gen_invoke.cc revision 9d46314a309aff327f9913789b5f61200c162609
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
17#include "dex/compiler_ir.h"
18#include "dex/frontend.h"
19#include "dex/quick/dex_file_method_inliner.h"
20#include "dex/quick/dex_file_to_method_inliner_map.h"
21#include "dex_file-inl.h"
22#include "entrypoints/quick/quick_entrypoints.h"
23#include "invoke_type.h"
24#include "mirror/array.h"
25#include "mirror/string.h"
26#include "mir_to_lir-inl.h"
27#include "x86/codegen_x86.h"
28
29namespace art {
30
31/*
32 * This source files contains "gen" codegen routines that should
33 * be applicable to most targets.  Only mid-level support utilities
34 * and "op" calls may be used here.
35 */
36
37void Mir2Lir::AddIntrinsicLaunchpad(CallInfo* info, LIR* branch, LIR* resume) {
38  class IntrinsicLaunchpadPath : public Mir2Lir::LIRSlowPath {
39   public:
40    IntrinsicLaunchpadPath(Mir2Lir* m2l, CallInfo* info, LIR* branch, LIR* resume = nullptr)
41        : LIRSlowPath(m2l, info->offset, branch, resume), info_(info) {
42    }
43
44    void Compile() {
45      m2l_->ResetRegPool();
46      m2l_->ResetDefTracking();
47      LIR* label = GenerateTargetLabel();
48      label->opcode = kPseudoIntrinsicRetry;
49      // NOTE: GenInvokeNoInline() handles MarkSafepointPC.
50      m2l_->GenInvokeNoInline(info_);
51      if (cont_ != nullptr) {
52        m2l_->OpUnconditionalBranch(cont_);
53      }
54    }
55
56   private:
57    CallInfo* const info_;
58  };
59
60  AddSlowPath(new (arena_) IntrinsicLaunchpadPath(this, info, branch, resume));
61}
62
63/*
64 * To save scheduling time, helper calls are broken into two parts: generation of
65 * the helper target address, and the actual call to the helper.  Because x86
66 * has a memory call operation, part 1 is a NOP for x86.  For other targets,
67 * load arguments between the two parts.
68 */
69RegStorage Mir2Lir::CallHelperSetup(ThreadOffset<4> helper_offset) {
70  return (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) ? RegStorage::InvalidReg() : LoadHelper(helper_offset);
71}
72
73/* NOTE: if r_tgt is a temp, it will be freed following use */
74LIR* Mir2Lir::CallHelper(RegStorage r_tgt, ThreadOffset<4> helper_offset, bool safepoint_pc,
75                         bool use_link) {
76  LIR* call_inst;
77  OpKind op = use_link ? kOpBlx : kOpBx;
78  if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) {
79    call_inst = OpThreadMem(op, helper_offset);
80  } else {
81    call_inst = OpReg(op, r_tgt);
82    FreeTemp(r_tgt);
83  }
84  if (safepoint_pc) {
85    MarkSafepointPC(call_inst);
86  }
87  return call_inst;
88}
89
90void Mir2Lir::CallRuntimeHelper(ThreadOffset<4> helper_offset, bool safepoint_pc) {
91  RegStorage r_tgt = CallHelperSetup(helper_offset);
92  ClobberCallerSave();
93  CallHelper(r_tgt, helper_offset, safepoint_pc);
94}
95
96void Mir2Lir::CallRuntimeHelperImm(ThreadOffset<4> helper_offset, int arg0, bool safepoint_pc) {
97  RegStorage r_tgt = CallHelperSetup(helper_offset);
98  LoadConstant(TargetReg(kArg0), arg0);
99  ClobberCallerSave();
100  CallHelper(r_tgt, helper_offset, safepoint_pc);
101}
102
103void Mir2Lir::CallRuntimeHelperReg(ThreadOffset<4> helper_offset, RegStorage arg0,
104                                   bool safepoint_pc) {
105  RegStorage r_tgt = CallHelperSetup(helper_offset);
106  OpRegCopy(TargetReg(kArg0), arg0);
107  ClobberCallerSave();
108  CallHelper(r_tgt, helper_offset, safepoint_pc);
109}
110
111void Mir2Lir::CallRuntimeHelperRegLocation(ThreadOffset<4> helper_offset, RegLocation arg0,
112                                           bool safepoint_pc) {
113  RegStorage r_tgt = CallHelperSetup(helper_offset);
114  if (arg0.wide == 0) {
115    LoadValueDirectFixed(arg0, TargetReg(kArg0));
116  } else {
117    RegStorage r_tmp = RegStorage::MakeRegPair(TargetReg(kArg0), TargetReg(kArg1));
118    LoadValueDirectWideFixed(arg0, r_tmp);
119  }
120  ClobberCallerSave();
121  CallHelper(r_tgt, helper_offset, safepoint_pc);
122}
123
124void Mir2Lir::CallRuntimeHelperImmImm(ThreadOffset<4> helper_offset, int arg0, int arg1,
125                                      bool safepoint_pc) {
126  RegStorage r_tgt = CallHelperSetup(helper_offset);
127  LoadConstant(TargetReg(kArg0), arg0);
128  LoadConstant(TargetReg(kArg1), arg1);
129  ClobberCallerSave();
130  CallHelper(r_tgt, helper_offset, safepoint_pc);
131}
132
133void Mir2Lir::CallRuntimeHelperImmRegLocation(ThreadOffset<4> helper_offset, int arg0,
134                                              RegLocation arg1, bool safepoint_pc) {
135  RegStorage r_tgt = CallHelperSetup(helper_offset);
136  if (arg1.wide == 0) {
137    LoadValueDirectFixed(arg1, TargetReg(kArg1));
138  } else {
139    RegStorage r_tmp = RegStorage::MakeRegPair(TargetReg(kArg1), TargetReg(kArg2));
140    LoadValueDirectWideFixed(arg1, r_tmp);
141  }
142  LoadConstant(TargetReg(kArg0), arg0);
143  ClobberCallerSave();
144  CallHelper(r_tgt, helper_offset, safepoint_pc);
145}
146
147void Mir2Lir::CallRuntimeHelperRegLocationImm(ThreadOffset<4> helper_offset, RegLocation arg0,
148                                              int arg1, bool safepoint_pc) {
149  RegStorage r_tgt = CallHelperSetup(helper_offset);
150  LoadValueDirectFixed(arg0, TargetReg(kArg0));
151  LoadConstant(TargetReg(kArg1), arg1);
152  ClobberCallerSave();
153  CallHelper(r_tgt, helper_offset, safepoint_pc);
154}
155
156void Mir2Lir::CallRuntimeHelperImmReg(ThreadOffset<4> helper_offset, int arg0, RegStorage arg1,
157                                      bool safepoint_pc) {
158  RegStorage r_tgt = CallHelperSetup(helper_offset);
159  OpRegCopy(TargetReg(kArg1), arg1);
160  LoadConstant(TargetReg(kArg0), arg0);
161  ClobberCallerSave();
162  CallHelper(r_tgt, helper_offset, safepoint_pc);
163}
164
165void Mir2Lir::CallRuntimeHelperRegImm(ThreadOffset<4> helper_offset, RegStorage arg0, int arg1,
166                                      bool safepoint_pc) {
167  RegStorage r_tgt = CallHelperSetup(helper_offset);
168  OpRegCopy(TargetReg(kArg0), arg0);
169  LoadConstant(TargetReg(kArg1), arg1);
170  ClobberCallerSave();
171  CallHelper(r_tgt, helper_offset, safepoint_pc);
172}
173
174void Mir2Lir::CallRuntimeHelperImmMethod(ThreadOffset<4> helper_offset, int arg0,
175                                         bool safepoint_pc) {
176  RegStorage r_tgt = CallHelperSetup(helper_offset);
177  LoadCurrMethodDirect(TargetReg(kArg1));
178  LoadConstant(TargetReg(kArg0), arg0);
179  ClobberCallerSave();
180  CallHelper(r_tgt, helper_offset, safepoint_pc);
181}
182
183void Mir2Lir::CallRuntimeHelperRegMethod(ThreadOffset<4> helper_offset, RegStorage arg0,
184                                         bool safepoint_pc) {
185  RegStorage r_tgt = CallHelperSetup(helper_offset);
186  DCHECK_NE(TargetReg(kArg1).GetReg(), arg0.GetReg());
187  if (TargetReg(kArg0) != arg0) {
188    OpRegCopy(TargetReg(kArg0), arg0);
189  }
190  LoadCurrMethodDirect(TargetReg(kArg1));
191  ClobberCallerSave();
192  CallHelper(r_tgt, helper_offset, safepoint_pc);
193}
194
195void Mir2Lir::CallRuntimeHelperRegMethodRegLocation(ThreadOffset<4> helper_offset, RegStorage arg0,
196                                                    RegLocation arg2, bool safepoint_pc) {
197  RegStorage r_tgt = CallHelperSetup(helper_offset);
198  DCHECK_NE(TargetReg(kArg1).GetReg(), arg0.GetReg());
199  if (TargetReg(kArg0) != arg0) {
200    OpRegCopy(TargetReg(kArg0), arg0);
201  }
202  LoadCurrMethodDirect(TargetReg(kArg1));
203  LoadValueDirectFixed(arg2, TargetReg(kArg2));
204  ClobberCallerSave();
205  CallHelper(r_tgt, helper_offset, safepoint_pc);
206}
207
208void Mir2Lir::CallRuntimeHelperRegLocationRegLocation(ThreadOffset<4> helper_offset,
209                                                      RegLocation arg0, RegLocation arg1,
210                                                      bool safepoint_pc) {
211  RegStorage r_tgt = CallHelperSetup(helper_offset);
212  if (arg0.wide == 0) {
213    LoadValueDirectFixed(arg0, arg0.fp ? TargetReg(kFArg0) : TargetReg(kArg0));
214    if (arg1.wide == 0) {
215      if (cu_->instruction_set == kMips) {
216        LoadValueDirectFixed(arg1, arg1.fp ? TargetReg(kFArg2) : TargetReg(kArg1));
217      } else {
218        LoadValueDirectFixed(arg1, TargetReg(kArg1));
219      }
220    } else {
221      if (cu_->instruction_set == kMips) {
222        RegStorage r_tmp;
223        if (arg1.fp) {
224          r_tmp = RegStorage::MakeRegPair(TargetReg(kFArg2), TargetReg(kFArg3));
225        } else {
226          r_tmp = RegStorage::MakeRegPair(TargetReg(kArg1), TargetReg(kArg2));
227        }
228        LoadValueDirectWideFixed(arg1, r_tmp);
229      } else {
230        RegStorage r_tmp = RegStorage::MakeRegPair(TargetReg(kArg1), TargetReg(kArg2));
231        LoadValueDirectWideFixed(arg1, r_tmp);
232      }
233    }
234  } else {
235    RegStorage r_tmp;
236    if (arg0.fp) {
237      r_tmp = RegStorage::MakeRegPair(TargetReg(kFArg0), TargetReg(kFArg1));
238    } else {
239      r_tmp = RegStorage::MakeRegPair(TargetReg(kArg0), TargetReg(kArg1));
240    }
241    LoadValueDirectWideFixed(arg0, r_tmp);
242    if (arg1.wide == 0) {
243      LoadValueDirectFixed(arg1, arg1.fp ? TargetReg(kFArg2) : TargetReg(kArg2));
244    } else {
245      RegStorage r_tmp;
246      if (arg1.fp) {
247        r_tmp = RegStorage::MakeRegPair(TargetReg(kFArg2), TargetReg(kFArg3));
248      } else {
249        r_tmp = RegStorage::MakeRegPair(TargetReg(kArg2), TargetReg(kArg3));
250      }
251      LoadValueDirectWideFixed(arg1, r_tmp);
252    }
253  }
254  ClobberCallerSave();
255  CallHelper(r_tgt, helper_offset, safepoint_pc);
256}
257
258void Mir2Lir::CopyToArgumentRegs(RegStorage arg0, RegStorage arg1) {
259  if (arg1.GetReg() == TargetReg(kArg0).GetReg()) {
260    if (arg0.GetReg() == TargetReg(kArg1).GetReg()) {
261      // Swap kArg0 and kArg1 with kArg2 as temp.
262      OpRegCopy(TargetReg(kArg2), arg1);
263      OpRegCopy(TargetReg(kArg0), arg0);
264      OpRegCopy(TargetReg(kArg1), TargetReg(kArg2));
265    } else {
266      OpRegCopy(TargetReg(kArg1), arg1);
267      OpRegCopy(TargetReg(kArg0), arg0);
268    }
269  } else {
270    OpRegCopy(TargetReg(kArg0), arg0);
271    OpRegCopy(TargetReg(kArg1), arg1);
272  }
273}
274
275void Mir2Lir::CallRuntimeHelperRegReg(ThreadOffset<4> helper_offset, RegStorage arg0,
276                                      RegStorage arg1, bool safepoint_pc) {
277  RegStorage r_tgt = CallHelperSetup(helper_offset);
278  CopyToArgumentRegs(arg0, arg1);
279  ClobberCallerSave();
280  CallHelper(r_tgt, helper_offset, safepoint_pc);
281}
282
283void Mir2Lir::CallRuntimeHelperRegRegImm(ThreadOffset<4> helper_offset, RegStorage arg0,
284                                         RegStorage arg1, int arg2, bool safepoint_pc) {
285  RegStorage r_tgt = CallHelperSetup(helper_offset);
286  CopyToArgumentRegs(arg0, arg1);
287  LoadConstant(TargetReg(kArg2), arg2);
288  ClobberCallerSave();
289  CallHelper(r_tgt, helper_offset, safepoint_pc);
290}
291
292void Mir2Lir::CallRuntimeHelperImmMethodRegLocation(ThreadOffset<4> helper_offset,
293                                                    int arg0, RegLocation arg2, bool safepoint_pc) {
294  RegStorage r_tgt = CallHelperSetup(helper_offset);
295  LoadValueDirectFixed(arg2, TargetReg(kArg2));
296  LoadCurrMethodDirect(TargetReg(kArg1));
297  LoadConstant(TargetReg(kArg0), arg0);
298  ClobberCallerSave();
299  CallHelper(r_tgt, helper_offset, safepoint_pc);
300}
301
302void Mir2Lir::CallRuntimeHelperImmMethodImm(ThreadOffset<4> helper_offset, int arg0,
303                                            int arg2, bool safepoint_pc) {
304  RegStorage r_tgt = CallHelperSetup(helper_offset);
305  LoadCurrMethodDirect(TargetReg(kArg1));
306  LoadConstant(TargetReg(kArg2), arg2);
307  LoadConstant(TargetReg(kArg0), arg0);
308  ClobberCallerSave();
309  CallHelper(r_tgt, helper_offset, safepoint_pc);
310}
311
312void Mir2Lir::CallRuntimeHelperImmRegLocationRegLocation(ThreadOffset<4> helper_offset,
313                                                         int arg0, RegLocation arg1,
314                                                         RegLocation arg2, bool safepoint_pc) {
315  RegStorage r_tgt = CallHelperSetup(helper_offset);
316  DCHECK_EQ(arg1.wide, 0U);
317  LoadValueDirectFixed(arg1, TargetReg(kArg1));
318  if (arg2.wide == 0) {
319    LoadValueDirectFixed(arg2, TargetReg(kArg2));
320  } else {
321    RegStorage r_tmp = RegStorage::MakeRegPair(TargetReg(kArg2), TargetReg(kArg3));
322    LoadValueDirectWideFixed(arg2, r_tmp);
323  }
324  LoadConstant(TargetReg(kArg0), arg0);
325  ClobberCallerSave();
326  CallHelper(r_tgt, helper_offset, safepoint_pc);
327}
328
329void Mir2Lir::CallRuntimeHelperRegLocationRegLocationRegLocation(ThreadOffset<4> helper_offset,
330                                                                 RegLocation arg0, RegLocation arg1,
331                                                                 RegLocation arg2,
332                                                                 bool safepoint_pc) {
333  RegStorage r_tgt = CallHelperSetup(helper_offset);
334  DCHECK_EQ(arg0.wide, 0U);
335  LoadValueDirectFixed(arg0, TargetReg(kArg0));
336  DCHECK_EQ(arg1.wide, 0U);
337  LoadValueDirectFixed(arg1, TargetReg(kArg1));
338  DCHECK_EQ(arg1.wide, 0U);
339  LoadValueDirectFixed(arg2, TargetReg(kArg2));
340  ClobberCallerSave();
341  CallHelper(r_tgt, helper_offset, safepoint_pc);
342}
343
344/*
345 * If there are any ins passed in registers that have not been promoted
346 * to a callee-save register, flush them to the frame.  Perform intial
347 * assignment of promoted arguments.
348 *
349 * ArgLocs is an array of location records describing the incoming arguments
350 * with one location record per word of argument.
351 */
352void Mir2Lir::FlushIns(RegLocation* ArgLocs, RegLocation rl_method) {
353  /*
354   * Dummy up a RegLocation for the incoming Method*
355   * It will attempt to keep kArg0 live (or copy it to home location
356   * if promoted).
357   */
358  RegLocation rl_src = rl_method;
359  rl_src.location = kLocPhysReg;
360  rl_src.reg = TargetReg(kArg0);
361  rl_src.home = false;
362  MarkLive(rl_src.reg, rl_src.s_reg_low);
363  StoreValue(rl_method, rl_src);
364  // If Method* has been promoted, explicitly flush
365  if (rl_method.location == kLocPhysReg) {
366    StoreWordDisp(TargetReg(kSp), 0, TargetReg(kArg0));
367  }
368
369  if (cu_->num_ins == 0) {
370    return;
371  }
372
373  int start_vreg = cu_->num_dalvik_registers - cu_->num_ins;
374  /*
375   * Copy incoming arguments to their proper home locations.
376   * NOTE: an older version of dx had an issue in which
377   * it would reuse static method argument registers.
378   * This could result in the same Dalvik virtual register
379   * being promoted to both core and fp regs. To account for this,
380   * we only copy to the corresponding promoted physical register
381   * if it matches the type of the SSA name for the incoming
382   * argument.  It is also possible that long and double arguments
383   * end up half-promoted.  In those cases, we must flush the promoted
384   * half to memory as well.
385   */
386  for (int i = 0; i < cu_->num_ins; i++) {
387    PromotionMap* v_map = &promotion_map_[start_vreg + i];
388    RegStorage reg = GetArgMappingToPhysicalReg(i);
389
390    if (reg.Valid()) {
391      // If arriving in register
392      bool need_flush = true;
393      RegLocation* t_loc = &ArgLocs[i];
394      if ((v_map->core_location == kLocPhysReg) && !t_loc->fp) {
395        OpRegCopy(RegStorage::Solo32(v_map->core_reg), reg);
396        need_flush = false;
397      } else if ((v_map->fp_location == kLocPhysReg) && t_loc->fp) {
398        OpRegCopy(RegStorage::Solo32(v_map->FpReg), reg);
399        need_flush = false;
400      } else {
401        need_flush = true;
402      }
403
404      // For wide args, force flush if not fully promoted
405      if (t_loc->wide) {
406        PromotionMap* p_map = v_map + (t_loc->high_word ? -1 : +1);
407        // Is only half promoted?
408        need_flush |= (p_map->core_location != v_map->core_location) ||
409            (p_map->fp_location != v_map->fp_location);
410        if ((cu_->instruction_set == kThumb2) && t_loc->fp && !need_flush) {
411          /*
412           * In Arm, a double is represented as a pair of consecutive single float
413           * registers starting at an even number.  It's possible that both Dalvik vRegs
414           * representing the incoming double were independently promoted as singles - but
415           * not in a form usable as a double.  If so, we need to flush - even though the
416           * incoming arg appears fully in register.  At this point in the code, both
417           * halves of the double are promoted.  Make sure they are in a usable form.
418           */
419          int lowreg_index = start_vreg + i + (t_loc->high_word ? -1 : 0);
420          int low_reg = promotion_map_[lowreg_index].FpReg;
421          int high_reg = promotion_map_[lowreg_index + 1].FpReg;
422          if (((low_reg & 0x1) != 0) || (high_reg != (low_reg + 1))) {
423            need_flush = true;
424          }
425        }
426      }
427      if (need_flush) {
428        StoreBaseDisp(TargetReg(kSp), SRegOffset(start_vreg + i), reg, kWord);
429      }
430    } else {
431      // If arriving in frame & promoted
432      if (v_map->core_location == kLocPhysReg) {
433        LoadWordDisp(TargetReg(kSp), SRegOffset(start_vreg + i),
434                     RegStorage::Solo32(v_map->core_reg));
435      }
436      if (v_map->fp_location == kLocPhysReg) {
437        LoadWordDisp(TargetReg(kSp), SRegOffset(start_vreg + i), RegStorage::Solo32(v_map->FpReg));
438      }
439    }
440  }
441}
442
443/*
444 * Bit of a hack here - in the absence of a real scheduling pass,
445 * emit the next instruction in static & direct invoke sequences.
446 */
447static int NextSDCallInsn(CompilationUnit* cu, CallInfo* info,
448                          int state, const MethodReference& target_method,
449                          uint32_t unused,
450                          uintptr_t direct_code, uintptr_t direct_method,
451                          InvokeType type) {
452  Mir2Lir* cg = static_cast<Mir2Lir*>(cu->cg.get());
453  if (direct_code != 0 && direct_method != 0) {
454    switch (state) {
455    case 0:  // Get the current Method* [sets kArg0]
456      if (direct_code != static_cast<unsigned int>(-1)) {
457        if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) {
458          cg->LoadConstant(cg->TargetReg(kInvokeTgt), direct_code);
459        }
460      } else if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) {
461        cg->LoadCodeAddress(target_method, type, kInvokeTgt);
462      }
463      if (direct_method != static_cast<unsigned int>(-1)) {
464        cg->LoadConstant(cg->TargetReg(kArg0), direct_method);
465      } else {
466        cg->LoadMethodAddress(target_method, type, kArg0);
467      }
468      break;
469    default:
470      return -1;
471    }
472  } else {
473    switch (state) {
474    case 0:  // Get the current Method* [sets kArg0]
475      // TUNING: we can save a reg copy if Method* has been promoted.
476      cg->LoadCurrMethodDirect(cg->TargetReg(kArg0));
477      break;
478    case 1:  // Get method->dex_cache_resolved_methods_
479      cg->LoadWordDisp(cg->TargetReg(kArg0),
480                       mirror::ArtMethod::DexCacheResolvedMethodsOffset().Int32Value(),
481                       cg->TargetReg(kArg0));
482      // Set up direct code if known.
483      if (direct_code != 0) {
484        if (direct_code != static_cast<unsigned int>(-1)) {
485          cg->LoadConstant(cg->TargetReg(kInvokeTgt), direct_code);
486        } else if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) {
487          CHECK_LT(target_method.dex_method_index, target_method.dex_file->NumMethodIds());
488          cg->LoadCodeAddress(target_method, type, kInvokeTgt);
489        }
490      }
491      break;
492    case 2:  // Grab target method*
493      CHECK_EQ(cu->dex_file, target_method.dex_file);
494      cg->LoadWordDisp(cg->TargetReg(kArg0),
495                       mirror::Array::DataOffset(sizeof(mirror::Object*)).Int32Value() +
496                       (target_method.dex_method_index * 4), cg->TargetReg(kArg0));
497      break;
498    case 3:  // Grab the code from the method*
499      if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) {
500        if (direct_code == 0) {
501          cg->LoadWordDisp(cg->TargetReg(kArg0),
502                           mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset().Int32Value(),
503                           cg->TargetReg(kInvokeTgt));
504        }
505        break;
506      }
507      // Intentional fallthrough for x86
508    default:
509      return -1;
510    }
511  }
512  return state + 1;
513}
514
515/*
516 * Bit of a hack here - in the absence of a real scheduling pass,
517 * emit the next instruction in a virtual invoke sequence.
518 * We can use kLr as a temp prior to target address loading
519 * Note also that we'll load the first argument ("this") into
520 * kArg1 here rather than the standard LoadArgRegs.
521 */
522static int NextVCallInsn(CompilationUnit* cu, CallInfo* info,
523                         int state, const MethodReference& target_method,
524                         uint32_t method_idx, uintptr_t unused, uintptr_t unused2,
525                         InvokeType unused3) {
526  Mir2Lir* cg = static_cast<Mir2Lir*>(cu->cg.get());
527  /*
528   * This is the fast path in which the target virtual method is
529   * fully resolved at compile time.
530   */
531  switch (state) {
532    case 0: {  // Get "this" [set kArg1]
533      RegLocation  rl_arg = info->args[0];
534      cg->LoadValueDirectFixed(rl_arg, cg->TargetReg(kArg1));
535      break;
536    }
537    case 1:  // Is "this" null? [use kArg1]
538      cg->GenNullCheck(cg->TargetReg(kArg1), info->opt_flags);
539      // get this->klass_ [use kArg1, set kInvokeTgt]
540      cg->LoadWordDisp(cg->TargetReg(kArg1), mirror::Object::ClassOffset().Int32Value(),
541                       cg->TargetReg(kInvokeTgt));
542      cg->MarkPossibleNullPointerException(info->opt_flags);
543      break;
544    case 2:  // Get this->klass_->vtable [usr kInvokeTgt, set kInvokeTgt]
545      cg->LoadWordDisp(cg->TargetReg(kInvokeTgt), mirror::Class::VTableOffset().Int32Value(),
546                       cg->TargetReg(kInvokeTgt));
547      break;
548    case 3:  // Get target method [use kInvokeTgt, set kArg0]
549      cg->LoadWordDisp(cg->TargetReg(kInvokeTgt), (method_idx * 4) +
550                       mirror::Array::DataOffset(sizeof(mirror::Object*)).Int32Value(),
551                       cg->TargetReg(kArg0));
552      break;
553    case 4:  // Get the compiled code address [uses kArg0, sets kInvokeTgt]
554      if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) {
555        cg->LoadWordDisp(cg->TargetReg(kArg0),
556                         mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset().Int32Value(),
557                         cg->TargetReg(kInvokeTgt));
558        break;
559      }
560      // Intentional fallthrough for X86
561    default:
562      return -1;
563  }
564  return state + 1;
565}
566
567/*
568 * Emit the next instruction in an invoke interface sequence. This will do a lookup in the
569 * class's IMT, calling either the actual method or art_quick_imt_conflict_trampoline if
570 * more than one interface method map to the same index. Note also that we'll load the first
571 * argument ("this") into kArg1 here rather than the standard LoadArgRegs.
572 */
573static int NextInterfaceCallInsn(CompilationUnit* cu, CallInfo* info, int state,
574                                 const MethodReference& target_method,
575                                 uint32_t method_idx, uintptr_t unused,
576                                 uintptr_t direct_method, InvokeType unused2) {
577  Mir2Lir* cg = static_cast<Mir2Lir*>(cu->cg.get());
578
579  switch (state) {
580    case 0:  // Set target method index in case of conflict [set kHiddenArg, kHiddenFpArg (x86)]
581      CHECK_LT(target_method.dex_method_index, target_method.dex_file->NumMethodIds());
582      cg->LoadConstant(cg->TargetReg(kHiddenArg), target_method.dex_method_index);
583      if (cu->instruction_set == kX86 || cu->instruction_set == kX86_64) {
584        cg->OpRegCopy(cg->TargetReg(kHiddenFpArg), cg->TargetReg(kHiddenArg));
585      }
586      break;
587    case 1: {  // Get "this" [set kArg1]
588      RegLocation  rl_arg = info->args[0];
589      cg->LoadValueDirectFixed(rl_arg, cg->TargetReg(kArg1));
590      break;
591    }
592    case 2:  // Is "this" null? [use kArg1]
593      cg->GenNullCheck(cg->TargetReg(kArg1), info->opt_flags);
594      // Get this->klass_ [use kArg1, set kInvokeTgt]
595      cg->LoadWordDisp(cg->TargetReg(kArg1), mirror::Object::ClassOffset().Int32Value(),
596                       cg->TargetReg(kInvokeTgt));
597      cg->MarkPossibleNullPointerException(info->opt_flags);
598      break;
599    case 3:  // Get this->klass_->imtable [use kInvokeTgt, set kInvokeTgt]
600      cg->LoadWordDisp(cg->TargetReg(kInvokeTgt), mirror::Class::ImTableOffset().Int32Value(),
601                       cg->TargetReg(kInvokeTgt));
602      break;
603    case 4:  // Get target method [use kInvokeTgt, set kArg0]
604      cg->LoadWordDisp(cg->TargetReg(kInvokeTgt), ((method_idx % ClassLinker::kImtSize) * 4) +
605                       mirror::Array::DataOffset(sizeof(mirror::Object*)).Int32Value(),
606                       cg->TargetReg(kArg0));
607      break;
608    case 5:  // Get the compiled code address [use kArg0, set kInvokeTgt]
609      if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) {
610        cg->LoadWordDisp(cg->TargetReg(kArg0),
611                         mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset().Int32Value(),
612                         cg->TargetReg(kInvokeTgt));
613        break;
614      }
615      // Intentional fallthrough for X86
616    default:
617      return -1;
618  }
619  return state + 1;
620}
621
622static int NextInvokeInsnSP(CompilationUnit* cu, CallInfo* info, ThreadOffset<4> trampoline,
623                            int state, const MethodReference& target_method,
624                            uint32_t method_idx) {
625  Mir2Lir* cg = static_cast<Mir2Lir*>(cu->cg.get());
626  /*
627   * This handles the case in which the base method is not fully
628   * resolved at compile time, we bail to a runtime helper.
629   */
630  if (state == 0) {
631    if (cu->instruction_set != kX86 && cu->instruction_set != kX86_64) {
632      // Load trampoline target
633      cg->LoadWordDisp(cg->TargetReg(kSelf), trampoline.Int32Value(), cg->TargetReg(kInvokeTgt));
634    }
635    // Load kArg0 with method index
636    CHECK_EQ(cu->dex_file, target_method.dex_file);
637    cg->LoadConstant(cg->TargetReg(kArg0), target_method.dex_method_index);
638    return 1;
639  }
640  return -1;
641}
642
643static int NextStaticCallInsnSP(CompilationUnit* cu, CallInfo* info,
644                                int state,
645                                const MethodReference& target_method,
646                                uint32_t unused, uintptr_t unused2,
647                                uintptr_t unused3, InvokeType unused4) {
648  ThreadOffset<4> trampoline = QUICK_ENTRYPOINT_OFFSET(4, pInvokeStaticTrampolineWithAccessCheck);
649  return NextInvokeInsnSP(cu, info, trampoline, state, target_method, 0);
650}
651
652static int NextDirectCallInsnSP(CompilationUnit* cu, CallInfo* info, int state,
653                                const MethodReference& target_method,
654                                uint32_t unused, uintptr_t unused2,
655                                uintptr_t unused3, InvokeType unused4) {
656  ThreadOffset<4> trampoline = QUICK_ENTRYPOINT_OFFSET(4, pInvokeDirectTrampolineWithAccessCheck);
657  return NextInvokeInsnSP(cu, info, trampoline, state, target_method, 0);
658}
659
660static int NextSuperCallInsnSP(CompilationUnit* cu, CallInfo* info, int state,
661                               const MethodReference& target_method,
662                               uint32_t unused, uintptr_t unused2,
663                               uintptr_t unused3, InvokeType unused4) {
664  ThreadOffset<4> trampoline = QUICK_ENTRYPOINT_OFFSET(4, pInvokeSuperTrampolineWithAccessCheck);
665  return NextInvokeInsnSP(cu, info, trampoline, state, target_method, 0);
666}
667
668static int NextVCallInsnSP(CompilationUnit* cu, CallInfo* info, int state,
669                           const MethodReference& target_method,
670                           uint32_t unused, uintptr_t unused2,
671                           uintptr_t unused3, InvokeType unused4) {
672  ThreadOffset<4> trampoline = QUICK_ENTRYPOINT_OFFSET(4, pInvokeVirtualTrampolineWithAccessCheck);
673  return NextInvokeInsnSP(cu, info, trampoline, state, target_method, 0);
674}
675
676static int NextInterfaceCallInsnWithAccessCheck(CompilationUnit* cu,
677                                                CallInfo* info, int state,
678                                                const MethodReference& target_method,
679                                                uint32_t unused, uintptr_t unused2,
680                                                uintptr_t unused3, InvokeType unused4) {
681  ThreadOffset<4> trampoline =
682      QUICK_ENTRYPOINT_OFFSET(4, pInvokeInterfaceTrampolineWithAccessCheck);
683  return NextInvokeInsnSP(cu, info, trampoline, state, target_method, 0);
684}
685
686int Mir2Lir::LoadArgRegs(CallInfo* info, int call_state,
687                         NextCallInsn next_call_insn,
688                         const MethodReference& target_method,
689                         uint32_t vtable_idx, uintptr_t direct_code,
690                         uintptr_t direct_method, InvokeType type, bool skip_this) {
691  int last_arg_reg = TargetReg(kArg3).GetReg();
692  int next_reg = TargetReg(kArg1).GetReg();
693  int next_arg = 0;
694  if (skip_this) {
695    next_reg++;
696    next_arg++;
697  }
698  for (; (next_reg <= last_arg_reg) && (next_arg < info->num_arg_words); next_reg++) {
699    RegLocation rl_arg = info->args[next_arg++];
700    rl_arg = UpdateRawLoc(rl_arg);
701    if (rl_arg.wide && (next_reg <= TargetReg(kArg2).GetReg())) {
702      RegStorage r_tmp(RegStorage::k64BitPair, next_reg, next_reg + 1);
703      LoadValueDirectWideFixed(rl_arg, r_tmp);
704      next_reg++;
705      next_arg++;
706    } else {
707      if (rl_arg.wide) {
708        rl_arg = NarrowRegLoc(rl_arg);
709        rl_arg.is_const = false;
710      }
711      LoadValueDirectFixed(rl_arg, RegStorage::Solo32(next_reg));
712    }
713    call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx,
714                                direct_code, direct_method, type);
715  }
716  return call_state;
717}
718
719/*
720 * Load up to 5 arguments, the first three of which will be in
721 * kArg1 .. kArg3.  On entry kArg0 contains the current method pointer,
722 * and as part of the load sequence, it must be replaced with
723 * the target method pointer.  Note, this may also be called
724 * for "range" variants if the number of arguments is 5 or fewer.
725 */
726int Mir2Lir::GenDalvikArgsNoRange(CallInfo* info,
727                                  int call_state, LIR** pcrLabel, NextCallInsn next_call_insn,
728                                  const MethodReference& target_method,
729                                  uint32_t vtable_idx, uintptr_t direct_code,
730                                  uintptr_t direct_method, InvokeType type, bool skip_this) {
731  RegLocation rl_arg;
732
733  /* If no arguments, just return */
734  if (info->num_arg_words == 0)
735    return call_state;
736
737  call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx,
738                              direct_code, direct_method, type);
739
740  DCHECK_LE(info->num_arg_words, 5);
741  if (info->num_arg_words > 3) {
742    int32_t next_use = 3;
743    // Detect special case of wide arg spanning arg3/arg4
744    RegLocation rl_use0 = info->args[0];
745    RegLocation rl_use1 = info->args[1];
746    RegLocation rl_use2 = info->args[2];
747    if (((!rl_use0.wide && !rl_use1.wide) || rl_use0.wide) && rl_use2.wide) {
748      RegStorage reg;
749      // Wide spans, we need the 2nd half of uses[2].
750      rl_arg = UpdateLocWide(rl_use2);
751      if (rl_arg.location == kLocPhysReg) {
752        reg = rl_arg.reg.GetHigh();
753      } else {
754        // kArg2 & rArg3 can safely be used here
755        reg = TargetReg(kArg3);
756        LoadWordDisp(TargetReg(kSp), SRegOffset(rl_arg.s_reg_low) + 4, reg);
757        call_state = next_call_insn(cu_, info, call_state, target_method,
758                                    vtable_idx, direct_code, direct_method, type);
759      }
760      StoreBaseDisp(TargetReg(kSp), (next_use + 1) * 4, reg, kWord);
761      call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx,
762                                  direct_code, direct_method, type);
763      next_use++;
764    }
765    // Loop through the rest
766    while (next_use < info->num_arg_words) {
767      RegStorage low_reg;
768      RegStorage high_reg;
769      rl_arg = info->args[next_use];
770      rl_arg = UpdateRawLoc(rl_arg);
771      if (rl_arg.location == kLocPhysReg) {
772        if (rl_arg.wide) {
773          low_reg = rl_arg.reg.GetLow();
774          high_reg = rl_arg.reg.GetHigh();
775        } else {
776          low_reg = rl_arg.reg;
777        }
778      } else {
779        low_reg = TargetReg(kArg2);
780        if (rl_arg.wide) {
781          high_reg = TargetReg(kArg3);
782          LoadValueDirectWideFixed(rl_arg, RegStorage::MakeRegPair(low_reg, high_reg));
783        } else {
784          LoadValueDirectFixed(rl_arg, low_reg);
785        }
786        call_state = next_call_insn(cu_, info, call_state, target_method,
787                                    vtable_idx, direct_code, direct_method, type);
788      }
789      int outs_offset = (next_use + 1) * 4;
790      if (rl_arg.wide) {
791        StoreBaseDispWide(TargetReg(kSp), outs_offset, RegStorage::MakeRegPair(low_reg, high_reg));
792        next_use += 2;
793      } else {
794        StoreWordDisp(TargetReg(kSp), outs_offset, low_reg);
795        next_use++;
796      }
797      call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx,
798                               direct_code, direct_method, type);
799    }
800  }
801
802  call_state = LoadArgRegs(info, call_state, next_call_insn,
803                           target_method, vtable_idx, direct_code, direct_method,
804                           type, skip_this);
805
806  if (pcrLabel) {
807    if (Runtime::Current()->ExplicitNullChecks()) {
808      *pcrLabel = GenExplicitNullCheck(TargetReg(kArg1), info->opt_flags);
809    } else {
810      *pcrLabel = nullptr;
811      // In lieu of generating a check for kArg1 being null, we need to
812      // perform a load when doing implicit checks.
813      RegStorage tmp = AllocTemp();
814      LoadWordDisp(TargetReg(kArg1), 0, tmp);
815      MarkPossibleNullPointerException(info->opt_flags);
816      FreeTemp(tmp);
817    }
818  }
819  return call_state;
820}
821
822/*
823 * May have 0+ arguments (also used for jumbo).  Note that
824 * source virtual registers may be in physical registers, so may
825 * need to be flushed to home location before copying.  This
826 * applies to arg3 and above (see below).
827 *
828 * Two general strategies:
829 *    If < 20 arguments
830 *       Pass args 3-18 using vldm/vstm block copy
831 *       Pass arg0, arg1 & arg2 in kArg1-kArg3
832 *    If 20+ arguments
833 *       Pass args arg19+ using memcpy block copy
834 *       Pass arg0, arg1 & arg2 in kArg1-kArg3
835 *
836 */
837int Mir2Lir::GenDalvikArgsRange(CallInfo* info, int call_state,
838                                LIR** pcrLabel, NextCallInsn next_call_insn,
839                                const MethodReference& target_method,
840                                uint32_t vtable_idx, uintptr_t direct_code, uintptr_t direct_method,
841                                InvokeType type, bool skip_this) {
842  // If we can treat it as non-range (Jumbo ops will use range form)
843  if (info->num_arg_words <= 5)
844    return GenDalvikArgsNoRange(info, call_state, pcrLabel,
845                                next_call_insn, target_method, vtable_idx,
846                                direct_code, direct_method, type, skip_this);
847  /*
848   * First load the non-register arguments.  Both forms expect all
849   * of the source arguments to be in their home frame location, so
850   * scan the s_reg names and flush any that have been promoted to
851   * frame backing storage.
852   */
853  // Scan the rest of the args - if in phys_reg flush to memory
854  for (int next_arg = 0; next_arg < info->num_arg_words;) {
855    RegLocation loc = info->args[next_arg];
856    if (loc.wide) {
857      loc = UpdateLocWide(loc);
858      if ((next_arg >= 2) && (loc.location == kLocPhysReg)) {
859        StoreBaseDispWide(TargetReg(kSp), SRegOffset(loc.s_reg_low), loc.reg);
860      }
861      next_arg += 2;
862    } else {
863      loc = UpdateLoc(loc);
864      if ((next_arg >= 3) && (loc.location == kLocPhysReg)) {
865        StoreBaseDisp(TargetReg(kSp), SRegOffset(loc.s_reg_low), loc.reg, kWord);
866      }
867      next_arg++;
868    }
869  }
870
871  // Logic below assumes that Method pointer is at offset zero from SP.
872  DCHECK_EQ(VRegOffset(static_cast<int>(kVRegMethodPtrBaseReg)), 0);
873
874  // The first 3 arguments are passed via registers.
875  // TODO: For 64-bit, instead of hardcoding 4 for Method* size, we should either
876  // get size of uintptr_t or size of object reference according to model being used.
877  int outs_offset = 4 /* Method* */ + (3 * sizeof(uint32_t));
878  int start_offset = SRegOffset(info->args[3].s_reg_low);
879  int regs_left_to_pass_via_stack = info->num_arg_words - 3;
880  DCHECK_GT(regs_left_to_pass_via_stack, 0);
881
882  if (cu_->instruction_set == kThumb2 && regs_left_to_pass_via_stack <= 16) {
883    // Use vldm/vstm pair using kArg3 as a temp
884    call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx,
885                             direct_code, direct_method, type);
886    OpRegRegImm(kOpAdd, TargetReg(kArg3), TargetReg(kSp), start_offset);
887    LIR* ld = OpVldm(TargetReg(kArg3), regs_left_to_pass_via_stack);
888    // TUNING: loosen barrier
889    ld->u.m.def_mask = ENCODE_ALL;
890    SetMemRefType(ld, true /* is_load */, kDalvikReg);
891    call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx,
892                             direct_code, direct_method, type);
893    OpRegRegImm(kOpAdd, TargetReg(kArg3), TargetReg(kSp), 4 /* Method* */ + (3 * 4));
894    call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx,
895                             direct_code, direct_method, type);
896    LIR* st = OpVstm(TargetReg(kArg3), regs_left_to_pass_via_stack);
897    SetMemRefType(st, false /* is_load */, kDalvikReg);
898    st->u.m.def_mask = ENCODE_ALL;
899    call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx,
900                             direct_code, direct_method, type);
901  } else if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) {
902    int current_src_offset = start_offset;
903    int current_dest_offset = outs_offset;
904
905    while (regs_left_to_pass_via_stack > 0) {
906      // This is based on the knowledge that the stack itself is 16-byte aligned.
907      bool src_is_16b_aligned = (current_src_offset & 0xF) == 0;
908      bool dest_is_16b_aligned = (current_dest_offset & 0xF) == 0;
909      size_t bytes_to_move;
910
911      /*
912       * The amount to move defaults to 32-bit. If there are 4 registers left to move, then do a
913       * a 128-bit move because we won't get the chance to try to aligned. If there are more than
914       * 4 registers left to move, consider doing a 128-bit only if either src or dest are aligned.
915       * We do this because we could potentially do a smaller move to align.
916       */
917      if (regs_left_to_pass_via_stack == 4 ||
918          (regs_left_to_pass_via_stack > 4 && (src_is_16b_aligned || dest_is_16b_aligned))) {
919        // Moving 128-bits via xmm register.
920        bytes_to_move = sizeof(uint32_t) * 4;
921
922        // Allocate a free xmm temp. Since we are working through the calling sequence,
923        // we expect to have an xmm temporary available.
924        RegStorage temp = AllocTempDouble();
925        CHECK_GT(temp.GetLowReg(), 0);
926
927        LIR* ld1 = nullptr;
928        LIR* ld2 = nullptr;
929        LIR* st1 = nullptr;
930        LIR* st2 = nullptr;
931
932        /*
933         * The logic is similar for both loads and stores. If we have 16-byte alignment,
934         * do an aligned move. If we have 8-byte alignment, then do the move in two
935         * parts. This approach prevents possible cache line splits. Finally, fall back
936         * to doing an unaligned move. In most cases we likely won't split the cache
937         * line but we cannot prove it and thus take a conservative approach.
938         */
939        bool src_is_8b_aligned = (current_src_offset & 0x7) == 0;
940        bool dest_is_8b_aligned = (current_dest_offset & 0x7) == 0;
941
942        if (src_is_16b_aligned) {
943          ld1 = OpMovRegMem(temp, TargetReg(kSp), current_src_offset, kMovA128FP);
944        } else if (src_is_8b_aligned) {
945          ld1 = OpMovRegMem(temp, TargetReg(kSp), current_src_offset, kMovLo128FP);
946          ld2 = OpMovRegMem(temp, TargetReg(kSp), current_src_offset + (bytes_to_move >> 1),
947                            kMovHi128FP);
948        } else {
949          ld1 = OpMovRegMem(temp, TargetReg(kSp), current_src_offset, kMovU128FP);
950        }
951
952        if (dest_is_16b_aligned) {
953          st1 = OpMovMemReg(TargetReg(kSp), current_dest_offset, temp, kMovA128FP);
954        } else if (dest_is_8b_aligned) {
955          st1 = OpMovMemReg(TargetReg(kSp), current_dest_offset, temp, kMovLo128FP);
956          st2 = OpMovMemReg(TargetReg(kSp), current_dest_offset + (bytes_to_move >> 1),
957                            temp, kMovHi128FP);
958        } else {
959          st1 = OpMovMemReg(TargetReg(kSp), current_dest_offset, temp, kMovU128FP);
960        }
961
962        // TODO If we could keep track of aliasing information for memory accesses that are wider
963        // than 64-bit, we wouldn't need to set up a barrier.
964        if (ld1 != nullptr) {
965          if (ld2 != nullptr) {
966            // For 64-bit load we can actually set up the aliasing information.
967            AnnotateDalvikRegAccess(ld1, current_src_offset >> 2, true, true);
968            AnnotateDalvikRegAccess(ld2, (current_src_offset + (bytes_to_move >> 1)) >> 2, true, true);
969          } else {
970            // Set barrier for 128-bit load.
971            SetMemRefType(ld1, true /* is_load */, kDalvikReg);
972            ld1->u.m.def_mask = ENCODE_ALL;
973          }
974        }
975        if (st1 != nullptr) {
976          if (st2 != nullptr) {
977            // For 64-bit store we can actually set up the aliasing information.
978            AnnotateDalvikRegAccess(st1, current_dest_offset >> 2, false, true);
979            AnnotateDalvikRegAccess(st2, (current_dest_offset + (bytes_to_move >> 1)) >> 2, false, true);
980          } else {
981            // Set barrier for 128-bit store.
982            SetMemRefType(st1, false /* is_load */, kDalvikReg);
983            st1->u.m.def_mask = ENCODE_ALL;
984          }
985        }
986
987        // Free the temporary used for the data movement.
988        // CLEANUP: temp is currently a bogus pair, elmiminate extra free when updated.
989        FreeTemp(temp.GetLow());
990        FreeTemp(temp.GetHigh());
991      } else {
992        // Moving 32-bits via general purpose register.
993        bytes_to_move = sizeof(uint32_t);
994
995        // Instead of allocating a new temp, simply reuse one of the registers being used
996        // for argument passing.
997        RegStorage temp = TargetReg(kArg3);
998
999        // Now load the argument VR and store to the outs.
1000        LoadWordDisp(TargetReg(kSp), current_src_offset, temp);
1001        StoreWordDisp(TargetReg(kSp), current_dest_offset, temp);
1002      }
1003
1004      current_src_offset += bytes_to_move;
1005      current_dest_offset += bytes_to_move;
1006      regs_left_to_pass_via_stack -= (bytes_to_move >> 2);
1007    }
1008  } else {
1009    // Generate memcpy
1010    OpRegRegImm(kOpAdd, TargetReg(kArg0), TargetReg(kSp), outs_offset);
1011    OpRegRegImm(kOpAdd, TargetReg(kArg1), TargetReg(kSp), start_offset);
1012    CallRuntimeHelperRegRegImm(QUICK_ENTRYPOINT_OFFSET(4, pMemcpy), TargetReg(kArg0),
1013                               TargetReg(kArg1), (info->num_arg_words - 3) * 4, false);
1014  }
1015
1016  call_state = LoadArgRegs(info, call_state, next_call_insn,
1017                           target_method, vtable_idx, direct_code, direct_method,
1018                           type, skip_this);
1019
1020  call_state = next_call_insn(cu_, info, call_state, target_method, vtable_idx,
1021                           direct_code, direct_method, type);
1022  if (pcrLabel) {
1023    if (Runtime::Current()->ExplicitNullChecks()) {
1024      *pcrLabel = GenExplicitNullCheck(TargetReg(kArg1), info->opt_flags);
1025    } else {
1026      *pcrLabel = nullptr;
1027      // In lieu of generating a check for kArg1 being null, we need to
1028      // perform a load when doing implicit checks.
1029      RegStorage tmp = AllocTemp();
1030      LoadWordDisp(TargetReg(kArg1), 0, tmp);
1031      MarkPossibleNullPointerException(info->opt_flags);
1032      FreeTemp(tmp);
1033    }
1034  }
1035  return call_state;
1036}
1037
1038RegLocation Mir2Lir::InlineTarget(CallInfo* info) {
1039  RegLocation res;
1040  if (info->result.location == kLocInvalid) {
1041    res = GetReturn(false);
1042  } else {
1043    res = info->result;
1044  }
1045  return res;
1046}
1047
1048RegLocation Mir2Lir::InlineTargetWide(CallInfo* info) {
1049  RegLocation res;
1050  if (info->result.location == kLocInvalid) {
1051    res = GetReturnWide(false);
1052  } else {
1053    res = info->result;
1054  }
1055  return res;
1056}
1057
1058bool Mir2Lir::GenInlinedCharAt(CallInfo* info) {
1059  if (cu_->instruction_set == kMips) {
1060    // TODO - add Mips implementation
1061    return false;
1062  }
1063  // Location of reference to data array
1064  int value_offset = mirror::String::ValueOffset().Int32Value();
1065  // Location of count
1066  int count_offset = mirror::String::CountOffset().Int32Value();
1067  // Starting offset within data array
1068  int offset_offset = mirror::String::OffsetOffset().Int32Value();
1069  // Start of char data with array_
1070  int data_offset = mirror::Array::DataOffset(sizeof(uint16_t)).Int32Value();
1071
1072  RegLocation rl_obj = info->args[0];
1073  RegLocation rl_idx = info->args[1];
1074  rl_obj = LoadValue(rl_obj, kCoreReg);
1075  // X86 wants to avoid putting a constant index into a register.
1076  if (!((cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64)&& rl_idx.is_const)) {
1077    rl_idx = LoadValue(rl_idx, kCoreReg);
1078  }
1079  RegStorage reg_max;
1080  GenNullCheck(rl_obj.reg, info->opt_flags);
1081  bool range_check = (!(info->opt_flags & MIR_IGNORE_RANGE_CHECK));
1082  LIR* range_check_branch = nullptr;
1083  RegStorage reg_off;
1084  RegStorage reg_ptr;
1085  if (cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64) {
1086    reg_off = AllocTemp();
1087    reg_ptr = AllocTemp();
1088    if (range_check) {
1089      reg_max = AllocTemp();
1090      LoadWordDisp(rl_obj.reg, count_offset, reg_max);
1091      MarkPossibleNullPointerException(info->opt_flags);
1092    }
1093    LoadWordDisp(rl_obj.reg, offset_offset, reg_off);
1094    MarkPossibleNullPointerException(info->opt_flags);
1095    LoadWordDisp(rl_obj.reg, value_offset, reg_ptr);
1096    if (range_check) {
1097      // Set up a launch pad to allow retry in case of bounds violation */
1098      OpRegReg(kOpCmp, rl_idx.reg, reg_max);
1099      FreeTemp(reg_max);
1100      range_check_branch = OpCondBranch(kCondUge, nullptr);
1101    }
1102    OpRegImm(kOpAdd, reg_ptr, data_offset);
1103  } else {
1104    if (range_check) {
1105      // On x86, we can compare to memory directly
1106      // Set up a launch pad to allow retry in case of bounds violation */
1107      if (rl_idx.is_const) {
1108        range_check_branch = OpCmpMemImmBranch(
1109            kCondUlt, RegStorage::InvalidReg(), rl_obj.reg, count_offset,
1110            mir_graph_->ConstantValue(rl_idx.orig_sreg), nullptr);
1111      } else {
1112        OpRegMem(kOpCmp, rl_idx.reg, rl_obj.reg, count_offset);
1113        range_check_branch = OpCondBranch(kCondUge, nullptr);
1114      }
1115    }
1116    reg_off = AllocTemp();
1117    reg_ptr = AllocTemp();
1118    LoadWordDisp(rl_obj.reg, offset_offset, reg_off);
1119    LoadWordDisp(rl_obj.reg, value_offset, reg_ptr);
1120  }
1121  if (rl_idx.is_const) {
1122    OpRegImm(kOpAdd, reg_off, mir_graph_->ConstantValue(rl_idx.orig_sreg));
1123  } else {
1124    OpRegReg(kOpAdd, reg_off, rl_idx.reg);
1125  }
1126  FreeTemp(rl_obj.reg);
1127  if (rl_idx.location == kLocPhysReg) {
1128    FreeTemp(rl_idx.reg);
1129  }
1130  RegLocation rl_dest = InlineTarget(info);
1131  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1132  if (cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64) {
1133    LoadBaseIndexed(reg_ptr, reg_off, rl_result.reg, 1, kUnsignedHalf);
1134  } else {
1135    LoadBaseIndexedDisp(reg_ptr, reg_off, 1, data_offset, rl_result.reg,
1136                        RegStorage::InvalidReg(), kUnsignedHalf, INVALID_SREG);
1137  }
1138  FreeTemp(reg_off);
1139  FreeTemp(reg_ptr);
1140  StoreValue(rl_dest, rl_result);
1141  if (range_check) {
1142    DCHECK(range_check_branch != nullptr);
1143    info->opt_flags |= MIR_IGNORE_NULL_CHECK;  // Record that we've already null checked.
1144    AddIntrinsicLaunchpad(info, range_check_branch);
1145  }
1146  return true;
1147}
1148
1149// Generates an inlined String.is_empty or String.length.
1150bool Mir2Lir::GenInlinedStringIsEmptyOrLength(CallInfo* info, bool is_empty) {
1151  if (cu_->instruction_set == kMips) {
1152    // TODO - add Mips implementation
1153    return false;
1154  }
1155  // dst = src.length();
1156  RegLocation rl_obj = info->args[0];
1157  rl_obj = LoadValue(rl_obj, kCoreReg);
1158  RegLocation rl_dest = InlineTarget(info);
1159  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1160  GenNullCheck(rl_obj.reg, info->opt_flags);
1161  LoadWordDisp(rl_obj.reg, mirror::String::CountOffset().Int32Value(), rl_result.reg);
1162  MarkPossibleNullPointerException(info->opt_flags);
1163  if (is_empty) {
1164    // dst = (dst == 0);
1165    if (cu_->instruction_set == kThumb2) {
1166      RegStorage t_reg = AllocTemp();
1167      OpRegReg(kOpNeg, t_reg, rl_result.reg);
1168      OpRegRegReg(kOpAdc, rl_result.reg, rl_result.reg, t_reg);
1169    } else {
1170      DCHECK(cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64);
1171      OpRegImm(kOpSub, rl_result.reg, 1);
1172      OpRegImm(kOpLsr, rl_result.reg, 31);
1173    }
1174  }
1175  StoreValue(rl_dest, rl_result);
1176  return true;
1177}
1178
1179bool Mir2Lir::GenInlinedReverseBytes(CallInfo* info, OpSize size) {
1180  if (cu_->instruction_set == kMips) {
1181    // TODO - add Mips implementation
1182    return false;
1183  }
1184  RegLocation rl_src_i = info->args[0];
1185  RegLocation rl_dest = (size == kLong) ? InlineTargetWide(info) : InlineTarget(info);  // result reg
1186  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1187  if (size == kLong) {
1188    RegLocation rl_i = LoadValueWide(rl_src_i, kCoreReg);
1189    RegStorage r_i_low = rl_i.reg.GetLow();
1190    if (rl_i.reg.GetLowReg() == rl_result.reg.GetLowReg()) {
1191      // First REV shall clobber rl_result.reg.GetReg(), save the value in a temp for the second REV.
1192      r_i_low = AllocTemp();
1193      OpRegCopy(r_i_low, rl_i.reg);
1194    }
1195    OpRegReg(kOpRev, rl_result.reg.GetLow(), rl_i.reg.GetHigh());
1196    OpRegReg(kOpRev, rl_result.reg.GetHigh(), r_i_low);
1197    if (rl_i.reg.GetLowReg() == rl_result.reg.GetLowReg()) {
1198      FreeTemp(r_i_low);
1199    }
1200    StoreValueWide(rl_dest, rl_result);
1201  } else {
1202    DCHECK(size == kWord || size == kSignedHalf);
1203    OpKind op = (size == kWord) ? kOpRev : kOpRevsh;
1204    RegLocation rl_i = LoadValue(rl_src_i, kCoreReg);
1205    OpRegReg(op, rl_result.reg, rl_i.reg);
1206    StoreValue(rl_dest, rl_result);
1207  }
1208  return true;
1209}
1210
1211bool Mir2Lir::GenInlinedAbsInt(CallInfo* info) {
1212  if (cu_->instruction_set == kMips) {
1213    // TODO - add Mips implementation
1214    return false;
1215  }
1216  RegLocation rl_src = info->args[0];
1217  rl_src = LoadValue(rl_src, kCoreReg);
1218  RegLocation rl_dest = InlineTarget(info);
1219  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1220  RegStorage sign_reg = AllocTemp();
1221  // abs(x) = y<=x>>31, (x+y)^y.
1222  OpRegRegImm(kOpAsr, sign_reg, rl_src.reg, 31);
1223  OpRegRegReg(kOpAdd, rl_result.reg, rl_src.reg, sign_reg);
1224  OpRegReg(kOpXor, rl_result.reg, sign_reg);
1225  StoreValue(rl_dest, rl_result);
1226  return true;
1227}
1228
1229bool Mir2Lir::GenInlinedAbsLong(CallInfo* info) {
1230  if (cu_->instruction_set == kMips) {
1231    // TODO - add Mips implementation
1232    return false;
1233  }
1234  RegLocation rl_src = info->args[0];
1235  rl_src = LoadValueWide(rl_src, kCoreReg);
1236  RegLocation rl_dest = InlineTargetWide(info);
1237  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1238
1239  // If on x86 or if we would clobber a register needed later, just copy the source first.
1240  if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64 || rl_result.reg.GetLowReg() == rl_src.reg.GetHighReg()) {
1241    OpRegCopyWide(rl_result.reg, rl_src.reg);
1242    if (rl_result.reg.GetLowReg() != rl_src.reg.GetLowReg() &&
1243        rl_result.reg.GetLowReg() != rl_src.reg.GetHighReg() &&
1244        rl_result.reg.GetHighReg() != rl_src.reg.GetLowReg() &&
1245        rl_result.reg.GetHighReg() != rl_src.reg.GetHighReg()) {
1246      // Reuse source registers to avoid running out of temps.
1247      FreeTemp(rl_src.reg);
1248    }
1249    rl_src = rl_result;
1250  }
1251
1252  // abs(x) = y<=x>>31, (x+y)^y.
1253  RegStorage sign_reg = AllocTemp();
1254  OpRegRegImm(kOpAsr, sign_reg, rl_src.reg.GetHigh(), 31);
1255  OpRegRegReg(kOpAdd, rl_result.reg.GetLow(), rl_src.reg.GetLow(), sign_reg);
1256  OpRegRegReg(kOpAdc, rl_result.reg.GetHigh(), rl_src.reg.GetHigh(), sign_reg);
1257  OpRegReg(kOpXor, rl_result.reg.GetLow(), sign_reg);
1258  OpRegReg(kOpXor, rl_result.reg.GetHigh(), sign_reg);
1259  StoreValueWide(rl_dest, rl_result);
1260  return true;
1261}
1262
1263bool Mir2Lir::GenInlinedAbsFloat(CallInfo* info) {
1264  if (cu_->instruction_set == kMips) {
1265    // TODO - add Mips implementation
1266    return false;
1267  }
1268  RegLocation rl_src = info->args[0];
1269  rl_src = LoadValue(rl_src, kCoreReg);
1270  RegLocation rl_dest = InlineTarget(info);
1271  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1272  OpRegRegImm(kOpAnd, rl_result.reg, rl_src.reg, 0x7fffffff);
1273  StoreValue(rl_dest, rl_result);
1274  return true;
1275}
1276
1277bool Mir2Lir::GenInlinedAbsDouble(CallInfo* info) {
1278  if (cu_->instruction_set == kMips) {
1279    // TODO - add Mips implementation
1280    return false;
1281  }
1282  RegLocation rl_src = info->args[0];
1283  rl_src = LoadValueWide(rl_src, kCoreReg);
1284  RegLocation rl_dest = InlineTargetWide(info);
1285  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1286  OpRegCopyWide(rl_result.reg, rl_src.reg);
1287  OpRegImm(kOpAnd, rl_result.reg.GetHigh(), 0x7fffffff);
1288  StoreValueWide(rl_dest, rl_result);
1289  return true;
1290}
1291
1292bool Mir2Lir::GenInlinedFloatCvt(CallInfo* info) {
1293  if (cu_->instruction_set == kMips) {
1294    // TODO - add Mips implementation
1295    return false;
1296  }
1297  RegLocation rl_src = info->args[0];
1298  RegLocation rl_dest = InlineTarget(info);
1299  StoreValue(rl_dest, rl_src);
1300  return true;
1301}
1302
1303bool Mir2Lir::GenInlinedDoubleCvt(CallInfo* info) {
1304  if (cu_->instruction_set == kMips) {
1305    // TODO - add Mips implementation
1306    return false;
1307  }
1308  RegLocation rl_src = info->args[0];
1309  RegLocation rl_dest = InlineTargetWide(info);
1310  StoreValueWide(rl_dest, rl_src);
1311  return true;
1312}
1313
1314/*
1315 * Fast String.indexOf(I) & (II).  Tests for simple case of char <= 0xFFFF,
1316 * otherwise bails to standard library code.
1317 */
1318bool Mir2Lir::GenInlinedIndexOf(CallInfo* info, bool zero_based) {
1319  if (cu_->instruction_set == kMips) {
1320    // TODO - add Mips implementation
1321    return false;
1322  }
1323  RegLocation rl_obj = info->args[0];
1324  RegLocation rl_char = info->args[1];
1325  if (rl_char.is_const && (mir_graph_->ConstantValue(rl_char) & ~0xFFFF) != 0) {
1326    // Code point beyond 0xFFFF. Punt to the real String.indexOf().
1327    return false;
1328  }
1329
1330  ClobberCallerSave();
1331  LockCallTemps();  // Using fixed registers
1332  RegStorage reg_ptr = TargetReg(kArg0);
1333  RegStorage reg_char = TargetReg(kArg1);
1334  RegStorage reg_start = TargetReg(kArg2);
1335
1336  LoadValueDirectFixed(rl_obj, reg_ptr);
1337  LoadValueDirectFixed(rl_char, reg_char);
1338  if (zero_based) {
1339    LoadConstant(reg_start, 0);
1340  } else {
1341    RegLocation rl_start = info->args[2];     // 3rd arg only present in III flavor of IndexOf.
1342    LoadValueDirectFixed(rl_start, reg_start);
1343  }
1344  RegStorage r_tgt = LoadHelper(QUICK_ENTRYPOINT_OFFSET(4, pIndexOf));
1345  GenExplicitNullCheck(reg_ptr, info->opt_flags);
1346  LIR* high_code_point_branch =
1347      rl_char.is_const ? nullptr : OpCmpImmBranch(kCondGt, reg_char, 0xFFFF, nullptr);
1348  // NOTE: not a safepoint
1349  OpReg(kOpBlx, r_tgt);
1350  if (!rl_char.is_const) {
1351    // Add the slow path for code points beyond 0xFFFF.
1352    DCHECK(high_code_point_branch != nullptr);
1353    LIR* resume_tgt = NewLIR0(kPseudoTargetLabel);
1354    info->opt_flags |= MIR_IGNORE_NULL_CHECK;  // Record that we've null checked.
1355    AddIntrinsicLaunchpad(info, high_code_point_branch, resume_tgt);
1356  } else {
1357    DCHECK_EQ(mir_graph_->ConstantValue(rl_char) & ~0xFFFF, 0);
1358    DCHECK(high_code_point_branch == nullptr);
1359  }
1360  RegLocation rl_return = GetReturn(false);
1361  RegLocation rl_dest = InlineTarget(info);
1362  StoreValue(rl_dest, rl_return);
1363  return true;
1364}
1365
1366/* Fast string.compareTo(Ljava/lang/string;)I. */
1367bool Mir2Lir::GenInlinedStringCompareTo(CallInfo* info) {
1368  if (cu_->instruction_set == kMips) {
1369    // TODO - add Mips implementation
1370    return false;
1371  }
1372  ClobberCallerSave();
1373  LockCallTemps();  // Using fixed registers
1374  RegStorage reg_this = TargetReg(kArg0);
1375  RegStorage reg_cmp = TargetReg(kArg1);
1376
1377  RegLocation rl_this = info->args[0];
1378  RegLocation rl_cmp = info->args[1];
1379  LoadValueDirectFixed(rl_this, reg_this);
1380  LoadValueDirectFixed(rl_cmp, reg_cmp);
1381  RegStorage r_tgt = (cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64) ?
1382      LoadHelper(QUICK_ENTRYPOINT_OFFSET(4, pStringCompareTo)) : RegStorage::InvalidReg();
1383  GenExplicitNullCheck(reg_this, info->opt_flags);
1384  info->opt_flags |= MIR_IGNORE_NULL_CHECK;  // Record that we've null checked.
1385  // TUNING: check if rl_cmp.s_reg_low is already null checked
1386  LIR* cmp_null_check_branch = OpCmpImmBranch(kCondEq, reg_cmp, 0, nullptr);
1387  AddIntrinsicLaunchpad(info, cmp_null_check_branch);
1388  // NOTE: not a safepoint
1389  if (cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64) {
1390    OpReg(kOpBlx, r_tgt);
1391  } else {
1392    OpThreadMem(kOpBlx, QUICK_ENTRYPOINT_OFFSET(4, pStringCompareTo));
1393  }
1394  RegLocation rl_return = GetReturn(false);
1395  RegLocation rl_dest = InlineTarget(info);
1396  StoreValue(rl_dest, rl_return);
1397  return true;
1398}
1399
1400bool Mir2Lir::GenInlinedCurrentThread(CallInfo* info) {
1401  RegLocation rl_dest = InlineTarget(info);
1402  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1403  ThreadOffset<4> offset = Thread::PeerOffset<4>();
1404  if (cu_->instruction_set == kThumb2 || cu_->instruction_set == kMips) {
1405    LoadWordDisp(TargetReg(kSelf), offset.Int32Value(), rl_result.reg);
1406  } else {
1407    CHECK(cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64);
1408    reinterpret_cast<X86Mir2Lir*>(this)->OpRegThreadMem(kOpMov, rl_result.reg.GetReg(), offset);
1409  }
1410  StoreValue(rl_dest, rl_result);
1411  return true;
1412}
1413
1414bool Mir2Lir::GenInlinedUnsafeGet(CallInfo* info,
1415                                  bool is_long, bool is_volatile) {
1416  if (cu_->instruction_set == kMips) {
1417    // TODO - add Mips implementation
1418    return false;
1419  }
1420  // Unused - RegLocation rl_src_unsafe = info->args[0];
1421  RegLocation rl_src_obj = info->args[1];  // Object
1422  RegLocation rl_src_offset = info->args[2];  // long low
1423  rl_src_offset = NarrowRegLoc(rl_src_offset);  // ignore high half in info->args[3]
1424  RegLocation rl_dest = is_long ? InlineTargetWide(info) : InlineTarget(info);  // result reg
1425
1426  RegLocation rl_object = LoadValue(rl_src_obj, kCoreReg);
1427  RegLocation rl_offset = LoadValue(rl_src_offset, kCoreReg);
1428  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1429  if (is_long) {
1430    if (cu_->instruction_set == kX86) {
1431      LoadBaseIndexedDisp(rl_object.reg, rl_offset.reg, 0, 0, rl_result.reg.GetLow(),
1432                          rl_result.reg.GetHigh(), kLong, INVALID_SREG);
1433    } else {
1434      RegStorage rl_temp_offset = AllocTemp();
1435      OpRegRegReg(kOpAdd, rl_temp_offset, rl_object.reg, rl_offset.reg);
1436      LoadBaseDispWide(rl_temp_offset, 0, rl_result.reg, INVALID_SREG);
1437      FreeTemp(rl_temp_offset.GetReg());
1438    }
1439  } else {
1440    LoadBaseIndexed(rl_object.reg, rl_offset.reg, rl_result.reg, 0, kWord);
1441  }
1442
1443  if (is_volatile) {
1444    // Without context sensitive analysis, we must issue the most conservative barriers.
1445    // In this case, either a load or store may follow so we issue both barriers.
1446    GenMemBarrier(kLoadLoad);
1447    GenMemBarrier(kLoadStore);
1448  }
1449
1450  if (is_long) {
1451    StoreValueWide(rl_dest, rl_result);
1452  } else {
1453    StoreValue(rl_dest, rl_result);
1454  }
1455  return true;
1456}
1457
1458bool Mir2Lir::GenInlinedUnsafePut(CallInfo* info, bool is_long,
1459                                  bool is_object, bool is_volatile, bool is_ordered) {
1460  if (cu_->instruction_set == kMips) {
1461    // TODO - add Mips implementation
1462    return false;
1463  }
1464  // Unused - RegLocation rl_src_unsafe = info->args[0];
1465  RegLocation rl_src_obj = info->args[1];  // Object
1466  RegLocation rl_src_offset = info->args[2];  // long low
1467  rl_src_offset = NarrowRegLoc(rl_src_offset);  // ignore high half in info->args[3]
1468  RegLocation rl_src_value = info->args[4];  // value to store
1469  if (is_volatile || is_ordered) {
1470    // There might have been a store before this volatile one so insert StoreStore barrier.
1471    GenMemBarrier(kStoreStore);
1472  }
1473  RegLocation rl_object = LoadValue(rl_src_obj, kCoreReg);
1474  RegLocation rl_offset = LoadValue(rl_src_offset, kCoreReg);
1475  RegLocation rl_value;
1476  if (is_long) {
1477    rl_value = LoadValueWide(rl_src_value, kCoreReg);
1478    if (cu_->instruction_set == kX86) {
1479      StoreBaseIndexedDisp(rl_object.reg, rl_offset.reg, 0, 0, rl_value.reg.GetLow(),
1480                           rl_value.reg.GetHigh(), kLong, INVALID_SREG);
1481    } else {
1482      RegStorage rl_temp_offset = AllocTemp();
1483      OpRegRegReg(kOpAdd, rl_temp_offset, rl_object.reg, rl_offset.reg);
1484      StoreBaseDispWide(rl_temp_offset, 0, rl_value.reg);
1485      FreeTemp(rl_temp_offset.GetReg());
1486    }
1487  } else {
1488    rl_value = LoadValue(rl_src_value, kCoreReg);
1489    StoreBaseIndexed(rl_object.reg, rl_offset.reg, rl_value.reg, 0, kWord);
1490  }
1491
1492  // Free up the temp early, to ensure x86 doesn't run out of temporaries in MarkGCCard.
1493  FreeTemp(rl_offset.reg.GetReg());
1494
1495  if (is_volatile) {
1496    // A load might follow the volatile store so insert a StoreLoad barrier.
1497    GenMemBarrier(kStoreLoad);
1498  }
1499  if (is_object) {
1500    MarkGCCard(rl_value.reg, rl_object.reg);
1501  }
1502  return true;
1503}
1504
1505void Mir2Lir::GenInvoke(CallInfo* info) {
1506  if ((info->opt_flags & MIR_INLINED) != 0) {
1507    // Already inlined but we may still need the null check.
1508    if (info->type != kStatic &&
1509        ((cu_->disable_opt & (1 << kNullCheckElimination)) != 0 ||
1510         (info->opt_flags & MIR_IGNORE_NULL_CHECK) == 0))  {
1511      RegLocation rl_obj = LoadValue(info->args[0], kCoreReg);
1512      GenNullCheck(rl_obj.reg);
1513    }
1514    return;
1515  }
1516  DCHECK(cu_->compiler_driver->GetMethodInlinerMap() != nullptr);
1517  if (cu_->compiler_driver->GetMethodInlinerMap()->GetMethodInliner(cu_->dex_file)
1518      ->GenIntrinsic(this, info)) {
1519    return;
1520  }
1521  GenInvokeNoInline(info);
1522}
1523
1524void Mir2Lir::GenInvokeNoInline(CallInfo* info) {
1525  int call_state = 0;
1526  LIR* null_ck;
1527  LIR** p_null_ck = NULL;
1528  NextCallInsn next_call_insn;
1529  FlushAllRegs();  /* Everything to home location */
1530  // Explicit register usage
1531  LockCallTemps();
1532
1533  const MirMethodLoweringInfo& method_info = mir_graph_->GetMethodLoweringInfo(info->mir);
1534  cu_->compiler_driver->ProcessedInvoke(method_info.GetInvokeType(), method_info.StatsFlags());
1535  InvokeType original_type = static_cast<InvokeType>(method_info.GetInvokeType());
1536  info->type = static_cast<InvokeType>(method_info.GetSharpType());
1537  bool fast_path = method_info.FastPath();
1538  bool skip_this;
1539  if (info->type == kInterface) {
1540    next_call_insn = fast_path ? NextInterfaceCallInsn : NextInterfaceCallInsnWithAccessCheck;
1541    skip_this = fast_path;
1542  } else if (info->type == kDirect) {
1543    if (fast_path) {
1544      p_null_ck = &null_ck;
1545    }
1546    next_call_insn = fast_path ? NextSDCallInsn : NextDirectCallInsnSP;
1547    skip_this = false;
1548  } else if (info->type == kStatic) {
1549    next_call_insn = fast_path ? NextSDCallInsn : NextStaticCallInsnSP;
1550    skip_this = false;
1551  } else if (info->type == kSuper) {
1552    DCHECK(!fast_path);  // Fast path is a direct call.
1553    next_call_insn = NextSuperCallInsnSP;
1554    skip_this = false;
1555  } else {
1556    DCHECK_EQ(info->type, kVirtual);
1557    next_call_insn = fast_path ? NextVCallInsn : NextVCallInsnSP;
1558    skip_this = fast_path;
1559  }
1560  MethodReference target_method = method_info.GetTargetMethod();
1561  if (!info->is_range) {
1562    call_state = GenDalvikArgsNoRange(info, call_state, p_null_ck,
1563                                      next_call_insn, target_method, method_info.VTableIndex(),
1564                                      method_info.DirectCode(), method_info.DirectMethod(),
1565                                      original_type, skip_this);
1566  } else {
1567    call_state = GenDalvikArgsRange(info, call_state, p_null_ck,
1568                                    next_call_insn, target_method, method_info.VTableIndex(),
1569                                    method_info.DirectCode(), method_info.DirectMethod(),
1570                                    original_type, skip_this);
1571  }
1572  // Finish up any of the call sequence not interleaved in arg loading
1573  while (call_state >= 0) {
1574    call_state = next_call_insn(cu_, info, call_state, target_method, method_info.VTableIndex(),
1575                                method_info.DirectCode(), method_info.DirectMethod(), original_type);
1576  }
1577  LIR* call_inst;
1578  if (cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64) {
1579    call_inst = OpReg(kOpBlx, TargetReg(kInvokeTgt));
1580  } else {
1581    if (fast_path) {
1582      if (method_info.DirectCode() == static_cast<uintptr_t>(-1)) {
1583        // We can have the linker fixup a call relative.
1584        call_inst =
1585          reinterpret_cast<X86Mir2Lir*>(this)->CallWithLinkerFixup(target_method, info->type);
1586      } else {
1587        call_inst = OpMem(kOpBlx, TargetReg(kArg0),
1588                          mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset().Int32Value());
1589      }
1590    } else {
1591      ThreadOffset<4> trampoline(-1);
1592      switch (info->type) {
1593      case kInterface:
1594        trampoline = QUICK_ENTRYPOINT_OFFSET(4, pInvokeInterfaceTrampolineWithAccessCheck);
1595        break;
1596      case kDirect:
1597        trampoline = QUICK_ENTRYPOINT_OFFSET(4, pInvokeDirectTrampolineWithAccessCheck);
1598        break;
1599      case kStatic:
1600        trampoline = QUICK_ENTRYPOINT_OFFSET(4, pInvokeStaticTrampolineWithAccessCheck);
1601        break;
1602      case kSuper:
1603        trampoline = QUICK_ENTRYPOINT_OFFSET(4, pInvokeSuperTrampolineWithAccessCheck);
1604        break;
1605      case kVirtual:
1606        trampoline = QUICK_ENTRYPOINT_OFFSET(4, pInvokeVirtualTrampolineWithAccessCheck);
1607        break;
1608      default:
1609        LOG(FATAL) << "Unexpected invoke type";
1610      }
1611      call_inst = OpThreadMem(kOpBlx, trampoline);
1612    }
1613  }
1614  MarkSafepointPC(call_inst);
1615
1616  ClobberCallerSave();
1617  if (info->result.location != kLocInvalid) {
1618    // We have a following MOVE_RESULT - do it now.
1619    if (info->result.wide) {
1620      RegLocation ret_loc = GetReturnWide(info->result.fp);
1621      StoreValueWide(info->result, ret_loc);
1622    } else {
1623      RegLocation ret_loc = GetReturn(info->result.fp);
1624      StoreValue(info->result, ret_loc);
1625    }
1626  }
1627}
1628
1629}  // namespace art
1630