XmmRegister> {
50 public:
51  InvokeRuntimeCallingConvention()
52      : CallingConvention(kRuntimeParameterCoreRegisters,
53                          kRuntimeParameterCoreRegistersLength,
54                          kRuntimeParameterFpuRegisters,
55                          kRuntimeParameterFpuRegistersLength) {}
56
57 private:
58  DISALLOW_COPY_AND_ASSIGN(InvokeRuntimeCallingConvention);
59};
60
61#define __ reinterpret_cast<X86Assembler*>(codegen->GetAssembler())->
62
63class SlowPathCodeX86 : public SlowPathCode {
64 public:
65  SlowPathCodeX86() : entry_label_(), exit_label_() {}
66
67  Label* GetEntryLabel() { return &entry_label_; }
68  Label* GetExitLabel() { return &exit_label_; }
69
70 private:
71  Label entry_label_;
72  Label exit_label_;
73
74  DISALLOW_COPY_AND_ASSIGN(SlowPathCodeX86);
75};
76
77class NullCheckSlowPathX86 : public SlowPathCodeX86 {
78 public:
79  explicit NullCheckSlowPathX86(HNullCheck* instruction) : instruction_(instruction) {}
80
81  void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
82    __ Bind(GetEntryLabel());
83    __ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pThrowNullPointer)));
84    codegen->RecordPcInfo(instruction_, instruction_->GetDexPc());
85  }
86
87 private:
88  HNullCheck* const instruction_;
89  DISALLOW_COPY_AND_ASSIGN(NullCheckSlowPathX86);
90};
91
92class DivZeroCheckSlowPathX86 : public SlowPathCodeX86 {
93 public:
94  explicit DivZeroCheckSlowPathX86(HDivZeroCheck* instruction) : instruction_(instruction) {}
95
96  void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
97    __ Bind(GetEntryLabel());
98    __ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pThrowDivZero)));
99    codegen->RecordPcInfo(instruction_, instruction_->GetDexPc());
100  }
101
102 private:
103  HDivZeroCheck* const instruction_;
104  DISALLOW_COPY_AND_ASSIGN(DivZeroCheckSlowPathX86);
105};
106
107class DivRemMinusOneSlowPathX86 : public SlowPathCodeX86 {
108 public:
109  explicit DivRemMinusOneSlowPathX86(Register reg, bool is_div) : reg_(reg), is_div_(is_div) {}
110
111  void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
112    __ Bind(GetEntryLabel());
113    if (is_div_) {
114      __ negl(reg_);
115    } else {
116      __ movl(reg_, Immediate(0));
117    }
118    __ jmp(GetExitLabel());
119  }
120
121 private:
122  Register reg_;
123  bool is_div_;
124  DISALLOW_COPY_AND_ASSIGN(DivRemMinusOneSlowPathX86);
125};
126
127class BoundsCheckSlowPathX86 : public SlowPathCodeX86 {
128 public:
129  BoundsCheckSlowPathX86(HBoundsCheck* instruction,
130                         Location index_location,
131                         Location length_location)
132      : instruction_(instruction),
133        index_location_(index_location),
134        length_location_(length_location) {}
135
136  void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
137    CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen);
138    __ Bind(GetEntryLabel());
139    // We're moving two locations to locations that could overlap, so we need a parallel
140    // move resolver.
141    InvokeRuntimeCallingConvention calling_convention;
142    x86_codegen->EmitParallelMoves(
143        index_location_,
144        Location::RegisterLocation(calling_convention.GetRegisterAt(0)),
145        length_location_,
146        Location::RegisterLocation(calling_convention.GetRegisterAt(1)));
147    __ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pThrowArrayBounds)));
148    codegen->RecordPcInfo(instruction_, instruction_->GetDexPc());
149  }
150
151 private:
152  HBoundsCheck* const instruction_;
153  const Location index_location_;
154  const Location length_location_;
155
156  DISALLOW_COPY_AND_ASSIGN(BoundsCheckSlowPathX86);
157};
158
159class SuspendCheckSlowPathX86 : public SlowPathCodeX86 {
160 public:
161  SuspendCheckSlowPathX86(HSuspendCheck* instruction, HBasicBlock* successor)
162      : instruction_(instruction), successor_(successor) {}
163
164  void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
165    CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen);
166    __ Bind(GetEntryLabel());
167    SaveLiveRegisters(codegen, instruction_->GetLocations());
168    __ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pTestSuspend)));
169    codegen->RecordPcInfo(instruction_, instruction_->GetDexPc());
170    RestoreLiveRegisters(codegen, instruction_->GetLocations());
171    if (successor_ == nullptr) {
172      __ jmp(GetReturnLabel());
173    } else {
174      __ jmp(x86_codegen->GetLabelOf(successor_));
175    }
176  }
177
178  Label* GetReturnLabel() {
179    DCHECK(successor_ == nullptr);
180    return &return_label_;
181  }
182
183 private:
184  HSuspendCheck* const instruction_;
185  HBasicBlock* const successor_;
186  Label return_label_;
187
188  DISALLOW_COPY_AND_ASSIGN(SuspendCheckSlowPathX86);
189};
190
191class LoadStringSlowPathX86 : public SlowPathCodeX86 {
192 public:
193  explicit LoadStringSlowPathX86(HLoadString* instruction) : instruction_(instruction) {}
194
195  void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
196    LocationSummary* locations = instruction_->GetLocations();
197    DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg()));
198
199    CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen);
200    __ Bind(GetEntryLabel());
201    SaveLiveRegisters(codegen, locations);
202
203    InvokeRuntimeCallingConvention calling_convention;
204    x86_codegen->LoadCurrentMethod(calling_convention.GetRegisterAt(1));
205    __ movl(calling_convention.GetRegisterAt(0), Immediate(instruction_->GetStringIndex()));
206    __ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pResolveString)));
207    RecordPcInfo(codegen, instruction_, instruction_->GetDexPc());
208    x86_codegen->Move32(locations->Out(), Location::RegisterLocation(EAX));
209    RestoreLiveRegisters(codegen, locations);
210
211    __ jmp(GetExitLabel());
212  }
213
214 private:
215  HLoadString* const instruction_;
216
217  DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathX86);
218};
219
220class LoadClassSlowPathX86 : public SlowPathCodeX86 {
221 public:
222  LoadClassSlowPathX86(HLoadClass* cls,
223                       HInstruction* at,
224                       uint32_t dex_pc,
225                       bool do_clinit)
226      : cls_(cls), at_(at), dex_pc_(dex_pc), do_clinit_(do_clinit) {
227    DCHECK(at->IsLoadClass() || at->IsClinitCheck());
228  }
229
230  void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
231    LocationSummary* locations = at_->GetLocations();
232    CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen);
233    __ Bind(GetEntryLabel());
234    SaveLiveRegisters(codegen, locations);
235
236    InvokeRuntimeCallingConvention calling_convention;
237    __ movl(calling_convention.GetRegisterAt(0), Immediate(cls_->GetTypeIndex()));
238    x86_codegen->LoadCurrentMethod(calling_convention.GetRegisterAt(1));
239    __ fs()->call(Address::Absolute(do_clinit_
240        ? QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pInitializeStaticStorage)
241        : QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pInitializeType)));
242    RecordPcInfo(codegen, at_, dex_pc_);
243
244    // Move the class to the desired location.
245    Location out = locations->Out();
246    if (out.IsValid()) {
247      DCHECK(out.IsRegister() && !locations->GetLiveRegisters()->ContainsCoreRegister(out.reg()));
248      x86_codegen->Move32(out, Location::RegisterLocation(EAX));
249    }
250
251    RestoreLiveRegisters(codegen, locations);
252    __ jmp(GetExitLabel());
253  }
254
255 private:
256  // The class this slow path will load.
257  HLoadClass* const cls_;
258
259  // The instruction where this slow path is happening.
260  // (Might be the load class or an initialization check).
261  HInstruction* const at_;
262
263  // The dex PC of `at_`.
264  const uint32_t dex_pc_;
265
266  // Whether to initialize the class.
267  const bool do_clinit_;
268
269  DISALLOW_COPY_AND_ASSIGN(LoadClassSlowPathX86);
270};
271
272class TypeCheckSlowPathX86 : public SlowPathCodeX86 {
273 public:
274  TypeCheckSlowPathX86(HInstruction* instruction,
275                       Location class_to_check,
276                       Location object_class,
277                       uint32_t dex_pc)
278      : instruction_(instruction),
279        class_to_check_(class_to_check),
280        object_class_(object_class),
281        dex_pc_(dex_pc) {}
282
283  void EmitNativeCode(CodeGenerator* codegen) OVERRIDE {
284    LocationSummary* locations = instruction_->GetLocations();
285    DCHECK(instruction_->IsCheckCast()
286           || !locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg()));
287
288    CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen);
289    __ Bind(GetEntryLabel());
290    SaveLiveRegisters(codegen, locations);
291
292    // We're moving two locations to locations that could overlap, so we need a parallel
293    // move resolver.
294    InvokeRuntimeCallingConvention calling_convention;
295    x86_codegen->EmitParallelMoves(
296        class_to_check_,
297        Location::RegisterLocation(calling_convention.GetRegisterAt(0)),
298        object_class_,
299        Location::RegisterLocation(calling_convention.GetRegisterAt(1)));
300
301    if (instruction_->IsInstanceOf()) {
302      __ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize,
303                                                              pInstanceofNonTrivial)));
304    } else {
305      DCHECK(instruction_->IsCheckCast());
306      __ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pCheckCast)));
307    }
308
309    RecordPcInfo(codegen, instruction_, dex_pc_);
310    if (instruction_->IsInstanceOf()) {
311      x86_codegen->Move32(locations->Out(), Location::RegisterLocation(EAX));
312    }
313    RestoreLiveRegisters(codegen, locations);
314
315    __ jmp(GetExitLabel());
316  }
317
318 private:
319  HInstruction* const instruction_;
320  const Location class_to_check_;
321  const Location object_class_;
322  const uint32_t dex_pc_;
323
324  DISALLOW_COPY_AND_ASSIGN(TypeCheckSlowPathX86);
325};
326
327#undef __
328#define __ reinterpret_cast<X86Assembler*>(GetAssembler())->
329
330inline Condition X86Condition(IfCondition cond) {
331  switch (cond) {
332    case kCondEQ: return kEqual;
333    case kCondNE: return kNotEqual;
334    case kCondLT: return kLess;
335    case kCondLE: return kLessEqual;
336    case kCondGT: return kGreater;
337    case kCondGE: return kGreaterEqual;
338    default:
339      LOG(FATAL) << "Unknown if condition";
340  }
341  return kEqual;
342}
343
344void CodeGeneratorX86::DumpCoreRegister(std::ostream& stream, int reg) const {
345  stream << X86ManagedRegister::FromCpuRegister(Register(reg));
346}
347
348void CodeGeneratorX86::DumpFloatingPointRegister(std::ostream& stream, int reg) const {
349  stream << X86ManagedRegister::FromXmmRegister(XmmRegister(reg));
350}
351
352size_t CodeGeneratorX86::SaveCoreRegister(size_t stack_index, uint32_t reg_id) {
353  __ movl(Address(ESP, stack_index), static_cast<Register>(reg_id));
354  return kX86WordSize;
355}
356
357size_t CodeGeneratorX86::RestoreCoreRegister(size_t stack_index, uint32_t reg_id) {
358  __ movl(static_cast<Register>(reg_id), Address(ESP, stack_index));
359  return kX86WordSize;
360}
361
362size_t CodeGeneratorX86::SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) {
363  __ movsd(Address(ESP, stack_index), XmmRegister(reg_id));
364  return GetFloatingPointSpillSlotSize();
365}
366
367size_t CodeGeneratorX86::RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) {
368  __ movsd(XmmRegister(reg_id), Address(ESP, stack_index));
369  return GetFloatingPointSpillSlotSize();
370}
371
372CodeGeneratorX86::CodeGeneratorX86(HGraph* graph, const CompilerOptions& compiler_options)
373    : CodeGenerator(graph,
374                    kNumberOfCpuRegisters,
375                    kNumberOfXmmRegisters,
376                    kNumberOfRegisterPairs,
377                    ComputeRegisterMask(reinterpret_cast<const int*>(kCoreCalleeSaves),
378                                        arraysize(kCoreCalleeSaves))
379                        | (1 << kFakeReturnRegister),
380                        0,
381                        compiler_options),
382      block_labels_(graph->GetArena(), 0),
383      location_builder_(graph, this),
384      instruction_visitor_(graph, this),
385      move_resolver_(graph->GetArena(), this) {
386  // Use a fake return address register to mimic Quick.
387  AddAllocatedRegister(Location::RegisterLocation(kFakeReturnRegister));
388}
389
390Location CodeGeneratorX86::AllocateFreeRegister(Primitive::Type type) const {
391  switch (type) {
392    case Primitive::kPrimLong: {
393      size_t reg = FindFreeEntry(blocked_register_pairs_, kNumberOfRegisterPairs);
394      X86ManagedRegister pair =
395          X86ManagedRegister::FromRegisterPair(static_cast<RegisterPair>(reg));
396      DCHECK(!blocked_core_registers_[pair.AsRegisterPairLow()]);
397      DCHECK(!blocked_core_registers_[pair.AsRegisterPairHigh()]);
398      blocked_core_registers_[pair.AsRegisterPairLow()] = true;
399      blocked_core_registers_[pair.AsRegisterPairHigh()] = true;
400      UpdateBlockedPairRegisters();
401      return Location::RegisterPairLocation(pair.AsRegisterPairLow(), pair.AsRegisterPairHigh());
402    }
403
404    case Primitive::kPrimByte:
405    case Primitive::kPrimBoolean:
406    case Primitive::kPrimChar:
407    case Primitive::kPrimShort:
408    case Primitive::kPrimInt:
409    case Primitive::kPrimNot: {
410      Register reg = static_cast<Register>(
411          FindFreeEntry(blocked_core_registers_, kNumberOfCpuRegisters));
412      // Block all register pairs that contain `reg`.
413      for (int i = 0; i < kNumberOfRegisterPairs; i++) {
414        X86ManagedRegister current =
415            X86ManagedRegister::FromRegisterPair(static_cast<RegisterPair>(i));
416        if (current.AsRegisterPairLow() == reg || current.AsRegisterPairHigh() == reg) {
417          blocked_register_pairs_[i] = true;
418        }
419      }
420      return Location::RegisterLocation(reg);
421    }
422
423    case Primitive::kPrimFloat:
424    case Primitive::kPrimDouble: {
425      return Location::FpuRegisterLocation(
426          FindFreeEntry(blocked_fpu_registers_, kNumberOfXmmRegisters));
427    }
428
429    case Primitive::kPrimVoid:
430      LOG(FATAL) << "Unreachable type " << type;
431  }
432
433  return Location();
434}
435
436void CodeGeneratorX86::SetupBlockedRegisters(bool is_baseline) const {
437  // Don't allocate the dalvik style register pair passing.
438  blocked_register_pairs_[ECX_EDX] = true;
439
440  // Stack register is always reserved.
441  blocked_core_registers_[ESP] = true;
442
443  if (is_baseline) {
444    blocked_core_registers_[EBP] = true;
445    blocked_core_registers_[ESI] = true;
446    blocked_core_registers_[EDI] = true;
447  }
448
449  UpdateBlockedPairRegisters();
450}
451
452void CodeGeneratorX86::UpdateBlockedPairRegisters() const {
453  for (int i = 0; i < kNumberOfRegisterPairs; i++) {
454    X86ManagedRegister current =
455        X86ManagedRegister::FromRegisterPair(static_cast<RegisterPair>(i));
456    if (blocked_core_registers_[current.AsRegisterPairLow()]
457        || blocked_core_registers_[current.AsRegisterPairHigh()]) {
458      blocked_register_pairs_[i] = true;
459    }
460  }
461}
462
463InstructionCodeGeneratorX86::InstructionCodeGeneratorX86(HGraph* graph, CodeGeneratorX86* codegen)
464      : HGraphVisitor(graph),
465        assembler_(codegen->GetAssembler()),
466        codegen_(codegen) {}
467
468void CodeGeneratorX86::GenerateFrameEntry() {
469  __ Bind(&frame_entry_label_);
470  bool skip_overflow_check =
471      IsLeafMethod() && !FrameNeedsStackCheck(GetFrameSize(), InstructionSet::kX86);
472  DCHECK(GetCompilerOptions().GetImplicitStackOverflowChecks());
473
474  if (!skip_overflow_check) {
475    __ testl(EAX, Address(ESP, -static_cast<int32_t>(GetStackOverflowReservedBytes(kX86))));
476    RecordPcInfo(nullptr, 0);
477  }
478
479  if (HasEmptyFrame()) {
480    return;
481  }
482
483  for (int i = arraysize(kCoreCalleeSaves) - 1; i >= 0; --i) {
484    Register reg = kCoreCalleeSaves[i];
485    if (allocated_registers_.ContainsCoreRegister(reg)) {
486      __ pushl(reg);
487    }
488  }
489
490  __ subl(ESP, Immediate(GetFrameSize() - FrameEntrySpillSize()));
491  __ movl(Address(ESP, kCurrentMethodStackOffset), EAX);
492}
493
494void CodeGeneratorX86::GenerateFrameExit() {
495  if (HasEmptyFrame()) {
496    return;
497  }
498
499  __ addl(ESP, Immediate(GetFrameSize() - FrameEntrySpillSize()));
500
501  for (size_t i = 0; i < arraysize(kCoreCalleeSaves); ++i) {
502    Register reg = kCoreCalleeSaves[i];
503    if (allocated_registers_.ContainsCoreRegister(reg)) {
504      __ popl(reg);
505    }
506  }
507}
508
509void CodeGeneratorX86::Bind(HBasicBlock* block) {
510  __ Bind(GetLabelOf(block));
511}
512
513void CodeGeneratorX86::LoadCurrentMethod(Register reg) {
514  DCHECK(RequiresCurrentMethod());
515  __ movl(reg, Address(ESP, kCurrentMethodStackOffset));
516}
517
518Location CodeGeneratorX86::GetStackLocation(HLoadLocal* load) const {
519  switch (load->GetType()) {
520    case Primitive::kPrimLong:
521    case Primitive::kPrimDouble:
522      return Location::DoubleStackSlot(GetStackSlot(load->GetLocal()));
523      break;
524
525    case Primitive::kPrimInt:
526    case Primitive::kPrimNot:
527    case Primitive::kPrimFloat:
528      return Location::StackSlot(GetStackSlot(load->GetLocal()));
529
530    case Primitive::kPrimBoolean:
531    case Primitive::kPrimByte:
532    case Primitive::kPrimChar:
533    case Primitive::kPrimShort:
534    case Primitive::kPrimVoid:
535      LOG(FATAL) << "Unexpected type " << load->GetType();
536  }
537
538  LOG(FATAL) << "Unreachable";
539  return Location();
540}
541
542Location InvokeDexCallingConventionVisitor::GetNextLocation(Primitive::Type type) {
543  switch (type) {
544    case Primitive::kPrimBoolean:
545    case Primitive::kPrimByte:
546    case Primitive::kPrimChar:
547    case Primitive::kPrimShort:
548    case Primitive::kPrimInt:
549    case Primitive::kPrimNot: {
550      uint32_t index = gp_index_++;
551      stack_index_++;
552      if (index < calling_convention.GetNumberOfRegisters()) {
553        return Location::RegisterLocation(calling_convention.GetRegisterAt(index));
554      } else {
555        return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index_ - 1));
556      }
557    }
558
559    case Primitive::kPrimLong: {
560      uint32_t index = gp_index_;
561      gp_index_ += 2;
562      stack_index_ += 2;
563      if (index + 1 < calling_convention.GetNumberOfRegisters()) {
564        X86ManagedRegister pair = X86ManagedRegister::FromRegisterPair(
565            calling_convention.GetRegisterPairAt(index));
566        return Location::RegisterPairLocation(pair.AsRegisterPairLow(), pair.AsRegisterPairHigh());
567      } else {
568        return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index_ - 2));
569      }
570    }
571
572    case Primitive::kPrimFloat: {
573      uint32_t index = fp_index_++;
574      stack_index_++;
575      if (index < calling_convention.GetNumberOfFpuRegisters()) {
576        return Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(index));
577      } else {
578        return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index_ - 1));
579      }
580    }
581
582    case Primitive::kPrimDouble: {
583      uint32_t index = fp_index_++;
584      stack_index_ += 2;
585      if (index < calling_convention.GetNumberOfFpuRegisters()) {
586        return Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(index));
587      } else {
588        return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index_ - 2));
589      }
590    }
591
592    case Primitive::kPrimVoid:
593      LOG(FATAL) << "Unexpected parameter type " << type;
594      break;
595  }
596  return Location();
597}
598
599void CodeGeneratorX86::Move32(Location destination, Location source) {
600  if (source.Equals(destination)) {
601    return;
602  }
603  if (destination.IsRegister()) {
604    if (source.IsRegister()) {
605      __ movl(destination.AsRegister<Register>(), source.AsRegister<Register>());
606    } else if (source.IsFpuRegister()) {
607      __ movd(destination.AsRegister<Register>(), source.AsFpuRegister<XmmRegister>());
608    } else {
609      DCHECK(source.IsStackSlot());
610      __ movl(destination.AsRegister<Register>(), Address(ESP, source.GetStackIndex()));
611    }
612  } else if (destination.IsFpuRegister()) {
613    if (source.IsRegister()) {
614      __ movd(destination.AsFpuRegister<XmmRegister>(), source.AsRegister<Register>());
615    } else if (source.IsFpuRegister()) {
616      __ movaps(destination.AsFpuRegister<XmmRegister>(), source.AsFpuRegister<XmmRegister>());
617    } else {
618      DCHECK(source.IsStackSlot());
619      __ movss(destination.AsFpuRegister<XmmRegister>(), Address(ESP, source.GetStackIndex()));
620    }
621  } else {
622    DCHECK(destination.IsStackSlot()) << destination;
623    if (source.IsRegister()) {
624      __ movl(Address(ESP, destination.GetStackIndex()), source.AsRegister<Register>());
625    } else if (source.IsFpuRegister()) {
626      __ movss(Address(ESP, destination.GetStackIndex()), source.AsFpuRegister<XmmRegister>());
627    } else if (source.IsConstant()) {
628      HConstant* constant = source.GetConstant();
629      int32_t value = GetInt32ValueOf(constant);
630      __ movl(Address(ESP, destination.GetStackIndex()), Immediate(value));
631    } else {
632      DCHECK(source.IsStackSlot());
633      __ pushl(Address(ESP, source.GetStackIndex()));
634      __ popl(Address(ESP, destination.GetStackIndex()));
635    }
636  }
637}
638
639void CodeGeneratorX86::Move64(Location destination, Location source) {
640  if (source.Equals(destination)) {
641    return;
642  }
643  if (destination.IsRegisterPair()) {
644    if (source.IsRegisterPair()) {
645      EmitParallelMoves(
646          Location::RegisterLocation(source.AsRegisterPairHigh<Register>()),
647          Location::RegisterLocation(destination.AsRegisterPairHigh<Register>()),
648          Location::RegisterLocation(source.AsRegisterPairLow<Register>()),
649          Location::RegisterLocation(destination.AsRegisterPairLow<Register>()));
650    } else if (source.IsFpuRegister()) {
651      LOG(FATAL) << "Unimplemented";
652    } else {
653      // No conflict possible, so just do the moves.
654      DCHECK(source.IsDoubleStackSlot());
655      __ movl(destination.AsRegisterPairLow<Register>(), Address(ESP, source.GetStackIndex()));
656      __ movl(destination.AsRegisterPairHigh<Register>(),
657              Address(ESP, source.GetHighStackIndex(kX86WordSize)));
658    }
659  } else if (destination.IsFpuRegister()) {
660    if (source.IsFpuRegister()) {
661      __ movaps(destination.AsFpuRegister<XmmRegister>(), source.AsFpuRegister<XmmRegister>());
662    } else if (source.IsDoubleStackSlot()) {
663      __ movsd(destination.AsFpuRegister<XmmRegister>(), Address(ESP, source.GetStackIndex()));
664    } else {
665      LOG(FATAL) << "Unimplemented";
666    }
667  } else {
668    DCHECK(destination.IsDoubleStackSlot()) << destination;
669    if (source.IsRegisterPair()) {
670      // No conflict possible, so just do the moves.
671      __ movl(Address(ESP, destination.GetStackIndex()), source.AsRegisterPairLow<Register>());
672      __ movl(Address(ESP, destination.GetHighStackIndex(kX86WordSize)),
673              source.AsRegisterPairHigh<Register>());
674    } else if (source.IsFpuRegister()) {
675      __ movsd(Address(ESP, destination.GetStackIndex()), source.AsFpuRegister<XmmRegister>());
676    } else if (source.IsConstant()) {
677      HConstant* constant = source.GetConstant();
678      int64_t value;
679      if (constant->IsLongConstant()) {
680        value = constant->AsLongConstant()->GetValue();
681      } else {
682        DCHECK(constant->IsDoubleConstant());
683        value = bit_cast<int64_t, double>(constant->AsDoubleConstant()->GetValue());
684      }
685      __ movl(Address(ESP, destination.GetStackIndex()), Immediate(Low32Bits(value)));
686      __ movl(Address(ESP, destination.GetHighStackIndex(kX86WordSize)), Immediate(High32Bits(value)));
687    } else {
688      DCHECK(source.IsDoubleStackSlot()) << source;
689      EmitParallelMoves(
690          Location::StackSlot(source.GetStackIndex()),
691          Location::StackSlot(destination.GetStackIndex()),
692          Location::StackSlot(source.GetHighStackIndex(kX86WordSize)),
693          Location::StackSlot(destination.GetHighStackIndex(kX86WordSize)));
694    }
695  }
696}
697
698void CodeGeneratorX86::Move(HInstruction* instruction, Location location, HInstruction* move_for) {
699  LocationSummary* locations = instruction->GetLocations();
700  if (locations != nullptr && locations->Out().Equals(location)) {
701    return;
702  }
703
704  if (locations != nullptr && locations->Out().IsConstant()) {
705    HConstant* const_to_move = locations->Out().GetConstant();
706    if (const_to_move->IsIntConstant() || const_to_move->IsNullConstant()) {
707      Immediate imm(GetInt32ValueOf(const_to_move));
708      if (location.IsRegister()) {
709        __ movl(location.AsRegister<Register>(), imm);
710      } else if (location.IsStackSlot()) {
711        __ movl(Address(ESP, location.GetStackIndex()), imm);
712      } else {
713        DCHECK(location.IsConstant());
714        DCHECK_EQ(location.GetConstant(), const_to_move);
715      }
716    } else if (const_to_move->IsLongConstant()) {
717      int64_t value = const_to_move->AsLongConstant()->GetValue();
718      if (location.IsRegisterPair()) {
719        __ movl(location.AsRegisterPairLow<Register>(), Immediate(Low32Bits(value)));
720        __ movl(location.AsRegisterPairHigh<Register>(), Immediate(High32Bits(value)));
721      } else if (location.IsDoubleStackSlot()) {
722        __ movl(Address(ESP, location.GetStackIndex()), Immediate(Low32Bits(value)));
723        __ movl(Address(ESP, location.GetHighStackIndex(kX86WordSize)),
724                Immediate(High32Bits(value)));
725      } else {
726        DCHECK(location.IsConstant());
727        DCHECK_EQ(location.GetConstant(), instruction);
728      }
729    }
730  } else if (instruction->IsTemporary()) {
731    Location temp_location = GetTemporaryLocation(instruction->AsTemporary());
732    if (temp_location.IsStackSlot()) {
733      Move32(location, temp_location);
734    } else {
735      DCHECK(temp_location.IsDoubleStackSlot());
736      Move64(location, temp_location);
737    }
738  } else if (instruction->IsLoadLocal()) {
739    int slot = GetStackSlot(instruction->AsLoadLocal()->GetLocal());
740    switch (instruction->GetType()) {
741      case Primitive::kPrimBoolean:
742      case Primitive::kPrimByte:
743      case Primitive::kPrimChar:
744      case Primitive::kPrimShort:
745      case Primitive::kPrimInt:
746      case Primitive::kPrimNot:
747      case Primitive::kPrimFloat:
748        Move32(location, Location::StackSlot(slot));
749        break;
750
751      case Primitive::kPrimLong:
752      case Primitive::kPrimDouble:
753        Move64(location, Location::DoubleStackSlot(slot));
754        break;
755
756      default:
757        LOG(FATAL) << "Unimplemented local type " << instruction->GetType();
758    }
759  } else {
760    DCHECK((instruction->GetNext() == move_for) || instruction->GetNext()->IsTemporary());
761    switch (instruction->GetType()) {
762      case Primitive::kPrimBoolean:
763      case Primitive::kPrimByte:
764      case Primitive::kPrimChar:
765      case Primitive::kPrimShort:
766      case Primitive::kPrimInt:
767      case Primitive::kPrimNot:
768      case Primitive::kPrimFloat:
769        Move32(location, locations->Out());
770        break;
771
772      case Primitive::kPrimLong:
773      case Primitive::kPrimDouble:
774        Move64(location, locations->Out());
775        break;
776
777      default:
778        LOG(FATAL) << "Unexpected type " << instruction->GetType();
779    }
780  }
781}
782
783void LocationsBuilderX86::VisitGoto(HGoto* got) {
784  got->SetLocations(nullptr);
785}
786
787void InstructionCodeGeneratorX86::VisitGoto(HGoto* got) {
788  HBasicBlock* successor = got->GetSuccessor();
789  DCHECK(!successor->IsExitBlock());
790
791  HBasicBlock* block = got->GetBlock();
792  HInstruction* previous = got->GetPrevious();
793
794  HLoopInformation* info = block->GetLoopInformation();
795  if (info != nullptr && info->IsBackEdge(*block) && info->HasSuspendCheck()) {
796    codegen_->ClearSpillSlotsFromLoopPhisInStackMap(info->GetSuspendCheck());
797    GenerateSuspendCheck(info->GetSuspendCheck(), successor);
798    return;
799  }
800
801  if (block->IsEntryBlock() && (previous != nullptr) && previous->IsSuspendCheck()) {
802    GenerateSuspendCheck(previous->AsSuspendCheck(), nullptr);
803  }
804  if (!codegen_->GoesToNextBlock(got->GetBlock(), successor)) {
805    __ jmp(codegen_->GetLabelOf(successor));
806  }
807}
808
809void LocationsBuilderX86::VisitExit(HExit* exit) {
810  exit->SetLocations(nullptr);
811}
812
813void InstructionCodeGeneratorX86::VisitExit(HExit* exit) {
814  UNUSED(exit);
815}
816
817void LocationsBuilderX86::VisitIf(HIf* if_instr) {
818  LocationSummary* locations =
819      new (GetGraph()->GetArena()) LocationSummary(if_instr, LocationSummary::kNoCall);
820  HInstruction* cond = if_instr->InputAt(0);
821  if (!cond->IsCondition() || cond->AsCondition()->NeedsMaterialization()) {
822    locations->SetInAt(0, Location::Any());
823  }
824}
825
826void InstructionCodeGeneratorX86::VisitIf(HIf* if_instr) {
827  HInstruction* cond = if_instr->InputAt(0);
828  if (cond->IsIntConstant()) {
829    // Constant condition, statically compared against 1.
830    int32_t cond_value = cond->AsIntConstant()->GetValue();
831    if (cond_value == 1) {
832      if (!codegen_->GoesToNextBlock(if_instr->GetBlock(),
833                                     if_instr->IfTrueSuccessor())) {
834        __ jmp(codegen_->GetLabelOf(if_instr->IfTrueSuccessor()));
835      }
836      return;
837    } else {
838      DCHECK_EQ(cond_value, 0);
839    }
840  } else {
841    bool materialized =
842        !cond->IsCondition() || cond->AsCondition()->NeedsMaterialization();
843    // Moves do not affect the eflags register, so if the condition is
844    // evaluated just before the if, we don't need to evaluate it
845    // again.
846    bool eflags_set = cond->IsCondition()
847        && cond->AsCondition()->IsBeforeWhenDisregardMoves(if_instr);
848    if (materialized) {
849      if (!eflags_set) {
850        // Materialized condition, compare against 0.
851        Location lhs = if_instr->GetLocations()->InAt(0);
852        if (lhs.IsRegister()) {
853          __ testl(lhs.AsRegister<Register>(), lhs.AsRegister<Register>());
854        } else {
855          __ cmpl(Address(ESP, lhs.GetStackIndex()), Immediate(0));
856        }
857        __ j(kNotEqual,  codegen_->GetLabelOf(if_instr->IfTrueSuccessor()));
858      } else {
859        __ j(X86Condition(cond->AsCondition()->GetCondition()),
860             codegen_->GetLabelOf(if_instr->IfTrueSuccessor()));
861      }
862    } else {
863      Location lhs = cond->GetLocations()->InAt(0);
864      Location rhs = cond->GetLocations()->InAt(1);
865      // LHS is guaranteed to be in a register (see
866      // LocationsBuilderX86::VisitCondition).
867      if (rhs.IsRegister()) {
868        __ cmpl(lhs.AsRegister<Register>(), rhs.AsRegister<Register>());
869      } else if (rhs.IsConstant()) {
870        int32_t constant = rhs.GetConstant()->AsIntConstant()->GetValue();
871        if (constant == 0) {
872          __ testl(lhs.AsRegister<Register>(), lhs.AsRegister<Register>());
873        } else {
874          __ cmpl(lhs.AsRegister<Register>(), Immediate(constant));
875        }
876      } else {
877        __ cmpl(lhs.AsRegister<Register>(), Address(ESP, rhs.GetStackIndex()));
878      }
879      __ j(X86Condition(cond->AsCondition()->GetCondition()),
880           codegen_->GetLabelOf(if_instr->IfTrueSuccessor()));
881    }
882  }
883  if (!codegen_->GoesToNextBlock(if_instr->GetBlock(),
884                                 if_instr->IfFalseSuccessor())) {
885    __ jmp(codegen_->GetLabelOf(if_instr->IfFalseSuccessor()));
886  }
887}
888
889void LocationsBuilderX86::VisitLocal(HLocal* local) {
890  local->SetLocations(nullptr);
891}
892
893void InstructionCodeGeneratorX86::VisitLocal(HLocal* local) {
894  DCHECK_EQ(local->GetBlock(), GetGraph()->GetEntryBlock());
895}
896
897void LocationsBuilderX86::VisitLoadLocal(HLoadLocal* local) {
898  local->SetLocations(nullptr);
899}
900
901void InstructionCodeGeneratorX86::VisitLoadLocal(HLoadLocal* load) {
902  // Nothing to do, this is driven by the code generator.
903  UNUSED(load);
904}
905
906void LocationsBuilderX86::VisitStoreLocal(HStoreLocal* store) {
907  LocationSummary* locations =
908      new (GetGraph()->GetArena()) LocationSummary(store, LocationSummary::kNoCall);
909  switch (store->InputAt(1)->GetType()) {
910    case Primitive::kPrimBoolean:
911    case Primitive::kPrimByte:
912    case Primitive::kPrimChar:
913    case Primitive::kPrimShort:
914    case Primitive::kPrimInt:
915    case Primitive::kPrimNot:
916    case Primitive::kPrimFloat:
917      locations->SetInAt(1, Location::StackSlot(codegen_->GetStackSlot(store->GetLocal())));
918      break;
919
920    case Primitive::kPrimLong:
921    case Primitive::kPrimDouble:
922      locations->SetInAt(1, Location::DoubleStackSlot(codegen_->GetStackSlot(store->GetLocal())));
923      break;
924
925    default:
926      LOG(FATAL) << "Unknown local type " << store->InputAt(1)->GetType();
927  }
928  store->SetLocations(locations);
929}
930
931void InstructionCodeGeneratorX86::VisitStoreLocal(HStoreLocal* store) {
932  UNUSED(store);
933}
934
935void LocationsBuilderX86::VisitCondition(HCondition* comp) {
936  LocationSummary* locations =
937      new (GetGraph()->GetArena()) LocationSummary(comp, LocationSummary::kNoCall);
938  locations->SetInAt(0, Location::RequiresRegister());
939  locations->SetInAt(1, Location::Any());
940  if (comp->NeedsMaterialization()) {
941    // We need a byte register.
942    locations->SetOut(Location::RegisterLocation(ECX));
943  }
944}
945
946void InstructionCodeGeneratorX86::VisitCondition(HCondition* comp) {
947  if (comp->NeedsMaterialization()) {
948    LocationSummary* locations = comp->GetLocations();
949    Register reg = locations->Out().AsRegister<Register>();
950    // Clear register: setcc only sets the low byte.
951    __ xorl(reg, reg);
952    Location lhs = locations->InAt(0);
953    Location rhs = locations->InAt(1);
954    if (rhs.IsRegister()) {
955      __ cmpl(lhs.AsRegister<Register>(), rhs.AsRegister<Register>());
956    } else if (rhs.IsConstant()) {
957      int32_t constant = CodeGenerator::GetInt32ValueOf(rhs.GetConstant());
958      if (constant == 0) {
959        __ testl(lhs.AsRegister<Register>(), lhs.AsRegister<Register>());
960      } else {
961      __ cmpl(lhs.AsRegister<Register>(), Immediate(constant));
962      }
963    } else {
964      __ cmpl(lhs.AsRegister<Register>(), Address(ESP, rhs.GetStackIndex()));
965    }
966    __ setb(X86Condition(comp->GetCondition()), reg);
967  }
968}
969
970void LocationsBuilderX86::VisitEqual(HEqual* comp) {
971  VisitCondition(comp);
972}
973
974void InstructionCodeGeneratorX86::VisitEqual(HEqual* comp) {
975  VisitCondition(comp);
976}
977
978void LocationsBuilderX86::VisitNotEqual(HNotEqual* comp) {
979  VisitCondition(comp);
980}
981
982void InstructionCodeGeneratorX86::VisitNotEqual(HNotEqual* comp) {
983  VisitCondition(comp);
984}
985
986void LocationsBuilderX86::VisitLessThan(HLessThan* comp) {
987  VisitCondition(comp);
988}
989
990void InstructionCodeGeneratorX86::VisitLessThan(HLessThan* comp) {
991  VisitCondition(comp);
992}
993
994void LocationsBuilderX86::VisitLessThanOrEqual(HLessThanOrEqual* comp) {
995  VisitCondition(comp);
996}
997
998void InstructionCodeGeneratorX86::VisitLessThanOrEqual(HLessThanOrEqual* comp) {
999  VisitCondition(comp);
1000}
1001
1002void LocationsBuilderX86::VisitGreaterThan(HGreaterThan* comp) {
1003  VisitCondition(comp);
1004}
1005
1006void InstructionCodeGeneratorX86::VisitGreaterThan(HGreaterThan* comp) {
1007  VisitCondition(comp);
1008}
1009
1010void LocationsBuilderX86::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) {
1011  VisitCondition(comp);
1012}
1013
1014void InstructionCodeGeneratorX86::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) {
1015  VisitCondition(comp);
1016}
1017
1018void LocationsBuilderX86::VisitIntConstant(HIntConstant* constant) {
1019  LocationSummary* locations =
1020      new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall);
1021  locations->SetOut(Location::ConstantLocation(constant));
1022}
1023
1024void InstructionCodeGeneratorX86::VisitIntConstant(HIntConstant* constant) {
1025  // Will be generated at use site.
1026  UNUSED(constant);
1027}
1028
1029void LocationsBuilderX86::VisitNullConstant(HNullConstant* constant) {
1030  LocationSummary* locations =
1031      new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall);
1032  locations->SetOut(Location::ConstantLocation(constant));
1033}
1034
1035void InstructionCodeGeneratorX86::VisitNullConstant(HNullConstant* constant) {
1036  // Will be generated at use site.
1037  UNUSED(constant);
1038}
1039
1040void LocationsBuilderX86::VisitLongConstant(HLongConstant* constant) {
1041  LocationSummary* locations =
1042      new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall);
1043  locations->SetOut(Location::ConstantLocation(constant));
1044}
1045
1046void InstructionCodeGeneratorX86::VisitLongConstant(HLongConstant* constant) {
1047  // Will be generated at use site.
1048  UNUSED(constant);
1049}
1050
1051void LocationsBuilderX86::VisitFloatConstant(HFloatConstant* constant) {
1052  LocationSummary* locations =
1053      new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall);
1054  locations->SetOut(Location::ConstantLocation(constant));
1055}
1056
1057void InstructionCodeGeneratorX86::VisitFloatConstant(HFloatConstant* constant) {
1058  // Will be generated at use site.
1059  UNUSED(constant);
1060}
1061
1062void LocationsBuilderX86::VisitDoubleConstant(HDoubleConstant* constant) {
1063  LocationSummary* locations =
1064      new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall);
1065  locations->SetOut(Location::ConstantLocation(constant));
1066}
1067
1068void InstructionCodeGeneratorX86::VisitDoubleConstant(HDoubleConstant* constant) {
1069  // Will be generated at use site.
1070  UNUSED(constant);
1071}
1072
1073void LocationsBuilderX86::VisitReturnVoid(HReturnVoid* ret) {
1074  ret->SetLocations(nullptr);
1075}
1076
1077void InstructionCodeGeneratorX86::VisitReturnVoid(HReturnVoid* ret) {
1078  UNUSED(ret);
1079  codegen_->GenerateFrameExit();
1080  __ ret();
1081}
1082
1083void LocationsBuilderX86::VisitReturn(HReturn* ret) {
1084  LocationSummary* locations =
1085      new (GetGraph()->GetArena()) LocationSummary(ret, LocationSummary::kNoCall);
1086  switch (ret->InputAt(0)->GetType()) {
1087    case Primitive::kPrimBoolean:
1088    case Primitive::kPrimByte:
1089    case Primitive::kPrimChar:
1090    case Primitive::kPrimShort:
1091    case Primitive::kPrimInt:
1092    case Primitive::kPrimNot:
1093      locations->SetInAt(0, Location::RegisterLocation(EAX));
1094      break;
1095
1096    case Primitive::kPrimLong:
1097      locations->SetInAt(
1098          0, Location::RegisterPairLocation(EAX, EDX));
1099      break;
1100
1101    case Primitive::kPrimFloat:
1102    case Primitive::kPrimDouble:
1103      locations->SetInAt(
1104          0, Location::FpuRegisterLocation(XMM0));
1105      break;
1106
1107    default:
1108      LOG(FATAL) << "Unknown return type " << ret->InputAt(0)->GetType();
1109  }
1110}
1111
1112void InstructionCodeGeneratorX86::VisitReturn(HReturn* ret) {
1113  if (kIsDebugBuild) {
1114    switch (ret->InputAt(0)->GetType()) {
1115      case Primitive::kPrimBoolean:
1116      case Primitive::kPrimByte:
1117      case Primitive::kPrimChar:
1118      case Primitive::kPrimShort:
1119      case Primitive::kPrimInt:
1120      case Primitive::kPrimNot:
1121        DCHECK_EQ(ret->GetLocations()->InAt(0).AsRegister<Register>(), EAX);
1122        break;
1123
1124      case Primitive::kPrimLong:
1125        DCHECK_EQ(ret->GetLocations()->InAt(0).AsRegisterPairLow<Register>(), EAX);
1126        DCHECK_EQ(ret->GetLocations()->InAt(0).AsRegisterPairHigh<Register>(), EDX);
1127        break;
1128
1129      case Primitive::kPrimFloat:
1130      case Primitive::kPrimDouble:
1131        DCHECK_EQ(ret->GetLocations()->InAt(0).AsFpuRegister<XmmRegister>(), XMM0);
1132        break;
1133
1134      default:
1135        LOG(FATAL) << "Unknown return type " << ret->InputAt(0)->GetType();
1136    }
1137  }
1138  codegen_->GenerateFrameExit();
1139  __ ret();
1140}
1141
1142void LocationsBuilderX86::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
1143  HandleInvoke(invoke);
1144}
1145
1146void InstructionCodeGeneratorX86::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
1147  Register temp = invoke->GetLocations()->GetTemp(0).AsRegister<Register>();
1148
1149  // TODO: Implement all kinds of calls:
1150  // 1) boot -> boot
1151  // 2) app -> boot
1152  // 3) app -> app
1153  //
1154  // Currently we implement the app -> app logic, which looks up in the resolve cache.
1155
1156  // temp = method;
1157  codegen_->LoadCurrentMethod(temp);
1158  if (!invoke->IsRecursive()) {
1159    // temp = temp->dex_cache_resolved_methods_;
1160    __ movl(temp, Address(temp, mirror::ArtMethod::DexCacheResolvedMethodsOffset().Int32Value()));
1161    // temp = temp[index_in_cache]
1162    __ movl(temp, Address(temp, CodeGenerator::GetCacheOffset(invoke->GetDexMethodIndex())));
1163    // (temp + offset_of_quick_compiled_code)()
1164    __ call(Address(
1165        temp, mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset(kX86WordSize).Int32Value()));
1166  } else {
1167    __ call(codegen_->GetFrameEntryLabel());
1168  }
1169
1170  DCHECK(!codegen_->IsLeafMethod());
1171  codegen_->RecordPcInfo(invoke, invoke->GetDexPc());
1172}
1173
1174void LocationsBuilderX86::VisitInvokeVirtual(HInvokeVirtual* invoke) {
1175  HandleInvoke(invoke);
1176}
1177
1178void LocationsBuilderX86::HandleInvoke(HInvoke* invoke) {
1179  LocationSummary* locations =
1180      new (GetGraph()->GetArena()) LocationSummary(invoke, LocationSummary::kCall);
1181  locations->AddTemp(Location::RegisterLocation(EAX));
1182
1183  InvokeDexCallingConventionVisitor calling_convention_visitor;
1184  for (size_t i = 0; i < invoke->InputCount(); i++) {
1185    HInstruction* input = invoke->InputAt(i);
1186    locations->SetInAt(i, calling_convention_visitor.GetNextLocation(input->GetType()));
1187  }
1188
1189  switch (invoke->GetType()) {
1190    case Primitive::kPrimBoolean:
1191    case Primitive::kPrimByte:
1192    case Primitive::kPrimChar:
1193    case Primitive::kPrimShort:
1194    case Primitive::kPrimInt:
1195    case Primitive::kPrimNot:
1196      locations->SetOut(Location::RegisterLocation(EAX));
1197      break;
1198
1199    case Primitive::kPrimLong:
1200      locations->SetOut(Location::RegisterPairLocation(EAX, EDX));
1201      break;
1202
1203    case Primitive::kPrimVoid:
1204      break;
1205
1206    case Primitive::kPrimDouble:
1207    case Primitive::kPrimFloat:
1208      locations->SetOut(Location::FpuRegisterLocation(XMM0));
1209      break;
1210  }
1211
1212  invoke->SetLocations(locations);
1213}
1214
1215void InstructionCodeGeneratorX86::VisitInvokeVirtual(HInvokeVirtual* invoke) {
1216  Register temp = invoke->GetLocations()->GetTemp(0).AsRegister<Register>();
1217  uint32_t method_offset = mirror::Class::EmbeddedVTableOffset().Uint32Value() +
1218          invoke->GetVTableIndex() * sizeof(mirror::Class::VTableEntry);
1219  LocationSummary* locations = invoke->GetLocations();
1220  Location receiver = locations->InAt(0);
1221  uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
1222  // temp = object->GetClass();
1223  if (receiver.IsStackSlot()) {
1224    __ movl(temp, Address(ESP, receiver.GetStackIndex()));
1225    __ movl(temp, Address(temp, class_offset));
1226  } else {
1227    __ movl(temp, Address(receiver.AsRegister<Register>(), class_offset));
1228  }
1229  codegen_->MaybeRecordImplicitNullCheck(invoke);
1230  // temp = temp->GetMethodAt(method_offset);
1231  __ movl(temp, Address(temp, method_offset));
1232  // call temp->GetEntryPoint();
1233  __ call(Address(
1234      temp, mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset(kX86WordSize).Int32Value()));
1235
1236  DCHECK(!codegen_->IsLeafMethod());
1237  codegen_->RecordPcInfo(invoke, invoke->GetDexPc());
1238}
1239
1240void LocationsBuilderX86::VisitInvokeInterface(HInvokeInterface* invoke) {
1241  HandleInvoke(invoke);
1242  // Add the hidden argument.
1243  invoke->GetLocations()->AddTemp(Location::FpuRegisterLocation(XMM7));
1244}
1245
1246void InstructionCodeGeneratorX86::VisitInvokeInterface(HInvokeInterface* invoke) {
1247  // TODO: b/18116999, our IMTs can miss an IncompatibleClassChangeError.
1248  Register temp = invoke->GetLocations()->GetTemp(0).AsRegister<Register>();
1249  uint32_t method_offset = mirror::Class::EmbeddedImTableOffset().Uint32Value() +
1250          (invoke->GetImtIndex() % mirror::Class::kImtSize) * sizeof(mirror::Class::ImTableEntry);
1251  LocationSummary* locations = invoke->GetLocations();
1252  Location receiver = locations->InAt(0);
1253  uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
1254
1255  // Set the hidden argument.
1256  __ movl(temp, Immediate(invoke->GetDexMethodIndex()));
1257  __ movd(invoke->GetLocations()->GetTemp(1).AsFpuRegister<XmmRegister>(), temp);
1258
1259  // temp = object->GetClass();
1260  if (receiver.IsStackSlot()) {
1261    __ movl(temp, Address(ESP, receiver.GetStackIndex()));
1262    __ movl(temp, Address(temp, class_offset));
1263  } else {
1264    __ movl(temp, Address(receiver.AsRegister<Register>(), class_offset));
1265  }
1266    codegen_->MaybeRecordImplicitNullCheck(invoke);
1267  // temp = temp->GetImtEntryAt(method_offset);
1268  __ movl(temp, Address(temp, method_offset));
1269  // call temp->GetEntryPoint();
1270  __ call(Address(temp, mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset(
1271      kX86WordSize).Int32Value()));
1272
1273  DCHECK(!codegen_->IsLeafMethod());
1274  codegen_->RecordPcInfo(invoke, invoke->GetDexPc());
1275}
1276
1277void LocationsBuilderX86::VisitNeg(HNeg* neg) {
1278  LocationSummary* locations =
1279      new (GetGraph()->GetArena()) LocationSummary(neg, LocationSummary::kNoCall);
1280  switch (neg->GetResultType()) {
1281    case Primitive::kPrimInt:
1282    case Primitive::kPrimLong:
1283      locations->SetInAt(0, Location::RequiresRegister());
1284      locations->SetOut(Location::SameAsFirstInput());
1285      break;
1286
1287    case Primitive::kPrimFloat:
1288      locations->SetInAt(0, Location::RequiresFpuRegister());
1289      locations->SetOut(Location::SameAsFirstInput());
1290      locations->AddTemp(Location::RequiresRegister());
1291      locations->AddTemp(Location::RequiresFpuRegister());
1292      break;
1293
1294    case Primitive::kPrimDouble:
1295      locations->SetInAt(0, Location::RequiresFpuRegister());
1296      locations->SetOut(Location::SameAsFirstInput());
1297      locations->AddTemp(Location::RequiresFpuRegister());
1298      break;
1299
1300    default:
1301      LOG(FATAL) << "Unexpected neg type " << neg->GetResultType();
1302  }
1303}
1304
1305void InstructionCodeGeneratorX86::VisitNeg(HNeg* neg) {
1306  LocationSummary* locations = neg->GetLocations();
1307  Location out = locations->Out();
1308  Location in = locations->InAt(0);
1309  switch (neg->GetResultType()) {
1310    case Primitive::kPrimInt:
1311      DCHECK(in.IsRegister());
1312      DCHECK(in.Equals(out));
1313      __ negl(out.AsRegister<Register>());
1314      break;
1315
1316    case Primitive::kPrimLong:
1317      DCHECK(in.IsRegisterPair());
1318      DCHECK(in.Equals(out));
1319      __ negl(out.AsRegisterPairLow<Register>());
1320      // Negation is similar to subtraction from zero.  The least
1321      // significant byte triggers a borrow when it is different from
1322      // zero; to take it into account, add 1 to the most significant
1323      // byte if the carry flag (CF) is set to 1 after the first NEGL
1324      // operation.
1325      __ adcl(out.AsRegisterPairHigh<Register>(), Immediate(0));
1326      __ negl(out.AsRegisterPairHigh<Register>());
1327      break;
1328
1329    case Primitive::kPrimFloat: {
1330      DCHECK(in.Equals(out));
1331      Register constant = locations->GetTemp(0).AsRegister<Register>();
1332      XmmRegister mask = locations->GetTemp(1).AsFpuRegister<XmmRegister>();
1333      // Implement float negation with an exclusive or with value
1334      // 0x80000000 (mask for bit 31, representing the sign of a
1335      // single-precision floating-point number).
1336      __ movl(constant, Immediate(INT32_C(0x80000000)));
1337      __ movd(mask, constant);
1338      __ xorps(out.AsFpuRegister<XmmRegister>(), mask);
1339      break;
1340    }
1341
1342    case Primitive::kPrimDouble: {
1343      DCHECK(in.Equals(out));
1344      XmmRegister mask = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
1345      // Implement double negation with an exclusive or with value
1346      // 0x8000000000000000 (mask for bit 63, representing the sign of
1347      // a double-precision floating-point number).
1348      __ LoadLongConstant(mask, INT64_C(0x8000000000000000));
1349      __ xorpd(out.AsFpuRegister<XmmRegister>(), mask);
1350      break;
1351    }
1352
1353    default:
1354      LOG(FATAL) << "Unexpected neg type " << neg->GetResultType();
1355  }
1356}
1357
1358void LocationsBuilderX86::VisitTypeConversion(HTypeConversion* conversion) {
1359  Primitive::Type result_type = conversion->GetResultType();
1360  Primitive::Type input_type = conversion->GetInputType();
1361  DCHECK_NE(result_type, input_type);
1362
1363  // The float-to-long and double-to-long type conversions rely on a
1364  // call to the runtime.
1365  LocationSummary::CallKind call_kind =
1366      ((input_type == Primitive::kPrimFloat || input_type == Primitive::kPrimDouble)
1367       && result_type == Primitive::kPrimLong)
1368      ? LocationSummary::kCall
1369      : LocationSummary::kNoCall;
1370  LocationSummary* locations =
1371      new (GetGraph()->GetArena()) LocationSummary(conversion, call_kind);
1372
1373  // The Java language does not allow treating boolean as an integral type but
1374  // our bit representation makes it safe.
1375
1376  switch (result_type) {
1377    case Primitive::kPrimByte:
1378      switch (input_type) {
1379        case Primitive::kPrimBoolean:
1380          // Boolean input is a result of code transformations.
1381        case Primitive::kPrimShort:
1382        case Primitive::kPrimInt:
1383        case Primitive::kPrimChar:
1384          // Processing a Dex `int-to-byte' instruction.
1385          locations->SetInAt(0, Location::ByteRegisterOrConstant(ECX, conversion->InputAt(0)));
1386          // Make the output overlap to please the register allocator. This greatly simplifies
1387          // the validation of the linear scan implementation
1388          locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
1389          break;
1390
1391        default:
1392          LOG(FATAL) << "Unexpected type conversion from " << input_type
1393                     << " to " << result_type;
1394      }
1395      break;
1396
1397    case Primitive::kPrimShort:
1398      switch (input_type) {
1399        case Primitive::kPrimBoolean:
1400          // Boolean input is a result of code transformations.
1401        case Primitive::kPrimByte:
1402        case Primitive::kPrimInt:
1403        case Primitive::kPrimChar:
1404          // Processing a Dex `int-to-short' instruction.
1405          locations->SetInAt(0, Location::Any());
1406          locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
1407          break;
1408
1409        default:
1410          LOG(FATAL) << "Unexpected type conversion from " << input_type
1411                     << " to " << result_type;
1412      }
1413      break;
1414
1415    case Primitive::kPrimInt:
1416      switch (input_type) {
1417        case Primitive::kPrimLong:
1418          // Processing a Dex `long-to-int' instruction.
1419          locations->SetInAt(0, Location::Any());
1420          locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
1421          break;
1422
1423        case Primitive::kPrimFloat:
1424          // Processing a Dex `float-to-int' instruction.
1425          locations->SetInAt(0, Location::RequiresFpuRegister());
1426          locations->SetOut(Location::RequiresRegister());
1427          locations->AddTemp(Location::RequiresFpuRegister());
1428          break;
1429
1430        case Primitive::kPrimDouble:
1431          // Processing a Dex `double-to-int' instruction.
1432          locations->SetInAt(0, Location::RequiresFpuRegister());
1433          locations->SetOut(Location::RequiresRegister());
1434          locations->AddTemp(Location::RequiresFpuRegister());
1435          break;
1436
1437        default:
1438          LOG(FATAL) << "Unexpected type conversion from " << input_type
1439                     << " to " << result_type;
1440      }
1441      break;
1442
1443    case Primitive::kPrimLong:
1444      switch (input_type) {
1445        case Primitive::kPrimBoolean:
1446          // Boolean input is a result of code transformations.
1447        case Primitive::kPrimByte:
1448        case Primitive::kPrimShort:
1449        case Primitive::kPrimInt:
1450        case Primitive::kPrimChar:
1451          // Processing a Dex `int-to-long' instruction.
1452          locations->SetInAt(0, Location::RegisterLocation(EAX));
1453          locations->SetOut(Location::RegisterPairLocation(EAX, EDX));
1454          break;
1455
1456        case Primitive::kPrimFloat:
1457        case Primitive::kPrimDouble: {
1458          // Processing a Dex `float-to-long' or 'double-to-long' instruction.
1459          InvokeRuntimeCallingConvention calling_convention;
1460          XmmRegister parameter = calling_convention.GetFpuRegisterAt(0);
1461          locations->SetInAt(0, Location::FpuRegisterLocation(parameter));
1462
1463          // The runtime helper puts the result in EAX, EDX.
1464          locations->SetOut(Location::RegisterPairLocation(EAX, EDX));
1465        }
1466        break;
1467
1468        default:
1469          LOG(FATAL) << "Unexpected type conversion from " << input_type
1470                     << " to " << result_type;
1471      }
1472      break;
1473
1474    case Primitive::kPrimChar:
1475      switch (input_type) {
1476        case Primitive::kPrimBoolean:
1477          // Boolean input is a result of code transformations.
1478        case Primitive::kPrimByte:
1479        case Primitive::kPrimShort:
1480        case Primitive::kPrimInt:
1481          // Processing a Dex `int-to-char' instruction.
1482          locations->SetInAt(0, Location::Any());
1483          locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
1484          break;
1485
1486        default:
1487          LOG(FATAL) << "Unexpected type conversion from " << input_type
1488                     << " to " << result_type;
1489      }
1490      break;
1491
1492    case Primitive::kPrimFloat:
1493      switch (input_type) {
1494        case Primitive::kPrimBoolean:
1495          // Boolean input is a result of code transformations.
1496        case Primitive::kPrimByte:
1497        case Primitive::kPrimShort:
1498        case Primitive::kPrimInt:
1499        case Primitive::kPrimChar:
1500          // Processing a Dex `int-to-float' instruction.
1501          locations->SetInAt(0, Location::RequiresRegister());
1502          locations->SetOut(Location::RequiresFpuRegister());
1503          break;
1504
1505        case Primitive::kPrimLong:
1506          // Processing a Dex `long-to-float' instruction.
1507          locations->SetInAt(0, Location::RequiresRegister());
1508          locations->SetOut(Location::RequiresFpuRegister());
1509          locations->AddTemp(Location::RequiresFpuRegister());
1510          locations->AddTemp(Location::RequiresFpuRegister());
1511          break;
1512
1513        case Primitive::kPrimDouble:
1514          // Processing a Dex `double-to-float' instruction.
1515          locations->SetInAt(0, Location::RequiresFpuRegister());
1516          locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
1517          break;
1518
1519        default:
1520          LOG(FATAL) << "Unexpected type conversion from " << input_type
1521                     << " to " << result_type;
1522      };
1523      break;
1524
1525    case Primitive::kPrimDouble:
1526      switch (input_type) {
1527        case Primitive::kPrimBoolean:
1528          // Boolean input is a result of code transformations.
1529        case Primitive::kPrimByte:
1530        case Primitive::kPrimShort:
1531        case Primitive::kPrimInt:
1532        case Primitive::kPrimChar:
1533          // Processing a Dex `int-to-double' instruction.
1534          locations->SetInAt(0, Location::RequiresRegister());
1535          locations->SetOut(Location::RequiresFpuRegister());
1536          break;
1537
1538        case Primitive::kPrimLong:
1539          // Processing a Dex `long-to-double' instruction.
1540          locations->SetInAt(0, Location::RequiresRegister());
1541          locations->SetOut(Location::RequiresFpuRegister());
1542          locations->AddTemp(Location::RequiresFpuRegister());
1543          locations->AddTemp(Location::RequiresFpuRegister());
1544          break;
1545
1546        case Primitive::kPrimFloat:
1547          // Processing a Dex `float-to-double' instruction.
1548          locations->SetInAt(0, Location::RequiresFpuRegister());
1549          locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
1550          break;
1551
1552        default:
1553          LOG(FATAL) << "Unexpected type conversion from " << input_type
1554                     << " to " << result_type;
1555      }
1556      break;
1557
1558    default:
1559      LOG(FATAL) << "Unexpected type conversion from " << input_type
1560                 << " to " << result_type;
1561  }
1562}
1563
1564void InstructionCodeGeneratorX86::VisitTypeConversion(HTypeConversion* conversion) {
1565  LocationSummary* locations = conversion->GetLocations();
1566  Location out = locations->Out();
1567  Location in = locations->InAt(0);
1568  Primitive::Type result_type = conversion->GetResultType();
1569  Primitive::Type input_type = conversion->GetInputType();
1570  DCHECK_NE(result_type, input_type);
1571  switch (result_type) {
1572    case Primitive::kPrimByte:
1573      switch (input_type) {
1574        case Primitive::kPrimBoolean:
1575          // Boolean input is a result of code transformations.
1576        case Primitive::kPrimShort:
1577        case Primitive::kPrimInt:
1578        case Primitive::kPrimChar:
1579          // Processing a Dex `int-to-byte' instruction.
1580          if (in.IsRegister()) {
1581            __ movsxb(out.AsRegister<Register>(), in.AsRegister<ByteRegister>());
1582          } else {
1583            DCHECK(in.GetConstant()->IsIntConstant());
1584            int32_t value = in.GetConstant()->AsIntConstant()->GetValue();
1585            __ movl(out.AsRegister<Register>(), Immediate(static_cast<int8_t>(value)));
1586          }
1587          break;
1588
1589        default:
1590          LOG(FATAL) << "Unexpected type conversion from " << input_type
1591                     << " to " << result_type;
1592      }
1593      break;
1594
1595    case Primitive::kPrimShort:
1596      switch (input_type) {
1597        case Primitive::kPrimBoolean:
1598          // Boolean input is a result of code transformations.
1599        case Primitive::kPrimByte:
1600        case Primitive::kPrimInt:
1601        case Primitive::kPrimChar:
1602          // Processing a Dex `int-to-short' instruction.
1603          if (in.IsRegister()) {
1604            __ movsxw(out.AsRegister<Register>(), in.AsRegister<Register>());
1605          } else if (in.IsStackSlot()) {
1606            __ movsxw(out.AsRegister<Register>(), Address(ESP, in.GetStackIndex()));
1607          } else {
1608            DCHECK(in.GetConstant()->IsIntConstant());
1609            int32_t value = in.GetConstant()->AsIntConstant()->GetValue();
1610            __ movl(out.AsRegister<Register>(), Immediate(static_cast<int16_t>(value)));
1611          }
1612          break;
1613
1614        default:
1615          LOG(FATAL) << "Unexpected type conversion from " << input_type
1616                     << " to " << result_type;
1617      }
1618      break;
1619
1620    case Primitive::kPrimInt:
1621      switch (input_type) {
1622        case Primitive::kPrimLong:
1623          // Processing a Dex `long-to-int' instruction.
1624          if (in.IsRegisterPair()) {
1625            __ movl(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>());
1626          } else if (in.IsDoubleStackSlot()) {
1627            __ movl(out.AsRegister<Register>(), Address(ESP, in.GetStackIndex()));
1628          } else {
1629            DCHECK(in.IsConstant());
1630            DCHECK(in.GetConstant()->IsLongConstant());
1631            int64_t value = in.GetConstant()->AsLongConstant()->GetValue();
1632            __ movl(out.AsRegister<Register>(), Immediate(static_cast<int32_t>(value)));
1633          }
1634          break;
1635
1636        case Primitive::kPrimFloat: {
1637          // Processing a Dex `float-to-int' instruction.
1638          XmmRegister input = in.AsFpuRegister<XmmRegister>();
1639          Register output = out.AsRegister<Register>();
1640          XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
1641          Label done, nan;
1642
1643          __ movl(output, Immediate(kPrimIntMax));
1644          // temp = int-to-float(output)
1645          __ cvtsi2ss(temp, output);
1646          // if input >= temp goto done
1647          __ comiss(input, temp);
1648          __ j(kAboveEqual, &done);
1649          // if input == NaN goto nan
1650          __ j(kUnordered, &nan);
1651          // output = float-to-int-truncate(input)
1652          __ cvttss2si(output, input);
1653          __ jmp(&done);
1654          __ Bind(&nan);
1655          //  output = 0
1656          __ xorl(output, output);
1657          __ Bind(&done);
1658          break;
1659        }
1660
1661        case Primitive::kPrimDouble: {
1662          // Processing a Dex `double-to-int' instruction.
1663          XmmRegister input = in.AsFpuRegister<XmmRegister>();
1664          Register output = out.AsRegister<Register>();
1665          XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
1666          Label done, nan;
1667
1668          __ movl(output, Immediate(kPrimIntMax));
1669          // temp = int-to-double(output)
1670          __ cvtsi2sd(temp, output);
1671          // if input >= temp goto done
1672          __ comisd(input, temp);
1673          __ j(kAboveEqual, &done);
1674          // if input == NaN goto nan
1675          __ j(kUnordered, &nan);
1676          // output = double-to-int-truncate(input)
1677          __ cvttsd2si(output, input);
1678          __ jmp(&done);
1679          __ Bind(&nan);
1680          //  output = 0
1681          __ xorl(output, output);
1682          __ Bind(&done);
1683          break;
1684        }
1685
1686        default:
1687          LOG(FATAL) << "Unexpected type conversion from " << input_type
1688                     << " to " << result_type;
1689      }
1690      break;
1691
1692    case Primitive::kPrimLong:
1693      switch (input_type) {
1694        case Primitive::kPrimBoolean:
1695          // Boolean input is a result of code transformations.
1696        case Primitive::kPrimByte:
1697        case Primitive::kPrimShort:
1698        case Primitive::kPrimInt:
1699        case Primitive::kPrimChar:
1700          // Processing a Dex `int-to-long' instruction.
1701          DCHECK_EQ(out.AsRegisterPairLow<Register>(), EAX);
1702          DCHECK_EQ(out.AsRegisterPairHigh<Register>(), EDX);
1703          DCHECK_EQ(in.AsRegister<Register>(), EAX);
1704          __ cdq();
1705          break;
1706
1707        case Primitive::kPrimFloat:
1708          // Processing a Dex `float-to-long' instruction.
1709          __ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pF2l)));
1710          codegen_->RecordPcInfo(conversion, conversion->GetDexPc());
1711          break;
1712
1713        case Primitive::kPrimDouble:
1714          // Processing a Dex `double-to-long' instruction.
1715          __ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pD2l)));
1716          codegen_->RecordPcInfo(conversion, conversion->GetDexPc());
1717          break;
1718
1719        default:
1720          LOG(FATAL) << "Unexpected type conversion from " << input_type
1721                     << " to " << result_type;
1722      }
1723      break;
1724
1725    case Primitive::kPrimChar:
1726      switch (input_type) {
1727        case Primitive::kPrimBoolean:
1728          // Boolean input is a result of code transformations.
1729        case Primitive::kPrimByte:
1730        case Primitive::kPrimShort:
1731        case Primitive::kPrimInt:
1732          // Processing a Dex `Process a Dex `int-to-char'' instruction.
1733          if (in.IsRegister()) {
1734            __ movzxw(out.AsRegister<Register>(), in.AsRegister<Register>());
1735          } else if (in.IsStackSlot()) {
1736            __ movzxw(out.AsRegister<Register>(), Address(ESP, in.GetStackIndex()));
1737          } else {
1738            DCHECK(in.GetConstant()->IsIntConstant());
1739            int32_t value = in.GetConstant()->AsIntConstant()->GetValue();
1740            __ movl(out.AsRegister<Register>(), Immediate(static_cast<uint16_t>(value)));
1741          }
1742          break;
1743
1744        default:
1745          LOG(FATAL) << "Unexpected type conversion from " << input_type
1746                     << " to " << result_type;
1747      }
1748      break;
1749
1750    case Primitive::kPrimFloat:
1751      switch (input_type) {
1752        case Primitive::kPrimBoolean:
1753          // Boolean input is a result of code transformations.
1754        case Primitive::kPrimByte:
1755        case Primitive::kPrimShort:
1756        case Primitive::kPrimInt:
1757        case Primitive::kPrimChar:
1758          // Processing a Dex `int-to-float' instruction.
1759          __ cvtsi2ss(out.AsFpuRegister<XmmRegister>(), in.AsRegister<Register>());
1760          break;
1761
1762        case Primitive::kPrimLong: {
1763          // Processing a Dex `long-to-float' instruction.
1764          Register low = in.AsRegisterPairLow<Register>();
1765          Register high = in.AsRegisterPairHigh<Register>();
1766          XmmRegister result = out.AsFpuRegister<XmmRegister>();
1767          XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
1768          XmmRegister constant = locations->GetTemp(1).AsFpuRegister<XmmRegister>();
1769
1770          // Operations use doubles for precision reasons (each 32-bit
1771          // half of a long fits in the 53-bit mantissa of a double,
1772          // but not in the 24-bit mantissa of a float).  This is
1773          // especially important for the low bits.  The result is
1774          // eventually converted to float.
1775
1776          // low = low - 2^31 (to prevent bit 31 of `low` to be
1777          // interpreted as a sign bit)
1778          __ subl(low, Immediate(0x80000000));
1779          // temp = int-to-double(high)
1780          __ cvtsi2sd(temp, high);
1781          // temp = temp * 2^32
1782          __ LoadLongConstant(constant, k2Pow32EncodingForDouble);
1783          __ mulsd(temp, constant);
1784          // result = int-to-double(low)
1785          __ cvtsi2sd(result, low);
1786          // result = result + 2^31 (restore the original value of `low`)
1787          __ LoadLongConstant(constant, k2Pow31EncodingForDouble);
1788          __ addsd(result, constant);
1789          // result = result + temp
1790          __ addsd(result, temp);
1791          // result = double-to-float(result)
1792          __ cvtsd2ss(result, result);
1793          // Restore low.
1794          __ addl(low, Immediate(0x80000000));
1795          break;
1796        }
1797
1798        case Primitive::kPrimDouble:
1799          // Processing a Dex `double-to-float' instruction.
1800          __ cvtsd2ss(out.AsFpuRegister<XmmRegister>(), in.AsFpuRegister<XmmRegister>());
1801          break;
1802
1803        default:
1804          LOG(FATAL) << "Unexpected type conversion from " << input_type
1805                     << " to " << result_type;
1806      };
1807      break;
1808
1809    case Primitive::kPrimDouble:
1810      switch (input_type) {
1811        case Primitive::kPrimBoolean:
1812          // Boolean input is a result of code transformations.
1813        case Primitive::kPrimByte:
1814        case Primitive::kPrimShort:
1815        case Primitive::kPrimInt:
1816        case Primitive::kPrimChar:
1817          // Processing a Dex `int-to-double' instruction.
1818          __ cvtsi2sd(out.AsFpuRegister<XmmRegister>(), in.AsRegister<Register>());
1819          break;
1820
1821        case Primitive::kPrimLong: {
1822          // Processing a Dex `long-to-double' instruction.
1823          Register low = in.AsRegisterPairLow<Register>();
1824          Register high = in.AsRegisterPairHigh<Register>();
1825          XmmRegister result = out.AsFpuRegister<XmmRegister>();
1826          XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
1827          XmmRegister constant = locations->GetTemp(1).AsFpuRegister<XmmRegister>();
1828
1829          // low = low - 2^31 (to prevent bit 31 of `low` to be
1830          // interpreted as a sign bit)
1831          __ subl(low, Immediate(0x80000000));
1832          // temp = int-to-double(high)
1833          __ cvtsi2sd(temp, high);
1834          // temp = temp * 2^32
1835          __ LoadLongConstant(constant, k2Pow32EncodingForDouble);
1836          __ mulsd(temp, constant);
1837          // result = int-to-double(low)
1838          __ cvtsi2sd(result, low);
1839          // result = result + 2^31 (restore the original value of `low`)
1840          __ LoadLongConstant(constant, k2Pow31EncodingForDouble);
1841          __ addsd(result, constant);
1842          // result = result + temp
1843          __ addsd(result, temp);
1844          // Restore low.
1845          __ addl(low, Immediate(0x80000000));
1846          break;
1847        }
1848
1849        case Primitive::kPrimFloat:
1850          // Processing a Dex `float-to-double' instruction.
1851          __ cvtss2sd(out.AsFpuRegister<XmmRegister>(), in.AsFpuRegister<XmmRegister>());
1852          break;
1853
1854        default:
1855          LOG(FATAL) << "Unexpected type conversion from " << input_type
1856                     << " to " << result_type;
1857      };
1858      break;
1859
1860    default:
1861      LOG(FATAL) << "Unexpected type conversion from " << input_type
1862                 << " to " << result_type;
1863  }
1864}
1865
1866void LocationsBuilderX86::VisitAdd(HAdd* add) {
1867  LocationSummary* locations =
1868      new (GetGraph()->GetArena()) LocationSummary(add, LocationSummary::kNoCall);
1869  switch (add->GetResultType()) {
1870    case Primitive::kPrimInt: {
1871      locations->SetInAt(0, Location::RequiresRegister());
1872      locations->SetInAt(1, Location::RegisterOrConstant(add->InputAt(1)));
1873      locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
1874      break;
1875    }
1876
1877    case Primitive::kPrimLong: {
1878      locations->SetInAt(0, Location::RequiresRegister());
1879      locations->SetInAt(1, Location::Any());
1880      locations->SetOut(Location::SameAsFirstInput());
1881      break;
1882    }
1883
1884    case Primitive::kPrimFloat:
1885    case Primitive::kPrimDouble: {
1886      locations->SetInAt(0, Location::RequiresFpuRegister());
1887      locations->SetInAt(1, Location::RequiresFpuRegister());
1888      locations->SetOut(Location::SameAsFirstInput());
1889      break;
1890    }
1891
1892    default:
1893      LOG(FATAL) << "Unexpected add type " << add->GetResultType();
1894      break;
1895  }
1896}
1897
1898void InstructionCodeGeneratorX86::VisitAdd(HAdd* add) {
1899  LocationSummary* locations = add->GetLocations();
1900  Location first = locations->InAt(0);
1901  Location second = locations->InAt(1);
1902  Location out = locations->Out();
1903
1904  switch (add->GetResultType()) {
1905    case Primitive::kPrimInt: {
1906      if (second.IsRegister()) {
1907        if (out.AsRegister<Register>() == first.AsRegister<Register>()) {
1908          __ addl(out.AsRegister<Register>(), second.AsRegister<Register>());
1909        } else {
1910          __ leal(out.AsRegister<Register>(), Address(
1911              first.AsRegister<Register>(), second.AsRegister<Register>(), TIMES_1, 0));
1912          }
1913      } else if (second.IsConstant()) {
1914        int32_t value = second.GetConstant()->AsIntConstant()->GetValue();
1915        if (out.AsRegister<Register>() == first.AsRegister<Register>()) {
1916          __ addl(out.AsRegister<Register>(), Immediate(value));
1917        } else {
1918          __ leal(out.AsRegister<Register>(), Address(first.AsRegister<Register>(), value));
1919        }
1920      } else {
1921        DCHECK(first.Equals(locations->Out()));
1922        __ addl(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex()));
1923      }
1924      break;
1925    }
1926
1927    case Primitive::kPrimLong: {
1928      if (second.IsRegisterPair()) {
1929        __ addl(first.AsRegisterPairLow<Register>(), second.AsRegisterPairLow<Register>());
1930        __ adcl(first.AsRegisterPairHigh<Register>(), second.AsRegisterPairHigh<Register>());
1931      } else if (second.IsDoubleStackSlot()) {
1932        __ addl(first.AsRegisterPairLow<Register>(), Address(ESP, second.GetStackIndex()));
1933        __ adcl(first.AsRegisterPairHigh<Register>(),
1934                Address(ESP, second.GetHighStackIndex(kX86WordSize)));
1935      } else {
1936        DCHECK(second.IsConstant()) << second;
1937        int64_t value = second.GetConstant()->AsLongConstant()->GetValue();
1938        __ addl(first.AsRegisterPairLow<Register>(), Immediate(Low32Bits(value)));
1939        __ adcl(first.AsRegisterPairHigh<Register>(), Immediate(High32Bits(value)));
1940      }
1941      break;
1942    }
1943
1944    case Primitive::kPrimFloat: {
1945      if (second.IsFpuRegister()) {
1946        __ addss(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
1947      }
1948      break;
1949    }
1950
1951    case Primitive::kPrimDouble: {
1952      if (second.IsFpuRegister()) {
1953        __ addsd(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
1954      }
1955      break;
1956    }
1957
1958    default:
1959      LOG(FATAL) << "Unexpected add type " << add->GetResultType();
1960  }
1961}
1962
1963void LocationsBuilderX86::VisitSub(HSub* sub) {
1964  LocationSummary* locations =
1965      new (GetGraph()->GetArena()) LocationSummary(sub, LocationSummary::kNoCall);
1966  switch (sub->GetResultType()) {
1967    case Primitive::kPrimInt:
1968    case Primitive::kPrimLong: {
1969      locations->SetInAt(0, Location::RequiresRegister());
1970      locations->SetInAt(1, Location::Any());
1971      locations->SetOut(Location::SameAsFirstInput());
1972      break;
1973    }
1974    case Primitive::kPrimFloat:
1975    case Primitive::kPrimDouble: {
1976      locations->SetInAt(0, Location::RequiresFpuRegister());
1977      locations->SetInAt(1, Location::RequiresFpuRegister());
1978      locations->SetOut(Location::SameAsFirstInput());
1979      break;
1980    }
1981
1982    default:
1983      LOG(FATAL) << "Unexpected sub type " << sub->GetResultType();
1984  }
1985}
1986
1987void InstructionCodeGeneratorX86::VisitSub(HSub* sub) {
1988  LocationSummary* locations = sub->GetLocations();
1989  Location first = locations->InAt(0);
1990  Location second = locations->InAt(1);
1991  DCHECK(first.Equals(locations->Out()));
1992  switch (sub->GetResultType()) {
1993    case Primitive::kPrimInt: {
1994      if (second.IsRegister()) {
1995        __ subl(first.AsRegister<Register>(), second.AsRegister<Register>());
1996      } else if (second.IsConstant()) {
1997        __ subl(first.AsRegister<Register>(),
1998                Immediate(second.GetConstant()->AsIntConstant()->GetValue()));
1999      } else {
2000        __ subl(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex()));
2001      }
2002      break;
2003    }
2004
2005    case Primitive::kPrimLong: {
2006      if (second.IsRegisterPair()) {
2007        __ subl(first.AsRegisterPairLow<Register>(), second.AsRegisterPairLow<Register>());
2008        __ sbbl(first.AsRegisterPairHigh<Register>(), second.AsRegisterPairHigh<Register>());
2009      } else if (second.IsDoubleStackSlot()) {
2010        __ subl(first.AsRegisterPairLow<Register>(), Address(ESP, second.GetStackIndex()));
2011        __ sbbl(first.AsRegisterPairHigh<Register>(),
2012                Address(ESP, second.GetHighStackIndex(kX86WordSize)));
2013      } else {
2014        DCHECK(second.IsConstant()) << second;
2015        int64_t value = second.GetConstant()->AsLongConstant()->GetValue();
2016        __ subl(first.AsRegisterPairLow<Register>(), Immediate(Low32Bits(value)));
2017        __ sbbl(first.AsRegisterPairHigh<Register>(), Immediate(High32Bits(value)));
2018      }
2019      break;
2020    }
2021
2022    case Primitive::kPrimFloat: {
2023      __ subss(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
2024      break;
2025    }
2026
2027    case Primitive::kPrimDouble: {
2028      __ subsd(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
2029      break;
2030    }
2031
2032    default:
2033      LOG(FATAL) << "Unexpected sub type " << sub->GetResultType();
2034  }
2035}
2036
2037void LocationsBuilderX86::VisitMul(HMul* mul) {
2038  LocationSummary* locations =
2039      new (GetGraph()->GetArena()) LocationSummary(mul, LocationSummary::kNoCall);
2040  switch (mul->GetResultType()) {
2041    case Primitive::kPrimInt:
2042      locations->SetInAt(0, Location::RequiresRegister());
2043      locations->SetInAt(1, Location::Any());
2044      locations->SetOut(Location::SameAsFirstInput());
2045      break;
2046    case Primitive::kPrimLong: {
2047      locations->SetInAt(0, Location::RequiresRegister());
2048      locations->SetInAt(1, Location::Any());
2049      locations->SetOut(Location::SameAsFirstInput());
2050      // Needed for imul on 32bits with 64bits output.
2051      locations->AddTemp(Location::RegisterLocation(EAX));
2052      locations->AddTemp(Location::RegisterLocation(EDX));
2053      break;
2054    }
2055    case Primitive::kPrimFloat:
2056    case Primitive::kPrimDouble: {
2057      locations->SetInAt(0, Location::RequiresFpuRegister());
2058      locations->SetInAt(1, Location::RequiresFpuRegister());
2059      locations->SetOut(Location::SameAsFirstInput());
2060      break;
2061    }
2062
2063    default:
2064      LOG(FATAL) << "Unexpected mul type " << mul->GetResultType();
2065  }
2066}
2067
2068void InstructionCodeGeneratorX86::VisitMul(HMul* mul) {
2069  LocationSummary* locations = mul->GetLocations();
2070  Location first = locations->InAt(0);
2071  Location second = locations->InAt(1);
2072  DCHECK(first.Equals(locations->Out()));
2073
2074  switch (mul->GetResultType()) {
2075    case Primitive::kPrimInt: {
2076      if (second.IsRegister()) {
2077        __ imull(first.AsRegister<Register>(), second.AsRegister<Register>());
2078      } else if (second.IsConstant()) {
2079        Immediate imm(second.GetConstant()->AsIntConstant()->GetValue());
2080        __ imull(first.AsRegister<Register>(), imm);
2081      } else {
2082        DCHECK(second.IsStackSlot());
2083        __ imull(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex()));
2084      }
2085      break;
2086    }
2087
2088    case Primitive::kPrimLong: {
2089      Register in1_hi = first.AsRegisterPairHigh<Register>();
2090      Register in1_lo = first.AsRegisterPairLow<Register>();
2091      Register eax = locations->GetTemp(0).AsRegister<Register>();
2092      Register edx = locations->GetTemp(1).AsRegister<Register>();
2093
2094      DCHECK_EQ(EAX, eax);
2095      DCHECK_EQ(EDX, edx);
2096
2097      // input: in1 - 64 bits, in2 - 64 bits.
2098      // output: in1
2099      // formula: in1.hi : in1.lo = (in1.lo * in2.hi + in1.hi * in2.lo)* 2^32 + in1.lo * in2.lo
2100      // parts: in1.hi = in1.lo * in2.hi + in1.hi * in2.lo + (in1.lo * in2.lo)[63:32]
2101      // parts: in1.lo = (in1.lo * in2.lo)[31:0]
2102      if (second.IsConstant()) {
2103        DCHECK(second.GetConstant()->IsLongConstant());
2104
2105        int64_t value = second.GetConstant()->AsLongConstant()->GetValue();
2106        int32_t low_value = Low32Bits(value);
2107        int32_t high_value = High32Bits(value);
2108        Immediate low(low_value);
2109        Immediate high(high_value);
2110
2111        __ movl(eax, high);
2112        // eax <- in1.lo * in2.hi
2113        __ imull(eax, in1_lo);
2114        // in1.hi <- in1.hi * in2.lo
2115        __ imull(in1_hi, low);
2116        // in1.hi <- in1.lo * in2.hi + in1.hi * in2.lo
2117        __ addl(in1_hi, eax);
2118        // move in2_lo to eax to prepare for double precision
2119        __ movl(eax, low);
2120        // edx:eax <- in1.lo * in2.lo
2121        __ mull(in1_lo);
2122        // in1.hi <- in2.hi * in1.lo +  in2.lo * in1.hi + (in1.lo * in2.lo)[63:32]
2123        __ addl(in1_hi, edx);
2124        // in1.lo <- (in1.lo * in2.lo)[31:0];
2125        __ movl(in1_lo, eax);
2126      } else if (second.IsRegisterPair()) {
2127        Register in2_hi = second.AsRegisterPairHigh<Register>();
2128        Register in2_lo = second.AsRegisterPairLow<Register>();
2129
2130        __ movl(eax, in2_hi);
2131        // eax <- in1.lo * in2.hi
2132        __ imull(eax, in1_lo);
2133        // in1.hi <- in1.hi * in2.lo
2134        __ imull(in1_hi, in2_lo);
2135        // in1.hi <- in1.lo * in2.hi + in1.hi * in2.lo
2136        __ addl(in1_hi, eax);
2137        // move in1_lo to eax to prepare for double precision
2138        __ movl(eax, in1_lo);
2139        // edx:eax <- in1.lo * in2.lo
2140        __ mull(in2_lo);
2141        // in1.hi <- in2.hi * in1.lo +  in2.lo * in1.hi + (in1.lo * in2.lo)[63:32]
2142        __ addl(in1_hi, edx);
2143        // in1.lo <- (in1.lo * in2.lo)[31:0];
2144        __ movl(in1_lo, eax);
2145      } else {
2146        DCHECK(second.IsDoubleStackSlot()) << second;
2147        Address in2_hi(ESP, second.GetHighStackIndex(kX86WordSize));
2148        Address in2_lo(ESP, second.GetStackIndex());
2149
2150        __ movl(eax, in2_hi);
2151        // eax <- in1.lo * in2.hi
2152        __ imull(eax, in1_lo);
2153        // in1.hi <- in1.hi * in2.lo
2154        __ imull(in1_hi, in2_lo);
2155        // in1.hi <- in1.lo * in2.hi + in1.hi * in2.lo
2156        __ addl(in1_hi, eax);
2157        // move in1_lo to eax to prepare for double precision
2158        __ movl(eax, in1_lo);
2159        // edx:eax <- in1.lo * in2.lo
2160        __ mull(in2_lo);
2161        // in1.hi <- in2.hi * in1.lo +  in2.lo * in1.hi + (in1.lo * in2.lo)[63:32]
2162        __ addl(in1_hi, edx);
2163        // in1.lo <- (in1.lo * in2.lo)[31:0];
2164        __ movl(in1_lo, eax);
2165      }
2166
2167      break;
2168    }
2169
2170    case Primitive::kPrimFloat: {
2171      __ mulss(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
2172      break;
2173    }
2174
2175    case Primitive::kPrimDouble: {
2176      __ mulsd(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
2177      break;
2178    }
2179
2180    default:
2181      LOG(FATAL) << "Unexpected mul type " << mul->GetResultType();
2182  }
2183}
2184
2185void InstructionCodeGeneratorX86::PushOntoFPStack(Location source, uint32_t temp_offset,
2186                                                  uint32_t stack_adjustment, bool is_float) {
2187  if (source.IsStackSlot()) {
2188    DCHECK(is_float);
2189    __ flds(Address(ESP, source.GetStackIndex() + stack_adjustment));
2190  } else if (source.IsDoubleStackSlot()) {
2191    DCHECK(!is_float);
2192    __ fldl(Address(ESP, source.GetStackIndex() + stack_adjustment));
2193  } else {
2194    // Write the value to the temporary location on the stack and load to FP stack.
2195    if (is_float) {
2196      Location stack_temp = Location::StackSlot(temp_offset);
2197      codegen_->Move32(stack_temp, source);
2198      __ flds(Address(ESP, temp_offset));
2199    } else {
2200      Location stack_temp = Location::DoubleStackSlot(temp_offset);
2201      codegen_->Move64(stack_temp, source);
2202      __ fldl(Address(ESP, temp_offset));
2203    }
2204  }
2205}
2206
2207void InstructionCodeGeneratorX86::GenerateRemFP(HRem *rem) {
2208  Primitive::Type type = rem->GetResultType();
2209  bool is_float = type == Primitive::kPrimFloat;
2210  size_t elem_size = Primitive::ComponentSize(type);
2211  LocationSummary* locations = rem->GetLocations();
2212  Location first = locations->InAt(0);
2213  Location second = locations->InAt(1);
2214  Location out = locations->Out();
2215
2216  // Create stack space for 2 elements.
2217  // TODO: enhance register allocator to ask for stack temporaries.
2218  __ subl(ESP, Immediate(2 * elem_size));
2219
2220  // Load the values to the FP stack in reverse order, using temporaries if needed.
2221  PushOntoFPStack(second, elem_size, 2 * elem_size, is_float);
2222  PushOntoFPStack(first, 0, 2 * elem_size, is_float);
2223
2224  // Loop doing FPREM until we stabilize.
2225  Label retry;
2226  __ Bind(&retry);
2227  __ fprem();
2228
2229  // Move FP status to AX.
2230  __ fstsw();
2231
2232  // And see if the argument reduction is complete. This is signaled by the
2233  // C2 FPU flag bit set to 0.
2234  __ andl(EAX, Immediate(kC2ConditionMask));
2235  __ j(kNotEqual, &retry);
2236
2237  // We have settled on the final value. Retrieve it into an XMM register.
2238  // Store FP top of stack to real stack.
2239  if (is_float) {
2240    __ fsts(Address(ESP, 0));
2241  } else {
2242    __ fstl(Address(ESP, 0));
2243  }
2244
2245  // Pop the 2 items from the FP stack.
2246  __ fucompp();
2247
2248  // Load the value from the stack into an XMM register.
2249  DCHECK(out.IsFpuRegister()) << out;
2250  if (is_float) {
2251    __ movss(out.AsFpuRegister<XmmRegister>(), Address(ESP, 0));
2252  } else {
2253    __ movsd(out.AsFpuRegister<XmmRegister>(), Address(ESP, 0));
2254  }
2255
2256  // And remove the temporary stack space we allocated.
2257  __ addl(ESP, Immediate(2 * elem_size));
2258}
2259
2260void InstructionCodeGeneratorX86::GenerateDivRemIntegral(HBinaryOperation* instruction) {
2261  DCHECK(instruction->IsDiv() || instruction->IsRem());
2262
2263  LocationSummary* locations = instruction->GetLocations();
2264  Location out = locations->Out();
2265  Location first = locations->InAt(0);
2266  Location second = locations->InAt(1);
2267  bool is_div = instruction->IsDiv();
2268
2269  switch (instruction->GetResultType()) {
2270    case Primitive::kPrimInt: {
2271      Register second_reg = second.AsRegister<Register>();
2272      DCHECK_EQ(EAX, first.AsRegister<Register>());
2273      DCHECK_EQ(is_div ? EAX : EDX, out.AsRegister<Register>());
2274
2275      SlowPathCodeX86* slow_path =
2276          new (GetGraph()->GetArena()) DivRemMinusOneSlowPathX86(out.AsRegister<Register>(),
2277                                                                 is_div);
2278      codegen_->AddSlowPath(slow_path);
2279
2280      // 0x80000000/-1 triggers an arithmetic exception!
2281      // Dividing by -1 is actually negation and -0x800000000 = 0x80000000 so
2282      // it's safe to just use negl instead of more complex comparisons.
2283
2284      __ cmpl(second_reg, Immediate(-1));
2285      __ j(kEqual, slow_path->GetEntryLabel());
2286
2287      // edx:eax <- sign-extended of eax
2288      __ cdq();
2289      // eax = quotient, edx = remainder
2290      __ idivl(second_reg);
2291
2292      __ Bind(slow_path->GetExitLabel());
2293      break;
2294    }
2295
2296    case Primitive::kPrimLong: {
2297      InvokeRuntimeCallingConvention calling_convention;
2298      DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegisterPairLow<Register>());
2299      DCHECK_EQ(calling_convention.GetRegisterAt(1), first.AsRegisterPairHigh<Register>());
2300      DCHECK_EQ(calling_convention.GetRegisterAt(2), second.AsRegisterPairLow<Register>());
2301      DCHECK_EQ(calling_convention.GetRegisterAt(3), second.AsRegisterPairHigh<Register>());
2302      DCHECK_EQ(EAX, out.AsRegisterPairLow<Register>());
2303      DCHECK_EQ(EDX, out.AsRegisterPairHigh<Register>());
2304
2305      if (is_div) {
2306        __ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pLdiv)));
2307      } else {
2308        __ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pLmod)));
2309      }
2310      uint32_t dex_pc = is_div
2311          ? instruction->AsDiv()->GetDexPc()
2312          : instruction->AsRem()->GetDexPc();
2313      codegen_->RecordPcInfo(instruction, dex_pc);
2314
2315      break;
2316    }
2317
2318    default:
2319      LOG(FATAL) << "Unexpected type for GenerateDivRemIntegral " << instruction->GetResultType();
2320  }
2321}
2322
2323void LocationsBuilderX86::VisitDiv(HDiv* div) {
2324  LocationSummary::CallKind call_kind = (div->GetResultType() == Primitive::kPrimLong)
2325      ? LocationSummary::kCall
2326      : LocationSummary::kNoCall;
2327  LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(div, call_kind);
2328
2329  switch (div->GetResultType()) {
2330    case Primitive::kPrimInt: {
2331      locations->SetInAt(0, Location::RegisterLocation(EAX));
2332      locations->SetInAt(1, Location::RequiresRegister());
2333      locations->SetOut(Location::SameAsFirstInput());
2334      // Intel uses edx:eax as the dividend.
2335      locations->AddTemp(Location::RegisterLocation(EDX));
2336      break;
2337    }
2338    case Primitive::kPrimLong: {
2339      InvokeRuntimeCallingConvention calling_convention;
2340      locations->SetInAt(0, Location::RegisterPairLocation(
2341          calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1)));
2342      locations->SetInAt(1, Location::RegisterPairLocation(
2343          calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3)));
2344      // Runtime helper puts the result in EAX, EDX.
2345      locations->SetOut(Location::RegisterPairLocation(EAX, EDX));
2346      break;
2347    }
2348    case Primitive::kPrimFloat:
2349    case Primitive::kPrimDouble: {
2350      locations->SetInAt(0, Location::RequiresFpuRegister());
2351      locations->SetInAt(1, Location::RequiresFpuRegister());
2352      locations->SetOut(Location::SameAsFirstInput());
2353      break;
2354    }
2355
2356    default:
2357      LOG(FATAL) << "Unexpected div type " << div->GetResultType();
2358  }
2359}
2360
2361void InstructionCodeGeneratorX86::VisitDiv(HDiv* div) {
2362  LocationSummary* locations = div->GetLocations();
2363  Location out = locations->Out();
2364  Location first = locations->InAt(0);
2365  Location second = locations->InAt(1);
2366
2367  switch (div->GetResultType()) {
2368    case Primitive::kPrimInt:
2369    case Primitive::kPrimLong: {
2370      GenerateDivRemIntegral(div);
2371      break;
2372    }
2373
2374    case Primitive::kPrimFloat: {
2375      DCHECK(first.Equals(out));
2376      __ divss(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
2377      break;
2378    }
2379
2380    case Primitive::kPrimDouble: {
2381      DCHECK(first.Equals(out));
2382      __ divsd(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>());
2383      break;
2384    }
2385
2386    default:
2387      LOG(FATAL) << "Unexpected div type " << div->GetResultType();
2388  }
2389}
2390
2391void LocationsBuilderX86::VisitRem(HRem* rem) {
2392  Primitive::Type type = rem->GetResultType();
2393  LocationSummary::CallKind call_kind = (rem->GetResultType() == Primitive::kPrimLong)
2394      ? LocationSummary::kCall
2395      : LocationSummary::kNoCall;
2396  LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(rem, call_kind);
2397
2398  switch (type) {
2399    case Primitive::kPrimInt: {
2400      locations->SetInAt(0, Location::RegisterLocation(EAX));
2401      locations->SetInAt(1, Location::RequiresRegister());
2402      locations->SetOut(Location::RegisterLocation(EDX));
2403      break;
2404    }
2405    case Primitive::kPrimLong: {
2406      InvokeRuntimeCallingConvention calling_convention;
2407      locations->SetInAt(0, Location::RegisterPairLocation(
2408          calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1)));
2409      locations->SetInAt(1, Location::RegisterPairLocation(
2410          calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3)));
2411      // Runtime helper puts the result in EAX, EDX.
2412      locations->SetOut(Location::RegisterPairLocation(EAX, EDX));
2413      break;
2414    }
2415    case Primitive::kPrimDouble:
2416    case Primitive::kPrimFloat: {
2417      locations->SetInAt(0, Location::Any());
2418      locations->SetInAt(1, Location::Any());
2419      locations->SetOut(Location::RequiresFpuRegister());
2420      locations->AddTemp(Location::RegisterLocation(EAX));
2421      break;
2422    }
2423
2424    default:
2425      LOG(FATAL) << "Unexpected rem type " << type;
2426  }
2427}
2428
2429void InstructionCodeGeneratorX86::VisitRem(HRem* rem) {
2430  Primitive::Type type = rem->GetResultType();
2431  switch (type) {
2432    case Primitive::kPrimInt:
2433    case Primitive::kPrimLong: {
2434      GenerateDivRemIntegral(rem);
2435      break;
2436    }
2437    case Primitive::kPrimFloat:
2438    case Primitive::kPrimDouble: {
2439      GenerateRemFP(rem);
2440      break;
2441    }
2442    default:
2443      LOG(FATAL) << "Unexpected rem type " << type;
2444  }
2445}
2446
2447void LocationsBuilderX86::VisitDivZeroCheck(HDivZeroCheck* instruction) {
2448  LocationSummary* locations =
2449      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall);
2450  switch (instruction->GetType()) {
2451    case Primitive::kPrimInt: {
2452      locations->SetInAt(0, Location::Any());
2453      break;
2454    }
2455    case Primitive::kPrimLong: {
2456      locations->SetInAt(0, Location::RegisterOrConstant(instruction->InputAt(0)));
2457      if (!instruction->IsConstant()) {
2458        locations->AddTemp(Location::RequiresRegister());
2459      }
2460      break;
2461    }
2462    default:
2463      LOG(FATAL) << "Unexpected type for HDivZeroCheck " << instruction->GetType();
2464  }
2465  if (instruction->HasUses()) {
2466    locations->SetOut(Location::SameAsFirstInput());
2467  }
2468}
2469
2470void InstructionCodeGeneratorX86::VisitDivZeroCheck(HDivZeroCheck* instruction) {
2471  SlowPathCodeX86* slow_path = new (GetGraph()->GetArena()) DivZeroCheckSlowPathX86(instruction);
2472  codegen_->AddSlowPath(slow_path);
2473
2474  LocationSummary* locations = instruction->GetLocations();
2475  Location value = locations->InAt(0);
2476
2477  switch (instruction->GetType()) {
2478    case Primitive::kPrimInt: {
2479      if (value.IsRegister()) {
2480        __ testl(value.AsRegister<Register>(), value.AsRegister<Register>());
2481        __ j(kEqual, slow_path->GetEntryLabel());
2482      } else if (value.IsStackSlot()) {
2483        __ cmpl(Address(ESP, value.GetStackIndex()), Immediate(0));
2484        __ j(kEqual, slow_path->GetEntryLabel());
2485      } else {
2486        DCHECK(value.IsConstant()) << value;
2487        if (value.GetConstant()->AsIntConstant()->GetValue() == 0) {
2488        __ jmp(slow_path->GetEntryLabel());
2489        }
2490      }
2491      break;
2492    }
2493    case Primitive::kPrimLong: {
2494      if (value.IsRegisterPair()) {
2495        Register temp = locations->GetTemp(0).AsRegister<Register>();
2496        __ movl(temp, value.AsRegisterPairLow<Register>());
2497        __ orl(temp, value.AsRegisterPairHigh<Register>());
2498        __ j(kEqual, slow_path->GetEntryLabel());
2499      } else {
2500        DCHECK(value.IsConstant()) << value;
2501        if (value.GetConstant()->AsLongConstant()->GetValue() == 0) {
2502          __ jmp(slow_path->GetEntryLabel());
2503        }
2504      }
2505      break;
2506    }
2507    default:
2508      LOG(FATAL) << "Unexpected type for HDivZeroCheck" << instruction->GetType();
2509  }
2510}
2511
2512void LocationsBuilderX86::HandleShift(HBinaryOperation* op) {
2513  DCHECK(op->IsShl() || op->IsShr() || op->IsUShr());
2514
2515  LocationSummary* locations =
2516      new (GetGraph()->GetArena()) LocationSummary(op, LocationSummary::kNoCall);
2517
2518  switch (op->GetResultType()) {
2519    case Primitive::kPrimInt: {
2520      locations->SetInAt(0, Location::RequiresRegister());
2521      // The shift count needs to be in CL.
2522      locations->SetInAt(1, Location::ByteRegisterOrConstant(ECX, op->InputAt(1)));
2523      locations->SetOut(Location::SameAsFirstInput());
2524      break;
2525    }
2526    case Primitive::kPrimLong: {
2527      locations->SetInAt(0, Location::RequiresRegister());
2528      // The shift count needs to be in CL.
2529      locations->SetInAt(1, Location::RegisterLocation(ECX));
2530      locations->SetOut(Location::SameAsFirstInput());
2531      break;
2532    }
2533    default:
2534      LOG(FATAL) << "Unexpected op type " << op->GetResultType();
2535  }
2536}
2537
2538void InstructionCodeGeneratorX86::HandleShift(HBinaryOperation* op) {
2539  DCHECK(op->IsShl() || op->IsShr() || op->IsUShr());
2540
2541  LocationSummary* locations = op->GetLocations();
2542  Location first = locations->InAt(0);
2543  Location second = locations->InAt(1);
2544  DCHECK(first.Equals(locations->Out()));
2545
2546  switch (op->GetResultType()) {
2547    case Primitive::kPrimInt: {
2548      Register first_reg = first.AsRegister<Register>();
2549      if (second.IsRegister()) {
2550        Register second_reg = second.AsRegister<Register>();
2551        DCHECK_EQ(ECX, second_reg);
2552        if (op->IsShl()) {
2553          __ shll(first_reg, second_reg);
2554        } else if (op->IsShr()) {
2555          __ sarl(first_reg, second_reg);
2556        } else {
2557          __ shrl(first_reg, second_reg);
2558        }
2559      } else {
2560        Immediate imm(second.GetConstant()->AsIntConstant()->GetValue() & kMaxIntShiftValue);
2561        if (op->IsShl()) {
2562          __ shll(first_reg, imm);
2563        } else if (op->IsShr()) {
2564          __ sarl(first_reg, imm);
2565        } else {
2566          __ shrl(first_reg, imm);
2567        }
2568      }
2569      break;
2570    }
2571    case Primitive::kPrimLong: {
2572      Register second_reg = second.AsRegister<Register>();
2573      DCHECK_EQ(ECX, second_reg);
2574      if (op->IsShl()) {
2575        GenerateShlLong(first, second_reg);
2576      } else if (op->IsShr()) {
2577        GenerateShrLong(first, second_reg);
2578      } else {
2579        GenerateUShrLong(first, second_reg);
2580      }
2581      break;
2582    }
2583    default:
2584      LOG(FATAL) << "Unexpected op type " << op->GetResultType();
2585  }
2586}
2587
2588void InstructionCodeGeneratorX86::GenerateShlLong(const Location& loc, Register shifter) {
2589  Label done;
2590  __ shld(loc.AsRegisterPairHigh<Register>(), loc.AsRegisterPairLow<Register>(), shifter);
2591  __ shll(loc.AsRegisterPairLow<Register>(), shifter);
2592  __ testl(shifter, Immediate(32));
2593  __ j(kEqual, &done);
2594  __ movl(loc.AsRegisterPairHigh<Register>(), loc.AsRegisterPairLow<Register>());
2595  __ movl(loc.AsRegisterPairLow<Register>(), Immediate(0));
2596  __ Bind(&done);
2597}
2598
2599void InstructionCodeGeneratorX86::GenerateShrLong(const Location& loc, Register shifter) {
2600  Label done;
2601  __ shrd(loc.AsRegisterPairLow<Register>(), loc.AsRegisterPairHigh<Register>(), shifter);
2602  __ sarl(loc.AsRegisterPairHigh<Register>(), shifter);
2603  __ testl(shifter, Immediate(32));
2604  __ j(kEqual, &done);
2605  __ movl(loc.AsRegisterPairLow<Register>(), loc.AsRegisterPairHigh<Register>());
2606  __ sarl(loc.AsRegisterPairHigh<Register>(), Immediate(31));
2607  __ Bind(&done);
2608}
2609
2610void InstructionCodeGeneratorX86::GenerateUShrLong(const Location& loc, Register shifter) {
2611  Label done;
2612  __ shrd(loc.AsRegisterPairLow<Register>(), loc.AsRegisterPairHigh<Register>(), shifter);
2613  __ shrl(loc.AsRegisterPairHigh<Register>(), shifter);
2614  __ testl(shifter, Immediate(32));
2615  __ j(kEqual, &done);
2616  __ movl(loc.AsRegisterPairLow<Register>(), loc.AsRegisterPairHigh<Register>());
2617  __ movl(loc.AsRegisterPairHigh<Register>(), Immediate(0));
2618  __ Bind(&done);
2619}
2620
2621void LocationsBuilderX86::VisitShl(HShl* shl) {
2622  HandleShift(shl);
2623}
2624
2625void InstructionCodeGeneratorX86::VisitShl(HShl* shl) {
2626  HandleShift(shl);
2627}
2628
2629void LocationsBuilderX86::VisitShr(HShr* shr) {
2630  HandleShift(shr);
2631}
2632
2633void InstructionCodeGeneratorX86::VisitShr(HShr* shr) {
2634  HandleShift(shr);
2635}
2636
2637void LocationsBuilderX86::VisitUShr(HUShr* ushr) {
2638  HandleShift(ushr);
2639}
2640
2641void InstructionCodeGeneratorX86::VisitUShr(HUShr* ushr) {
2642  HandleShift(ushr);
2643}
2644
2645void LocationsBuilderX86::VisitNewInstance(HNewInstance* instruction) {
2646  LocationSummary* locations =
2647      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall);
2648  locations->SetOut(Location::RegisterLocation(EAX));
2649  InvokeRuntimeCallingConvention calling_convention;
2650  locations->AddTemp(Location::RegisterLocation(calling_convention.GetRegisterAt(0)));
2651  locations->AddTemp(Location::RegisterLocation(calling_convention.GetRegisterAt(1)));
2652}
2653
2654void InstructionCodeGeneratorX86::VisitNewInstance(HNewInstance* instruction) {
2655  InvokeRuntimeCallingConvention calling_convention;
2656  codegen_->LoadCurrentMethod(calling_convention.GetRegisterAt(1));
2657  __ movl(calling_convention.GetRegisterAt(0), Immediate(instruction->GetTypeIndex()));
2658
2659  __ fs()->call(Address::Absolute(GetThreadOffset<kX86WordSize>(instruction->GetEntrypoint())));
2660
2661  codegen_->RecordPcInfo(instruction, instruction->GetDexPc());
2662  DCHECK(!codegen_->IsLeafMethod());
2663}
2664
2665void LocationsBuilderX86::VisitNewArray(HNewArray* instruction) {
2666  LocationSummary* locations =
2667      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall);
2668  locations->SetOut(Location::RegisterLocation(EAX));
2669  InvokeRuntimeCallingConvention calling_convention;
2670  locations->AddTemp(Location::RegisterLocation(calling_convention.GetRegisterAt(0)));
2671  locations->AddTemp(Location::RegisterLocation(calling_convention.GetRegisterAt(2)));
2672  locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(1)));
2673}
2674
2675void InstructionCodeGeneratorX86::VisitNewArray(HNewArray* instruction) {
2676  InvokeRuntimeCallingConvention calling_convention;
2677  codegen_->LoadCurrentMethod(calling_convention.GetRegisterAt(2));
2678  __ movl(calling_convention.GetRegisterAt(0), Immediate(instruction->GetTypeIndex()));
2679
2680  __ fs()->call(Address::Absolute(GetThreadOffset<kX86WordSize>(instruction->GetEntrypoint())));
2681
2682  codegen_->RecordPcInfo(instruction, instruction->GetDexPc());
2683  DCHECK(!codegen_->IsLeafMethod());
2684}
2685
2686void LocationsBuilderX86::VisitParameterValue(HParameterValue* instruction) {
2687  LocationSummary* locations =
2688      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall);
2689  Location location = parameter_visitor_.GetNextLocation(instruction->GetType());
2690  if (location.IsStackSlot()) {
2691    location = Location::StackSlot(location.GetStackIndex() + codegen_->GetFrameSize());
2692  } else if (location.IsDoubleStackSlot()) {
2693    location = Location::DoubleStackSlot(location.GetStackIndex() + codegen_->GetFrameSize());
2694  }
2695  locations->SetOut(location);
2696}
2697
2698void InstructionCodeGeneratorX86::VisitParameterValue(HParameterValue* instruction) {
2699  UNUSED(instruction);
2700}
2701
2702void LocationsBuilderX86::VisitNot(HNot* not_) {
2703  LocationSummary* locations =
2704      new (GetGraph()->GetArena()) LocationSummary(not_, LocationSummary::kNoCall);
2705  locations->SetInAt(0, Location::RequiresRegister());
2706  locations->SetOut(Location::SameAsFirstInput());
2707}
2708
2709void InstructionCodeGeneratorX86::VisitNot(HNot* not_) {
2710  LocationSummary* locations = not_->GetLocations();
2711  Location in = locations->InAt(0);
2712  Location out = locations->Out();
2713  DCHECK(in.Equals(out));
2714  switch (not_->GetResultType()) {
2715    case Primitive::kPrimInt:
2716      __ notl(out.AsRegister<Register>());
2717      break;
2718
2719    case Primitive::kPrimLong:
2720      __ notl(out.AsRegisterPairLow<Register>());
2721      __ notl(out.AsRegisterPairHigh<Register>());
2722      break;
2723
2724    default:
2725      LOG(FATAL) << "Unimplemented type for not operation " << not_->GetResultType();
2726  }
2727}
2728
2729void LocationsBuilderX86::VisitCompare(HCompare* compare) {
2730  LocationSummary* locations =
2731      new (GetGraph()->GetArena()) LocationSummary(compare, LocationSummary::kNoCall);
2732  switch (compare->InputAt(0)->GetType()) {
2733    case Primitive::kPrimLong: {
2734      locations->SetInAt(0, Location::RequiresRegister());
2735      locations->SetInAt(1, Location::Any());
2736      locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
2737      break;
2738    }
2739    case Primitive::kPrimFloat:
2740    case Primitive::kPrimDouble: {
2741      locations->SetInAt(0, Location::RequiresFpuRegister());
2742      locations->SetInAt(1, Location::RequiresFpuRegister());
2743      locations->SetOut(Location::RequiresRegister());
2744      break;
2745    }
2746    default:
2747      LOG(FATAL) << "Unexpected type for compare operation " << compare->InputAt(0)->GetType();
2748  }
2749}
2750
2751void InstructionCodeGeneratorX86::VisitCompare(HCompare* compare) {
2752  LocationSummary* locations = compare->GetLocations();
2753  Register out = locations->Out().AsRegister<Register>();
2754  Location left = locations->InAt(0);
2755  Location right = locations->InAt(1);
2756
2757  Label less, greater, done;
2758  switch (compare->InputAt(0)->GetType()) {
2759    case Primitive::kPrimLong: {
2760      Register left_low = left.AsRegisterPairLow<Register>();
2761      Register left_high = left.AsRegisterPairHigh<Register>();
2762      int32_t val_low = 0;
2763      int32_t val_high = 0;
2764      bool right_is_const = false;
2765
2766      if (right.IsConstant()) {
2767        DCHECK(right.GetConstant()->IsLongConstant());
2768        right_is_const = true;
2769        int64_t val = right.GetConstant()->AsLongConstant()->GetValue();
2770        val_low = Low32Bits(val);
2771        val_high = High32Bits(val);
2772      }
2773
2774      if (right.IsRegisterPair()) {
2775        __ cmpl(left_high, right.AsRegisterPairHigh<Register>());
2776      } else if (right.IsDoubleStackSlot()) {
2777        __ cmpl(left_high, Address(ESP, right.GetHighStackIndex(kX86WordSize)));
2778      } else {
2779        DCHECK(right_is_const) << right;
2780        if (val_high == 0) {
2781          __ testl(left_high, left_high);
2782        } else {
2783          __ cmpl(left_high, Immediate(val_high));
2784        }
2785      }
2786      __ j(kLess, &less);  // Signed compare.
2787      __ j(kGreater, &greater);  // Signed compare.
2788      if (right.IsRegisterPair()) {
2789        __ cmpl(left_low, right.AsRegisterPairLow<Register>());
2790      } else if (right.IsDoubleStackSlot()) {
2791        __ cmpl(left_low, Address(ESP, right.GetStackIndex()));
2792      } else {
2793        DCHECK(right_is_const) << right;
2794        if (val_low == 0) {
2795          __ testl(left_low, left_low);
2796        } else {
2797          __ cmpl(left_low, Immediate(val_low));
2798        }
2799      }
2800      break;
2801    }
2802    case Primitive::kPrimFloat: {
2803      __ ucomiss(left.AsFpuRegister<XmmRegister>(), right.AsFpuRegister<XmmRegister>());
2804      __ j(kUnordered, compare->IsGtBias() ? &greater : &less);
2805      break;
2806    }
2807    case Primitive::kPrimDouble: {
2808      __ ucomisd(left.AsFpuRegister<XmmRegister>(), right.AsFpuRegister<XmmRegister>());
2809      __ j(kUnordered, compare->IsGtBias() ? &greater : &less);
2810      break;
2811    }
2812    default:
2813      LOG(FATAL) << "Unexpected type for compare operation " << compare->InputAt(0)->GetType();
2814  }
2815  __ movl(out, Immediate(0));
2816  __ j(kEqual, &done);
2817  __ j(kBelow, &less);  // kBelow is for CF (unsigned & floats).
2818
2819  __ Bind(&greater);
2820  __ movl(out, Immediate(1));
2821  __ jmp(&done);
2822
2823  __ Bind(&less);
2824  __ movl(out, Immediate(-1));
2825
2826  __ Bind(&done);
2827}
2828
2829void LocationsBuilderX86::VisitPhi(HPhi* instruction) {
2830  LocationSummary* locations =
2831      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall);
2832  for (size_t i = 0, e = instruction->InputCount(); i < e; ++i) {
2833    locations->SetInAt(i, Location::Any());
2834  }
2835  locations->SetOut(Location::Any());
2836}
2837
2838void InstructionCodeGeneratorX86::VisitPhi(HPhi* instruction) {
2839  UNUSED(instruction);
2840  LOG(FATAL) << "Unreachable";
2841}
2842
2843void InstructionCodeGeneratorX86::GenerateMemoryBarrier(MemBarrierKind kind) {
2844  /*
2845   * According to the JSR-133 Cookbook, for x86 only StoreLoad/AnyAny barriers need memory fence.
2846   * All other barriers (LoadAny, AnyStore, StoreStore) are nops due to the x86 memory model.
2847   * For those cases, all we need to ensure is that there is a scheduling barrier in place.
2848   */
2849  switch (kind) {
2850    case MemBarrierKind::kAnyAny: {
2851      __ mfence();
2852      break;
2853    }
2854    case MemBarrierKind::kAnyStore:
2855    case MemBarrierKind::kLoadAny:
2856    case MemBarrierKind::kStoreStore: {
2857      // nop
2858      break;
2859    }
2860    default:
2861      LOG(FATAL) << "Unexpected memory barrier " << kind;
2862  }
2863}
2864
2865
2866void CodeGeneratorX86::MarkGCCard(Register temp, Register card, Register object, Register value) {
2867  Label is_null;
2868  __ testl(value, value);
2869  __ j(kEqual, &is_null);
2870  __ fs()->movl(card, Address::Absolute(Thread::CardTableOffset<kX86WordSize>().Int32Value()));
2871  __ movl(temp, object);
2872  __ shrl(temp, Immediate(gc::accounting::CardTable::kCardShift));
2873  __ movb(Address(temp, card, TIMES_1, 0),
2874          X86ManagedRegister::FromCpuRegister(card).AsByteRegister());
2875  __ Bind(&is_null);
2876}
2877
2878void LocationsBuilderX86::HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info) {
2879  DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet());
2880  LocationSummary* locations =
2881      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall);
2882  locations->SetInAt(0, Location::RequiresRegister());
2883
2884  // The output overlaps in case of long: we don't want the low move to overwrite
2885  // the object's location.
2886  locations->SetOut(Location::RequiresRegister(),
2887      (instruction->GetType() == Primitive::kPrimLong) ? Location::kOutputOverlap
2888                                                       : Location::kNoOutputOverlap);
2889
2890  if (field_info.IsVolatile() && (field_info.GetFieldType() == Primitive::kPrimLong)) {
2891    // Long values can be loaded atomically into an XMM using movsd.
2892    // So we use an XMM register as a temp to achieve atomicity (first load the temp into the XMM
2893    // and then copy the XMM into the output 32bits at a time).
2894    locations->AddTemp(Location::RequiresFpuRegister());
2895  }
2896}
2897
2898void InstructionCodeGeneratorX86::HandleFieldGet(HInstruction* instruction,
2899                                                 const FieldInfo& field_info) {
2900  DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet());
2901
2902  LocationSummary* locations = instruction->GetLocations();
2903  Register base = locations->InAt(0).AsRegister<Register>();
2904  Location out = locations->Out();
2905  bool is_volatile = field_info.IsVolatile();
2906  Primitive::Type field_type = field_info.GetFieldType();
2907  uint32_t offset = field_info.GetFieldOffset().Uint32Value();
2908
2909  switch (field_type) {
2910    case Primitive::kPrimBoolean: {
2911      __ movzxb(out.AsRegister<Register>(), Address(base, offset));
2912      break;
2913    }
2914
2915    case Primitive::kPrimByte: {
2916      __ movsxb(out.AsRegister<Register>(), Address(base, offset));
2917      break;
2918    }
2919
2920    case Primitive::kPrimShort: {
2921      __ movsxw(out.AsRegister<Register>(), Address(base, offset));
2922      break;
2923    }
2924
2925    case Primitive::kPrimChar: {
2926      __ movzxw(out.AsRegister<Register>(), Address(base, offset));
2927      break;
2928    }
2929
2930    case Primitive::kPrimInt:
2931    case Primitive::kPrimNot: {
2932      __ movl(out.AsRegister<Register>(), Address(base, offset));
2933      break;
2934    }
2935
2936    case Primitive::kPrimLong: {
2937      if (is_volatile) {
2938        XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
2939        __ movsd(temp, Address(base, offset));
2940        codegen_->MaybeRecordImplicitNullCheck(instruction);
2941        __ movd(out.AsRegisterPairLow<Register>(), temp);
2942        __ psrlq(temp, Immediate(32));
2943        __ movd(out.AsRegisterPairHigh<Register>(), temp);
2944      } else {
2945        DCHECK_NE(base, out.AsRegisterPairLow<Register>());
2946        __ movl(out.AsRegisterPairLow<Register>(), Address(base, offset));
2947        codegen_->MaybeRecordImplicitNullCheck(instruction);
2948        __ movl(out.AsRegisterPairHigh<Register>(), Address(base, kX86WordSize + offset));
2949      }
2950      break;
2951    }
2952
2953    case Primitive::kPrimFloat: {
2954      __ movss(out.AsFpuRegister<XmmRegister>(), Address(base, offset));
2955      break;
2956    }
2957
2958    case Primitive::kPrimDouble: {
2959      __ movsd(out.AsFpuRegister<XmmRegister>(), Address(base, offset));
2960      break;
2961    }
2962
2963    case Primitive::kPrimVoid:
2964      LOG(FATAL) << "Unreachable type " << field_type;
2965      UNREACHABLE();
2966  }
2967
2968  // Longs are handled in the switch.
2969  if (field_type != Primitive::kPrimLong) {
2970    codegen_->MaybeRecordImplicitNullCheck(instruction);
2971  }
2972
2973  if (is_volatile) {
2974    GenerateMemoryBarrier(MemBarrierKind::kLoadAny);
2975  }
2976}
2977
2978void LocationsBuilderX86::HandleFieldSet(HInstruction* instruction, const FieldInfo& field_info) {
2979  DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet());
2980
2981  LocationSummary* locations =
2982      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall);
2983  locations->SetInAt(0, Location::RequiresRegister());
2984  bool is_volatile = field_info.IsVolatile();
2985  Primitive::Type field_type = field_info.GetFieldType();
2986  bool is_byte_type = (field_type == Primitive::kPrimBoolean)
2987    || (field_type == Primitive::kPrimByte);
2988
2989  // The register allocator does not support multiple
2990  // inputs that die at entry with one in a specific register.
2991  if (is_byte_type) {
2992    // Ensure the value is in a byte register.
2993    locations->SetInAt(1, Location::RegisterLocation(EAX));
2994  } else {
2995    locations->SetInAt(1, Location::RequiresRegister());
2996  }
2997  // Temporary registers for the write barrier.
2998  if (CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1))) {
2999    locations->AddTemp(Location::RequiresRegister());
3000    // Ensure the card is in a byte register.
3001    locations->AddTemp(Location::RegisterLocation(ECX));
3002  } else if (is_volatile && (field_type == Primitive::kPrimLong)) {
3003    // 64bits value can be atomically written to an address with movsd and an XMM register.
3004    // We need two XMM registers because there's no easier way to (bit) copy a register pair
3005    // into a single XMM register (we copy each pair part into the XMMs and then interleave them).
3006    // NB: We could make the register allocator understand fp_reg <-> core_reg moves but given the
3007    // isolated cases when we need this it isn't worth adding the extra complexity.
3008    locations->AddTemp(Location::RequiresFpuRegister());
3009    locations->AddTemp(Location::RequiresFpuRegister());
3010  }
3011}
3012
3013void InstructionCodeGeneratorX86::HandleFieldSet(HInstruction* instruction,
3014                                                 const FieldInfo& field_info) {
3015  DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet());
3016
3017  LocationSummary* locations = instruction->GetLocations();
3018  Register base = locations->InAt(0).AsRegister<Register>();
3019  Location value = locations->InAt(1);
3020  bool is_volatile = field_info.IsVolatile();
3021  Primitive::Type field_type = field_info.GetFieldType();
3022  uint32_t offset = field_info.GetFieldOffset().Uint32Value();
3023
3024  if (is_volatile) {
3025    GenerateMemoryBarrier(MemBarrierKind::kAnyStore);
3026  }
3027
3028  switch (field_type) {
3029    case Primitive::kPrimBoolean:
3030    case Primitive::kPrimByte: {
3031      __ movb(Address(base, offset), value.AsRegister<ByteRegister>());
3032      break;
3033    }
3034
3035    case Primitive::kPrimShort:
3036    case Primitive::kPrimChar: {
3037      __ movw(Address(base, offset), value.AsRegister<Register>());
3038      break;
3039    }
3040
3041    case Primitive::kPrimInt:
3042    case Primitive::kPrimNot: {
3043      __ movl(Address(base, offset), value.AsRegister<Register>());
3044      break;
3045    }
3046
3047    case Primitive::kPrimLong: {
3048      if (is_volatile) {
3049        XmmRegister temp1 = locations->GetTemp(0).AsFpuRegister<XmmRegister>();
3050        XmmRegister temp2 = locations->GetTemp(1).AsFpuRegister<XmmRegister>();
3051        __ movd(temp1, value.AsRegisterPairLow<Register>());
3052        __ movd(temp2, value.AsRegisterPairHigh<Register>());
3053        __ punpckldq(temp1, temp2);
3054        __ movsd(Address(base, offset), temp1);
3055        codegen_->MaybeRecordImplicitNullCheck(instruction);
3056      } else {
3057        __ movl(Address(base, offset), value.AsRegisterPairLow<Register>());
3058        codegen_->MaybeRecordImplicitNullCheck(instruction);
3059        __ movl(Address(base, kX86WordSize + offset), value.AsRegisterPairHigh<Register>());
3060      }
3061      break;
3062    }
3063
3064    case Primitive::kPrimFloat: {
3065      __ movss(Address(base, offset), value.AsFpuRegister<XmmRegister>());
3066      break;
3067    }
3068
3069    case Primitive::kPrimDouble: {
3070      __ movsd(Address(base, offset), value.AsFpuRegister<XmmRegister>());
3071      break;
3072    }
3073
3074    case Primitive::kPrimVoid:
3075      LOG(FATAL) << "Unreachable type " << field_type;
3076      UNREACHABLE();
3077  }
3078
3079  // Longs are handled in the switch.
3080  if (field_type != Primitive::kPrimLong) {
3081    codegen_->MaybeRecordImplicitNullCheck(instruction);
3082  }
3083
3084  if (CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1))) {
3085    Register temp = locations->GetTemp(0).AsRegister<Register>();
3086    Register card = locations->GetTemp(1).AsRegister<Register>();
3087    codegen_->MarkGCCard(temp, card, base, value.AsRegister<Register>());
3088  }
3089
3090  if (is_volatile) {
3091    GenerateMemoryBarrier(MemBarrierKind::kAnyAny);
3092  }
3093}
3094
3095void LocationsBuilderX86::VisitStaticFieldGet(HStaticFieldGet* instruction) {
3096  HandleFieldGet(instruction, instruction->GetFieldInfo());
3097}
3098
3099void InstructionCodeGeneratorX86::VisitStaticFieldGet(HStaticFieldGet* instruction) {
3100  HandleFieldGet(instruction, instruction->GetFieldInfo());
3101}
3102
3103void LocationsBuilderX86::VisitStaticFieldSet(HStaticFieldSet* instruction) {
3104  HandleFieldSet(instruction, instruction->GetFieldInfo());
3105}
3106
3107void InstructionCodeGeneratorX86::VisitStaticFieldSet(HStaticFieldSet* instruction) {
3108  HandleFieldSet(instruction, instruction->GetFieldInfo());
3109}
3110
3111void LocationsBuilderX86::VisitInstanceFieldSet(HInstanceFieldSet* instruction) {
3112  HandleFieldSet(instruction, instruction->GetFieldInfo());
3113}
3114
3115void InstructionCodeGeneratorX86::VisitInstanceFieldSet(HInstanceFieldSet* instruction) {
3116  HandleFieldSet(instruction, instruction->GetFieldInfo());
3117}
3118
3119void LocationsBuilderX86::VisitInstanceFieldGet(HInstanceFieldGet* instruction) {
3120  HandleFieldGet(instruction, instruction->GetFieldInfo());
3121}
3122
3123void InstructionCodeGeneratorX86::VisitInstanceFieldGet(HInstanceFieldGet* instruction) {
3124  HandleFieldGet(instruction, instruction->GetFieldInfo());
3125}
3126
3127void LocationsBuilderX86::VisitNullCheck(HNullCheck* instruction) {
3128  LocationSummary* locations =
3129      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall);
3130  Location loc = codegen_->GetCompilerOptions().GetImplicitNullChecks()
3131      ? Location::RequiresRegister()
3132      : Location::Any();
3133  locations->SetInAt(0, loc);
3134  if (instruction->HasUses()) {
3135    locations->SetOut(Location::SameAsFirstInput());
3136  }
3137}
3138
3139void InstructionCodeGeneratorX86::GenerateImplicitNullCheck(HNullCheck* instruction) {
3140  if (codegen_->CanMoveNullCheckToUser(instruction)) {
3141    return;
3142  }
3143  LocationSummary* locations = instruction->GetLocations();
3144  Location obj = locations->InAt(0);
3145
3146  __ testl(EAX, Address(obj.AsRegister<Register>(), 0));
3147  codegen_->RecordPcInfo(instruction, instruction->GetDexPc());
3148}
3149
3150void InstructionCodeGeneratorX86::GenerateExplicitNullCheck(HNullCheck* instruction) {
3151  SlowPathCodeX86* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathX86(instruction);
3152  codegen_->AddSlowPath(slow_path);
3153
3154  LocationSummary* locations = instruction->GetLocations();
3155  Location obj = locations->InAt(0);
3156
3157  if (obj.IsRegister()) {
3158    __ cmpl(obj.AsRegister<Register>(), Immediate(0));
3159  } else if (obj.IsStackSlot()) {
3160    __ cmpl(Address(ESP, obj.GetStackIndex()), Immediate(0));
3161  } else {
3162    DCHECK(obj.IsConstant()) << obj;
3163    DCHECK_EQ(obj.GetConstant()->AsIntConstant()->GetValue(), 0);
3164    __ jmp(slow_path->GetEntryLabel());
3165    return;
3166  }
3167  __ j(kEqual, slow_path->GetEntryLabel());
3168}
3169
3170void InstructionCodeGeneratorX86::VisitNullCheck(HNullCheck* instruction) {
3171  if (codegen_->GetCompilerOptions().GetImplicitNullChecks()) {
3172    GenerateImplicitNullCheck(instruction);
3173  } else {
3174    GenerateExplicitNullCheck(instruction);
3175  }
3176}
3177
3178void LocationsBuilderX86::VisitArrayGet(HArrayGet* instruction) {
3179  LocationSummary* locations =
3180      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall);
3181  locations->SetInAt(0, Location::RequiresRegister());
3182  locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1)));
3183  // The output overlaps in case of long: we don't want the low move to overwrite
3184  // the array's location.
3185  locations->SetOut(Location::RequiresRegister(),
3186      (instruction->GetType() == Primitive::kPrimLong) ? Location::kOutputOverlap
3187                                                       : Location::kNoOutputOverlap);
3188}
3189
3190void InstructionCodeGeneratorX86::VisitArrayGet(HArrayGet* instruction) {
3191  LocationSummary* locations = instruction->GetLocations();
3192  Register obj = locations->InAt(0).AsRegister<Register>();
3193  Location index = locations->InAt(1);
3194
3195  Primitive::Type type = instruction->GetType();
3196  switch (type) {
3197    case Primitive::kPrimBoolean: {
3198      uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint8_t)).Uint32Value();
3199      Register out = locations->Out().AsRegister<Register>();
3200      if (index.IsConstant()) {
3201        __ movzxb(out, Address(obj,
3202            (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_1) + data_offset));
3203      } else {
3204        __ movzxb(out, Address(obj, index.AsRegister<Register>(), TIMES_1, data_offset));
3205      }
3206      break;
3207    }
3208
3209    case Primitive::kPrimByte: {
3210      uint32_t data_offset = mirror::Array::DataOffset(sizeof(int8_t)).Uint32Value();
3211      Register out = locations->Out().AsRegister<Register>();
3212      if (index.IsConstant()) {
3213        __ movsxb(out, Address(obj,
3214            (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_1) + data_offset));
3215      } else {
3216        __ movsxb(out, Address(obj, index.AsRegister<Register>(), TIMES_1, data_offset));
3217      }
3218      break;
3219    }
3220
3221    case Primitive::kPrimShort: {
3222      uint32_t data_offset = mirror::Array::DataOffset(sizeof(int16_t)).Uint32Value();
3223      Register out = locations->Out().AsRegister<Register>();
3224      if (index.IsConstant()) {
3225        __ movsxw(out, Address(obj,
3226            (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_2) + data_offset));
3227      } else {
3228        __ movsxw(out, Address(obj, index.AsRegister<Register>(), TIMES_2, data_offset));
3229      }
3230      break;
3231    }
3232
3233    case Primitive::kPrimChar: {
3234      uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint16_t)).Uint32Value();
3235      Register out = locations->Out().AsRegister<Register>();
3236      if (index.IsConstant()) {
3237        __ movzxw(out, Address(obj,
3238            (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_2) + data_offset));
3239      } else {
3240        __ movzxw(out, Address(obj, index.AsRegister<Register>(), TIMES_2, data_offset));
3241      }
3242      break;
3243    }
3244
3245    case Primitive::kPrimInt:
3246    case Primitive::kPrimNot: {
3247      uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value();
3248      Register out = locations->Out().AsRegister<Register>();
3249      if (index.IsConstant()) {
3250        __ movl(out, Address(obj,
3251            (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset));
3252      } else {
3253        __ movl(out, Address(obj, index.AsRegister<Register>(), TIMES_4, data_offset));
3254      }
3255      break;
3256    }
3257
3258    case Primitive::kPrimLong: {
3259      uint32_t data_offset = mirror::Array::DataOffset(sizeof(int64_t)).Uint32Value();
3260      Location out = locations->Out();
3261      DCHECK_NE(obj, out.AsRegisterPairLow<Register>());
3262      if (index.IsConstant()) {
3263        size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset;
3264        __ movl(out.AsRegisterPairLow<Register>(), Address(obj, offset));
3265        codegen_->MaybeRecordImplicitNullCheck(instruction);
3266        __ movl(out.AsRegisterPairHigh<Register>(), Address(obj, offset + kX86WordSize));
3267      } else {
3268        __ movl(out.AsRegisterPairLow<Register>(),
3269                Address(obj, index.AsRegister<Register>(), TIMES_8, data_offset));
3270        codegen_->MaybeRecordImplicitNullCheck(instruction);
3271        __ movl(out.AsRegisterPairHigh<Register>(),
3272                Address(obj, index.AsRegister<Register>(), TIMES_8, data_offset + kX86WordSize));
3273      }
3274      break;
3275    }
3276
3277    case Primitive::kPrimFloat: {
3278      uint32_t data_offset = mirror::Array::DataOffset(sizeof(float)).Uint32Value();
3279      XmmRegister out = locations->Out().AsFpuRegister<XmmRegister>();
3280      if (index.IsConstant()) {
3281        __ movss(out, Address(obj,
3282            (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset));
3283      } else {
3284        __ movss(out, Address(obj, index.AsRegister<Register>(), TIMES_4, data_offset));
3285      }
3286      break;
3287    }
3288
3289    case Primitive::kPrimDouble: {
3290      uint32_t data_offset = mirror::Array::DataOffset(sizeof(double)).Uint32Value();
3291      XmmRegister out = locations->Out().AsFpuRegister<XmmRegister>();
3292      if (index.IsConstant()) {
3293        __ movsd(out, Address(obj,
3294            (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset));
3295      } else {
3296        __ movsd(out, Address(obj, index.AsRegister<Register>(), TIMES_8, data_offset));
3297      }
3298      break;
3299    }
3300
3301    case Primitive::kPrimVoid:
3302      LOG(FATAL) << "Unreachable type " << type;
3303      UNREACHABLE();
3304  }
3305
3306  if (type != Primitive::kPrimLong) {
3307    codegen_->MaybeRecordImplicitNullCheck(instruction);
3308  }
3309}
3310
3311void LocationsBuilderX86::VisitArraySet(HArraySet* instruction) {
3312  // This location builder might end up asking to up to four registers, which is
3313  // not currently possible for baseline. The situation in which we need four
3314  // registers cannot be met by baseline though, because it has not run any
3315  // optimization.
3316
3317  Primitive::Type value_type = instruction->GetComponentType();
3318  bool needs_write_barrier =
3319      CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue());
3320
3321  bool needs_runtime_call = instruction->NeedsTypeCheck();
3322
3323  LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(
3324      instruction,
3325      needs_runtime_call ? LocationSummary::kCall : LocationSummary::kNoCall);
3326
3327  if (needs_runtime_call) {
3328    InvokeRuntimeCallingConvention calling_convention;
3329    locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0)));
3330    locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1)));
3331    locations->SetInAt(2, Location::RegisterLocation(calling_convention.GetRegisterAt(2)));
3332  } else {
3333    bool is_byte_type = (value_type == Primitive::kPrimBoolean)
3334        || (value_type == Primitive::kPrimByte);
3335    // We need the inputs to be different than the output in case of long operation.
3336    // In case of a byte operation, the register allocator does not support multiple
3337    // inputs that die at entry with one in a specific register.
3338    locations->SetInAt(0, Location::RequiresRegister());
3339    locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1)));
3340    if (is_byte_type) {
3341      // Ensure the value is in a byte register.
3342      locations->SetInAt(2, Location::ByteRegisterOrConstant(EAX, instruction->InputAt(2)));
3343    } else {
3344      locations->SetInAt(2, Location::RegisterOrConstant(instruction->InputAt(2)));
3345    }
3346    // Temporary registers for the write barrier.
3347    if (needs_write_barrier) {
3348      locations->AddTemp(Location::RequiresRegister());
3349      // Ensure the card is in a byte register.
3350      locations->AddTemp(Location::RegisterLocation(ECX));
3351    }
3352  }
3353}
3354
3355void InstructionCodeGeneratorX86::VisitArraySet(HArraySet* instruction) {
3356  LocationSummary* locations = instruction->GetLocations();
3357  Register obj = locations->InAt(0).AsRegister<Register>();
3358  Location index = locations->InAt(1);
3359  Location value = locations->InAt(2);
3360  Primitive::Type value_type = instruction->GetComponentType();
3361  bool needs_runtime_call = locations->WillCall();
3362  bool needs_write_barrier =
3363      CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue());
3364
3365  switch (value_type) {
3366    case Primitive::kPrimBoolean:
3367    case Primitive::kPrimByte: {
3368      uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint8_t)).Uint32Value();
3369      if (index.IsConstant()) {
3370        size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_1) + data_offset;
3371        if (value.IsRegister()) {
3372          __ movb(Address(obj, offset), value.AsRegister<ByteRegister>());
3373        } else {
3374          __ movb(Address(obj, offset),
3375                  Immediate(value.GetConstant()->AsIntConstant()->GetValue()));
3376        }
3377      } else {
3378        if (value.IsRegister()) {
3379          __ movb(Address(obj, index.AsRegister<Register>(), TIMES_1, data_offset),
3380                  value.AsRegister<ByteRegister>());
3381        } else {
3382          __ movb(Address(obj, index.AsRegister<Register>(), TIMES_1, data_offset),
3383                  Immediate(value.GetConstant()->AsIntConstant()->GetValue()));
3384        }
3385      }
3386      codegen_->MaybeRecordImplicitNullCheck(instruction);
3387      break;
3388    }
3389
3390    case Primitive::kPrimShort:
3391    case Primitive::kPrimChar: {
3392      uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint16_t)).Uint32Value();
3393      if (index.IsConstant()) {
3394        size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_2) + data_offset;
3395        if (value.IsRegister()) {
3396          __ movw(Address(obj, offset), value.AsRegister<Register>());
3397        } else {
3398          __ movw(Address(obj, offset),
3399                  Immediate(value.GetConstant()->AsIntConstant()->GetValue()));
3400        }
3401      } else {
3402        if (value.IsRegister()) {
3403          __ movw(Address(obj, index.AsRegister<Register>(), TIMES_2, data_offset),
3404                  value.AsRegister<Register>());
3405        } else {
3406          __ movw(Address(obj, index.AsRegister<Register>(), TIMES_2, data_offset),
3407                  Immediate(value.GetConstant()->AsIntConstant()->GetValue()));
3408        }
3409      }
3410      codegen_->MaybeRecordImplicitNullCheck(instruction);
3411      break;
3412    }
3413
3414    case Primitive::kPrimInt:
3415    case Primitive::kPrimNot: {
3416      if (!needs_runtime_call) {
3417        uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value();
3418        if (index.IsConstant()) {
3419          size_t offset =
3420              (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset;
3421          if (value.IsRegister()) {
3422            __ movl(Address(obj, offset), value.AsRegister<Register>());
3423          } else {
3424            DCHECK(value.IsConstant()) << value;
3425            __ movl(Address(obj, offset),
3426                    Immediate(value.GetConstant()->AsIntConstant()->GetValue()));
3427          }
3428        } else {
3429          DCHECK(index.IsRegister()) << index;
3430          if (value.IsRegister()) {
3431            __ movl(Address(obj, index.AsRegister<Register>(), TIMES_4, data_offset),
3432                    value.AsRegister<Register>());
3433          } else {
3434            DCHECK(value.IsConstant()) << value;
3435            __ movl(Address(obj, index.AsRegister<Register>(), TIMES_4, data_offset),
3436                    Immediate(value.GetConstant()->AsIntConstant()->GetValue()));
3437          }
3438        }
3439        codegen_->MaybeRecordImplicitNullCheck(instruction);
3440
3441        if (needs_write_barrier) {
3442          Register temp = locations->GetTemp(0).AsRegister<Register>();
3443          Register card = locations->GetTemp(1).AsRegister<Register>();
3444          codegen_->MarkGCCard(temp, card, obj, value.AsRegister<Register>());
3445        }
3446      } else {
3447        DCHECK_EQ(value_type, Primitive::kPrimNot);
3448        DCHECK(!codegen_->IsLeafMethod());
3449        __ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pAputObject)));
3450        codegen_->RecordPcInfo(instruction, instruction->GetDexPc());
3451      }
3452      break;
3453    }
3454
3455    case Primitive::kPrimLong: {
3456      uint32_t data_offset = mirror::Array::DataOffset(sizeof(int64_t)).Uint32Value();
3457      if (index.IsConstant()) {
3458        size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset;
3459        if (value.IsRegisterPair()) {
3460          __ movl(Address(obj, offset), value.AsRegisterPairLow<Register>());
3461          codegen_->MaybeRecordImplicitNullCheck(instruction);
3462          __ movl(Address(obj, offset + kX86WordSize), value.AsRegisterPairHigh<Register>());
3463        } else {
3464          DCHECK(value.IsConstant());
3465          int64_t val = value.GetConstant()->AsLongConstant()->GetValue();
3466          __ movl(Address(obj, offset), Immediate(Low32Bits(val)));
3467          codegen_->MaybeRecordImplicitNullCheck(instruction);
3468          __ movl(Address(obj, offset + kX86WordSize), Immediate(High32Bits(val)));
3469        }
3470      } else {
3471        if (value.IsRegisterPair()) {
3472          __ movl(Address(obj, index.AsRegister<Register>(), TIMES_8, data_offset),
3473                  value.AsRegisterPairLow<Register>());
3474          codegen_->MaybeRecordImplicitNullCheck(instruction);
3475          __ movl(Address(obj, index.AsRegister<Register>(), TIMES_8, data_offset + kX86WordSize),
3476                  value.AsRegisterPairHigh<Register>());
3477        } else {
3478          DCHECK(value.IsConstant());
3479          int64_t val = value.GetConstant()->AsLongConstant()->GetValue();
3480          __ movl(Address(obj, index.AsRegister<Register>(), TIMES_8, data_offset),
3481                  Immediate(Low32Bits(val)));
3482          codegen_->MaybeRecordImplicitNullCheck(instruction);
3483          __ movl(Address(obj, index.AsRegister<Register>(), TIMES_8, data_offset + kX86WordSize),
3484                  Immediate(High32Bits(val)));
3485        }
3486      }
3487      break;
3488    }
3489
3490    case Primitive::kPrimFloat: {
3491      uint32_t data_offset = mirror::Array::DataOffset(sizeof(float)).Uint32Value();
3492      DCHECK(value.IsFpuRegister());
3493      if (index.IsConstant()) {
3494        size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset;
3495        __ movss(Address(obj, offset), value.AsFpuRegister<XmmRegister>());
3496      } else {
3497        __ movss(Address(obj, index.AsRegister<Register>(), TIMES_4, data_offset),
3498                value.AsFpuRegister<XmmRegister>());
3499      }
3500      break;
3501    }
3502
3503    case Primitive::kPrimDouble: {
3504      uint32_t data_offset = mirror::Array::DataOffset(sizeof(double)).Uint32Value();
3505      DCHECK(value.IsFpuRegister());
3506      if (index.IsConstant()) {
3507        size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset;
3508        __ movsd(Address(obj, offset), value.AsFpuRegister<XmmRegister>());
3509      } else {
3510        __ movsd(Address(obj, index.AsRegister<Register>(), TIMES_8, data_offset),
3511                value.AsFpuRegister<XmmRegister>());
3512      }
3513      break;
3514    }
3515
3516    case Primitive::kPrimVoid:
3517      LOG(FATAL) << "Unreachable type " << instruction->GetType();
3518      UNREACHABLE();
3519  }
3520}
3521
3522void LocationsBuilderX86::VisitArrayLength(HArrayLength* instruction) {
3523  LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction);
3524  locations->SetInAt(0, Location::RequiresRegister());
3525  locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
3526  instruction->SetLocations(locations);
3527}
3528
3529void InstructionCodeGeneratorX86::VisitArrayLength(HArrayLength* instruction) {
3530  LocationSummary* locations = instruction->GetLocations();
3531  uint32_t offset = mirror::Array::LengthOffset().Uint32Value();
3532  Register obj = locations->InAt(0).AsRegister<Register>();
3533  Register out = locations->Out().AsRegister<Register>();
3534  __ movl(out, Address(obj, offset));
3535  codegen_->MaybeRecordImplicitNullCheck(instruction);
3536}
3537
3538void LocationsBuilderX86::VisitBoundsCheck(HBoundsCheck* instruction) {
3539  LocationSummary* locations =
3540      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall);
3541  locations->SetInAt(0, Location::RegisterOrConstant(instruction->InputAt(0)));
3542  locations->SetInAt(1, Location::RequiresRegister());
3543  if (instruction->HasUses()) {
3544    locations->SetOut(Location::SameAsFirstInput());
3545  }
3546}
3547
3548void InstructionCodeGeneratorX86::VisitBoundsCheck(HBoundsCheck* instruction) {
3549  LocationSummary* locations = instruction->GetLocations();
3550  Location index_loc = locations->InAt(0);
3551  Location length_loc = locations->InAt(1);
3552  SlowPathCodeX86* slow_path =
3553    new (GetGraph()->GetArena()) BoundsCheckSlowPathX86(instruction, index_loc, length_loc);
3554  codegen_->AddSlowPath(slow_path);
3555
3556  Register length = length_loc.AsRegister<Register>();
3557  if (index_loc.IsConstant()) {
3558    int32_t value = CodeGenerator::GetInt32ValueOf(index_loc.GetConstant());
3559    __ cmpl(length, Immediate(value));
3560  } else {
3561    __ cmpl(length, index_loc.AsRegister<Register>());
3562  }
3563  __ j(kBelowEqual, slow_path->GetEntryLabel());
3564}
3565
3566void LocationsBuilderX86::VisitTemporary(HTemporary* temp) {
3567  temp->SetLocations(nullptr);
3568}
3569
3570void InstructionCodeGeneratorX86::VisitTemporary(HTemporary* temp) {
3571  // Nothing to do, this is driven by the code generator.
3572  UNUSED(temp);
3573}
3574
3575void LocationsBuilderX86::VisitParallelMove(HParallelMove* instruction) {
3576  UNUSED(instruction);
3577  LOG(FATAL) << "Unreachable";
3578}
3579
3580void InstructionCodeGeneratorX86::VisitParallelMove(HParallelMove* instruction) {
3581  codegen_->GetMoveResolver()->EmitNativeCode(instruction);
3582}
3583
3584void LocationsBuilderX86::VisitSuspendCheck(HSuspendCheck* instruction) {
3585  new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnSlowPath);
3586}
3587
3588void InstructionCodeGeneratorX86::VisitSuspendCheck(HSuspendCheck* instruction) {
3589  HBasicBlock* block = instruction->GetBlock();
3590  if (block->GetLoopInformation() != nullptr) {
3591    DCHECK(block->GetLoopInformation()->GetSuspendCheck() == instruction);
3592    // The back edge will generate the suspend check.
3593    return;
3594  }
3595  if (block->IsEntryBlock() && instruction->GetNext()->IsGoto()) {
3596    // The goto will generate the suspend check.
3597    return;
3598  }
3599  GenerateSuspendCheck(instruction, nullptr);
3600}
3601
3602void InstructionCodeGeneratorX86::GenerateSuspendCheck(HSuspendCheck* instruction,
3603                                                       HBasicBlock* successor) {
3604  SuspendCheckSlowPathX86* slow_path =
3605      new (GetGraph()->GetArena()) SuspendCheckSlowPathX86(instruction, successor);
3606  codegen_->AddSlowPath(slow_path);
3607  __ fs()->cmpw(Address::Absolute(
3608      Thread::ThreadFlagsOffset<kX86WordSize>().Int32Value()), Immediate(0));
3609  if (successor == nullptr) {
3610    __ j(kNotEqual, slow_path->GetEntryLabel());
3611    __ Bind(slow_path->GetReturnLabel());
3612  } else {
3613    __ j(kEqual, codegen_->GetLabelOf(successor));
3614    __ jmp(slow_path->GetEntryLabel());
3615  }
3616}
3617
3618X86Assembler* ParallelMoveResolverX86::GetAssembler() const {
3619  return codegen_->GetAssembler();
3620}
3621
3622void ParallelMoveResolverX86::MoveMemoryToMemory32(int dst, int src) {
3623  ScratchRegisterScope ensure_scratch(
3624      this, kNoRegister, EAX, codegen_->GetNumberOfCoreRegisters());
3625  Register temp_reg = static_cast<Register>(ensure_scratch.GetRegister());
3626  int stack_offset = ensure_scratch.IsSpilled() ? kX86WordSize : 0;
3627  __ movl(temp_reg, Address(ESP, src + stack_offset));
3628  __ movl(Address(ESP, dst + stack_offset), temp_reg);
3629}
3630
3631void ParallelMoveResolverX86::MoveMemoryToMemory64(int dst, int src) {
3632  ScratchRegisterScope ensure_scratch(
3633      this, kNoRegister, EAX, codegen_->GetNumberOfCoreRegisters());
3634  Register temp_reg = static_cast<Register>(ensure_scratch.GetRegister());
3635  int stack_offset = ensure_scratch.IsSpilled() ? kX86WordSize : 0;
3636  __ movl(temp_reg, Address(ESP, src + stack_offset));
3637  __ movl(Address(ESP, dst + stack_offset), temp_reg);
3638  __ movl(temp_reg, Address(ESP, src + stack_offset + kX86WordSize));
3639  __ movl(Address(ESP, dst + stack_offset + kX86WordSize), temp_reg);
3640}
3641
3642void ParallelMoveResolverX86::EmitMove(size_t index) {
3643  MoveOperands* move = moves_.Get(index);
3644  Location source = move->GetSource();
3645  Location destination = move->GetDestination();
3646
3647  if (source.IsRegister()) {
3648    if (destination.IsRegister()) {
3649      __ movl(destination.AsRegister<Register>(), source.AsRegister<Register>());
3650    } else {
3651      DCHECK(destination.IsStackSlot());
3652      __ movl(Address(ESP, destination.GetStackIndex()), source.AsRegister<Register>());
3653    }
3654  } else if (source.IsFpuRegister()) {
3655    if (destination.IsFpuRegister()) {
3656      __ movaps(destination.AsFpuRegister<XmmRegister>(), source.AsFpuRegister<XmmRegister>());
3657    } else if (destination.IsStackSlot()) {
3658      __ movss(Address(ESP, destination.GetStackIndex()), source.AsFpuRegister<XmmRegister>());
3659    } else {
3660      DCHECK(destination.IsDoubleStackSlot());
3661      __ movsd(Address(ESP, destination.GetStackIndex()), source.AsFpuRegister<XmmRegister>());
3662    }
3663  } else if (source.IsStackSlot()) {
3664    if (destination.IsRegister()) {
3665      __ movl(destination.AsRegister<Register>(), Address(ESP, source.GetStackIndex()));
3666    } else if (destination.IsFpuRegister()) {
3667      __ movss(destination.AsFpuRegister<XmmRegister>(), Address(ESP, source.GetStackIndex()));
3668    } else {
3669      DCHECK(destination.IsStackSlot());
3670      MoveMemoryToMemory32(destination.GetStackIndex(), source.GetStackIndex());
3671    }
3672  } else if (source.IsDoubleStackSlot()) {
3673    if (destination.IsFpuRegister()) {
3674      __ movsd(destination.AsFpuRegister<XmmRegister>(), Address(ESP, source.GetStackIndex()));
3675    } else {
3676      DCHECK(destination.IsDoubleStackSlot()) << destination;
3677      MoveMemoryToMemory64(destination.GetStackIndex(), source.GetStackIndex());
3678    }
3679  } else if (source.IsConstant()) {
3680    HConstant* constant = source.GetConstant();
3681    if (constant->IsIntConstant() || constant->IsNullConstant()) {
3682      int32_t value = CodeGenerator::GetInt32ValueOf(constant);
3683      if (destination.IsRegister()) {
3684        if (value == 0) {
3685          __ xorl(destination.AsRegister<Register>(), destination.AsRegister<Register>());
3686        } else {
3687          __ movl(destination.AsRegister<Register>(), Immediate(value));
3688        }
3689      } else {
3690        DCHECK(destination.IsStackSlot()) << destination;
3691        __ movl(Address(ESP, destination.GetStackIndex()), Immediate(value));
3692      }
3693    } else if (constant->IsFloatConstant()) {
3694      float fp_value = constant->AsFloatConstant()->GetValue();
3695      int32_t value = bit_cast<int32_t, float>(fp_value);
3696      Immediate imm(value);
3697      if (destination.IsFpuRegister()) {
3698        XmmRegister dest = destination.AsFpuRegister<XmmRegister>();
3699        if (value == 0) {
3700          // Easy handling of 0.0.
3701          __ xorps(dest, dest);
3702        } else {
3703          ScratchRegisterScope ensure_scratch(
3704              this, kNoRegister, EAX, codegen_->GetNumberOfCoreRegisters());
3705          Register temp = static_cast<Register>(ensure_scratch.GetRegister());
3706          __ movl(temp, Immediate(value));
3707          __ movd(dest, temp);
3708        }
3709      } else {
3710        DCHECK(destination.IsStackSlot()) << destination;
3711        __ movl(Address(ESP, destination.GetStackIndex()), imm);
3712      }
3713    } else if (constant->IsLongConstant()) {
3714      int64_t value = constant->AsLongConstant()->GetValue();
3715      int32_t low_value = Low32Bits(value);
3716      int32_t high_value = High32Bits(value);
3717      Immediate low(low_value);
3718      Immediate high(high_value);
3719      if (destination.IsDoubleStackSlot()) {
3720        __ movl(Address(ESP, destination.GetStackIndex()), low);
3721        __ movl(Address(ESP, destination.GetHighStackIndex(kX86WordSize)), high);
3722      } else {
3723        __ movl(destination.AsRegisterPairLow<Register>(), low);
3724        __ movl(destination.AsRegisterPairHigh<Register>(), high);
3725      }
3726    } else {
3727      DCHECK(constant->IsDoubleConstant());
3728      double dbl_value = constant->AsDoubleConstant()->GetValue();
3729      int64_t value = bit_cast<int64_t, double>(dbl_value);
3730      int32_t low_value = Low32Bits(value);
3731      int32_t high_value = High32Bits(value);
3732      Immediate low(low_value);
3733      Immediate high(high_value);
3734      if (destination.IsFpuRegister()) {
3735        XmmRegister dest = destination.AsFpuRegister<XmmRegister>();
3736        if (value == 0) {
3737          // Easy handling of 0.0.
3738          __ xorpd(dest, dest);
3739        } else {
3740          __ pushl(high);
3741          __ pushl(low);
3742          __ movsd(dest, Address(ESP, 0));
3743          __ addl(ESP, Immediate(8));
3744        }
3745      } else {
3746        DCHECK(destination.IsDoubleStackSlot()) << destination;
3747        __ movl(Address(ESP, destination.GetStackIndex()), low);
3748        __ movl(Address(ESP, destination.GetHighStackIndex(kX86WordSize)), high);
3749      }
3750    }
3751  } else {
3752    LOG(FATAL) << "Unimplemented move: " << destination << " <- " << source;
3753  }
3754}
3755
3756void ParallelMoveResolverX86::Exchange(Register reg, int mem) {
3757  Register suggested_scratch = reg == EAX ? EBX : EAX;
3758  ScratchRegisterScope ensure_scratch(
3759      this, reg, suggested_scratch, codegen_->GetNumberOfCoreRegisters());
3760
3761  int stack_offset = ensure_scratch.IsSpilled() ? kX86WordSize : 0;
3762  __ movl(static_cast<Register>(ensure_scratch.GetRegister()), Address(ESP, mem + stack_offset));
3763  __ movl(Address(ESP, mem + stack_offset), reg);
3764  __ movl(reg, static_cast<Register>(ensure_scratch.GetRegister()));
3765}
3766
3767void ParallelMoveResolverX86::Exchange32(XmmRegister reg, int mem) {
3768  ScratchRegisterScope ensure_scratch(
3769      this, kNoRegister, EAX, codegen_->GetNumberOfCoreRegisters());
3770
3771  Register temp_reg = static_cast<Register>(ensure_scratch.GetRegister());
3772  int stack_offset = ensure_scratch.IsSpilled() ? kX86WordSize : 0;
3773  __ movl(temp_reg, Address(ESP, mem + stack_offset));
3774  __ movss(Address(ESP, mem + stack_offset), reg);
3775  __ movd(reg, temp_reg);
3776}
3777
3778void ParallelMoveResolverX86::Exchange(int mem1, int mem2) {
3779  ScratchRegisterScope ensure_scratch1(
3780      this, kNoRegister, EAX, codegen_->GetNumberOfCoreRegisters());
3781
3782  Register suggested_scratch = ensure_scratch1.GetRegister() == EAX ? EBX : EAX;
3783  ScratchRegisterScope ensure_scratch2(
3784      this, ensure_scratch1.GetRegister(), suggested_scratch, codegen_->GetNumberOfCoreRegisters());
3785
3786  int stack_offset = ensure_scratch1.IsSpilled() ? kX86WordSize : 0;
3787  stack_offset += ensure_scratch2.IsSpilled() ? kX86WordSize : 0;
3788  __ movl(static_cast<Register>(ensure_scratch1.GetRegister()), Address(ESP, mem1 + stack_offset));
3789  __ movl(static_cast<Register>(ensure_scratch2.GetRegister()), Address(ESP, mem2 + stack_offset));
3790  __ movl(Address(ESP, mem2 + stack_offset), static_cast<Register>(ensure_scratch1.GetRegister()));
3791  __ movl(Address(ESP, mem1 + stack_offset), static_cast<Register>(ensure_scratch2.GetRegister()));
3792}
3793
3794void ParallelMoveResolverX86::EmitSwap(size_t index) {
3795  MoveOperands* move = moves_.Get(index);
3796  Location source = move->GetSource();
3797  Location destination = move->GetDestination();
3798
3799  if (source.IsRegister() && destination.IsRegister()) {
3800    __ xchgl(destination.AsRegister<Register>(), source.AsRegister<Register>());
3801  } else if (source.IsRegister() && destination.IsStackSlot()) {
3802    Exchange(source.AsRegister<Register>(), destination.GetStackIndex());
3803  } else if (source.IsStackSlot() && destination.IsRegister()) {
3804    Exchange(destination.AsRegister<Register>(), source.GetStackIndex());
3805  } else if (source.IsStackSlot() && destination.IsStackSlot()) {
3806    Exchange(destination.GetStackIndex(), source.GetStackIndex());
3807  } else if (source.IsFpuRegister() && destination.IsFpuRegister()) {
3808    // Use XOR Swap algorithm to avoid a temporary.
3809    DCHECK_NE(source.reg(), destination.reg());
3810    __ xorpd(destination.AsFpuRegister<XmmRegister>(), source.AsFpuRegister<XmmRegister>());
3811    __ xorpd(source.AsFpuRegister<XmmRegister>(), destination.AsFpuRegister<XmmRegister>());
3812    __ xorpd(destination.AsFpuRegister<XmmRegister>(), source.AsFpuRegister<XmmRegister>());
3813  } else if (source.IsFpuRegister() && destination.IsStackSlot()) {
3814    Exchange32(source.AsFpuRegister<XmmRegister>(), destination.GetStackIndex());
3815  } else if (destination.IsFpuRegister() && source.IsStackSlot()) {
3816    Exchange32(destination.AsFpuRegister<XmmRegister>(), source.GetStackIndex());
3817  } else if (source.IsFpuRegister() && destination.IsDoubleStackSlot()) {
3818    // Take advantage of the 16 bytes in the XMM register.
3819    XmmRegister reg = source.AsFpuRegister<XmmRegister>();
3820    Address stack(ESP, destination.GetStackIndex());
3821    // Load the double into the high doubleword.
3822    __ movhpd(reg, stack);
3823
3824    // Store the low double into the destination.
3825    __ movsd(stack, reg);
3826
3827    // Move the high double to the low double.
3828    __ psrldq(reg, Immediate(8));
3829  } else if (destination.IsFpuRegister() && source.IsDoubleStackSlot()) {
3830    // Take advantage of the 16 bytes in the XMM register.
3831    XmmRegister reg = destination.AsFpuRegister<XmmRegister>();
3832    Address stack(ESP, source.GetStackIndex());
3833    // Load the double into the high doubleword.
3834    __ movhpd(reg, stack);
3835
3836    // Store the low double into the destination.
3837    __ movsd(stack, reg);
3838
3839    // Move the high double to the low double.
3840    __ psrldq(reg, Immediate(8));
3841  } else if (destination.IsDoubleStackSlot() && source.IsDoubleStackSlot()) {
3842    Exchange(destination.GetStackIndex(), source.GetStackIndex());
3843    Exchange(destination.GetHighStackIndex(kX86WordSize), source.GetHighStackIndex(kX86WordSize));
3844  } else {
3845    LOG(FATAL) << "Unimplemented: source: " << source << ", destination: " << destination;
3846  }
3847}
3848
3849void ParallelMoveResolverX86::SpillScratch(int reg) {
3850  __ pushl(static_cast<Register>(reg));
3851}
3852
3853void ParallelMoveResolverX86::RestoreScratch(int reg) {
3854  __ popl(static_cast<Register>(reg));
3855}
3856
3857void LocationsBuilderX86::VisitLoadClass(HLoadClass* cls) {
3858  LocationSummary::CallKind call_kind = cls->CanCallRuntime()
3859      ? LocationSummary::kCallOnSlowPath
3860      : LocationSummary::kNoCall;
3861  LocationSummary* locations =
3862      new (GetGraph()->GetArena()) LocationSummary(cls, call_kind);
3863  locations->SetOut(Location::RequiresRegister());
3864}
3865
3866void InstructionCodeGeneratorX86::VisitLoadClass(HLoadClass* cls) {
3867  Register out = cls->GetLocations()->Out().AsRegister<Register>();
3868  if (cls->IsReferrersClass()) {
3869    DCHECK(!cls->CanCallRuntime());
3870    DCHECK(!cls->MustGenerateClinitCheck());
3871    codegen_->LoadCurrentMethod(out);
3872    __ movl(out, Address(out, mirror::ArtMethod::DeclaringClassOffset().Int32Value()));
3873  } else {
3874    DCHECK(cls->CanCallRuntime());
3875    codegen_->LoadCurrentMethod(out);
3876    __ movl(out, Address(out, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value()));
3877    __ movl(out, Address(out, CodeGenerator::GetCacheOffset(cls->GetTypeIndex())));
3878
3879    SlowPathCodeX86* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathX86(
3880        cls, cls, cls->GetDexPc(), cls->MustGenerateClinitCheck());
3881    codegen_->AddSlowPath(slow_path);
3882    __ testl(out, out);
3883    __ j(kEqual, slow_path->GetEntryLabel());
3884    if (cls->MustGenerateClinitCheck()) {
3885      GenerateClassInitializationCheck(slow_path, out);
3886    } else {
3887      __ Bind(slow_path->GetExitLabel());
3888    }
3889  }
3890}
3891
3892void LocationsBuilderX86::VisitClinitCheck(HClinitCheck* check) {
3893  LocationSummary* locations =
3894      new (GetGraph()->GetArena()) LocationSummary(check, LocationSummary::kCallOnSlowPath);
3895  locations->SetInAt(0, Location::RequiresRegister());
3896  if (check->HasUses()) {
3897    locations->SetOut(Location::SameAsFirstInput());
3898  }
3899}
3900
3901void InstructionCodeGeneratorX86::VisitClinitCheck(HClinitCheck* check) {
3902  // We assume the class to not be null.
3903  SlowPathCodeX86* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathX86(
3904      check->GetLoadClass(), check, check->GetDexPc(), true);
3905  codegen_->AddSlowPath(slow_path);
3906  GenerateClassInitializationCheck(slow_path,
3907                                   check->GetLocations()->InAt(0).AsRegister<Register>());
3908}
3909
3910void InstructionCodeGeneratorX86::GenerateClassInitializationCheck(
3911    SlowPathCodeX86* slow_path, Register class_reg) {
3912  __ cmpl(Address(class_reg,  mirror::Class::StatusOffset().Int32Value()),
3913          Immediate(mirror::Class::kStatusInitialized));
3914  __ j(kLess, slow_path->GetEntryLabel());
3915  __ Bind(slow_path->GetExitLabel());
3916  // No need for memory fence, thanks to the X86 memory model.
3917}
3918
3919void LocationsBuilderX86::VisitLoadString(HLoadString* load) {
3920  LocationSummary* locations =
3921      new (GetGraph()->GetArena()) LocationSummary(load, LocationSummary::kCallOnSlowPath);
3922  locations->SetOut(Location::RequiresRegister());
3923}
3924
3925void InstructionCodeGeneratorX86::VisitLoadString(HLoadString* load) {
3926  SlowPathCodeX86* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathX86(load);
3927  codegen_->AddSlowPath(slow_path);
3928
3929  Register out = load->GetLocations()->Out().AsRegister<Register>();
3930  codegen_->LoadCurrentMethod(out);
3931  __ movl(out, Address(out, mirror::ArtMethod::DeclaringClassOffset().Int32Value()));
3932  __ movl(out, Address(out, mirror::Class::DexCacheStringsOffset().Int32Value()));
3933  __ movl(out, Address(out, CodeGenerator::GetCacheOffset(load->GetStringIndex())));
3934  __ testl(out, out);
3935  __ j(kEqual, slow_path->GetEntryLabel());
3936  __ Bind(slow_path->GetExitLabel());
3937}
3938
3939void LocationsBuilderX86::VisitLoadException(HLoadException* load) {
3940  LocationSummary* locations =
3941      new (GetGraph()->GetArena()) LocationSummary(load, LocationSummary::kNoCall);
3942  locations->SetOut(Location::RequiresRegister());
3943}
3944
3945void InstructionCodeGeneratorX86::VisitLoadException(HLoadException* load) {
3946  Address address = Address::Absolute(Thread::ExceptionOffset<kX86WordSize>().Int32Value());
3947  __ fs()->movl(load->GetLocations()->Out().AsRegister<Register>(), address);
3948  __ fs()->movl(address, Immediate(0));
3949}
3950
3951void LocationsBuilderX86::VisitThrow(HThrow* instruction) {
3952  LocationSummary* locations =
3953      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall);
3954  InvokeRuntimeCallingConvention calling_convention;
3955  locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0)));
3956}
3957
3958void InstructionCodeGeneratorX86::VisitThrow(HThrow* instruction) {
3959  __ fs()->call(Address::Absolute(QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pDeliverException)));
3960  codegen_->RecordPcInfo(instruction, instruction->GetDexPc());
3961}
3962
3963void LocationsBuilderX86::VisitInstanceOf(HInstanceOf* instruction) {
3964  LocationSummary::CallKind call_kind = instruction->IsClassFinal()
3965      ? LocationSummary::kNoCall
3966      : LocationSummary::kCallOnSlowPath;
3967  LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind);
3968  locations->SetInAt(0, Location::RequiresRegister());
3969  locations->SetInAt(1, Location::Any());
3970  locations->SetOut(Location::RequiresRegister());
3971}
3972
3973void InstructionCodeGeneratorX86::VisitInstanceOf(HInstanceOf* instruction) {
3974  LocationSummary* locations = instruction->GetLocations();
3975  Register obj = locations->InAt(0).AsRegister<Register>();
3976  Location cls = locations->InAt(1);
3977  Register out = locations->Out().AsRegister<Register>();
3978  uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
3979  Label done, zero;
3980  SlowPathCodeX86* slow_path = nullptr;
3981
3982  // Return 0 if `obj` is null.
3983  // TODO: avoid this check if we know obj is not null.
3984  __ testl(obj, obj);
3985  __ j(kEqual, &zero);
3986  __ movl(out, Address(obj, class_offset));
3987  // Compare the class of `obj` with `cls`.
3988  if (cls.IsRegister()) {
3989    __ cmpl(out, cls.AsRegister<Register>());
3990  } else {
3991    DCHECK(cls.IsStackSlot()) << cls;
3992    __ cmpl(out, Address(ESP, cls.GetStackIndex()));
3993  }
3994
3995  if (instruction->IsClassFinal()) {
3996    // Classes must be equal for the instanceof to succeed.
3997    __ j(kNotEqual, &zero);
3998    __ movl(out, Immediate(1));
3999    __ jmp(&done);
4000  } else {
4001    // If the classes are not equal, we go into a slow path.
4002    DCHECK(locations->OnlyCallsOnSlowPath());
4003    slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathX86(
4004        instruction, locations->InAt(1), locations->Out(), instruction->GetDexPc());
4005    codegen_->AddSlowPath(slow_path);
4006    __ j(kNotEqual, slow_path->GetEntryLabel());
4007    __ movl(out, Immediate(1));
4008    __ jmp(&done);
4009  }
4010  __ Bind(&zero);
4011  __ movl(out, Immediate(0));
4012  if (slow_path != nullptr) {
4013    __ Bind(slow_path->GetExitLabel());
4014  }
4015  __ Bind(&done);
4016}
4017
4018void LocationsBuilderX86::VisitCheckCast(HCheckCast* instruction) {
4019  LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(
4020      instruction, LocationSummary::kCallOnSlowPath);
4021  locations->SetInAt(0, Location::RequiresRegister());
4022  locations->SetInAt(1, Location::Any());
4023  locations->AddTemp(Location::RequiresRegister());
4024}
4025
4026void InstructionCodeGeneratorX86::VisitCheckCast(HCheckCast* instruction) {
4027  LocationSummary* locations = instruction->GetLocations();
4028  Register obj = locations->InAt(0).AsRegister<Register>();
4029  Location cls = locations->InAt(1);
4030  Register temp = locations->GetTemp(0).AsRegister<Register>();
4031  uint32_t class_offset = mirror::Object::ClassOffset().Int32Value();
4032  SlowPathCodeX86* slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathX86(
4033      instruction, locations->InAt(1), locations->GetTemp(0), instruction->GetDexPc());
4034  codegen_->AddSlowPath(slow_path);
4035
4036  // TODO: avoid this check if we know obj is not null.
4037  __ testl(obj, obj);
4038  __ j(kEqual, slow_path->GetExitLabel());
4039  __ movl(temp, Address(obj, class_offset));
4040
4041  // Compare the class of `obj` with `cls`.
4042  if (cls.IsRegister()) {
4043    __ cmpl(temp, cls.AsRegister<Register>());
4044  } else {
4045    DCHECK(cls.IsStackSlot()) << cls;
4046    __ cmpl(temp, Address(ESP, cls.GetStackIndex()));
4047  }
4048
4049  __ j(kNotEqual, slow_path->GetEntryLabel());
4050  __ Bind(slow_path->GetExitLabel());
4051}
4052
4053void LocationsBuilderX86::VisitMonitorOperation(HMonitorOperation* instruction) {
4054  LocationSummary* locations =
4055      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall);
4056  InvokeRuntimeCallingConvention calling_convention;
4057  locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0)));
4058}
4059
4060void InstructionCodeGeneratorX86::VisitMonitorOperation(HMonitorOperation* instruction) {
4061  __ fs()->call(Address::Absolute(instruction->IsEnter()
4062        ? QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pLockObject)
4063        : QUICK_ENTRYPOINT_OFFSET(kX86WordSize, pUnlockObject)));
4064  codegen_->RecordPcInfo(instruction, instruction->GetDexPc());
4065}
4066
4067void LocationsBuilderX86::VisitAnd(HAnd* instruction) { HandleBitwiseOperation(instruction); }
4068void LocationsBuilderX86::VisitOr(HOr* instruction) { HandleBitwiseOperation(instruction); }
4069void LocationsBuilderX86::VisitXor(HXor* instruction) { HandleBitwiseOperation(instruction); }
4070
4071void LocationsBuilderX86::HandleBitwiseOperation(HBinaryOperation* instruction) {
4072  LocationSummary* locations =
4073      new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall);
4074  DCHECK(instruction->GetResultType() == Primitive::kPrimInt
4075         || instruction->GetResultType() == Primitive::kPrimLong);
4076  locations->SetInAt(0, Location::RequiresRegister());
4077  locations->SetInAt(1, Location::Any());
4078  locations->SetOut(Location::SameAsFirstInput());
4079}
4080
4081void InstructionCodeGeneratorX86::VisitAnd(HAnd* instruction) {
4082  HandleBitwiseOperation(instruction);
4083}
4084
4085void InstructionCodeGeneratorX86::VisitOr(HOr* instruction) {
4086  HandleBitwiseOperation(instruction);
4087}
4088
4089void InstructionCodeGeneratorX86::VisitXor(HXor* instruction) {
4090  HandleBitwiseOperation(instruction);
4091}
4092
4093void InstructionCodeGeneratorX86::HandleBitwiseOperation(HBinaryOperation* instruction) {
4094  LocationSummary* locations = instruction->GetLocations();
4095  Location first = locations->InAt(0);
4096  Location second = locations->InAt(1);
4097  DCHECK(first.Equals(locations->Out()));
4098
4099  if (instruction->GetResultType() == Primitive::kPrimInt) {
4100    if (second.IsRegister()) {
4101      if (instruction->IsAnd()) {
4102        __ andl(first.AsRegister<Register>(), second.AsRegister<Register>());
4103      } else if (instruction->IsOr()) {
4104        __ orl(first.AsRegister<Register>(), second.AsRegister<Register>());
4105      } else {
4106        DCHECK(instruction->IsXor());
4107        __ xorl(first.AsRegister<Register>(), second.AsRegister<Register>());
4108      }
4109    } else if (second.IsConstant()) {
4110      if (instruction->IsAnd()) {
4111        __ andl(first.AsRegister<Register>(),
4112                Immediate(second.GetConstant()->AsIntConstant()->GetValue()));
4113      } else if (instruction->IsOr()) {
4114        __ orl(first.AsRegister<Register>(),
4115               Immediate(second.GetConstant()->AsIntConstant()->GetValue()));
4116      } else {
4117        DCHECK(instruction->IsXor());
4118        __ xorl(first.AsRegister<Register>(),
4119                Immediate(second.GetConstant()->AsIntConstant()->GetValue()));
4120      }
4121    } else {
4122      if (instruction->IsAnd()) {
4123        __ andl(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex()));
4124      } else if (instruction->IsOr()) {
4125        __ orl(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex()));
4126      } else {
4127        DCHECK(instruction->IsXor());
4128        __ xorl(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex()));
4129      }
4130    }
4131  } else {
4132    DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong);
4133    if (second.IsRegisterPair()) {
4134      if (instruction->IsAnd()) {
4135        __ andl(first.AsRegisterPairLow<Register>(), second.AsRegisterPairLow<Register>());
4136        __ andl(first.AsRegisterPairHigh<Register>(), second.AsRegisterPairHigh<Register>());
4137      } else if (instruction->IsOr()) {
4138        __ orl(first.AsRegisterPairLow<Register>(), second.AsRegisterPairLow<Register>());
4139        __ orl(first.AsRegisterPairHigh<Register>(), second.AsRegisterPairHigh<Register>());
4140      } else {
4141        DCHECK(instruction->IsXor());
4142        __ xorl(first.AsRegisterPairLow<Register>(), second.AsRegisterPairLow<Register>());
4143        __ xorl(first.AsRegisterPairHigh<Register>(), second.AsRegisterPairHigh<Register>());
4144      }
4145    } else if (second.IsDoubleStackSlot()) {
4146      if (instruction->IsAnd()) {
4147        __ andl(first.AsRegisterPairLow<Register>(), Address(ESP, second.GetStackIndex()));
4148        __ andl(first.AsRegisterPairHigh<Register>(),
4149                Address(ESP, second.GetHighStackIndex(kX86WordSize)));
4150      } else if (instruction->IsOr()) {
4151        __ orl(first.AsRegisterPairLow<Register>(), Address(ESP, second.GetStackIndex()));
4152        __ orl(first.AsRegisterPairHigh<Register>(),
4153                Address(ESP, second.GetHighStackIndex(kX86WordSize)));
4154      } else {
4155        DCHECK(instruction->IsXor());
4156        __ xorl(first.AsRegisterPairLow<Register>(), Address(ESP, second.GetStackIndex()));
4157        __ xorl(first.AsRegisterPairHigh<Register>(),
4158                Address(ESP, second.GetHighStackIndex(kX86WordSize)));
4159      }
4160    } else {
4161      DCHECK(second.IsConstant()) << second;
4162      int64_t value = second.GetConstant()->AsLongConstant()->GetValue();
4163      int32_t low_value = Low32Bits(value);
4164      int32_t high_value = High32Bits(value);
4165      Immediate low(low_value);
4166      Immediate high(high_value);
4167      Register first_low = first.AsRegisterPairLow<Register>();
4168      Register first_high = first.AsRegisterPairHigh<Register>();
4169      if (instruction->IsAnd()) {
4170        if (low_value == 0) {
4171          __ xorl(first_low, first_low);
4172        } else if (low_value != -1) {
4173          __ andl(first_low, low);
4174        }
4175        if (high_value == 0) {
4176          __ xorl(first_high, first_high);
4177        } else if (high_value != -1) {
4178          __ andl(first_high, high);
4179        }
4180      } else if (instruction->IsOr()) {
4181        if (low_value != 0) {
4182          __ orl(first_low, low);
4183        }
4184        if (high_value != 0) {
4185          __ orl(first_high, high);
4186        }
4187      } else {
4188        DCHECK(instruction->IsXor());
4189        if (low_value != 0) {
4190          __ xorl(first_low, low);
4191        }
4192        if (high_value != 0) {
4193          __ xorl(first_high, high);
4194        }
4195      }
4196    }
4197  }
4198}
4199
4200void LocationsBuilderX86::VisitBoundType(HBoundType* instruction) {
4201  // Nothing to do, this should be removed during prepare for register allocator.
4202  UNUSED(instruction);
4203  LOG(FATAL) << "Unreachable";
4204}
4205
4206void InstructionCodeGeneratorX86::VisitBoundType(HBoundType* instruction) {
4207  // Nothing to do, this should be removed during prepare for register allocator.
4208  UNUSED(instruction);
4209  LOG(FATAL) << "Unreachable";
4210}
4211
4212}  // namespace x86
4213}  // namespace art
4214