gen_common.cc revision 2689fbad6b5ec1ae8f8c8791a80c6fd3cf24144d
1/*
2 * Copyright (C) 2012 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#include "dex/compiler_ir.h"
17#include "dex/compiler_internals.h"
18#include "dex/quick/arm/arm_lir.h"
19#include "dex/quick/mir_to_lir-inl.h"
20#include "entrypoints/quick/quick_entrypoints.h"
21#include "mirror/array.h"
22#include "mirror/object_array-inl.h"
23#include "mirror/object-inl.h"
24#include "verifier/method_verifier.h"
25#include <functional>
26
27namespace art {
28
29// Shortcuts to repeatedly used long types.
30typedef mirror::ObjectArray<mirror::Object> ObjArray;
31typedef mirror::ObjectArray<mirror::Class> ClassArray;
32
33/*
34 * This source files contains "gen" codegen routines that should
35 * be applicable to most targets.  Only mid-level support utilities
36 * and "op" calls may be used here.
37 */
38
39/*
40 * Generate a kPseudoBarrier marker to indicate the boundary of special
41 * blocks.
42 */
43void Mir2Lir::GenBarrier() {
44  LIR* barrier = NewLIR0(kPseudoBarrier);
45  /* Mark all resources as being clobbered */
46  DCHECK(!barrier->flags.use_def_invalid);
47  barrier->u.m.def_mask = &kEncodeAll;
48}
49
50void Mir2Lir::GenDivZeroException() {
51  LIR* branch = OpUnconditionalBranch(nullptr);
52  AddDivZeroCheckSlowPath(branch);
53}
54
55void Mir2Lir::GenDivZeroCheck(ConditionCode c_code) {
56  LIR* branch = OpCondBranch(c_code, nullptr);
57  AddDivZeroCheckSlowPath(branch);
58}
59
60void Mir2Lir::GenDivZeroCheck(RegStorage reg) {
61  LIR* branch = OpCmpImmBranch(kCondEq, reg, 0, nullptr);
62  AddDivZeroCheckSlowPath(branch);
63}
64
65void Mir2Lir::AddDivZeroCheckSlowPath(LIR* branch) {
66  class DivZeroCheckSlowPath : public Mir2Lir::LIRSlowPath {
67   public:
68    DivZeroCheckSlowPath(Mir2Lir* m2l, LIR* branch)
69        : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), branch) {
70    }
71
72    void Compile() OVERRIDE {
73      m2l_->ResetRegPool();
74      m2l_->ResetDefTracking();
75      GenerateTargetLabel(kPseudoThrowTarget);
76      if (m2l_->cu_->target64) {
77        m2l_->CallRuntimeHelper(QUICK_ENTRYPOINT_OFFSET(8, pThrowDivZero), true);
78      } else {
79        m2l_->CallRuntimeHelper(QUICK_ENTRYPOINT_OFFSET(4, pThrowDivZero), true);
80      }
81    }
82  };
83
84  AddSlowPath(new (arena_) DivZeroCheckSlowPath(this, branch));
85}
86
87void Mir2Lir::GenArrayBoundsCheck(RegStorage index, RegStorage length) {
88  class ArrayBoundsCheckSlowPath : public Mir2Lir::LIRSlowPath {
89   public:
90    ArrayBoundsCheckSlowPath(Mir2Lir* m2l, LIR* branch, RegStorage index, RegStorage length)
91        : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), branch),
92          index_(index), length_(length) {
93    }
94
95    void Compile() OVERRIDE {
96      m2l_->ResetRegPool();
97      m2l_->ResetDefTracking();
98      GenerateTargetLabel(kPseudoThrowTarget);
99      if (m2l_->cu_->target64) {
100        m2l_->CallRuntimeHelperRegReg(QUICK_ENTRYPOINT_OFFSET(8, pThrowArrayBounds),
101                                      index_, length_, true);
102      } else {
103        m2l_->CallRuntimeHelperRegReg(QUICK_ENTRYPOINT_OFFSET(4, pThrowArrayBounds),
104                                      index_, length_, true);
105      }
106    }
107
108   private:
109    const RegStorage index_;
110    const RegStorage length_;
111  };
112
113  LIR* branch = OpCmpBranch(kCondUge, index, length, nullptr);
114  AddSlowPath(new (arena_) ArrayBoundsCheckSlowPath(this, branch, index, length));
115}
116
117void Mir2Lir::GenArrayBoundsCheck(int index, RegStorage length) {
118  class ArrayBoundsCheckSlowPath : public Mir2Lir::LIRSlowPath {
119   public:
120    ArrayBoundsCheckSlowPath(Mir2Lir* m2l, LIR* branch, int index, RegStorage length)
121        : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), branch),
122          index_(index), length_(length) {
123    }
124
125    void Compile() OVERRIDE {
126      m2l_->ResetRegPool();
127      m2l_->ResetDefTracking();
128      GenerateTargetLabel(kPseudoThrowTarget);
129
130      m2l_->OpRegCopy(m2l_->TargetReg(kArg1), length_);
131      m2l_->LoadConstant(m2l_->TargetReg(kArg0), index_);
132      if (m2l_->cu_->target64) {
133        m2l_->CallRuntimeHelperRegReg(QUICK_ENTRYPOINT_OFFSET(8, pThrowArrayBounds),
134                                      m2l_->TargetReg(kArg0), m2l_->TargetReg(kArg1), true);
135      } else {
136        m2l_->CallRuntimeHelperRegReg(QUICK_ENTRYPOINT_OFFSET(4, pThrowArrayBounds),
137                                      m2l_->TargetReg(kArg0), m2l_->TargetReg(kArg1), true);
138      }
139    }
140
141   private:
142    const int32_t index_;
143    const RegStorage length_;
144  };
145
146  LIR* branch = OpCmpImmBranch(kCondLs, length, index, nullptr);
147  AddSlowPath(new (arena_) ArrayBoundsCheckSlowPath(this, branch, index, length));
148}
149
150LIR* Mir2Lir::GenNullCheck(RegStorage reg) {
151  class NullCheckSlowPath : public Mir2Lir::LIRSlowPath {
152   public:
153    NullCheckSlowPath(Mir2Lir* m2l, LIR* branch)
154        : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), branch) {
155    }
156
157    void Compile() OVERRIDE {
158      m2l_->ResetRegPool();
159      m2l_->ResetDefTracking();
160      GenerateTargetLabel(kPseudoThrowTarget);
161      if (m2l_->cu_->target64) {
162        m2l_->CallRuntimeHelper(QUICK_ENTRYPOINT_OFFSET(8, pThrowNullPointer), true);
163      } else {
164        m2l_->CallRuntimeHelper(QUICK_ENTRYPOINT_OFFSET(4, pThrowNullPointer), true);
165      }
166    }
167  };
168
169  LIR* branch = OpCmpImmBranch(kCondEq, reg, 0, nullptr);
170  AddSlowPath(new (arena_) NullCheckSlowPath(this, branch));
171  return branch;
172}
173
174/* Perform null-check on a register.  */
175LIR* Mir2Lir::GenNullCheck(RegStorage m_reg, int opt_flags) {
176  if (cu_->compiler_driver->GetCompilerOptions().GetExplicitNullChecks()) {
177    return GenExplicitNullCheck(m_reg, opt_flags);
178  }
179  return nullptr;
180}
181
182/* Perform an explicit null-check on a register.  */
183LIR* Mir2Lir::GenExplicitNullCheck(RegStorage m_reg, int opt_flags) {
184  if (!(cu_->disable_opt & (1 << kNullCheckElimination)) && (opt_flags & MIR_IGNORE_NULL_CHECK)) {
185    return NULL;
186  }
187  return GenNullCheck(m_reg);
188}
189
190void Mir2Lir::MarkPossibleNullPointerException(int opt_flags) {
191  if (!cu_->compiler_driver->GetCompilerOptions().GetExplicitNullChecks()) {
192    if (!(cu_->disable_opt & (1 << kNullCheckElimination)) && (opt_flags & MIR_IGNORE_NULL_CHECK)) {
193      return;
194    }
195    MarkSafepointPC(last_lir_insn_);
196  }
197}
198
199void Mir2Lir::MarkPossibleNullPointerExceptionAfter(int opt_flags, LIR* after) {
200  if (!cu_->compiler_driver->GetCompilerOptions().GetExplicitNullChecks()) {
201    if (!(cu_->disable_opt & (1 << kNullCheckElimination)) && (opt_flags & MIR_IGNORE_NULL_CHECK)) {
202      return;
203    }
204    MarkSafepointPCAfter(after);
205  }
206}
207
208void Mir2Lir::MarkPossibleStackOverflowException() {
209  if (!cu_->compiler_driver->GetCompilerOptions().GetExplicitStackOverflowChecks()) {
210    MarkSafepointPC(last_lir_insn_);
211  }
212}
213
214void Mir2Lir::ForceImplicitNullCheck(RegStorage reg, int opt_flags) {
215  if (!cu_->compiler_driver->GetCompilerOptions().GetExplicitNullChecks()) {
216    if (!(cu_->disable_opt & (1 << kNullCheckElimination)) && (opt_flags & MIR_IGNORE_NULL_CHECK)) {
217      return;
218    }
219    // Force an implicit null check by performing a memory operation (load) from the given
220    // register with offset 0.  This will cause a signal if the register contains 0 (null).
221    RegStorage tmp = AllocTemp();
222    // TODO: for Mips, would be best to use rZERO as the bogus register target.
223    LIR* load = Load32Disp(reg, 0, tmp);
224    FreeTemp(tmp);
225    MarkSafepointPC(load);
226  }
227}
228
229void Mir2Lir::GenCompareAndBranch(Instruction::Code opcode, RegLocation rl_src1,
230                                  RegLocation rl_src2, LIR* taken,
231                                  LIR* fall_through) {
232  DCHECK(!rl_src1.fp);
233  DCHECK(!rl_src2.fp);
234  ConditionCode cond;
235  switch (opcode) {
236    case Instruction::IF_EQ:
237      cond = kCondEq;
238      break;
239    case Instruction::IF_NE:
240      cond = kCondNe;
241      break;
242    case Instruction::IF_LT:
243      cond = kCondLt;
244      break;
245    case Instruction::IF_GE:
246      cond = kCondGe;
247      break;
248    case Instruction::IF_GT:
249      cond = kCondGt;
250      break;
251    case Instruction::IF_LE:
252      cond = kCondLe;
253      break;
254    default:
255      cond = static_cast<ConditionCode>(0);
256      LOG(FATAL) << "Unexpected opcode " << opcode;
257  }
258
259  // Normalize such that if either operand is constant, src2 will be constant
260  if (rl_src1.is_const) {
261    RegLocation rl_temp = rl_src1;
262    rl_src1 = rl_src2;
263    rl_src2 = rl_temp;
264    cond = FlipComparisonOrder(cond);
265  }
266
267  rl_src1 = LoadValue(rl_src1);
268  // Is this really an immediate comparison?
269  if (rl_src2.is_const) {
270    // If it's already live in a register or not easily materialized, just keep going
271    RegLocation rl_temp = UpdateLoc(rl_src2);
272    if ((rl_temp.location == kLocDalvikFrame) &&
273        InexpensiveConstantInt(mir_graph_->ConstantValue(rl_src2))) {
274      // OK - convert this to a compare immediate and branch
275      OpCmpImmBranch(cond, rl_src1.reg, mir_graph_->ConstantValue(rl_src2), taken);
276      return;
277    }
278  }
279  rl_src2 = LoadValue(rl_src2);
280  OpCmpBranch(cond, rl_src1.reg, rl_src2.reg, taken);
281}
282
283void Mir2Lir::GenCompareZeroAndBranch(Instruction::Code opcode, RegLocation rl_src, LIR* taken,
284                                      LIR* fall_through) {
285  ConditionCode cond;
286  DCHECK(!rl_src.fp);
287  rl_src = LoadValue(rl_src);
288  switch (opcode) {
289    case Instruction::IF_EQZ:
290      cond = kCondEq;
291      break;
292    case Instruction::IF_NEZ:
293      cond = kCondNe;
294      break;
295    case Instruction::IF_LTZ:
296      cond = kCondLt;
297      break;
298    case Instruction::IF_GEZ:
299      cond = kCondGe;
300      break;
301    case Instruction::IF_GTZ:
302      cond = kCondGt;
303      break;
304    case Instruction::IF_LEZ:
305      cond = kCondLe;
306      break;
307    default:
308      cond = static_cast<ConditionCode>(0);
309      LOG(FATAL) << "Unexpected opcode " << opcode;
310  }
311  OpCmpImmBranch(cond, rl_src.reg, 0, taken);
312}
313
314void Mir2Lir::GenIntToLong(RegLocation rl_dest, RegLocation rl_src) {
315  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
316  if (rl_src.location == kLocPhysReg) {
317    OpRegCopy(rl_result.reg, rl_src.reg);
318  } else {
319    LoadValueDirect(rl_src, rl_result.reg.GetLow());
320  }
321  OpRegRegImm(kOpAsr, rl_result.reg.GetHigh(), rl_result.reg.GetLow(), 31);
322  StoreValueWide(rl_dest, rl_result);
323}
324
325void Mir2Lir::GenIntNarrowing(Instruction::Code opcode, RegLocation rl_dest,
326                              RegLocation rl_src) {
327  rl_src = LoadValue(rl_src, kCoreReg);
328  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
329  OpKind op = kOpInvalid;
330  switch (opcode) {
331    case Instruction::INT_TO_BYTE:
332      op = kOp2Byte;
333      break;
334    case Instruction::INT_TO_SHORT:
335       op = kOp2Short;
336       break;
337    case Instruction::INT_TO_CHAR:
338       op = kOp2Char;
339       break;
340    default:
341      LOG(ERROR) << "Bad int conversion type";
342  }
343  OpRegReg(op, rl_result.reg, rl_src.reg);
344  StoreValue(rl_dest, rl_result);
345}
346
347template <size_t pointer_size>
348static void GenNewArrayImpl(Mir2Lir* mir_to_lir, CompilationUnit* cu,
349                            uint32_t type_idx, RegLocation rl_dest,
350                            RegLocation rl_src) {
351  mir_to_lir->FlushAllRegs();  /* Everything to home location */
352  ThreadOffset<pointer_size> func_offset(-1);
353  const DexFile* dex_file = cu->dex_file;
354  CompilerDriver* driver = cu->compiler_driver;
355  if (cu->compiler_driver->CanAccessTypeWithoutChecks(cu->method_idx, *dex_file,
356                                                      type_idx)) {
357    bool is_type_initialized;  // Ignored as an array does not have an initializer.
358    bool use_direct_type_ptr;
359    uintptr_t direct_type_ptr;
360    bool is_finalizable;
361    if (kEmbedClassInCode &&
362        driver->CanEmbedTypeInCode(*dex_file, type_idx, &is_type_initialized, &use_direct_type_ptr,
363                                   &direct_type_ptr, &is_finalizable)) {
364      // The fast path.
365      if (!use_direct_type_ptr) {
366        mir_to_lir->LoadClassType(type_idx, kArg0);
367        func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pAllocArrayResolved);
368        mir_to_lir->CallRuntimeHelperRegMethodRegLocation(func_offset, mir_to_lir->TargetReg(kArg0),
369                                                          rl_src, true);
370      } else {
371        // Use the direct pointer.
372        func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pAllocArrayResolved);
373        mir_to_lir->CallRuntimeHelperImmMethodRegLocation(func_offset, direct_type_ptr, rl_src,
374                                                          true);
375      }
376    } else {
377      // The slow path.
378      func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pAllocArray);
379      mir_to_lir->CallRuntimeHelperImmMethodRegLocation(func_offset, type_idx, rl_src, true);
380    }
381    DCHECK_NE(func_offset.Int32Value(), -1);
382  } else {
383    func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pAllocArrayWithAccessCheck);
384    mir_to_lir->CallRuntimeHelperImmMethodRegLocation(func_offset, type_idx, rl_src, true);
385  }
386  RegLocation rl_result = mir_to_lir->GetReturn(kRefReg);
387  mir_to_lir->StoreValue(rl_dest, rl_result);
388}
389
390/*
391 * Let helper function take care of everything.  Will call
392 * Array::AllocFromCode(type_idx, method, count);
393 * Note: AllocFromCode will handle checks for errNegativeArraySize.
394 */
395void Mir2Lir::GenNewArray(uint32_t type_idx, RegLocation rl_dest,
396                          RegLocation rl_src) {
397  if (cu_->target64) {
398    GenNewArrayImpl<8>(this, cu_, type_idx, rl_dest, rl_src);
399  } else {
400    GenNewArrayImpl<4>(this, cu_, type_idx, rl_dest, rl_src);
401  }
402}
403
404template <size_t pointer_size>
405static void GenFilledNewArrayCall(Mir2Lir* mir_to_lir, CompilationUnit* cu, int elems, int type_idx) {
406  ThreadOffset<pointer_size> func_offset(-1);
407  if (cu->compiler_driver->CanAccessTypeWithoutChecks(cu->method_idx, *cu->dex_file,
408                                                      type_idx)) {
409    func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pCheckAndAllocArray);
410  } else {
411    func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pCheckAndAllocArrayWithAccessCheck);
412  }
413  mir_to_lir->CallRuntimeHelperImmMethodImm(func_offset, type_idx, elems, true);
414}
415
416/*
417 * Similar to GenNewArray, but with post-allocation initialization.
418 * Verifier guarantees we're dealing with an array class.  Current
419 * code throws runtime exception "bad Filled array req" for 'D' and 'J'.
420 * Current code also throws internal unimp if not 'L', '[' or 'I'.
421 */
422void Mir2Lir::GenFilledNewArray(CallInfo* info) {
423  int elems = info->num_arg_words;
424  int type_idx = info->index;
425  FlushAllRegs();  /* Everything to home location */
426  if (cu_->target64) {
427    GenFilledNewArrayCall<8>(this, cu_, elems, type_idx);
428  } else {
429    GenFilledNewArrayCall<4>(this, cu_, elems, type_idx);
430  }
431  FreeTemp(TargetReg(kArg2));
432  FreeTemp(TargetReg(kArg1));
433  /*
434   * NOTE: the implicit target for Instruction::FILLED_NEW_ARRAY is the
435   * return region.  Because AllocFromCode placed the new array
436   * in kRet0, we'll just lock it into place.  When debugger support is
437   * added, it may be necessary to additionally copy all return
438   * values to a home location in thread-local storage
439   */
440  LockTemp(TargetReg(kRet0));
441
442  // TODO: use the correct component size, currently all supported types
443  // share array alignment with ints (see comment at head of function)
444  size_t component_size = sizeof(int32_t);
445
446  // Having a range of 0 is legal
447  if (info->is_range && (elems > 0)) {
448    /*
449     * Bit of ugliness here.  We're going generate a mem copy loop
450     * on the register range, but it is possible that some regs
451     * in the range have been promoted.  This is unlikely, but
452     * before generating the copy, we'll just force a flush
453     * of any regs in the source range that have been promoted to
454     * home location.
455     */
456    for (int i = 0; i < elems; i++) {
457      RegLocation loc = UpdateLoc(info->args[i]);
458      if (loc.location == kLocPhysReg) {
459        ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
460        Store32Disp(TargetReg(kSp), SRegOffset(loc.s_reg_low), loc.reg);
461      }
462    }
463    /*
464     * TUNING note: generated code here could be much improved, but
465     * this is an uncommon operation and isn't especially performance
466     * critical.
467     */
468    // This is addressing the stack, which may be out of the 4G area.
469    RegStorage r_src = AllocTempRef();
470    RegStorage r_dst = AllocTempRef();
471    RegStorage r_idx = AllocTempRef();  // Not really a reference, but match src/dst.
472    RegStorage r_val;
473    switch (cu_->instruction_set) {
474      case kThumb2:
475      case kArm64:
476        r_val = TargetReg(kLr);
477        break;
478      case kX86:
479      case kX86_64:
480        FreeTemp(TargetReg(kRet0));
481        r_val = AllocTemp();
482        break;
483      case kMips:
484        r_val = AllocTemp();
485        break;
486      default: LOG(FATAL) << "Unexpected instruction set: " << cu_->instruction_set;
487    }
488    // Set up source pointer
489    RegLocation rl_first = info->args[0];
490    OpRegRegImm(kOpAdd, r_src, TargetReg(kSp), SRegOffset(rl_first.s_reg_low));
491    // Set up the target pointer
492    OpRegRegImm(kOpAdd, r_dst, TargetReg(kRet0),
493                mirror::Array::DataOffset(component_size).Int32Value());
494    // Set up the loop counter (known to be > 0)
495    LoadConstant(r_idx, elems - 1);
496    // Generate the copy loop.  Going backwards for convenience
497    LIR* target = NewLIR0(kPseudoTargetLabel);
498    // Copy next element
499    {
500      ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
501      LoadBaseIndexed(r_src, r_idx, r_val, 2, k32);
502      // NOTE: No dalvik register annotation, local optimizations will be stopped
503      // by the loop boundaries.
504    }
505    StoreBaseIndexed(r_dst, r_idx, r_val, 2, k32);
506    FreeTemp(r_val);
507    OpDecAndBranch(kCondGe, r_idx, target);
508    if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) {
509      // Restore the target pointer
510      OpRegRegImm(kOpAdd, TargetReg(kRet0), r_dst,
511                  -mirror::Array::DataOffset(component_size).Int32Value());
512    }
513  } else if (!info->is_range) {
514    // TUNING: interleave
515    for (int i = 0; i < elems; i++) {
516      RegLocation rl_arg = LoadValue(info->args[i], kCoreReg);
517      Store32Disp(TargetReg(kRet0),
518                  mirror::Array::DataOffset(component_size).Int32Value() + i * 4, rl_arg.reg);
519      // If the LoadValue caused a temp to be allocated, free it
520      if (IsTemp(rl_arg.reg)) {
521        FreeTemp(rl_arg.reg);
522      }
523    }
524  }
525  if (info->result.location != kLocInvalid) {
526    StoreValue(info->result, GetReturn(kRefReg));
527  }
528}
529
530//
531// Slow path to ensure a class is initialized for sget/sput.
532//
533class StaticFieldSlowPath : public Mir2Lir::LIRSlowPath {
534 public:
535  StaticFieldSlowPath(Mir2Lir* m2l, LIR* unresolved, LIR* uninit, LIR* cont, int storage_index,
536                      RegStorage r_base) :
537    LIRSlowPath(m2l, m2l->GetCurrentDexPc(), unresolved, cont), uninit_(uninit),
538               storage_index_(storage_index), r_base_(r_base) {
539  }
540
541  void Compile() {
542    LIR* unresolved_target = GenerateTargetLabel();
543    uninit_->target = unresolved_target;
544    if (cu_->target64) {
545      m2l_->CallRuntimeHelperImm(QUICK_ENTRYPOINT_OFFSET(8, pInitializeStaticStorage),
546                                 storage_index_, true);
547    } else {
548      m2l_->CallRuntimeHelperImm(QUICK_ENTRYPOINT_OFFSET(4, pInitializeStaticStorage),
549                                 storage_index_, true);
550    }
551    // Copy helper's result into r_base, a no-op on all but MIPS.
552    m2l_->OpRegCopy(r_base_,  m2l_->TargetReg(kRet0));
553
554    m2l_->OpUnconditionalBranch(cont_);
555  }
556
557 private:
558  LIR* const uninit_;
559  const int storage_index_;
560  const RegStorage r_base_;
561};
562
563template <size_t pointer_size>
564static void GenSputCall(Mir2Lir* mir_to_lir, bool is_long_or_double, bool is_object,
565                        const MirSFieldLoweringInfo* field_info, RegLocation rl_src) {
566  ThreadOffset<pointer_size> setter_offset =
567      is_long_or_double ? QUICK_ENTRYPOINT_OFFSET(pointer_size, pSet64Static)
568          : (is_object ? QUICK_ENTRYPOINT_OFFSET(pointer_size, pSetObjStatic)
569              : QUICK_ENTRYPOINT_OFFSET(pointer_size, pSet32Static));
570  mir_to_lir->CallRuntimeHelperImmRegLocation(setter_offset, field_info->FieldIndex(), rl_src,
571                                              true);
572}
573
574void Mir2Lir::GenSput(MIR* mir, RegLocation rl_src, bool is_long_or_double,
575                      bool is_object) {
576  const MirSFieldLoweringInfo& field_info = mir_graph_->GetSFieldLoweringInfo(mir);
577  cu_->compiler_driver->ProcessedStaticField(field_info.FastPut(), field_info.IsReferrersClass());
578  OpSize store_size = LoadStoreOpSize(is_long_or_double, is_object);
579  if (!SLOW_FIELD_PATH && field_info.FastPut() &&
580      (!field_info.IsVolatile() || SupportsVolatileLoadStore(store_size))) {
581    DCHECK_GE(field_info.FieldOffset().Int32Value(), 0);
582    RegStorage r_base;
583    if (field_info.IsReferrersClass()) {
584      // Fast path, static storage base is this method's class
585      RegLocation rl_method = LoadCurrMethod();
586      r_base = AllocTempRef();
587      LoadRefDisp(rl_method.reg, mirror::ArtMethod::DeclaringClassOffset().Int32Value(), r_base,
588                  kNotVolatile);
589      if (IsTemp(rl_method.reg)) {
590        FreeTemp(rl_method.reg);
591      }
592    } else {
593      // Medium path, static storage base in a different class which requires checks that the other
594      // class is initialized.
595      // TODO: remove initialized check now that we are initializing classes in the compiler driver.
596      DCHECK_NE(field_info.StorageIndex(), DexFile::kDexNoIndex);
597      // May do runtime call so everything to home locations.
598      FlushAllRegs();
599      // Using fixed register to sync with possible call to runtime support.
600      RegStorage r_method = TargetReg(kArg1);
601      LockTemp(r_method);
602      LoadCurrMethodDirect(r_method);
603      r_base = TargetReg(kArg0);
604      LockTemp(r_base);
605      LoadRefDisp(r_method, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(), r_base,
606                  kNotVolatile);
607      int32_t offset_of_field = ObjArray::OffsetOfElement(field_info.StorageIndex()).Int32Value();
608      LoadRefDisp(r_base, offset_of_field, r_base, kNotVolatile);
609      // r_base now points at static storage (Class*) or NULL if the type is not yet resolved.
610      if (!field_info.IsInitialized() &&
611          (mir->optimization_flags & MIR_IGNORE_CLINIT_CHECK) == 0) {
612        // Check if r_base is NULL or a not yet initialized class.
613
614        // The slow path is invoked if the r_base is NULL or the class pointed
615        // to by it is not initialized.
616        LIR* unresolved_branch = OpCmpImmBranch(kCondEq, r_base, 0, NULL);
617        RegStorage r_tmp = TargetReg(kArg2);
618        LockTemp(r_tmp);
619        LIR* uninit_branch = OpCmpMemImmBranch(kCondLt, r_tmp, r_base,
620                                          mirror::Class::StatusOffset().Int32Value(),
621                                          mirror::Class::kStatusInitialized, NULL);
622        LIR* cont = NewLIR0(kPseudoTargetLabel);
623
624        AddSlowPath(new (arena_) StaticFieldSlowPath(this, unresolved_branch, uninit_branch, cont,
625                                                     field_info.StorageIndex(), r_base));
626
627        FreeTemp(r_tmp);
628        // Ensure load of status and load of value don't re-order.
629        GenMemBarrier(kLoadLoad);
630      }
631      FreeTemp(r_method);
632    }
633    // rBase now holds static storage base
634    RegisterClass reg_class = RegClassForFieldLoadStore(store_size, field_info.IsVolatile());
635    if (is_long_or_double) {
636      rl_src = LoadValueWide(rl_src, reg_class);
637    } else {
638      rl_src = LoadValue(rl_src, reg_class);
639    }
640    if (is_object) {
641      StoreRefDisp(r_base, field_info.FieldOffset().Int32Value(), rl_src.reg,
642                   field_info.IsVolatile() ? kVolatile : kNotVolatile);
643    } else {
644      StoreBaseDisp(r_base, field_info.FieldOffset().Int32Value(), rl_src.reg, store_size,
645                    field_info.IsVolatile() ? kVolatile : kNotVolatile);
646    }
647    if (is_object && !mir_graph_->IsConstantNullRef(rl_src)) {
648      MarkGCCard(rl_src.reg, r_base);
649    }
650    FreeTemp(r_base);
651  } else {
652    FlushAllRegs();  // Everything to home locations
653    if (cu_->target64) {
654      GenSputCall<8>(this, is_long_or_double, is_object, &field_info, rl_src);
655    } else {
656      GenSputCall<4>(this, is_long_or_double, is_object, &field_info, rl_src);
657    }
658  }
659}
660
661template <size_t pointer_size>
662static void GenSgetCall(Mir2Lir* mir_to_lir, bool is_long_or_double, bool is_object,
663                        const MirSFieldLoweringInfo* field_info) {
664  ThreadOffset<pointer_size> getter_offset =
665      is_long_or_double ? QUICK_ENTRYPOINT_OFFSET(pointer_size, pGet64Static)
666          : (is_object ? QUICK_ENTRYPOINT_OFFSET(pointer_size, pGetObjStatic)
667              : QUICK_ENTRYPOINT_OFFSET(pointer_size, pGet32Static));
668  mir_to_lir->CallRuntimeHelperImm(getter_offset, field_info->FieldIndex(), true);
669}
670
671void Mir2Lir::GenSget(MIR* mir, RegLocation rl_dest,
672                      bool is_long_or_double, bool is_object) {
673  const MirSFieldLoweringInfo& field_info = mir_graph_->GetSFieldLoweringInfo(mir);
674  cu_->compiler_driver->ProcessedStaticField(field_info.FastGet(), field_info.IsReferrersClass());
675  OpSize load_size = LoadStoreOpSize(is_long_or_double, is_object);
676  if (!SLOW_FIELD_PATH && field_info.FastGet() &&
677      (!field_info.IsVolatile() || SupportsVolatileLoadStore(load_size))) {
678    DCHECK_GE(field_info.FieldOffset().Int32Value(), 0);
679    RegStorage r_base;
680    if (field_info.IsReferrersClass()) {
681      // Fast path, static storage base is this method's class
682      RegLocation rl_method  = LoadCurrMethod();
683      r_base = AllocTempRef();
684      LoadRefDisp(rl_method.reg, mirror::ArtMethod::DeclaringClassOffset().Int32Value(), r_base,
685                  kNotVolatile);
686    } else {
687      // Medium path, static storage base in a different class which requires checks that the other
688      // class is initialized
689      DCHECK_NE(field_info.StorageIndex(), DexFile::kDexNoIndex);
690      // May do runtime call so everything to home locations.
691      FlushAllRegs();
692      // Using fixed register to sync with possible call to runtime support.
693      RegStorage r_method = TargetReg(kArg1);
694      LockTemp(r_method);
695      LoadCurrMethodDirect(r_method);
696      r_base = TargetReg(kArg0);
697      LockTemp(r_base);
698      LoadRefDisp(r_method, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(), r_base,
699                  kNotVolatile);
700      int32_t offset_of_field = ObjArray::OffsetOfElement(field_info.StorageIndex()).Int32Value();
701      LoadRefDisp(r_base, offset_of_field, r_base, kNotVolatile);
702      // r_base now points at static storage (Class*) or NULL if the type is not yet resolved.
703      if (!field_info.IsInitialized() &&
704          (mir->optimization_flags & MIR_IGNORE_CLINIT_CHECK) == 0) {
705        // Check if r_base is NULL or a not yet initialized class.
706
707        // The slow path is invoked if the r_base is NULL or the class pointed
708        // to by it is not initialized.
709        LIR* unresolved_branch = OpCmpImmBranch(kCondEq, r_base, 0, NULL);
710        RegStorage r_tmp = TargetReg(kArg2);
711        LockTemp(r_tmp);
712        LIR* uninit_branch = OpCmpMemImmBranch(kCondLt, r_tmp, r_base,
713                                          mirror::Class::StatusOffset().Int32Value(),
714                                          mirror::Class::kStatusInitialized, NULL);
715        LIR* cont = NewLIR0(kPseudoTargetLabel);
716
717        AddSlowPath(new (arena_) StaticFieldSlowPath(this, unresolved_branch, uninit_branch, cont,
718                                                     field_info.StorageIndex(), r_base));
719
720        FreeTemp(r_tmp);
721        // Ensure load of status and load of value don't re-order.
722        GenMemBarrier(kLoadLoad);
723      }
724      FreeTemp(r_method);
725    }
726    // r_base now holds static storage base
727    RegisterClass reg_class = RegClassForFieldLoadStore(load_size, field_info.IsVolatile());
728    RegLocation rl_result = EvalLoc(rl_dest, reg_class, true);
729
730    int field_offset = field_info.FieldOffset().Int32Value();
731    if (is_object) {
732      LoadRefDisp(r_base, field_offset, rl_result.reg, field_info.IsVolatile() ? kVolatile :
733          kNotVolatile);
734    } else {
735      LoadBaseDisp(r_base, field_offset, rl_result.reg, load_size, field_info.IsVolatile() ?
736          kVolatile : kNotVolatile);
737    }
738    FreeTemp(r_base);
739
740    if (is_long_or_double) {
741      StoreValueWide(rl_dest, rl_result);
742    } else {
743      StoreValue(rl_dest, rl_result);
744    }
745  } else {
746    FlushAllRegs();  // Everything to home locations
747    if (cu_->target64) {
748      GenSgetCall<8>(this, is_long_or_double, is_object, &field_info);
749    } else {
750      GenSgetCall<4>(this, is_long_or_double, is_object, &field_info);
751    }
752    if (is_long_or_double) {
753      RegLocation rl_result = GetReturnWide(LocToRegClass(rl_dest));
754      StoreValueWide(rl_dest, rl_result);
755    } else {
756      RegLocation rl_result = GetReturn(LocToRegClass(rl_dest));
757      StoreValue(rl_dest, rl_result);
758    }
759  }
760}
761
762// Generate code for all slow paths.
763void Mir2Lir::HandleSlowPaths() {
764  int n = slow_paths_.Size();
765  for (int i = 0; i < n; ++i) {
766    LIRSlowPath* slowpath = slow_paths_.Get(i);
767    slowpath->Compile();
768  }
769  slow_paths_.Reset();
770}
771
772template <size_t pointer_size>
773static void GenIgetCall(Mir2Lir* mir_to_lir, bool is_long_or_double, bool is_object,
774                        const MirIFieldLoweringInfo* field_info, RegLocation rl_obj) {
775  ThreadOffset<pointer_size> getter_offset =
776      is_long_or_double ? QUICK_ENTRYPOINT_OFFSET(pointer_size, pGet64Instance)
777          : (is_object ? QUICK_ENTRYPOINT_OFFSET(pointer_size, pGetObjInstance)
778              : QUICK_ENTRYPOINT_OFFSET(pointer_size, pGet32Instance));
779  mir_to_lir->CallRuntimeHelperImmRegLocation(getter_offset, field_info->FieldIndex(), rl_obj,
780                                              true);
781}
782
783void Mir2Lir::GenIGet(MIR* mir, int opt_flags, OpSize size,
784                      RegLocation rl_dest, RegLocation rl_obj, bool is_long_or_double,
785                      bool is_object) {
786  const MirIFieldLoweringInfo& field_info = mir_graph_->GetIFieldLoweringInfo(mir);
787  cu_->compiler_driver->ProcessedInstanceField(field_info.FastGet());
788  OpSize load_size = LoadStoreOpSize(is_long_or_double, is_object);
789  if (!SLOW_FIELD_PATH && field_info.FastGet() &&
790      (!field_info.IsVolatile() || SupportsVolatileLoadStore(load_size))) {
791    RegisterClass reg_class = RegClassForFieldLoadStore(load_size, field_info.IsVolatile());
792    DCHECK_GE(field_info.FieldOffset().Int32Value(), 0);
793    rl_obj = LoadValue(rl_obj, kRefReg);
794    GenNullCheck(rl_obj.reg, opt_flags);
795    RegLocation rl_result = EvalLoc(rl_dest, reg_class, true);
796    int field_offset = field_info.FieldOffset().Int32Value();
797    LIR* load_lir;
798    if (is_object) {
799      load_lir = LoadRefDisp(rl_obj.reg, field_offset, rl_result.reg, field_info.IsVolatile() ?
800          kVolatile : kNotVolatile);
801    } else {
802      load_lir = LoadBaseDisp(rl_obj.reg, field_offset, rl_result.reg, load_size,
803                              field_info.IsVolatile() ? kVolatile : kNotVolatile);
804    }
805    MarkPossibleNullPointerExceptionAfter(opt_flags, load_lir);
806    if (is_long_or_double) {
807      StoreValueWide(rl_dest, rl_result);
808    } else {
809      StoreValue(rl_dest, rl_result);
810    }
811  } else {
812    if (cu_->target64) {
813      GenIgetCall<8>(this, is_long_or_double, is_object, &field_info, rl_obj);
814    } else {
815      GenIgetCall<4>(this, is_long_or_double, is_object, &field_info, rl_obj);
816    }
817    if (is_long_or_double) {
818      RegLocation rl_result = GetReturnWide(LocToRegClass(rl_dest));
819      StoreValueWide(rl_dest, rl_result);
820    } else {
821      RegLocation rl_result = GetReturn(LocToRegClass(rl_dest));
822      StoreValue(rl_dest, rl_result);
823    }
824  }
825}
826
827template <size_t pointer_size>
828static void GenIputCall(Mir2Lir* mir_to_lir, bool is_long_or_double, bool is_object,
829                        const MirIFieldLoweringInfo* field_info, RegLocation rl_obj,
830                        RegLocation rl_src) {
831  ThreadOffset<pointer_size> setter_offset =
832      is_long_or_double ? QUICK_ENTRYPOINT_OFFSET(pointer_size, pSet64Instance)
833          : (is_object ? QUICK_ENTRYPOINT_OFFSET(pointer_size, pSetObjInstance)
834              : QUICK_ENTRYPOINT_OFFSET(pointer_size, pSet32Instance));
835  mir_to_lir->CallRuntimeHelperImmRegLocationRegLocation(setter_offset, field_info->FieldIndex(),
836                                                         rl_obj, rl_src, true);
837}
838
839void Mir2Lir::GenIPut(MIR* mir, int opt_flags, OpSize size,
840                      RegLocation rl_src, RegLocation rl_obj, bool is_long_or_double,
841                      bool is_object) {
842  const MirIFieldLoweringInfo& field_info = mir_graph_->GetIFieldLoweringInfo(mir);
843  cu_->compiler_driver->ProcessedInstanceField(field_info.FastPut());
844  OpSize store_size = LoadStoreOpSize(is_long_or_double, is_object);
845  if (!SLOW_FIELD_PATH && field_info.FastPut() &&
846      (!field_info.IsVolatile() || SupportsVolatileLoadStore(store_size))) {
847    RegisterClass reg_class = RegClassForFieldLoadStore(store_size, field_info.IsVolatile());
848    DCHECK_GE(field_info.FieldOffset().Int32Value(), 0);
849    rl_obj = LoadValue(rl_obj, kRefReg);
850    if (is_long_or_double) {
851      rl_src = LoadValueWide(rl_src, reg_class);
852    } else {
853      rl_src = LoadValue(rl_src, reg_class);
854    }
855    GenNullCheck(rl_obj.reg, opt_flags);
856    int field_offset = field_info.FieldOffset().Int32Value();
857    LIR* store;
858    if (is_object) {
859      store = StoreRefDisp(rl_obj.reg, field_offset, rl_src.reg, field_info.IsVolatile() ?
860          kVolatile : kNotVolatile);
861    } else {
862      store = StoreBaseDisp(rl_obj.reg, field_offset, rl_src.reg, store_size,
863                            field_info.IsVolatile() ? kVolatile : kNotVolatile);
864    }
865    MarkPossibleNullPointerExceptionAfter(opt_flags, store);
866    if (is_object && !mir_graph_->IsConstantNullRef(rl_src)) {
867      MarkGCCard(rl_src.reg, rl_obj.reg);
868    }
869  } else {
870    if (cu_->target64) {
871      GenIputCall<8>(this, is_long_or_double, is_object, &field_info, rl_obj, rl_src);
872    } else {
873      GenIputCall<4>(this, is_long_or_double, is_object, &field_info, rl_obj, rl_src);
874    }
875  }
876}
877
878template <size_t pointer_size>
879static void GenArrayObjPutCall(Mir2Lir* mir_to_lir, bool needs_range_check, bool needs_null_check,
880                               RegLocation rl_array, RegLocation rl_index, RegLocation rl_src) {
881  ThreadOffset<pointer_size> helper = needs_range_check
882        ? (needs_null_check ? QUICK_ENTRYPOINT_OFFSET(pointer_size, pAputObjectWithNullAndBoundCheck)
883                            : QUICK_ENTRYPOINT_OFFSET(pointer_size, pAputObjectWithBoundCheck))
884        : QUICK_ENTRYPOINT_OFFSET(pointer_size, pAputObject);
885  mir_to_lir->CallRuntimeHelperRegLocationRegLocationRegLocation(helper, rl_array, rl_index, rl_src,
886                                                                 true);
887}
888
889void Mir2Lir::GenArrayObjPut(int opt_flags, RegLocation rl_array, RegLocation rl_index,
890                             RegLocation rl_src) {
891  bool needs_range_check = !(opt_flags & MIR_IGNORE_RANGE_CHECK);
892  bool needs_null_check = !((cu_->disable_opt & (1 << kNullCheckElimination)) &&
893      (opt_flags & MIR_IGNORE_NULL_CHECK));
894  if (cu_->target64) {
895    GenArrayObjPutCall<8>(this, needs_range_check, needs_null_check, rl_array, rl_index, rl_src);
896  } else {
897    GenArrayObjPutCall<4>(this, needs_range_check, needs_null_check, rl_array, rl_index, rl_src);
898  }
899}
900
901void Mir2Lir::GenConstClass(uint32_t type_idx, RegLocation rl_dest) {
902  RegLocation rl_method = LoadCurrMethod();
903  DCHECK(!cu_->target64 || rl_method.reg.Is64Bit());
904  RegStorage res_reg = AllocTempRef();
905  RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true);
906  if (!cu_->compiler_driver->CanAccessTypeWithoutChecks(cu_->method_idx,
907                                                   *cu_->dex_file,
908                                                   type_idx)) {
909    // Call out to helper which resolves type and verifies access.
910    // Resolved type returned in kRet0.
911    if (cu_->target64) {
912      CallRuntimeHelperImmReg(QUICK_ENTRYPOINT_OFFSET(8, pInitializeTypeAndVerifyAccess),
913                              type_idx, rl_method.reg, true);
914    } else {
915      CallRuntimeHelperImmReg(QUICK_ENTRYPOINT_OFFSET(4, pInitializeTypeAndVerifyAccess),
916                              type_idx, rl_method.reg, true);
917    }
918    RegLocation rl_result = GetReturn(kRefReg);
919    StoreValue(rl_dest, rl_result);
920  } else {
921    // We're don't need access checks, load type from dex cache
922    int32_t dex_cache_offset =
923        mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value();
924    LoadRefDisp(rl_method.reg, dex_cache_offset, res_reg, kNotVolatile);
925    int32_t offset_of_type = ClassArray::OffsetOfElement(type_idx).Int32Value();
926    LoadRefDisp(res_reg, offset_of_type, rl_result.reg, kNotVolatile);
927    if (!cu_->compiler_driver->CanAssumeTypeIsPresentInDexCache(*cu_->dex_file,
928        type_idx) || SLOW_TYPE_PATH) {
929      // Slow path, at runtime test if type is null and if so initialize
930      FlushAllRegs();
931      LIR* branch = OpCmpImmBranch(kCondEq, rl_result.reg, 0, NULL);
932      LIR* cont = NewLIR0(kPseudoTargetLabel);
933
934      // Object to generate the slow path for class resolution.
935      class SlowPath : public LIRSlowPath {
936       public:
937        SlowPath(Mir2Lir* m2l, LIR* fromfast, LIR* cont, const int type_idx,
938                 const RegLocation& rl_method, const RegLocation& rl_result) :
939                   LIRSlowPath(m2l, m2l->GetCurrentDexPc(), fromfast, cont), type_idx_(type_idx),
940                   rl_method_(rl_method), rl_result_(rl_result) {
941        }
942
943        void Compile() {
944          GenerateTargetLabel();
945
946          if (cu_->target64) {
947            m2l_->CallRuntimeHelperImmReg(QUICK_ENTRYPOINT_OFFSET(8, pInitializeType), type_idx_,
948                                          rl_method_.reg, true);
949          } else {
950            m2l_->CallRuntimeHelperImmReg(QUICK_ENTRYPOINT_OFFSET(4, pInitializeType), type_idx_,
951                                                      rl_method_.reg, true);
952          }
953          m2l_->OpRegCopy(rl_result_.reg,  m2l_->TargetReg(kRet0));
954
955          m2l_->OpUnconditionalBranch(cont_);
956        }
957
958       private:
959        const int type_idx_;
960        const RegLocation rl_method_;
961        const RegLocation rl_result_;
962      };
963
964      // Add to list for future.
965      AddSlowPath(new (arena_) SlowPath(this, branch, cont, type_idx, rl_method, rl_result));
966
967      StoreValue(rl_dest, rl_result);
968     } else {
969      // Fast path, we're done - just store result
970      StoreValue(rl_dest, rl_result);
971    }
972  }
973}
974
975void Mir2Lir::GenConstString(uint32_t string_idx, RegLocation rl_dest) {
976  /* NOTE: Most strings should be available at compile time */
977  int32_t offset_of_string = mirror::ObjectArray<mirror::String>::OffsetOfElement(string_idx).
978                                                                                      Int32Value();
979  if (!cu_->compiler_driver->CanAssumeStringIsPresentInDexCache(
980      *cu_->dex_file, string_idx) || SLOW_STRING_PATH) {
981    // slow path, resolve string if not in dex cache
982    FlushAllRegs();
983    LockCallTemps();  // Using explicit registers
984
985    // If the Method* is already in a register, we can save a copy.
986    RegLocation rl_method = mir_graph_->GetMethodLoc();
987    RegStorage r_method;
988    if (rl_method.location == kLocPhysReg) {
989      // A temp would conflict with register use below.
990      DCHECK(!IsTemp(rl_method.reg));
991      r_method = rl_method.reg;
992    } else {
993      r_method = TargetReg(kArg2);
994      LoadCurrMethodDirect(r_method);
995    }
996    LoadRefDisp(r_method, mirror::ArtMethod::DexCacheStringsOffset().Int32Value(),
997                TargetReg(kArg0), kNotVolatile);
998
999    // Might call out to helper, which will return resolved string in kRet0
1000    LoadRefDisp(TargetReg(kArg0), offset_of_string, TargetReg(kRet0), kNotVolatile);
1001    LIR* fromfast = OpCmpImmBranch(kCondEq, TargetReg(kRet0), 0, NULL);
1002    LIR* cont = NewLIR0(kPseudoTargetLabel);
1003
1004    {
1005      // Object to generate the slow path for string resolution.
1006      class SlowPath : public LIRSlowPath {
1007       public:
1008        SlowPath(Mir2Lir* m2l, LIR* fromfast, LIR* cont, RegStorage r_method, int32_t string_idx) :
1009          LIRSlowPath(m2l, m2l->GetCurrentDexPc(), fromfast, cont),
1010          r_method_(r_method), string_idx_(string_idx) {
1011        }
1012
1013        void Compile() {
1014          GenerateTargetLabel();
1015          if (cu_->target64) {
1016            m2l_->CallRuntimeHelperRegImm(QUICK_ENTRYPOINT_OFFSET(8, pResolveString),
1017                                          r_method_, string_idx_, true);
1018          } else {
1019            m2l_->CallRuntimeHelperRegImm(QUICK_ENTRYPOINT_OFFSET(4, pResolveString),
1020                                          r_method_, string_idx_, true);
1021          }
1022          m2l_->OpUnconditionalBranch(cont_);
1023        }
1024
1025       private:
1026         const RegStorage r_method_;
1027         const int32_t string_idx_;
1028      };
1029
1030      AddSlowPath(new (arena_) SlowPath(this, fromfast, cont, r_method, string_idx));
1031    }
1032
1033    GenBarrier();
1034    StoreValue(rl_dest, GetReturn(kRefReg));
1035  } else {
1036    RegLocation rl_method = LoadCurrMethod();
1037    RegStorage res_reg = AllocTempRef();
1038    RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true);
1039    LoadRefDisp(rl_method.reg, mirror::ArtMethod::DexCacheStringsOffset().Int32Value(), res_reg,
1040                kNotVolatile);
1041    LoadRefDisp(res_reg, offset_of_string, rl_result.reg, kNotVolatile);
1042    StoreValue(rl_dest, rl_result);
1043  }
1044}
1045
1046template <size_t pointer_size>
1047static void GenNewInstanceImpl(Mir2Lir* mir_to_lir, CompilationUnit* cu, uint32_t type_idx,
1048                               RegLocation rl_dest) {
1049  mir_to_lir->FlushAllRegs();  /* Everything to home location */
1050  // alloc will always check for resolution, do we also need to verify
1051  // access because the verifier was unable to?
1052  ThreadOffset<pointer_size> func_offset(-1);
1053  const DexFile* dex_file = cu->dex_file;
1054  CompilerDriver* driver = cu->compiler_driver;
1055  if (driver->CanAccessInstantiableTypeWithoutChecks(
1056      cu->method_idx, *dex_file, type_idx)) {
1057    bool is_type_initialized;
1058    bool use_direct_type_ptr;
1059    uintptr_t direct_type_ptr;
1060    bool is_finalizable;
1061    if (kEmbedClassInCode &&
1062        driver->CanEmbedTypeInCode(*dex_file, type_idx, &is_type_initialized, &use_direct_type_ptr,
1063                                   &direct_type_ptr, &is_finalizable) &&
1064                                   !is_finalizable) {
1065      // The fast path.
1066      if (!use_direct_type_ptr) {
1067        mir_to_lir->LoadClassType(type_idx, kArg0);
1068        if (!is_type_initialized) {
1069          func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pAllocObjectResolved);
1070          mir_to_lir->CallRuntimeHelperRegMethod(func_offset, mir_to_lir->TargetReg(kArg0), true);
1071        } else {
1072          func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pAllocObjectInitialized);
1073          mir_to_lir->CallRuntimeHelperRegMethod(func_offset, mir_to_lir->TargetReg(kArg0), true);
1074        }
1075      } else {
1076        // Use the direct pointer.
1077        if (!is_type_initialized) {
1078          func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pAllocObjectResolved);
1079          mir_to_lir->CallRuntimeHelperImmMethod(func_offset, direct_type_ptr, true);
1080        } else {
1081          func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pAllocObjectInitialized);
1082          mir_to_lir->CallRuntimeHelperImmMethod(func_offset, direct_type_ptr, true);
1083        }
1084      }
1085    } else {
1086      // The slow path.
1087      DCHECK_EQ(func_offset.Int32Value(), -1);
1088      func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pAllocObject);
1089      mir_to_lir->CallRuntimeHelperImmMethod(func_offset, type_idx, true);
1090    }
1091    DCHECK_NE(func_offset.Int32Value(), -1);
1092  } else {
1093    func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pAllocObjectWithAccessCheck);
1094    mir_to_lir->CallRuntimeHelperImmMethod(func_offset, type_idx, true);
1095  }
1096  RegLocation rl_result = mir_to_lir->GetReturn(kRefReg);
1097  mir_to_lir->StoreValue(rl_dest, rl_result);
1098}
1099
1100/*
1101 * Let helper function take care of everything.  Will
1102 * call Class::NewInstanceFromCode(type_idx, method);
1103 */
1104void Mir2Lir::GenNewInstance(uint32_t type_idx, RegLocation rl_dest) {
1105  if (cu_->target64) {
1106    GenNewInstanceImpl<8>(this, cu_, type_idx, rl_dest);
1107  } else {
1108    GenNewInstanceImpl<4>(this, cu_, type_idx, rl_dest);
1109  }
1110}
1111
1112void Mir2Lir::GenThrow(RegLocation rl_src) {
1113  FlushAllRegs();
1114  if (cu_->target64) {
1115    CallRuntimeHelperRegLocation(QUICK_ENTRYPOINT_OFFSET(8, pDeliverException), rl_src, true);
1116  } else {
1117    CallRuntimeHelperRegLocation(QUICK_ENTRYPOINT_OFFSET(4, pDeliverException), rl_src, true);
1118  }
1119}
1120
1121// For final classes there are no sub-classes to check and so we can answer the instance-of
1122// question with simple comparisons.
1123void Mir2Lir::GenInstanceofFinal(bool use_declaring_class, uint32_t type_idx, RegLocation rl_dest,
1124                                 RegLocation rl_src) {
1125  // X86 has its own implementation.
1126  DCHECK(cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64);
1127
1128  RegLocation object = LoadValue(rl_src, kRefReg);
1129  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1130  RegStorage result_reg = rl_result.reg;
1131  if (result_reg == object.reg) {
1132    result_reg = AllocTypedTemp(false, kCoreReg);
1133  }
1134  LoadConstant(result_reg, 0);     // assume false
1135  LIR* null_branchover = OpCmpImmBranch(kCondEq, object.reg, 0, NULL);
1136
1137  RegStorage check_class = AllocTypedTemp(false, kRefReg);
1138  RegStorage object_class = AllocTypedTemp(false, kRefReg);
1139
1140  LoadCurrMethodDirect(check_class);
1141  if (use_declaring_class) {
1142    LoadRefDisp(check_class, mirror::ArtMethod::DeclaringClassOffset().Int32Value(), check_class,
1143                kNotVolatile);
1144    LoadRefDisp(object.reg,  mirror::Object::ClassOffset().Int32Value(), object_class,
1145                kNotVolatile);
1146  } else {
1147    LoadRefDisp(check_class, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(),
1148                check_class, kNotVolatile);
1149    LoadRefDisp(object.reg,  mirror::Object::ClassOffset().Int32Value(), object_class,
1150                kNotVolatile);
1151    int32_t offset_of_type = ClassArray::OffsetOfElement(type_idx).Int32Value();
1152    LoadRefDisp(check_class, offset_of_type, check_class, kNotVolatile);
1153  }
1154
1155  LIR* ne_branchover = NULL;
1156  // FIXME: what should we be comparing here? compressed or decompressed references?
1157  if (cu_->instruction_set == kThumb2) {
1158    OpRegReg(kOpCmp, check_class, object_class);  // Same?
1159    LIR* it = OpIT(kCondEq, "");   // if-convert the test
1160    LoadConstant(result_reg, 1);     // .eq case - load true
1161    OpEndIT(it);
1162  } else {
1163    ne_branchover = OpCmpBranch(kCondNe, check_class, object_class, NULL);
1164    LoadConstant(result_reg, 1);     // eq case - load true
1165  }
1166  LIR* target = NewLIR0(kPseudoTargetLabel);
1167  null_branchover->target = target;
1168  if (ne_branchover != NULL) {
1169    ne_branchover->target = target;
1170  }
1171  FreeTemp(object_class);
1172  FreeTemp(check_class);
1173  if (IsTemp(result_reg)) {
1174    OpRegCopy(rl_result.reg, result_reg);
1175    FreeTemp(result_reg);
1176  }
1177  StoreValue(rl_dest, rl_result);
1178}
1179
1180void Mir2Lir::GenInstanceofCallingHelper(bool needs_access_check, bool type_known_final,
1181                                         bool type_known_abstract, bool use_declaring_class,
1182                                         bool can_assume_type_is_in_dex_cache,
1183                                         uint32_t type_idx, RegLocation rl_dest,
1184                                         RegLocation rl_src) {
1185  // X86 has its own implementation.
1186  DCHECK(cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64);
1187
1188  FlushAllRegs();
1189  // May generate a call - use explicit registers
1190  LockCallTemps();
1191  LoadCurrMethodDirect(TargetReg(kArg1));  // kArg1 <= current Method*
1192  RegStorage class_reg = TargetReg(kArg2);  // kArg2 will hold the Class*
1193  if (needs_access_check) {
1194    // Check we have access to type_idx and if not throw IllegalAccessError,
1195    // returns Class* in kArg0
1196    if (cu_->target64) {
1197      CallRuntimeHelperImm(QUICK_ENTRYPOINT_OFFSET(8, pInitializeTypeAndVerifyAccess),
1198                           type_idx, true);
1199    } else {
1200      CallRuntimeHelperImm(QUICK_ENTRYPOINT_OFFSET(4, pInitializeTypeAndVerifyAccess),
1201                           type_idx, true);
1202    }
1203    OpRegCopy(class_reg, TargetReg(kRet0));  // Align usage with fast path
1204    LoadValueDirectFixed(rl_src, TargetReg(kArg0));  // kArg0 <= ref
1205  } else if (use_declaring_class) {
1206    LoadValueDirectFixed(rl_src, TargetReg(kArg0));  // kArg0 <= ref
1207    LoadRefDisp(TargetReg(kArg1), mirror::ArtMethod::DeclaringClassOffset().Int32Value(),
1208                class_reg, kNotVolatile);
1209  } else {
1210    // Load dex cache entry into class_reg (kArg2)
1211    LoadValueDirectFixed(rl_src, TargetReg(kArg0));  // kArg0 <= ref
1212    LoadRefDisp(TargetReg(kArg1), mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(),
1213                class_reg, kNotVolatile);
1214    int32_t offset_of_type = ClassArray::OffsetOfElement(type_idx).Int32Value();
1215    LoadRefDisp(class_reg, offset_of_type, class_reg, kNotVolatile);
1216    if (!can_assume_type_is_in_dex_cache) {
1217      // Need to test presence of type in dex cache at runtime
1218      LIR* hop_branch = OpCmpImmBranch(kCondNe, class_reg, 0, NULL);
1219      // Not resolved
1220      // Call out to helper, which will return resolved type in kRet0
1221      if (cu_->target64) {
1222        CallRuntimeHelperImm(QUICK_ENTRYPOINT_OFFSET(8, pInitializeType), type_idx, true);
1223      } else {
1224        CallRuntimeHelperImm(QUICK_ENTRYPOINT_OFFSET(4, pInitializeType), type_idx, true);
1225      }
1226      OpRegCopy(TargetReg(kArg2), TargetReg(kRet0));  // Align usage with fast path
1227      LoadValueDirectFixed(rl_src, TargetReg(kArg0));  /* reload Ref */
1228      // Rejoin code paths
1229      LIR* hop_target = NewLIR0(kPseudoTargetLabel);
1230      hop_branch->target = hop_target;
1231    }
1232  }
1233  /* kArg0 is ref, kArg2 is class. If ref==null, use directly as bool result */
1234  RegLocation rl_result = GetReturn(kRefReg);
1235  if (cu_->instruction_set == kMips) {
1236    // On MIPS rArg0 != rl_result, place false in result if branch is taken.
1237    LoadConstant(rl_result.reg, 0);
1238  }
1239  LIR* branch1 = OpCmpImmBranch(kCondEq, TargetReg(kArg0), 0, NULL);
1240
1241  /* load object->klass_ */
1242  DCHECK_EQ(mirror::Object::ClassOffset().Int32Value(), 0);
1243  LoadRefDisp(TargetReg(kArg0), mirror::Object::ClassOffset().Int32Value(), TargetReg(kArg1),
1244              kNotVolatile);
1245  /* kArg0 is ref, kArg1 is ref->klass_, kArg2 is class */
1246  LIR* branchover = NULL;
1247  if (type_known_final) {
1248    // rl_result == ref == null == 0.
1249    if (cu_->instruction_set == kThumb2) {
1250      OpRegReg(kOpCmp, TargetReg(kArg1), TargetReg(kArg2));  // Same?
1251      LIR* it = OpIT(kCondEq, "E");   // if-convert the test
1252      LoadConstant(rl_result.reg, 1);     // .eq case - load true
1253      LoadConstant(rl_result.reg, 0);     // .ne case - load false
1254      OpEndIT(it);
1255    } else {
1256      LoadConstant(rl_result.reg, 0);     // ne case - load false
1257      branchover = OpCmpBranch(kCondNe, TargetReg(kArg1), TargetReg(kArg2), NULL);
1258      LoadConstant(rl_result.reg, 1);     // eq case - load true
1259    }
1260  } else {
1261    if (cu_->instruction_set == kThumb2) {
1262      RegStorage r_tgt = cu_->target64 ?
1263          LoadHelper(QUICK_ENTRYPOINT_OFFSET(8, pInstanceofNonTrivial)) :
1264          LoadHelper(QUICK_ENTRYPOINT_OFFSET(4, pInstanceofNonTrivial));
1265      LIR* it = nullptr;
1266      if (!type_known_abstract) {
1267      /* Uses conditional nullification */
1268        OpRegReg(kOpCmp, TargetReg(kArg1), TargetReg(kArg2));  // Same?
1269        it = OpIT(kCondEq, "EE");   // if-convert the test
1270        LoadConstant(TargetReg(kArg0), 1);     // .eq case - load true
1271      }
1272      OpRegCopy(TargetReg(kArg0), TargetReg(kArg2));    // .ne case - arg0 <= class
1273      OpReg(kOpBlx, r_tgt);    // .ne case: helper(class, ref->class)
1274      if (it != nullptr) {
1275        OpEndIT(it);
1276      }
1277      FreeTemp(r_tgt);
1278    } else {
1279      if (!type_known_abstract) {
1280        /* Uses branchovers */
1281        LoadConstant(rl_result.reg, 1);     // assume true
1282        branchover = OpCmpBranch(kCondEq, TargetReg(kArg1), TargetReg(kArg2), NULL);
1283      }
1284      RegStorage r_tgt = cu_->target64 ?
1285          LoadHelper(QUICK_ENTRYPOINT_OFFSET(8, pInstanceofNonTrivial)) :
1286          LoadHelper(QUICK_ENTRYPOINT_OFFSET(4, pInstanceofNonTrivial));
1287      OpRegCopy(TargetReg(kArg0), TargetReg(kArg2));    // .ne case - arg0 <= class
1288      OpReg(kOpBlx, r_tgt);    // .ne case: helper(class, ref->class)
1289      FreeTemp(r_tgt);
1290    }
1291  }
1292  // TODO: only clobber when type isn't final?
1293  ClobberCallerSave();
1294  /* branch targets here */
1295  LIR* target = NewLIR0(kPseudoTargetLabel);
1296  StoreValue(rl_dest, rl_result);
1297  branch1->target = target;
1298  if (branchover != NULL) {
1299    branchover->target = target;
1300  }
1301}
1302
1303void Mir2Lir::GenInstanceof(uint32_t type_idx, RegLocation rl_dest, RegLocation rl_src) {
1304  bool type_known_final, type_known_abstract, use_declaring_class;
1305  bool needs_access_check = !cu_->compiler_driver->CanAccessTypeWithoutChecks(cu_->method_idx,
1306                                                                              *cu_->dex_file,
1307                                                                              type_idx,
1308                                                                              &type_known_final,
1309                                                                              &type_known_abstract,
1310                                                                              &use_declaring_class);
1311  bool can_assume_type_is_in_dex_cache = !needs_access_check &&
1312      cu_->compiler_driver->CanAssumeTypeIsPresentInDexCache(*cu_->dex_file, type_idx);
1313
1314  if ((use_declaring_class || can_assume_type_is_in_dex_cache) && type_known_final) {
1315    GenInstanceofFinal(use_declaring_class, type_idx, rl_dest, rl_src);
1316  } else {
1317    GenInstanceofCallingHelper(needs_access_check, type_known_final, type_known_abstract,
1318                               use_declaring_class, can_assume_type_is_in_dex_cache,
1319                               type_idx, rl_dest, rl_src);
1320  }
1321}
1322
1323void Mir2Lir::GenCheckCast(uint32_t insn_idx, uint32_t type_idx, RegLocation rl_src) {
1324  bool type_known_final, type_known_abstract, use_declaring_class;
1325  bool needs_access_check = !cu_->compiler_driver->CanAccessTypeWithoutChecks(cu_->method_idx,
1326                                                                              *cu_->dex_file,
1327                                                                              type_idx,
1328                                                                              &type_known_final,
1329                                                                              &type_known_abstract,
1330                                                                              &use_declaring_class);
1331  // Note: currently type_known_final is unused, as optimizing will only improve the performance
1332  // of the exception throw path.
1333  DexCompilationUnit* cu = mir_graph_->GetCurrentDexCompilationUnit();
1334  if (!needs_access_check && cu_->compiler_driver->IsSafeCast(cu, insn_idx)) {
1335    // Verifier type analysis proved this check cast would never cause an exception.
1336    return;
1337  }
1338  FlushAllRegs();
1339  // May generate a call - use explicit registers
1340  LockCallTemps();
1341  LoadCurrMethodDirect(TargetReg(kArg1));  // kArg1 <= current Method*
1342  RegStorage class_reg = TargetReg(kArg2);  // kArg2 will hold the Class*
1343  if (needs_access_check) {
1344    // Check we have access to type_idx and if not throw IllegalAccessError,
1345    // returns Class* in kRet0
1346    // InitializeTypeAndVerifyAccess(idx, method)
1347    if (cu_->target64) {
1348      CallRuntimeHelperImmReg(QUICK_ENTRYPOINT_OFFSET(8, pInitializeTypeAndVerifyAccess),
1349                              type_idx, TargetReg(kArg1), true);
1350    } else {
1351      CallRuntimeHelperImmReg(QUICK_ENTRYPOINT_OFFSET(4, pInitializeTypeAndVerifyAccess),
1352                              type_idx, TargetReg(kArg1), true);
1353    }
1354    OpRegCopy(class_reg, TargetReg(kRet0));  // Align usage with fast path
1355  } else if (use_declaring_class) {
1356    LoadRefDisp(TargetReg(kArg1), mirror::ArtMethod::DeclaringClassOffset().Int32Value(),
1357                class_reg, kNotVolatile);
1358  } else {
1359    // Load dex cache entry into class_reg (kArg2)
1360    LoadRefDisp(TargetReg(kArg1), mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(),
1361                class_reg, kNotVolatile);
1362    int32_t offset_of_type = ClassArray::OffsetOfElement(type_idx).Int32Value();
1363    LoadRefDisp(class_reg, offset_of_type, class_reg, kNotVolatile);
1364    if (!cu_->compiler_driver->CanAssumeTypeIsPresentInDexCache(*cu_->dex_file, type_idx)) {
1365      // Need to test presence of type in dex cache at runtime
1366      LIR* hop_branch = OpCmpImmBranch(kCondEq, class_reg, 0, NULL);
1367      LIR* cont = NewLIR0(kPseudoTargetLabel);
1368
1369      // Slow path to initialize the type.  Executed if the type is NULL.
1370      class SlowPath : public LIRSlowPath {
1371       public:
1372        SlowPath(Mir2Lir* m2l, LIR* fromfast, LIR* cont, const int type_idx,
1373                 const RegStorage class_reg) :
1374                   LIRSlowPath(m2l, m2l->GetCurrentDexPc(), fromfast, cont), type_idx_(type_idx),
1375                   class_reg_(class_reg) {
1376        }
1377
1378        void Compile() {
1379          GenerateTargetLabel();
1380
1381          // Call out to helper, which will return resolved type in kArg0
1382          // InitializeTypeFromCode(idx, method)
1383          if (m2l_->cu_->target64) {
1384            m2l_->CallRuntimeHelperImmReg(QUICK_ENTRYPOINT_OFFSET(8, pInitializeType), type_idx_,
1385                                          m2l_->TargetReg(kArg1), true);
1386          } else {
1387            m2l_->CallRuntimeHelperImmReg(QUICK_ENTRYPOINT_OFFSET(4, pInitializeType), type_idx_,
1388                                                      m2l_->TargetReg(kArg1), true);
1389          }
1390          m2l_->OpRegCopy(class_reg_, m2l_->TargetReg(kRet0));  // Align usage with fast path
1391          m2l_->OpUnconditionalBranch(cont_);
1392        }
1393
1394       public:
1395        const int type_idx_;
1396        const RegStorage class_reg_;
1397      };
1398
1399      AddSlowPath(new (arena_) SlowPath(this, hop_branch, cont, type_idx, class_reg));
1400    }
1401  }
1402  // At this point, class_reg (kArg2) has class
1403  LoadValueDirectFixed(rl_src, TargetReg(kArg0));  // kArg0 <= ref
1404
1405  // Slow path for the case where the classes are not equal.  In this case we need
1406  // to call a helper function to do the check.
1407  class SlowPath : public LIRSlowPath {
1408   public:
1409    SlowPath(Mir2Lir* m2l, LIR* fromfast, LIR* cont, bool load):
1410               LIRSlowPath(m2l, m2l->GetCurrentDexPc(), fromfast, cont), load_(load) {
1411    }
1412
1413    void Compile() {
1414      GenerateTargetLabel();
1415
1416      if (load_) {
1417        m2l_->LoadRefDisp(m2l_->TargetReg(kArg0), mirror::Object::ClassOffset().Int32Value(),
1418                          m2l_->TargetReg(kArg1), kNotVolatile);
1419      }
1420      if (m2l_->cu_->target64) {
1421        m2l_->CallRuntimeHelperRegReg(QUICK_ENTRYPOINT_OFFSET(8, pCheckCast), m2l_->TargetReg(kArg2),
1422                                      m2l_->TargetReg(kArg1), true);
1423      } else {
1424        m2l_->CallRuntimeHelperRegReg(QUICK_ENTRYPOINT_OFFSET(4, pCheckCast), m2l_->TargetReg(kArg2),
1425                                              m2l_->TargetReg(kArg1), true);
1426      }
1427
1428      m2l_->OpUnconditionalBranch(cont_);
1429    }
1430
1431   private:
1432    const bool load_;
1433  };
1434
1435  if (type_known_abstract) {
1436    // Easier case, run slow path if target is non-null (slow path will load from target)
1437    LIR* branch = OpCmpImmBranch(kCondNe, TargetReg(kArg0), 0, NULL);
1438    LIR* cont = NewLIR0(kPseudoTargetLabel);
1439    AddSlowPath(new (arena_) SlowPath(this, branch, cont, true));
1440  } else {
1441    // Harder, more common case.  We need to generate a forward branch over the load
1442    // if the target is null.  If it's non-null we perform the load and branch to the
1443    // slow path if the classes are not equal.
1444
1445    /* Null is OK - continue */
1446    LIR* branch1 = OpCmpImmBranch(kCondEq, TargetReg(kArg0), 0, NULL);
1447    /* load object->klass_ */
1448    DCHECK_EQ(mirror::Object::ClassOffset().Int32Value(), 0);
1449    LoadRefDisp(TargetReg(kArg0), mirror::Object::ClassOffset().Int32Value(), TargetReg(kArg1),
1450                kNotVolatile);
1451
1452    LIR* branch2 = OpCmpBranch(kCondNe, TargetReg(kArg1), class_reg, NULL);
1453    LIR* cont = NewLIR0(kPseudoTargetLabel);
1454
1455    // Add the slow path that will not perform load since this is already done.
1456    AddSlowPath(new (arena_) SlowPath(this, branch2, cont, false));
1457
1458    // Set the null check to branch to the continuation.
1459    branch1->target = cont;
1460  }
1461}
1462
1463void Mir2Lir::GenLong3Addr(OpKind first_op, OpKind second_op, RegLocation rl_dest,
1464                           RegLocation rl_src1, RegLocation rl_src2) {
1465  RegLocation rl_result;
1466  if (cu_->instruction_set == kThumb2) {
1467    /*
1468     * NOTE:  This is the one place in the code in which we might have
1469     * as many as six live temporary registers.  There are 5 in the normal
1470     * set for Arm.  Until we have spill capabilities, temporarily add
1471     * lr to the temp set.  It is safe to do this locally, but note that
1472     * lr is used explicitly elsewhere in the code generator and cannot
1473     * normally be used as a general temp register.
1474     */
1475    MarkTemp(TargetReg(kLr));   // Add lr to the temp pool
1476    FreeTemp(TargetReg(kLr));   // and make it available
1477  }
1478  rl_src1 = LoadValueWide(rl_src1, kCoreReg);
1479  rl_src2 = LoadValueWide(rl_src2, kCoreReg);
1480  rl_result = EvalLoc(rl_dest, kCoreReg, true);
1481  // The longs may overlap - use intermediate temp if so
1482  if ((rl_result.reg.GetLowReg() == rl_src1.reg.GetHighReg()) || (rl_result.reg.GetLowReg() == rl_src2.reg.GetHighReg())) {
1483    RegStorage t_reg = AllocTemp();
1484    OpRegRegReg(first_op, t_reg, rl_src1.reg.GetLow(), rl_src2.reg.GetLow());
1485    OpRegRegReg(second_op, rl_result.reg.GetHigh(), rl_src1.reg.GetHigh(), rl_src2.reg.GetHigh());
1486    OpRegCopy(rl_result.reg.GetLow(), t_reg);
1487    FreeTemp(t_reg);
1488  } else {
1489    OpRegRegReg(first_op, rl_result.reg.GetLow(), rl_src1.reg.GetLow(), rl_src2.reg.GetLow());
1490    OpRegRegReg(second_op, rl_result.reg.GetHigh(), rl_src1.reg.GetHigh(), rl_src2.reg.GetHigh());
1491  }
1492  /*
1493   * NOTE: If rl_dest refers to a frame variable in a large frame, the
1494   * following StoreValueWide might need to allocate a temp register.
1495   * To further work around the lack of a spill capability, explicitly
1496   * free any temps from rl_src1 & rl_src2 that aren't still live in rl_result.
1497   * Remove when spill is functional.
1498   */
1499  FreeRegLocTemps(rl_result, rl_src1);
1500  FreeRegLocTemps(rl_result, rl_src2);
1501  StoreValueWide(rl_dest, rl_result);
1502  if (cu_->instruction_set == kThumb2) {
1503    Clobber(TargetReg(kLr));
1504    UnmarkTemp(TargetReg(kLr));  // Remove lr from the temp pool
1505  }
1506}
1507
1508
1509template <size_t pointer_size>
1510static void GenShiftOpLongCall(Mir2Lir* mir_to_lir, Instruction::Code opcode, RegLocation rl_src1,
1511                               RegLocation rl_shift) {
1512  ThreadOffset<pointer_size> func_offset(-1);
1513
1514  switch (opcode) {
1515    case Instruction::SHL_LONG:
1516    case Instruction::SHL_LONG_2ADDR:
1517      func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pShlLong);
1518      break;
1519    case Instruction::SHR_LONG:
1520    case Instruction::SHR_LONG_2ADDR:
1521      func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pShrLong);
1522      break;
1523    case Instruction::USHR_LONG:
1524    case Instruction::USHR_LONG_2ADDR:
1525      func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pUshrLong);
1526      break;
1527    default:
1528      LOG(FATAL) << "Unexpected case";
1529  }
1530  mir_to_lir->FlushAllRegs();   /* Send everything to home location */
1531  mir_to_lir->CallRuntimeHelperRegLocationRegLocation(func_offset, rl_src1, rl_shift, false);
1532}
1533
1534void Mir2Lir::GenShiftOpLong(Instruction::Code opcode, RegLocation rl_dest,
1535                             RegLocation rl_src1, RegLocation rl_shift) {
1536  if (cu_->target64) {
1537    GenShiftOpLongCall<8>(this, opcode, rl_src1, rl_shift);
1538  } else {
1539    GenShiftOpLongCall<4>(this, opcode, rl_src1, rl_shift);
1540  }
1541  RegLocation rl_result = GetReturnWide(kCoreReg);
1542  StoreValueWide(rl_dest, rl_result);
1543}
1544
1545
1546void Mir2Lir::GenArithOpInt(Instruction::Code opcode, RegLocation rl_dest,
1547                            RegLocation rl_src1, RegLocation rl_src2) {
1548  DCHECK(cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64);
1549  OpKind op = kOpBkpt;
1550  bool is_div_rem = false;
1551  bool check_zero = false;
1552  bool unary = false;
1553  RegLocation rl_result;
1554  bool shift_op = false;
1555  switch (opcode) {
1556    case Instruction::NEG_INT:
1557      op = kOpNeg;
1558      unary = true;
1559      break;
1560    case Instruction::NOT_INT:
1561      op = kOpMvn;
1562      unary = true;
1563      break;
1564    case Instruction::ADD_INT:
1565    case Instruction::ADD_INT_2ADDR:
1566      op = kOpAdd;
1567      break;
1568    case Instruction::SUB_INT:
1569    case Instruction::SUB_INT_2ADDR:
1570      op = kOpSub;
1571      break;
1572    case Instruction::MUL_INT:
1573    case Instruction::MUL_INT_2ADDR:
1574      op = kOpMul;
1575      break;
1576    case Instruction::DIV_INT:
1577    case Instruction::DIV_INT_2ADDR:
1578      check_zero = true;
1579      op = kOpDiv;
1580      is_div_rem = true;
1581      break;
1582    /* NOTE: returns in kArg1 */
1583    case Instruction::REM_INT:
1584    case Instruction::REM_INT_2ADDR:
1585      check_zero = true;
1586      op = kOpRem;
1587      is_div_rem = true;
1588      break;
1589    case Instruction::AND_INT:
1590    case Instruction::AND_INT_2ADDR:
1591      op = kOpAnd;
1592      break;
1593    case Instruction::OR_INT:
1594    case Instruction::OR_INT_2ADDR:
1595      op = kOpOr;
1596      break;
1597    case Instruction::XOR_INT:
1598    case Instruction::XOR_INT_2ADDR:
1599      op = kOpXor;
1600      break;
1601    case Instruction::SHL_INT:
1602    case Instruction::SHL_INT_2ADDR:
1603      shift_op = true;
1604      op = kOpLsl;
1605      break;
1606    case Instruction::SHR_INT:
1607    case Instruction::SHR_INT_2ADDR:
1608      shift_op = true;
1609      op = kOpAsr;
1610      break;
1611    case Instruction::USHR_INT:
1612    case Instruction::USHR_INT_2ADDR:
1613      shift_op = true;
1614      op = kOpLsr;
1615      break;
1616    default:
1617      LOG(FATAL) << "Invalid word arith op: " << opcode;
1618  }
1619  if (!is_div_rem) {
1620    if (unary) {
1621      rl_src1 = LoadValue(rl_src1, kCoreReg);
1622      rl_result = EvalLoc(rl_dest, kCoreReg, true);
1623      OpRegReg(op, rl_result.reg, rl_src1.reg);
1624    } else {
1625      if ((shift_op) && (cu_->instruction_set != kArm64)) {
1626        rl_src2 = LoadValue(rl_src2, kCoreReg);
1627        RegStorage t_reg = AllocTemp();
1628        OpRegRegImm(kOpAnd, t_reg, rl_src2.reg, 31);
1629        rl_src1 = LoadValue(rl_src1, kCoreReg);
1630        rl_result = EvalLoc(rl_dest, kCoreReg, true);
1631        OpRegRegReg(op, rl_result.reg, rl_src1.reg, t_reg);
1632        FreeTemp(t_reg);
1633      } else {
1634        rl_src1 = LoadValue(rl_src1, kCoreReg);
1635        rl_src2 = LoadValue(rl_src2, kCoreReg);
1636        rl_result = EvalLoc(rl_dest, kCoreReg, true);
1637        OpRegRegReg(op, rl_result.reg, rl_src1.reg, rl_src2.reg);
1638      }
1639    }
1640    StoreValue(rl_dest, rl_result);
1641  } else {
1642    bool done = false;      // Set to true if we happen to find a way to use a real instruction.
1643    if (cu_->instruction_set == kMips || cu_->instruction_set == kArm64) {
1644      rl_src1 = LoadValue(rl_src1, kCoreReg);
1645      rl_src2 = LoadValue(rl_src2, kCoreReg);
1646      if (check_zero) {
1647        GenDivZeroCheck(rl_src2.reg);
1648      }
1649      rl_result = GenDivRem(rl_dest, rl_src1.reg, rl_src2.reg, op == kOpDiv);
1650      done = true;
1651    } else if (cu_->instruction_set == kThumb2) {
1652      if (cu_->GetInstructionSetFeatures().HasDivideInstruction()) {
1653        // Use ARM SDIV instruction for division.  For remainder we also need to
1654        // calculate using a MUL and subtract.
1655        rl_src1 = LoadValue(rl_src1, kCoreReg);
1656        rl_src2 = LoadValue(rl_src2, kCoreReg);
1657        if (check_zero) {
1658          GenDivZeroCheck(rl_src2.reg);
1659        }
1660        rl_result = GenDivRem(rl_dest, rl_src1.reg, rl_src2.reg, op == kOpDiv);
1661        done = true;
1662      }
1663    }
1664
1665    // If we haven't already generated the code use the callout function.
1666    if (!done) {
1667      FlushAllRegs();   /* Send everything to home location */
1668      LoadValueDirectFixed(rl_src2, TargetReg(kArg1));
1669      RegStorage r_tgt = cu_->target64 ?
1670          CallHelperSetup(QUICK_ENTRYPOINT_OFFSET(8, pIdivmod)) :
1671          CallHelperSetup(QUICK_ENTRYPOINT_OFFSET(4, pIdivmod));
1672      LoadValueDirectFixed(rl_src1, TargetReg(kArg0));
1673      if (check_zero) {
1674        GenDivZeroCheck(TargetReg(kArg1));
1675      }
1676      // NOTE: callout here is not a safepoint.
1677      if (cu_->target64) {
1678        CallHelper(r_tgt, QUICK_ENTRYPOINT_OFFSET(8, pIdivmod), false /* not a safepoint */);
1679      } else {
1680        CallHelper(r_tgt, QUICK_ENTRYPOINT_OFFSET(4, pIdivmod), false /* not a safepoint */);
1681      }
1682      if (op == kOpDiv)
1683        rl_result = GetReturn(kCoreReg);
1684      else
1685        rl_result = GetReturnAlt();
1686    }
1687    StoreValue(rl_dest, rl_result);
1688  }
1689}
1690
1691/*
1692 * The following are the first-level codegen routines that analyze the format
1693 * of each bytecode then either dispatch special purpose codegen routines
1694 * or produce corresponding Thumb instructions directly.
1695 */
1696
1697// Returns true if no more than two bits are set in 'x'.
1698static bool IsPopCountLE2(unsigned int x) {
1699  x &= x - 1;
1700  return (x & (x - 1)) == 0;
1701}
1702
1703// Returns true if it added instructions to 'cu' to divide 'rl_src' by 'lit'
1704// and store the result in 'rl_dest'.
1705bool Mir2Lir::HandleEasyDivRem(Instruction::Code dalvik_opcode, bool is_div,
1706                               RegLocation rl_src, RegLocation rl_dest, int lit) {
1707  if ((lit < 2) || ((cu_->instruction_set != kThumb2) && !IsPowerOfTwo(lit))) {
1708    return false;
1709  }
1710  // No divide instruction for Arm, so check for more special cases
1711  if ((cu_->instruction_set == kThumb2) && !IsPowerOfTwo(lit)) {
1712    return SmallLiteralDivRem(dalvik_opcode, is_div, rl_src, rl_dest, lit);
1713  }
1714  int k = LowestSetBit(lit);
1715  if (k >= 30) {
1716    // Avoid special cases.
1717    return false;
1718  }
1719  rl_src = LoadValue(rl_src, kCoreReg);
1720  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1721  if (is_div) {
1722    RegStorage t_reg = AllocTemp();
1723    if (lit == 2) {
1724      // Division by 2 is by far the most common division by constant.
1725      OpRegRegImm(kOpLsr, t_reg, rl_src.reg, 32 - k);
1726      OpRegRegReg(kOpAdd, t_reg, t_reg, rl_src.reg);
1727      OpRegRegImm(kOpAsr, rl_result.reg, t_reg, k);
1728    } else {
1729      OpRegRegImm(kOpAsr, t_reg, rl_src.reg, 31);
1730      OpRegRegImm(kOpLsr, t_reg, t_reg, 32 - k);
1731      OpRegRegReg(kOpAdd, t_reg, t_reg, rl_src.reg);
1732      OpRegRegImm(kOpAsr, rl_result.reg, t_reg, k);
1733    }
1734  } else {
1735    RegStorage t_reg1 = AllocTemp();
1736    RegStorage t_reg2 = AllocTemp();
1737    if (lit == 2) {
1738      OpRegRegImm(kOpLsr, t_reg1, rl_src.reg, 32 - k);
1739      OpRegRegReg(kOpAdd, t_reg2, t_reg1, rl_src.reg);
1740      OpRegRegImm(kOpAnd, t_reg2, t_reg2, lit -1);
1741      OpRegRegReg(kOpSub, rl_result.reg, t_reg2, t_reg1);
1742    } else {
1743      OpRegRegImm(kOpAsr, t_reg1, rl_src.reg, 31);
1744      OpRegRegImm(kOpLsr, t_reg1, t_reg1, 32 - k);
1745      OpRegRegReg(kOpAdd, t_reg2, t_reg1, rl_src.reg);
1746      OpRegRegImm(kOpAnd, t_reg2, t_reg2, lit - 1);
1747      OpRegRegReg(kOpSub, rl_result.reg, t_reg2, t_reg1);
1748    }
1749  }
1750  StoreValue(rl_dest, rl_result);
1751  return true;
1752}
1753
1754// Returns true if it added instructions to 'cu' to multiply 'rl_src' by 'lit'
1755// and store the result in 'rl_dest'.
1756bool Mir2Lir::HandleEasyMultiply(RegLocation rl_src, RegLocation rl_dest, int lit) {
1757  if (lit < 0) {
1758    return false;
1759  }
1760  if (lit == 0) {
1761    RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1762    LoadConstant(rl_result.reg, 0);
1763    StoreValue(rl_dest, rl_result);
1764    return true;
1765  }
1766  if (lit == 1) {
1767    rl_src = LoadValue(rl_src, kCoreReg);
1768    RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1769    OpRegCopy(rl_result.reg, rl_src.reg);
1770    StoreValue(rl_dest, rl_result);
1771    return true;
1772  }
1773  // There is RegRegRegShift on Arm, so check for more special cases
1774  if (cu_->instruction_set == kThumb2) {
1775    return EasyMultiply(rl_src, rl_dest, lit);
1776  }
1777  // Can we simplify this multiplication?
1778  bool power_of_two = false;
1779  bool pop_count_le2 = false;
1780  bool power_of_two_minus_one = false;
1781  if (IsPowerOfTwo(lit)) {
1782    power_of_two = true;
1783  } else if (IsPopCountLE2(lit)) {
1784    pop_count_le2 = true;
1785  } else if (IsPowerOfTwo(lit + 1)) {
1786    power_of_two_minus_one = true;
1787  } else {
1788    return false;
1789  }
1790  rl_src = LoadValue(rl_src, kCoreReg);
1791  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1792  if (power_of_two) {
1793    // Shift.
1794    OpRegRegImm(kOpLsl, rl_result.reg, rl_src.reg, LowestSetBit(lit));
1795  } else if (pop_count_le2) {
1796    // Shift and add and shift.
1797    int first_bit = LowestSetBit(lit);
1798    int second_bit = LowestSetBit(lit ^ (1 << first_bit));
1799    GenMultiplyByTwoBitMultiplier(rl_src, rl_result, lit, first_bit, second_bit);
1800  } else {
1801    // Reverse subtract: (src << (shift + 1)) - src.
1802    DCHECK(power_of_two_minus_one);
1803    // TUNING: rsb dst, src, src lsl#LowestSetBit(lit + 1)
1804    RegStorage t_reg = AllocTemp();
1805    OpRegRegImm(kOpLsl, t_reg, rl_src.reg, LowestSetBit(lit + 1));
1806    OpRegRegReg(kOpSub, rl_result.reg, t_reg, rl_src.reg);
1807  }
1808  StoreValue(rl_dest, rl_result);
1809  return true;
1810}
1811
1812void Mir2Lir::GenArithOpIntLit(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src,
1813                               int lit) {
1814  RegLocation rl_result;
1815  OpKind op = static_cast<OpKind>(0);    /* Make gcc happy */
1816  int shift_op = false;
1817  bool is_div = false;
1818
1819  switch (opcode) {
1820    case Instruction::RSUB_INT_LIT8:
1821    case Instruction::RSUB_INT: {
1822      rl_src = LoadValue(rl_src, kCoreReg);
1823      rl_result = EvalLoc(rl_dest, kCoreReg, true);
1824      if (cu_->instruction_set == kThumb2) {
1825        OpRegRegImm(kOpRsub, rl_result.reg, rl_src.reg, lit);
1826      } else {
1827        OpRegReg(kOpNeg, rl_result.reg, rl_src.reg);
1828        OpRegImm(kOpAdd, rl_result.reg, lit);
1829      }
1830      StoreValue(rl_dest, rl_result);
1831      return;
1832    }
1833
1834    case Instruction::SUB_INT:
1835    case Instruction::SUB_INT_2ADDR:
1836      lit = -lit;
1837      // Intended fallthrough
1838    case Instruction::ADD_INT:
1839    case Instruction::ADD_INT_2ADDR:
1840    case Instruction::ADD_INT_LIT8:
1841    case Instruction::ADD_INT_LIT16:
1842      op = kOpAdd;
1843      break;
1844    case Instruction::MUL_INT:
1845    case Instruction::MUL_INT_2ADDR:
1846    case Instruction::MUL_INT_LIT8:
1847    case Instruction::MUL_INT_LIT16: {
1848      if (HandleEasyMultiply(rl_src, rl_dest, lit)) {
1849        return;
1850      }
1851      op = kOpMul;
1852      break;
1853    }
1854    case Instruction::AND_INT:
1855    case Instruction::AND_INT_2ADDR:
1856    case Instruction::AND_INT_LIT8:
1857    case Instruction::AND_INT_LIT16:
1858      op = kOpAnd;
1859      break;
1860    case Instruction::OR_INT:
1861    case Instruction::OR_INT_2ADDR:
1862    case Instruction::OR_INT_LIT8:
1863    case Instruction::OR_INT_LIT16:
1864      op = kOpOr;
1865      break;
1866    case Instruction::XOR_INT:
1867    case Instruction::XOR_INT_2ADDR:
1868    case Instruction::XOR_INT_LIT8:
1869    case Instruction::XOR_INT_LIT16:
1870      op = kOpXor;
1871      break;
1872    case Instruction::SHL_INT_LIT8:
1873    case Instruction::SHL_INT:
1874    case Instruction::SHL_INT_2ADDR:
1875      lit &= 31;
1876      shift_op = true;
1877      op = kOpLsl;
1878      break;
1879    case Instruction::SHR_INT_LIT8:
1880    case Instruction::SHR_INT:
1881    case Instruction::SHR_INT_2ADDR:
1882      lit &= 31;
1883      shift_op = true;
1884      op = kOpAsr;
1885      break;
1886    case Instruction::USHR_INT_LIT8:
1887    case Instruction::USHR_INT:
1888    case Instruction::USHR_INT_2ADDR:
1889      lit &= 31;
1890      shift_op = true;
1891      op = kOpLsr;
1892      break;
1893
1894    case Instruction::DIV_INT:
1895    case Instruction::DIV_INT_2ADDR:
1896    case Instruction::DIV_INT_LIT8:
1897    case Instruction::DIV_INT_LIT16:
1898    case Instruction::REM_INT:
1899    case Instruction::REM_INT_2ADDR:
1900    case Instruction::REM_INT_LIT8:
1901    case Instruction::REM_INT_LIT16: {
1902      if (lit == 0) {
1903        GenDivZeroException();
1904        return;
1905      }
1906      if ((opcode == Instruction::DIV_INT) ||
1907          (opcode == Instruction::DIV_INT_2ADDR) ||
1908          (opcode == Instruction::DIV_INT_LIT8) ||
1909          (opcode == Instruction::DIV_INT_LIT16)) {
1910        is_div = true;
1911      } else {
1912        is_div = false;
1913      }
1914      if (HandleEasyDivRem(opcode, is_div, rl_src, rl_dest, lit)) {
1915        return;
1916      }
1917
1918      bool done = false;
1919      if (cu_->instruction_set == kMips || cu_->instruction_set == kArm64) {
1920        rl_src = LoadValue(rl_src, kCoreReg);
1921        rl_result = GenDivRemLit(rl_dest, rl_src.reg, lit, is_div);
1922        done = true;
1923      } else if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) {
1924        rl_result = GenDivRemLit(rl_dest, rl_src, lit, is_div);
1925        done = true;
1926      } else if (cu_->instruction_set == kThumb2) {
1927        if (cu_->GetInstructionSetFeatures().HasDivideInstruction()) {
1928          // Use ARM SDIV instruction for division.  For remainder we also need to
1929          // calculate using a MUL and subtract.
1930          rl_src = LoadValue(rl_src, kCoreReg);
1931          rl_result = GenDivRemLit(rl_dest, rl_src.reg, lit, is_div);
1932          done = true;
1933        }
1934      }
1935
1936      if (!done) {
1937        FlushAllRegs();   /* Everything to home location. */
1938        LoadValueDirectFixed(rl_src, TargetReg(kArg0));
1939        Clobber(TargetReg(kArg0));
1940        if (cu_->target64) {
1941          CallRuntimeHelperRegImm(QUICK_ENTRYPOINT_OFFSET(8, pIdivmod), TargetReg(kArg0), lit,
1942                                  false);
1943        } else {
1944          CallRuntimeHelperRegImm(QUICK_ENTRYPOINT_OFFSET(4, pIdivmod), TargetReg(kArg0), lit,
1945                                  false);
1946        }
1947        if (is_div)
1948          rl_result = GetReturn(kCoreReg);
1949        else
1950          rl_result = GetReturnAlt();
1951      }
1952      StoreValue(rl_dest, rl_result);
1953      return;
1954    }
1955    default:
1956      LOG(FATAL) << "Unexpected opcode " << opcode;
1957  }
1958  rl_src = LoadValue(rl_src, kCoreReg);
1959  rl_result = EvalLoc(rl_dest, kCoreReg, true);
1960  // Avoid shifts by literal 0 - no support in Thumb.  Change to copy.
1961  if (shift_op && (lit == 0)) {
1962    OpRegCopy(rl_result.reg, rl_src.reg);
1963  } else {
1964    OpRegRegImm(op, rl_result.reg, rl_src.reg, lit);
1965  }
1966  StoreValue(rl_dest, rl_result);
1967}
1968
1969template <size_t pointer_size>
1970static void GenArithOpLongImpl(Mir2Lir* mir_to_lir, CompilationUnit* cu, Instruction::Code opcode,
1971                               RegLocation rl_dest, RegLocation rl_src1, RegLocation rl_src2) {
1972  RegLocation rl_result;
1973  OpKind first_op = kOpBkpt;
1974  OpKind second_op = kOpBkpt;
1975  bool call_out = false;
1976  bool check_zero = false;
1977  ThreadOffset<pointer_size> func_offset(-1);
1978  int ret_reg = mir_to_lir->TargetReg(kRet0).GetReg();
1979
1980  switch (opcode) {
1981    case Instruction::NOT_LONG:
1982      if (cu->instruction_set == kArm64 || cu->instruction_set == kX86_64) {
1983        mir_to_lir->GenNotLong(rl_dest, rl_src2);
1984        return;
1985      }
1986      rl_src2 = mir_to_lir->LoadValueWide(rl_src2, kCoreReg);
1987      rl_result = mir_to_lir->EvalLoc(rl_dest, kCoreReg, true);
1988      // Check for destructive overlap
1989      if (rl_result.reg.GetLowReg() == rl_src2.reg.GetHighReg()) {
1990        RegStorage t_reg = mir_to_lir->AllocTemp();
1991        mir_to_lir->OpRegCopy(t_reg, rl_src2.reg.GetHigh());
1992        mir_to_lir->OpRegReg(kOpMvn, rl_result.reg.GetLow(), rl_src2.reg.GetLow());
1993        mir_to_lir->OpRegReg(kOpMvn, rl_result.reg.GetHigh(), t_reg);
1994        mir_to_lir->FreeTemp(t_reg);
1995      } else {
1996        mir_to_lir->OpRegReg(kOpMvn, rl_result.reg.GetLow(), rl_src2.reg.GetLow());
1997        mir_to_lir->OpRegReg(kOpMvn, rl_result.reg.GetHigh(), rl_src2.reg.GetHigh());
1998      }
1999      mir_to_lir->StoreValueWide(rl_dest, rl_result);
2000      return;
2001    case Instruction::ADD_LONG:
2002    case Instruction::ADD_LONG_2ADDR:
2003      if (cu->instruction_set != kThumb2) {
2004        mir_to_lir->GenAddLong(opcode, rl_dest, rl_src1, rl_src2);
2005        return;
2006      }
2007      first_op = kOpAdd;
2008      second_op = kOpAdc;
2009      break;
2010    case Instruction::SUB_LONG:
2011    case Instruction::SUB_LONG_2ADDR:
2012      if (cu->instruction_set != kThumb2) {
2013        mir_to_lir->GenSubLong(opcode, rl_dest, rl_src1, rl_src2);
2014        return;
2015      }
2016      first_op = kOpSub;
2017      second_op = kOpSbc;
2018      break;
2019    case Instruction::MUL_LONG:
2020    case Instruction::MUL_LONG_2ADDR:
2021      if (cu->instruction_set != kMips) {
2022        mir_to_lir->GenMulLong(opcode, rl_dest, rl_src1, rl_src2);
2023        return;
2024      } else {
2025        call_out = true;
2026        ret_reg = mir_to_lir->TargetReg(kRet0).GetReg();
2027        func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pLmul);
2028      }
2029      break;
2030    case Instruction::DIV_LONG:
2031    case Instruction::DIV_LONG_2ADDR:
2032      if (cu->instruction_set == kArm64 || cu->instruction_set == kX86_64) {
2033        mir_to_lir->GenDivRemLong(opcode, rl_dest, rl_src1, rl_src2, /*is_div*/ true);
2034        return;
2035      }
2036      call_out = true;
2037      check_zero = true;
2038      ret_reg = mir_to_lir->TargetReg(kRet0).GetReg();
2039      func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pLdiv);
2040      break;
2041    case Instruction::REM_LONG:
2042    case Instruction::REM_LONG_2ADDR:
2043      if (cu->instruction_set == kArm64 || cu->instruction_set == kX86_64) {
2044        mir_to_lir->GenDivRemLong(opcode, rl_dest, rl_src1, rl_src2, /*is_div*/ false);
2045        return;
2046      }
2047      call_out = true;
2048      check_zero = true;
2049      func_offset = QUICK_ENTRYPOINT_OFFSET(pointer_size, pLmod);
2050      /* NOTE - for Arm, result is in kArg2/kArg3 instead of kRet0/kRet1 */
2051      ret_reg = (cu->instruction_set == kThumb2) ? mir_to_lir->TargetReg(kArg2).GetReg() :
2052          mir_to_lir->TargetReg(kRet0).GetReg();
2053      break;
2054    case Instruction::AND_LONG_2ADDR:
2055    case Instruction::AND_LONG:
2056      if (cu->instruction_set == kX86 || cu->instruction_set == kX86_64 ||
2057          cu->instruction_set == kArm64) {
2058        return mir_to_lir->GenAndLong(opcode, rl_dest, rl_src1, rl_src2);
2059      }
2060      first_op = kOpAnd;
2061      second_op = kOpAnd;
2062      break;
2063    case Instruction::OR_LONG:
2064    case Instruction::OR_LONG_2ADDR:
2065      if (cu->instruction_set == kX86 || cu->instruction_set == kX86_64 ||
2066          cu->instruction_set == kArm64) {
2067        mir_to_lir->GenOrLong(opcode, rl_dest, rl_src1, rl_src2);
2068        return;
2069      }
2070      first_op = kOpOr;
2071      second_op = kOpOr;
2072      break;
2073    case Instruction::XOR_LONG:
2074    case Instruction::XOR_LONG_2ADDR:
2075      if (cu->instruction_set == kX86 || cu->instruction_set == kX86_64 ||
2076          cu->instruction_set == kArm64) {
2077        mir_to_lir->GenXorLong(opcode, rl_dest, rl_src1, rl_src2);
2078        return;
2079      }
2080      first_op = kOpXor;
2081      second_op = kOpXor;
2082      break;
2083    case Instruction::NEG_LONG: {
2084      mir_to_lir->GenNegLong(rl_dest, rl_src2);
2085      return;
2086    }
2087    default:
2088      LOG(FATAL) << "Invalid long arith op";
2089  }
2090  if (!call_out) {
2091    mir_to_lir->GenLong3Addr(first_op, second_op, rl_dest, rl_src1, rl_src2);
2092  } else {
2093    mir_to_lir->FlushAllRegs();   /* Send everything to home location */
2094    if (check_zero) {
2095      RegStorage r_tmp1 = RegStorage::MakeRegPair(mir_to_lir->TargetReg(kArg0),
2096                                                  mir_to_lir->TargetReg(kArg1));
2097      RegStorage r_tmp2 = RegStorage::MakeRegPair(mir_to_lir->TargetReg(kArg2),
2098                                                  mir_to_lir->TargetReg(kArg3));
2099      mir_to_lir->LoadValueDirectWideFixed(rl_src2, r_tmp2);
2100      RegStorage r_tgt = mir_to_lir->CallHelperSetup(func_offset);
2101      mir_to_lir->GenDivZeroCheckWide(RegStorage::MakeRegPair(mir_to_lir->TargetReg(kArg2),
2102                                                              mir_to_lir->TargetReg(kArg3)));
2103      mir_to_lir->LoadValueDirectWideFixed(rl_src1, r_tmp1);
2104      // NOTE: callout here is not a safepoint
2105      mir_to_lir->CallHelper(r_tgt, func_offset, false /* not safepoint */);
2106    } else {
2107      mir_to_lir->CallRuntimeHelperRegLocationRegLocation(func_offset, rl_src1, rl_src2, false);
2108    }
2109    // Adjust return regs in to handle case of rem returning kArg2/kArg3
2110    if (ret_reg == mir_to_lir->TargetReg(kRet0).GetReg())
2111      rl_result = mir_to_lir->GetReturnWide(kCoreReg);
2112    else
2113      rl_result = mir_to_lir->GetReturnWideAlt();
2114    mir_to_lir->StoreValueWide(rl_dest, rl_result);
2115  }
2116}
2117
2118void Mir2Lir::GenArithOpLong(Instruction::Code opcode, RegLocation rl_dest,
2119                             RegLocation rl_src1, RegLocation rl_src2) {
2120  if (cu_->target64) {
2121    GenArithOpLongImpl<8>(this, cu_, opcode, rl_dest, rl_src1, rl_src2);
2122  } else {
2123    GenArithOpLongImpl<4>(this, cu_, opcode, rl_dest, rl_src1, rl_src2);
2124  }
2125}
2126
2127void Mir2Lir::GenConst(RegLocation rl_dest, int value) {
2128  RegLocation rl_result = EvalLoc(rl_dest, kAnyReg, true);
2129  LoadConstantNoClobber(rl_result.reg, value);
2130  StoreValue(rl_dest, rl_result);
2131  if (value == 0) {
2132    Workaround7250540(rl_dest, rl_result.reg);
2133  }
2134}
2135
2136template <size_t pointer_size>
2137void Mir2Lir::GenConversionCall(ThreadOffset<pointer_size> func_offset,
2138                                RegLocation rl_dest, RegLocation rl_src) {
2139  /*
2140   * Don't optimize the register usage since it calls out to support
2141   * functions
2142   */
2143  DCHECK_EQ(pointer_size, GetInstructionSetPointerSize(cu_->instruction_set));
2144
2145  FlushAllRegs();   /* Send everything to home location */
2146  CallRuntimeHelperRegLocation(func_offset, rl_src, false);
2147  if (rl_dest.wide) {
2148    RegLocation rl_result;
2149    rl_result = GetReturnWide(LocToRegClass(rl_dest));
2150    StoreValueWide(rl_dest, rl_result);
2151  } else {
2152    RegLocation rl_result;
2153    rl_result = GetReturn(LocToRegClass(rl_dest));
2154    StoreValue(rl_dest, rl_result);
2155  }
2156}
2157template void Mir2Lir::GenConversionCall(ThreadOffset<4> func_offset,
2158                                         RegLocation rl_dest, RegLocation rl_src);
2159template void Mir2Lir::GenConversionCall(ThreadOffset<8> func_offset,
2160                                         RegLocation rl_dest, RegLocation rl_src);
2161
2162class SuspendCheckSlowPath : public Mir2Lir::LIRSlowPath {
2163 public:
2164  SuspendCheckSlowPath(Mir2Lir* m2l, LIR* branch, LIR* cont)
2165      : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), branch, cont) {
2166  }
2167
2168  void Compile() OVERRIDE {
2169    m2l_->ResetRegPool();
2170    m2l_->ResetDefTracking();
2171    GenerateTargetLabel(kPseudoSuspendTarget);
2172    if (cu_->target64) {
2173      m2l_->CallRuntimeHelper(QUICK_ENTRYPOINT_OFFSET(8, pTestSuspend), true);
2174    } else {
2175      m2l_->CallRuntimeHelper(QUICK_ENTRYPOINT_OFFSET(4, pTestSuspend), true);
2176    }
2177    if (cont_ != nullptr) {
2178      m2l_->OpUnconditionalBranch(cont_);
2179    }
2180  }
2181};
2182
2183/* Check if we need to check for pending suspend request */
2184void Mir2Lir::GenSuspendTest(int opt_flags) {
2185  if (cu_->compiler_driver->GetCompilerOptions().GetExplicitSuspendChecks()) {
2186    if (NO_SUSPEND || (opt_flags & MIR_IGNORE_SUSPEND_CHECK)) {
2187      return;
2188    }
2189    FlushAllRegs();
2190    LIR* branch = OpTestSuspend(NULL);
2191    LIR* cont = NewLIR0(kPseudoTargetLabel);
2192    AddSlowPath(new (arena_) SuspendCheckSlowPath(this, branch, cont));
2193  } else {
2194    if (NO_SUSPEND || (opt_flags & MIR_IGNORE_SUSPEND_CHECK)) {
2195      return;
2196    }
2197    FlushAllRegs();     // TODO: needed?
2198    LIR* inst = CheckSuspendUsingLoad();
2199    MarkSafepointPC(inst);
2200  }
2201}
2202
2203/* Check if we need to check for pending suspend request */
2204void Mir2Lir::GenSuspendTestAndBranch(int opt_flags, LIR* target) {
2205  if (cu_->compiler_driver->GetCompilerOptions().GetExplicitSuspendChecks()) {
2206    if (NO_SUSPEND || (opt_flags & MIR_IGNORE_SUSPEND_CHECK)) {
2207      OpUnconditionalBranch(target);
2208      return;
2209    }
2210    OpTestSuspend(target);
2211    FlushAllRegs();
2212    LIR* branch = OpUnconditionalBranch(nullptr);
2213    AddSlowPath(new (arena_) SuspendCheckSlowPath(this, branch, target));
2214  } else {
2215    // For the implicit suspend check, just perform the trigger
2216    // load and branch to the target.
2217    if (NO_SUSPEND || (opt_flags & MIR_IGNORE_SUSPEND_CHECK)) {
2218      OpUnconditionalBranch(target);
2219      return;
2220    }
2221    FlushAllRegs();
2222    LIR* inst = CheckSuspendUsingLoad();
2223    MarkSafepointPC(inst);
2224    OpUnconditionalBranch(target);
2225  }
2226}
2227
2228/* Call out to helper assembly routine that will null check obj and then lock it. */
2229void Mir2Lir::GenMonitorEnter(int opt_flags, RegLocation rl_src) {
2230  FlushAllRegs();
2231  if (cu_->target64) {
2232    CallRuntimeHelperRegLocation(QUICK_ENTRYPOINT_OFFSET(8, pLockObject), rl_src, true);
2233  } else {
2234    CallRuntimeHelperRegLocation(QUICK_ENTRYPOINT_OFFSET(4, pLockObject), rl_src, true);
2235  }
2236}
2237
2238/* Call out to helper assembly routine that will null check obj and then unlock it. */
2239void Mir2Lir::GenMonitorExit(int opt_flags, RegLocation rl_src) {
2240  FlushAllRegs();
2241  if (cu_->target64) {
2242    CallRuntimeHelperRegLocation(QUICK_ENTRYPOINT_OFFSET(8, pUnlockObject), rl_src, true);
2243  } else {
2244    CallRuntimeHelperRegLocation(QUICK_ENTRYPOINT_OFFSET(4, pUnlockObject), rl_src, true);
2245  }
2246}
2247
2248/* Generic code for generating a wide constant into a VR. */
2249void Mir2Lir::GenConstWide(RegLocation rl_dest, int64_t value) {
2250  RegLocation rl_result = EvalLoc(rl_dest, kAnyReg, true);
2251  LoadConstantWide(rl_result.reg, value);
2252  StoreValueWide(rl_dest, rl_result);
2253}
2254
2255}  // namespace art
2256