gen_common.cc revision af6925b7fe5dc5a3c8d52ee3370e86e75400f873
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 "arch/arm/instruction_set_features_arm.h"
17#include "dex/compiler_ir.h"
18#include "dex/compiler_internals.h"
19#include "dex/quick/arm/arm_lir.h"
20#include "dex/quick/mir_to_lir-inl.h"
21#include "entrypoints/quick/quick_entrypoints.h"
22#include "mirror/array.h"
23#include "mirror/object_array-inl.h"
24#include "mirror/object-inl.h"
25#include "mirror/object_reference.h"
26#include "verifier/method_verifier.h"
27#include <functional>
28
29namespace art {
30
31// Shortcuts to repeatedly used long types.
32typedef mirror::ObjectArray<mirror::Object> ObjArray;
33typedef mirror::ObjectArray<mirror::Class> ClassArray;
34
35/*
36 * This source files contains "gen" codegen routines that should
37 * be applicable to most targets.  Only mid-level support utilities
38 * and "op" calls may be used here.
39 */
40
41/*
42 * Generate a kPseudoBarrier marker to indicate the boundary of special
43 * blocks.
44 */
45void Mir2Lir::GenBarrier() {
46  LIR* barrier = NewLIR0(kPseudoBarrier);
47  /* Mark all resources as being clobbered */
48  DCHECK(!barrier->flags.use_def_invalid);
49  barrier->u.m.def_mask = &kEncodeAll;
50}
51
52void Mir2Lir::GenDivZeroException() {
53  LIR* branch = OpUnconditionalBranch(nullptr);
54  AddDivZeroCheckSlowPath(branch);
55}
56
57void Mir2Lir::GenDivZeroCheck(ConditionCode c_code) {
58  LIR* branch = OpCondBranch(c_code, nullptr);
59  AddDivZeroCheckSlowPath(branch);
60}
61
62void Mir2Lir::GenDivZeroCheck(RegStorage reg) {
63  LIR* branch = OpCmpImmBranch(kCondEq, reg, 0, nullptr);
64  AddDivZeroCheckSlowPath(branch);
65}
66
67void Mir2Lir::AddDivZeroCheckSlowPath(LIR* branch) {
68  class DivZeroCheckSlowPath : public Mir2Lir::LIRSlowPath {
69   public:
70    DivZeroCheckSlowPath(Mir2Lir* m2l, LIR* branch_in)
71        : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), branch_in) {
72    }
73
74    void Compile() OVERRIDE {
75      m2l_->ResetRegPool();
76      m2l_->ResetDefTracking();
77      GenerateTargetLabel(kPseudoThrowTarget);
78      m2l_->CallRuntimeHelper(kQuickThrowDivZero, true);
79    }
80  };
81
82  AddSlowPath(new (arena_) DivZeroCheckSlowPath(this, branch));
83}
84
85void Mir2Lir::GenArrayBoundsCheck(RegStorage index, RegStorage length) {
86  class ArrayBoundsCheckSlowPath : public Mir2Lir::LIRSlowPath {
87   public:
88    ArrayBoundsCheckSlowPath(Mir2Lir* m2l, LIR* branch_in, RegStorage index_in,
89                             RegStorage length_in)
90        : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), branch_in),
91          index_(index_in), length_(length_in) {
92    }
93
94    void Compile() OVERRIDE {
95      m2l_->ResetRegPool();
96      m2l_->ResetDefTracking();
97      GenerateTargetLabel(kPseudoThrowTarget);
98      m2l_->CallRuntimeHelperRegReg(kQuickThrowArrayBounds, index_, length_, true);
99    }
100
101   private:
102    const RegStorage index_;
103    const RegStorage length_;
104  };
105
106  LIR* branch = OpCmpBranch(kCondUge, index, length, nullptr);
107  AddSlowPath(new (arena_) ArrayBoundsCheckSlowPath(this, branch, index, length));
108}
109
110void Mir2Lir::GenArrayBoundsCheck(int index, RegStorage length) {
111  class ArrayBoundsCheckSlowPath : public Mir2Lir::LIRSlowPath {
112   public:
113    ArrayBoundsCheckSlowPath(Mir2Lir* m2l, LIR* branch_in, int index_in, RegStorage length_in)
114        : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), branch_in),
115          index_(index_in), length_(length_in) {
116    }
117
118    void Compile() OVERRIDE {
119      m2l_->ResetRegPool();
120      m2l_->ResetDefTracking();
121      GenerateTargetLabel(kPseudoThrowTarget);
122
123      RegStorage arg1_32 = m2l_->TargetReg(kArg1, kNotWide);
124      RegStorage arg0_32 = m2l_->TargetReg(kArg0, kNotWide);
125
126      m2l_->OpRegCopy(arg1_32, length_);
127      m2l_->LoadConstant(arg0_32, index_);
128      m2l_->CallRuntimeHelperRegReg(kQuickThrowArrayBounds, arg0_32, arg1_32, true);
129    }
130
131   private:
132    const int32_t index_;
133    const RegStorage length_;
134  };
135
136  LIR* branch = OpCmpImmBranch(kCondLs, length, index, nullptr);
137  AddSlowPath(new (arena_) ArrayBoundsCheckSlowPath(this, branch, index, length));
138}
139
140LIR* Mir2Lir::GenNullCheck(RegStorage reg) {
141  class NullCheckSlowPath : public Mir2Lir::LIRSlowPath {
142   public:
143    NullCheckSlowPath(Mir2Lir* m2l, LIR* branch)
144        : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), branch) {
145    }
146
147    void Compile() OVERRIDE {
148      m2l_->ResetRegPool();
149      m2l_->ResetDefTracking();
150      GenerateTargetLabel(kPseudoThrowTarget);
151      m2l_->CallRuntimeHelper(kQuickThrowNullPointer, true);
152    }
153  };
154
155  LIR* branch = OpCmpImmBranch(kCondEq, reg, 0, nullptr);
156  AddSlowPath(new (arena_) NullCheckSlowPath(this, branch));
157  return branch;
158}
159
160/* Perform null-check on a register.  */
161LIR* Mir2Lir::GenNullCheck(RegStorage m_reg, int opt_flags) {
162  if (!cu_->compiler_driver->GetCompilerOptions().GetImplicitNullChecks()) {
163    return GenExplicitNullCheck(m_reg, opt_flags);
164  }
165  // If null check has not been eliminated, reset redundant store tracking.
166  if ((opt_flags & MIR_IGNORE_NULL_CHECK) == 0) {
167    ResetDefTracking();
168  }
169  return nullptr;
170}
171
172/* Perform an explicit null-check on a register.  */
173LIR* Mir2Lir::GenExplicitNullCheck(RegStorage m_reg, int opt_flags) {
174  if (!(cu_->disable_opt & (1 << kNullCheckElimination)) && (opt_flags & MIR_IGNORE_NULL_CHECK)) {
175    return NULL;
176  }
177  return GenNullCheck(m_reg);
178}
179
180void Mir2Lir::MarkPossibleNullPointerException(int opt_flags) {
181  if (cu_->compiler_driver->GetCompilerOptions().GetImplicitNullChecks()) {
182    if (!(cu_->disable_opt & (1 << kNullCheckElimination)) && (opt_flags & MIR_IGNORE_NULL_CHECK)) {
183      return;
184    }
185    // Insert after last instruction.
186    MarkSafepointPC(last_lir_insn_);
187  }
188}
189
190void Mir2Lir::MarkPossibleNullPointerExceptionAfter(int opt_flags, LIR* after) {
191  if (cu_->compiler_driver->GetCompilerOptions().GetImplicitNullChecks()) {
192    if (!(cu_->disable_opt & (1 << kNullCheckElimination)) && (opt_flags & MIR_IGNORE_NULL_CHECK)) {
193      return;
194    }
195    MarkSafepointPCAfter(after);
196  }
197}
198
199void Mir2Lir::MarkPossibleStackOverflowException() {
200  if (cu_->compiler_driver->GetCompilerOptions().GetImplicitStackOverflowChecks()) {
201    MarkSafepointPC(last_lir_insn_);
202  }
203}
204
205void Mir2Lir::ForceImplicitNullCheck(RegStorage reg, int opt_flags) {
206  if (cu_->compiler_driver->GetCompilerOptions().GetImplicitNullChecks()) {
207    if (!(cu_->disable_opt & (1 << kNullCheckElimination)) && (opt_flags & MIR_IGNORE_NULL_CHECK)) {
208      return;
209    }
210    // Force an implicit null check by performing a memory operation (load) from the given
211    // register with offset 0.  This will cause a signal if the register contains 0 (null).
212    RegStorage tmp = AllocTemp();
213    // TODO: for Mips, would be best to use rZERO as the bogus register target.
214    LIR* load = Load32Disp(reg, 0, tmp);
215    FreeTemp(tmp);
216    MarkSafepointPC(load);
217  }
218}
219
220void Mir2Lir::GenCompareAndBranch(Instruction::Code opcode, RegLocation rl_src1,
221                                  RegLocation rl_src2, LIR* taken) {
222  ConditionCode cond;
223  RegisterClass reg_class = (rl_src1.ref || rl_src2.ref) ? kRefReg : kCoreReg;
224  switch (opcode) {
225    case Instruction::IF_EQ:
226      cond = kCondEq;
227      break;
228    case Instruction::IF_NE:
229      cond = kCondNe;
230      break;
231    case Instruction::IF_LT:
232      cond = kCondLt;
233      break;
234    case Instruction::IF_GE:
235      cond = kCondGe;
236      break;
237    case Instruction::IF_GT:
238      cond = kCondGt;
239      break;
240    case Instruction::IF_LE:
241      cond = kCondLe;
242      break;
243    default:
244      cond = static_cast<ConditionCode>(0);
245      LOG(FATAL) << "Unexpected opcode " << opcode;
246  }
247
248  // Normalize such that if either operand is constant, src2 will be constant
249  if (rl_src1.is_const) {
250    RegLocation rl_temp = rl_src1;
251    rl_src1 = rl_src2;
252    rl_src2 = rl_temp;
253    cond = FlipComparisonOrder(cond);
254  }
255
256  rl_src1 = LoadValue(rl_src1, reg_class);
257  // Is this really an immediate comparison?
258  if (rl_src2.is_const) {
259    // If it's already live in a register or not easily materialized, just keep going
260    RegLocation rl_temp = UpdateLoc(rl_src2);
261    int32_t constant_value = mir_graph_->ConstantValue(rl_src2);
262    if ((rl_temp.location == kLocDalvikFrame) &&
263        InexpensiveConstantInt(constant_value, opcode)) {
264      // OK - convert this to a compare immediate and branch
265      OpCmpImmBranch(cond, rl_src1.reg, mir_graph_->ConstantValue(rl_src2), taken);
266      return;
267    }
268
269    // It's also commonly more efficient to have a test against zero with Eq/Ne. This is not worse
270    // for x86, and allows a cbz/cbnz for Arm and Mips. At the same time, it works around a register
271    // mismatch for 64b systems, where a reference is compared against null, as dex bytecode uses
272    // the 32b literal 0 for null.
273    if (constant_value == 0 && (cond == kCondEq || cond == kCondNe)) {
274      // Use the OpCmpImmBranch and ignore the value in the register.
275      OpCmpImmBranch(cond, rl_src1.reg, 0, taken);
276      return;
277    }
278  }
279
280  rl_src2 = LoadValue(rl_src2, reg_class);
281  OpCmpBranch(cond, rl_src1.reg, rl_src2.reg, taken);
282}
283
284void Mir2Lir::GenCompareZeroAndBranch(Instruction::Code opcode, RegLocation rl_src, LIR* taken) {
285  ConditionCode cond;
286  RegisterClass reg_class = rl_src.ref ? kRefReg : kCoreReg;
287  rl_src = LoadValue(rl_src, reg_class);
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
347/*
348 * Let helper function take care of everything.  Will call
349 * Array::AllocFromCode(type_idx, method, count);
350 * Note: AllocFromCode will handle checks for errNegativeArraySize.
351 */
352void Mir2Lir::GenNewArray(uint32_t type_idx, RegLocation rl_dest,
353                          RegLocation rl_src) {
354  FlushAllRegs();  /* Everything to home location */
355  const DexFile* dex_file = cu_->dex_file;
356  CompilerDriver* driver = cu_->compiler_driver;
357  if (cu_->compiler_driver->CanAccessTypeWithoutChecks(cu_->method_idx, *dex_file, type_idx)) {
358    bool is_type_initialized;  // Ignored as an array does not have an initializer.
359    bool use_direct_type_ptr;
360    uintptr_t direct_type_ptr;
361    bool is_finalizable;
362    if (kEmbedClassInCode &&
363        driver->CanEmbedTypeInCode(*dex_file, type_idx, &is_type_initialized, &use_direct_type_ptr,
364                                   &direct_type_ptr, &is_finalizable)) {
365      // The fast path.
366      if (!use_direct_type_ptr) {
367        LoadClassType(*dex_file, type_idx, kArg0);
368        CallRuntimeHelperRegMethodRegLocation(kQuickAllocArrayResolved, TargetReg(kArg0, kNotWide),
369                                              rl_src, true);
370      } else {
371        // Use the direct pointer.
372        CallRuntimeHelperImmMethodRegLocation(kQuickAllocArrayResolved, direct_type_ptr, rl_src,
373                                              true);
374      }
375    } else {
376      // The slow path.
377      CallRuntimeHelperImmMethodRegLocation(kQuickAllocArray, type_idx, rl_src, true);
378    }
379  } else {
380    CallRuntimeHelperImmMethodRegLocation(kQuickAllocArrayWithAccessCheck, type_idx, rl_src, true);
381  }
382  StoreValue(rl_dest, GetReturn(kRefReg));
383}
384
385/*
386 * Similar to GenNewArray, but with post-allocation initialization.
387 * Verifier guarantees we're dealing with an array class.  Current
388 * code throws runtime exception "bad Filled array req" for 'D' and 'J'.
389 * Current code also throws internal unimp if not 'L', '[' or 'I'.
390 */
391void Mir2Lir::GenFilledNewArray(CallInfo* info) {
392  int elems = info->num_arg_words;
393  int type_idx = info->index;
394  FlushAllRegs();  /* Everything to home location */
395  QuickEntrypointEnum target;
396  if (cu_->compiler_driver->CanAccessTypeWithoutChecks(cu_->method_idx, *cu_->dex_file,
397                                                       type_idx)) {
398    target = kQuickCheckAndAllocArray;
399  } else {
400    target = kQuickCheckAndAllocArrayWithAccessCheck;
401  }
402  CallRuntimeHelperImmMethodImm(target, type_idx, elems, true);
403  FreeTemp(TargetReg(kArg2, kNotWide));
404  FreeTemp(TargetReg(kArg1, kNotWide));
405  /*
406   * NOTE: the implicit target for Instruction::FILLED_NEW_ARRAY is the
407   * return region.  Because AllocFromCode placed the new array
408   * in kRet0, we'll just lock it into place.  When debugger support is
409   * added, it may be necessary to additionally copy all return
410   * values to a home location in thread-local storage
411   */
412  RegStorage ref_reg = TargetReg(kRet0, kRef);
413  LockTemp(ref_reg);
414
415  // TODO: use the correct component size, currently all supported types
416  // share array alignment with ints (see comment at head of function)
417  size_t component_size = sizeof(int32_t);
418
419  if (elems > 5) {
420    DCHECK(info->is_range);  // Non-range insn can't encode more than 5 elems.
421    /*
422     * Bit of ugliness here.  We're going generate a mem copy loop
423     * on the register range, but it is possible that some regs
424     * in the range have been promoted.  This is unlikely, but
425     * before generating the copy, we'll just force a flush
426     * of any regs in the source range that have been promoted to
427     * home location.
428     */
429    for (int i = 0; i < elems; i++) {
430      RegLocation loc = UpdateLoc(info->args[i]);
431      if (loc.location == kLocPhysReg) {
432        ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
433        if (loc.ref) {
434          StoreRefDisp(TargetPtrReg(kSp), SRegOffset(loc.s_reg_low), loc.reg, kNotVolatile);
435        } else {
436          Store32Disp(TargetPtrReg(kSp), SRegOffset(loc.s_reg_low), loc.reg);
437        }
438      }
439    }
440    /*
441     * TUNING note: generated code here could be much improved, but
442     * this is an uncommon operation and isn't especially performance
443     * critical.
444     */
445    // This is addressing the stack, which may be out of the 4G area.
446    RegStorage r_src = AllocTempRef();
447    RegStorage r_dst = AllocTempRef();
448    RegStorage r_idx = AllocTempRef();  // Not really a reference, but match src/dst.
449    RegStorage r_val;
450    switch (cu_->instruction_set) {
451      case kThumb2:
452      case kArm64:
453        r_val = TargetReg(kLr, kNotWide);
454        break;
455      case kX86:
456      case kX86_64:
457        FreeTemp(ref_reg);
458        r_val = AllocTemp();
459        break;
460      case kMips:
461        r_val = AllocTemp();
462        break;
463      default: LOG(FATAL) << "Unexpected instruction set: " << cu_->instruction_set;
464    }
465    // Set up source pointer
466    RegLocation rl_first = info->args[0];
467    OpRegRegImm(kOpAdd, r_src, TargetPtrReg(kSp), SRegOffset(rl_first.s_reg_low));
468    // Set up the target pointer
469    OpRegRegImm(kOpAdd, r_dst, ref_reg,
470                mirror::Array::DataOffset(component_size).Int32Value());
471    // Set up the loop counter (known to be > 0)
472    LoadConstant(r_idx, elems - 1);
473    // Generate the copy loop.  Going backwards for convenience
474    LIR* loop_head_target = NewLIR0(kPseudoTargetLabel);
475    // Copy next element
476    {
477      ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
478      LoadBaseIndexed(r_src, r_idx, r_val, 2, k32);
479      // NOTE: No dalvik register annotation, local optimizations will be stopped
480      // by the loop boundaries.
481    }
482    StoreBaseIndexed(r_dst, r_idx, r_val, 2, k32);
483    FreeTemp(r_val);
484    OpDecAndBranch(kCondGe, r_idx, loop_head_target);
485    if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) {
486      // Restore the target pointer
487      OpRegRegImm(kOpAdd, ref_reg, r_dst,
488                  -mirror::Array::DataOffset(component_size).Int32Value());
489    }
490    FreeTemp(r_idx);
491    FreeTemp(r_dst);
492    FreeTemp(r_src);
493  } else {
494    DCHECK_LE(elems, 5);  // Usually but not necessarily non-range.
495    // TUNING: interleave
496    for (int i = 0; i < elems; i++) {
497      RegLocation rl_arg;
498      if (info->args[i].ref) {
499        rl_arg = LoadValue(info->args[i], kRefReg);
500        StoreRefDisp(ref_reg,
501                    mirror::Array::DataOffset(component_size).Int32Value() + i * 4, rl_arg.reg,
502                    kNotVolatile);
503      } else {
504        rl_arg = LoadValue(info->args[i], kCoreReg);
505        Store32Disp(ref_reg,
506                    mirror::Array::DataOffset(component_size).Int32Value() + i * 4, rl_arg.reg);
507      }
508      // If the LoadValue caused a temp to be allocated, free it
509      if (IsTemp(rl_arg.reg)) {
510        FreeTemp(rl_arg.reg);
511      }
512    }
513  }
514  if (elems != 0 && info->args[0].ref) {
515    // If there is at least one potentially non-null value, unconditionally mark the GC card.
516    for (int i = 0; i < elems; i++) {
517      if (!mir_graph_->IsConstantNullRef(info->args[i])) {
518        UnconditionallyMarkGCCard(ref_reg);
519        break;
520      }
521    }
522  }
523  if (info->result.location != kLocInvalid) {
524    StoreValue(info->result, GetReturn(kRefReg));
525  }
526}
527
528/*
529 * Array data table format:
530 *  ushort ident = 0x0300   magic value
531 *  ushort width            width of each element in the table
532 *  uint   size             number of elements in the table
533 *  ubyte  data[size*width] table of data values (may contain a single-byte
534 *                          padding at the end)
535 *
536 * Total size is 4+(width * size + 1)/2 16-bit code units.
537 */
538void Mir2Lir::GenFillArrayData(MIR* mir, DexOffset table_offset, RegLocation rl_src) {
539  if (kIsDebugBuild) {
540    const uint16_t* table = mir_graph_->GetTable(mir, table_offset);
541    const Instruction::ArrayDataPayload* payload =
542        reinterpret_cast<const Instruction::ArrayDataPayload*>(table);
543    CHECK_EQ(payload->ident, static_cast<uint16_t>(Instruction::kArrayDataSignature));
544  }
545  uint32_t table_offset_from_start = mir->offset + static_cast<int32_t>(table_offset);
546  CallRuntimeHelperImmRegLocation(kQuickHandleFillArrayData, table_offset_from_start, rl_src, true);
547}
548
549//
550// Slow path to ensure a class is initialized for sget/sput.
551//
552class StaticFieldSlowPath : public Mir2Lir::LIRSlowPath {
553 public:
554  // There are up to two branches to the static field slow path, the "unresolved" when the type
555  // entry in the dex cache is null, and the "uninit" when the class is not yet initialized.
556  // At least one will be non-null here, otherwise we wouldn't generate the slow path.
557  StaticFieldSlowPath(Mir2Lir* m2l, LIR* unresolved, LIR* uninit, LIR* cont, int storage_index,
558                      RegStorage r_base)
559      : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), unresolved != nullptr ? unresolved : uninit, cont),
560        second_branch_(unresolved != nullptr ? uninit : nullptr),
561        storage_index_(storage_index), r_base_(r_base) {
562  }
563
564  void Compile() {
565    LIR* target = GenerateTargetLabel();
566    if (second_branch_ != nullptr) {
567      second_branch_->target = target;
568    }
569    m2l_->CallRuntimeHelperImm(kQuickInitializeStaticStorage, storage_index_, true);
570    // Copy helper's result into r_base, a no-op on all but MIPS.
571    m2l_->OpRegCopy(r_base_,  m2l_->TargetReg(kRet0, kRef));
572
573    m2l_->OpUnconditionalBranch(cont_);
574  }
575
576 private:
577  // Second branch to the slow path, or null if there's only one branch.
578  LIR* const second_branch_;
579
580  const int storage_index_;
581  const RegStorage r_base_;
582};
583
584void Mir2Lir::GenSput(MIR* mir, RegLocation rl_src, OpSize size) {
585  const MirSFieldLoweringInfo& field_info = mir_graph_->GetSFieldLoweringInfo(mir);
586  DCHECK_EQ(SPutMemAccessType(mir->dalvikInsn.opcode), field_info.MemAccessType());
587  cu_->compiler_driver->ProcessedStaticField(field_info.FastPut(), field_info.IsReferrersClass());
588  if (!SLOW_FIELD_PATH && field_info.FastPut()) {
589    DCHECK_GE(field_info.FieldOffset().Int32Value(), 0);
590    RegStorage r_base;
591    if (field_info.IsReferrersClass()) {
592      // Fast path, static storage base is this method's class
593      RegLocation rl_method = LoadCurrMethod();
594      r_base = AllocTempRef();
595      LoadRefDisp(rl_method.reg, mirror::ArtMethod::DeclaringClassOffset().Int32Value(), r_base,
596                  kNotVolatile);
597      if (IsTemp(rl_method.reg)) {
598        FreeTemp(rl_method.reg);
599      }
600    } else {
601      // Medium path, static storage base in a different class which requires checks that the other
602      // class is initialized.
603      // TODO: remove initialized check now that we are initializing classes in the compiler driver.
604      DCHECK_NE(field_info.StorageIndex(), DexFile::kDexNoIndex);
605      // May do runtime call so everything to home locations.
606      FlushAllRegs();
607      // Using fixed register to sync with possible call to runtime support.
608      RegStorage r_method = TargetReg(kArg1, kRef);
609      LockTemp(r_method);
610      LoadCurrMethodDirect(r_method);
611      r_base = TargetReg(kArg0, kRef);
612      LockTemp(r_base);
613      LoadRefDisp(r_method, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(), r_base,
614                  kNotVolatile);
615      int32_t offset_of_field = ObjArray::OffsetOfElement(field_info.StorageIndex()).Int32Value();
616      LoadRefDisp(r_base, offset_of_field, r_base, kNotVolatile);
617      // r_base now points at static storage (Class*) or NULL if the type is not yet resolved.
618      LIR* unresolved_branch = nullptr;
619      if (!field_info.IsClassInDexCache() &&
620          (mir->optimization_flags & MIR_CLASS_IS_IN_DEX_CACHE) == 0) {
621        // Check if r_base is NULL.
622        unresolved_branch = OpCmpImmBranch(kCondEq, r_base, 0, NULL);
623      }
624      LIR* uninit_branch = nullptr;
625      if (!field_info.IsClassInitialized() &&
626          (mir->optimization_flags & MIR_CLASS_IS_INITIALIZED) == 0) {
627        // Check if r_base is not yet initialized class.
628        RegStorage r_tmp = TargetReg(kArg2, kNotWide);
629        LockTemp(r_tmp);
630        uninit_branch = OpCmpMemImmBranch(kCondLt, r_tmp, r_base,
631                                          mirror::Class::StatusOffset().Int32Value(),
632                                          mirror::Class::kStatusInitialized, nullptr, nullptr);
633        FreeTemp(r_tmp);
634      }
635      if (unresolved_branch != nullptr || uninit_branch != nullptr) {
636        // The slow path is invoked if the r_base is NULL or the class pointed
637        // to by it is not initialized.
638        LIR* cont = NewLIR0(kPseudoTargetLabel);
639        AddSlowPath(new (arena_) StaticFieldSlowPath(this, unresolved_branch, uninit_branch, cont,
640                                                     field_info.StorageIndex(), r_base));
641
642        if (uninit_branch != nullptr) {
643          // Ensure load of status and store of value don't re-order.
644          // TODO: Presumably the actual value store is control-dependent on the status load,
645          // and will thus not be reordered in any case, since stores are never speculated.
646          // Does later code "know" that the class is now initialized?  If so, we still
647          // need the barrier to guard later static loads.
648          GenMemBarrier(kLoadAny);
649        }
650      }
651      FreeTemp(r_method);
652    }
653    // rBase now holds static storage base
654    RegisterClass reg_class = RegClassForFieldLoadStore(size, field_info.IsVolatile());
655    if (IsWide(size)) {
656      rl_src = LoadValueWide(rl_src, reg_class);
657    } else {
658      rl_src = LoadValue(rl_src, reg_class);
659    }
660    if (IsRef(size)) {
661      StoreRefDisp(r_base, field_info.FieldOffset().Int32Value(), rl_src.reg,
662                   field_info.IsVolatile() ? kVolatile : kNotVolatile);
663    } else {
664      StoreBaseDisp(r_base, field_info.FieldOffset().Int32Value(), rl_src.reg, size,
665                    field_info.IsVolatile() ? kVolatile : kNotVolatile);
666    }
667    if (IsRef(size) && !mir_graph_->IsConstantNullRef(rl_src)) {
668      MarkGCCard(rl_src.reg, r_base);
669    }
670    FreeTemp(r_base);
671  } else {
672    FlushAllRegs();  // Everything to home locations
673    QuickEntrypointEnum target;
674    switch (size) {
675      case kReference:
676        target = kQuickSetObjStatic;
677        break;
678      case k64:
679      case kDouble:
680        target = kQuickSet64Static;
681        break;
682      case k32:
683      case kSingle:
684        target = kQuickSet32Static;
685        break;
686      case kSignedHalf:
687      case kUnsignedHalf:
688        target = kQuickSet16Static;
689        break;
690      case kSignedByte:
691      case kUnsignedByte:
692        target = kQuickSet8Static;
693        break;
694      case kWord:  // Intentional fallthrough.
695      default:
696        LOG(FATAL) << "Can't determine entrypoint for: " << size;
697        target = kQuickSet32Static;
698    }
699    CallRuntimeHelperImmRegLocation(target, field_info.FieldIndex(), rl_src, true);
700  }
701}
702
703void Mir2Lir::GenSget(MIR* mir, RegLocation rl_dest, OpSize size, Primitive::Type type) {
704  const MirSFieldLoweringInfo& field_info = mir_graph_->GetSFieldLoweringInfo(mir);
705  DCHECK_EQ(SGetMemAccessType(mir->dalvikInsn.opcode), field_info.MemAccessType());
706  cu_->compiler_driver->ProcessedStaticField(field_info.FastGet(), field_info.IsReferrersClass());
707
708  if (!SLOW_FIELD_PATH && field_info.FastGet()) {
709    DCHECK_GE(field_info.FieldOffset().Int32Value(), 0);
710    RegStorage r_base;
711    if (field_info.IsReferrersClass()) {
712      // Fast path, static storage base is this method's class
713      RegLocation rl_method  = LoadCurrMethod();
714      r_base = AllocTempRef();
715      LoadRefDisp(rl_method.reg, mirror::ArtMethod::DeclaringClassOffset().Int32Value(), r_base,
716                  kNotVolatile);
717    } else {
718      // Medium path, static storage base in a different class which requires checks that the other
719      // class is initialized
720      DCHECK_NE(field_info.StorageIndex(), DexFile::kDexNoIndex);
721      // May do runtime call so everything to home locations.
722      FlushAllRegs();
723      // Using fixed register to sync with possible call to runtime support.
724      RegStorage r_method = TargetReg(kArg1, kRef);
725      LockTemp(r_method);
726      LoadCurrMethodDirect(r_method);
727      r_base = TargetReg(kArg0, kRef);
728      LockTemp(r_base);
729      LoadRefDisp(r_method, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(), r_base,
730                  kNotVolatile);
731      int32_t offset_of_field = ObjArray::OffsetOfElement(field_info.StorageIndex()).Int32Value();
732      LoadRefDisp(r_base, offset_of_field, r_base, kNotVolatile);
733      // r_base now points at static storage (Class*) or NULL if the type is not yet resolved.
734      LIR* unresolved_branch = nullptr;
735      if (!field_info.IsClassInDexCache() &&
736          (mir->optimization_flags & MIR_CLASS_IS_IN_DEX_CACHE) == 0) {
737        // Check if r_base is NULL.
738        unresolved_branch = OpCmpImmBranch(kCondEq, r_base, 0, NULL);
739      }
740      LIR* uninit_branch = nullptr;
741      if (!field_info.IsClassInitialized() &&
742          (mir->optimization_flags & MIR_CLASS_IS_INITIALIZED) == 0) {
743        // Check if r_base is not yet initialized class.
744        RegStorage r_tmp = TargetReg(kArg2, kNotWide);
745        LockTemp(r_tmp);
746        uninit_branch = OpCmpMemImmBranch(kCondLt, r_tmp, r_base,
747                                          mirror::Class::StatusOffset().Int32Value(),
748                                          mirror::Class::kStatusInitialized, nullptr, nullptr);
749        FreeTemp(r_tmp);
750      }
751      if (unresolved_branch != nullptr || uninit_branch != nullptr) {
752        // The slow path is invoked if the r_base is NULL or the class pointed
753        // to by it is not initialized.
754        LIR* cont = NewLIR0(kPseudoTargetLabel);
755        AddSlowPath(new (arena_) StaticFieldSlowPath(this, unresolved_branch, uninit_branch, cont,
756                                                     field_info.StorageIndex(), r_base));
757
758        if (uninit_branch != nullptr) {
759          // Ensure load of status and load of value don't re-order.
760          GenMemBarrier(kLoadAny);
761        }
762      }
763      FreeTemp(r_method);
764    }
765    // r_base now holds static storage base
766    RegisterClass reg_class = RegClassForFieldLoadStore(size, field_info.IsVolatile());
767    RegLocation rl_result = EvalLoc(rl_dest, reg_class, true);
768
769    int field_offset = field_info.FieldOffset().Int32Value();
770    if (IsRef(size)) {
771      // TODO: DCHECK?
772      LoadRefDisp(r_base, field_offset, rl_result.reg, field_info.IsVolatile() ? kVolatile :
773          kNotVolatile);
774    } else {
775      LoadBaseDisp(r_base, field_offset, rl_result.reg, size, field_info.IsVolatile() ?
776          kVolatile : kNotVolatile);
777    }
778    FreeTemp(r_base);
779
780    if (IsWide(size)) {
781      StoreValueWide(rl_dest, rl_result);
782    } else {
783      StoreValue(rl_dest, rl_result);
784    }
785  } else {
786    DCHECK(SizeMatchesTypeForEntrypoint(size, type));
787    FlushAllRegs();  // Everything to home locations
788    QuickEntrypointEnum target;
789    switch (type) {
790      case Primitive::kPrimNot:
791        target = kQuickGetObjStatic;
792        break;
793      case Primitive::kPrimLong:
794      case Primitive::kPrimDouble:
795        target = kQuickGet64Static;
796        break;
797      case Primitive::kPrimInt:
798      case Primitive::kPrimFloat:
799        target = kQuickGet32Static;
800        break;
801      case Primitive::kPrimShort:
802        target = kQuickGetShortStatic;
803        break;
804      case Primitive::kPrimChar:
805        target = kQuickGetCharStatic;
806        break;
807      case Primitive::kPrimByte:
808        target = kQuickGetByteStatic;
809        break;
810      case Primitive::kPrimBoolean:
811        target = kQuickGetBooleanStatic;
812        break;
813      case Primitive::kPrimVoid:  // Intentional fallthrough.
814      default:
815        LOG(FATAL) << "Can't determine entrypoint for: " << type;
816        target = kQuickGet32Static;
817    }
818    CallRuntimeHelperImm(target, field_info.FieldIndex(), true);
819
820    // FIXME: pGetXXStatic always return an int or int64 regardless of rl_dest.fp.
821    if (IsWide(size)) {
822      RegLocation rl_result = GetReturnWide(kCoreReg);
823      StoreValueWide(rl_dest, rl_result);
824    } else {
825      RegLocation rl_result = GetReturn(rl_dest.ref ? kRefReg : kCoreReg);
826      StoreValue(rl_dest, rl_result);
827    }
828  }
829}
830
831// Generate code for all slow paths.
832void Mir2Lir::HandleSlowPaths() {
833  // We should check slow_paths_.Size() every time, because a new slow path
834  // may be created during slowpath->Compile().
835  for (LIRSlowPath* slowpath : slow_paths_) {
836    slowpath->Compile();
837  }
838  slow_paths_.clear();
839}
840
841void Mir2Lir::GenIGet(MIR* mir, int opt_flags, OpSize size, Primitive::Type type,
842                      RegLocation rl_dest, RegLocation rl_obj) {
843  const MirIFieldLoweringInfo& field_info = mir_graph_->GetIFieldLoweringInfo(mir);
844  DCHECK_EQ(IGetMemAccessType(mir->dalvikInsn.opcode), field_info.MemAccessType());
845  cu_->compiler_driver->ProcessedInstanceField(field_info.FastGet());
846  if (!SLOW_FIELD_PATH && field_info.FastGet()) {
847    RegisterClass reg_class = RegClassForFieldLoadStore(size, field_info.IsVolatile());
848    // A load of the class will lead to an iget with offset 0.
849    DCHECK_GE(field_info.FieldOffset().Int32Value(), 0);
850    rl_obj = LoadValue(rl_obj, kRefReg);
851    GenNullCheck(rl_obj.reg, opt_flags);
852    RegLocation rl_result = EvalLoc(rl_dest, reg_class, true);
853    int field_offset = field_info.FieldOffset().Int32Value();
854    LIR* load_lir;
855    if (IsRef(size)) {
856      load_lir = LoadRefDisp(rl_obj.reg, field_offset, rl_result.reg, field_info.IsVolatile() ?
857          kVolatile : kNotVolatile);
858    } else {
859      load_lir = LoadBaseDisp(rl_obj.reg, field_offset, rl_result.reg, size,
860                              field_info.IsVolatile() ? kVolatile : kNotVolatile);
861    }
862    MarkPossibleNullPointerExceptionAfter(opt_flags, load_lir);
863    if (IsWide(size)) {
864      StoreValueWide(rl_dest, rl_result);
865    } else {
866      StoreValue(rl_dest, rl_result);
867    }
868  } else {
869    DCHECK(SizeMatchesTypeForEntrypoint(size, type));
870    QuickEntrypointEnum target;
871    switch (type) {
872      case Primitive::kPrimNot:
873        target = kQuickGetObjInstance;
874        break;
875      case Primitive::kPrimLong:
876      case Primitive::kPrimDouble:
877        target = kQuickGet64Instance;
878        break;
879      case Primitive::kPrimFloat:
880      case Primitive::kPrimInt:
881        target = kQuickGet32Instance;
882        break;
883      case Primitive::kPrimShort:
884        target = kQuickGetShortInstance;
885        break;
886      case Primitive::kPrimChar:
887        target = kQuickGetCharInstance;
888        break;
889      case Primitive::kPrimByte:
890        target = kQuickGetByteInstance;
891        break;
892      case Primitive::kPrimBoolean:
893        target = kQuickGetBooleanInstance;
894        break;
895      case Primitive::kPrimVoid:  // Intentional fallthrough.
896      default:
897        LOG(FATAL) << "Can't determine entrypoint for: " << type;
898        target = kQuickGet32Instance;
899    }
900    // Second argument of pGetXXInstance is always a reference.
901    DCHECK_EQ(static_cast<unsigned int>(rl_obj.wide), 0U);
902    CallRuntimeHelperImmRegLocation(target, field_info.FieldIndex(), rl_obj, true);
903
904    // FIXME: pGetXXInstance always return an int or int64 regardless of rl_dest.fp.
905    if (IsWide(size)) {
906      RegLocation rl_result = GetReturnWide(kCoreReg);
907      StoreValueWide(rl_dest, rl_result);
908    } else {
909      RegLocation rl_result = GetReturn(rl_dest.ref ? kRefReg : kCoreReg);
910      StoreValue(rl_dest, rl_result);
911    }
912  }
913}
914
915void Mir2Lir::GenIPut(MIR* mir, int opt_flags, OpSize size,
916                      RegLocation rl_src, RegLocation rl_obj) {
917  const MirIFieldLoweringInfo& field_info = mir_graph_->GetIFieldLoweringInfo(mir);
918  DCHECK_EQ(IPutMemAccessType(mir->dalvikInsn.opcode), field_info.MemAccessType());
919  cu_->compiler_driver->ProcessedInstanceField(field_info.FastPut());
920  if (!SLOW_FIELD_PATH && field_info.FastPut()) {
921    RegisterClass reg_class = RegClassForFieldLoadStore(size, field_info.IsVolatile());
922    // Dex code never writes to the class field.
923    DCHECK_GE(static_cast<uint32_t>(field_info.FieldOffset().Int32Value()),
924              sizeof(mirror::HeapReference<mirror::Class>));
925    rl_obj = LoadValue(rl_obj, kRefReg);
926    if (IsWide(size)) {
927      rl_src = LoadValueWide(rl_src, reg_class);
928    } else {
929      rl_src = LoadValue(rl_src, reg_class);
930    }
931    GenNullCheck(rl_obj.reg, opt_flags);
932    int field_offset = field_info.FieldOffset().Int32Value();
933    LIR* store;
934    if (IsRef(size)) {
935      store = StoreRefDisp(rl_obj.reg, field_offset, rl_src.reg, field_info.IsVolatile() ?
936          kVolatile : kNotVolatile);
937    } else {
938      store = StoreBaseDisp(rl_obj.reg, field_offset, rl_src.reg, size,
939                            field_info.IsVolatile() ? kVolatile : kNotVolatile);
940    }
941    MarkPossibleNullPointerExceptionAfter(opt_flags, store);
942    if (IsRef(size) && !mir_graph_->IsConstantNullRef(rl_src)) {
943      MarkGCCard(rl_src.reg, rl_obj.reg);
944    }
945  } else {
946    QuickEntrypointEnum target;
947    switch (size) {
948      case kReference:
949        target = kQuickSetObjInstance;
950        break;
951      case k64:
952      case kDouble:
953        target = kQuickSet64Instance;
954        break;
955      case k32:
956      case kSingle:
957        target = kQuickSet32Instance;
958        break;
959      case kSignedHalf:
960      case kUnsignedHalf:
961        target = kQuickSet16Instance;
962        break;
963      case kSignedByte:
964      case kUnsignedByte:
965        target = kQuickSet8Instance;
966        break;
967      case kWord:  // Intentional fallthrough.
968      default:
969        LOG(FATAL) << "Can't determine entrypoint for: " << size;
970        target = kQuickSet32Instance;
971    }
972    CallRuntimeHelperImmRegLocationRegLocation(target, field_info.FieldIndex(), rl_obj, rl_src,
973                                               true);
974  }
975}
976
977void Mir2Lir::GenArrayObjPut(int opt_flags, RegLocation rl_array, RegLocation rl_index,
978                             RegLocation rl_src) {
979  bool needs_range_check = !(opt_flags & MIR_IGNORE_RANGE_CHECK);
980  bool needs_null_check = !((cu_->disable_opt & (1 << kNullCheckElimination)) &&
981      (opt_flags & MIR_IGNORE_NULL_CHECK));
982  QuickEntrypointEnum target = needs_range_check
983        ? (needs_null_check ? kQuickAputObjectWithNullAndBoundCheck
984                            : kQuickAputObjectWithBoundCheck)
985        : kQuickAputObject;
986  CallRuntimeHelperRegLocationRegLocationRegLocation(target, rl_array, rl_index, rl_src, true);
987}
988
989void Mir2Lir::GenConstClass(uint32_t type_idx, RegLocation rl_dest) {
990  RegLocation rl_method = LoadCurrMethod();
991  CheckRegLocation(rl_method);
992  RegStorage res_reg = AllocTempRef();
993  if (!cu_->compiler_driver->CanAccessTypeWithoutChecks(cu_->method_idx,
994                                                        *cu_->dex_file,
995                                                        type_idx)) {
996    // Call out to helper which resolves type and verifies access.
997    // Resolved type returned in kRet0.
998    CallRuntimeHelperImmReg(kQuickInitializeTypeAndVerifyAccess, type_idx, rl_method.reg, true);
999    RegLocation rl_result = GetReturn(kRefReg);
1000    StoreValue(rl_dest, rl_result);
1001  } else {
1002    RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true);
1003    // We're don't need access checks, load type from dex cache
1004    int32_t dex_cache_offset =
1005        mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value();
1006    LoadRefDisp(rl_method.reg, dex_cache_offset, res_reg, kNotVolatile);
1007    int32_t offset_of_type = ClassArray::OffsetOfElement(type_idx).Int32Value();
1008    LoadRefDisp(res_reg, offset_of_type, rl_result.reg, kNotVolatile);
1009    if (!cu_->compiler_driver->CanAssumeTypeIsPresentInDexCache(*cu_->dex_file,
1010        type_idx) || SLOW_TYPE_PATH) {
1011      // Slow path, at runtime test if type is null and if so initialize
1012      FlushAllRegs();
1013      LIR* branch = OpCmpImmBranch(kCondEq, rl_result.reg, 0, NULL);
1014      LIR* cont = NewLIR0(kPseudoTargetLabel);
1015
1016      // Object to generate the slow path for class resolution.
1017      class SlowPath : public LIRSlowPath {
1018       public:
1019        SlowPath(Mir2Lir* m2l, LIR* fromfast, LIR* cont_in, const int type_idx_in,
1020                 const RegLocation& rl_method_in, const RegLocation& rl_result_in) :
1021                   LIRSlowPath(m2l, m2l->GetCurrentDexPc(), fromfast, cont_in),
1022                   type_idx_(type_idx_in), rl_method_(rl_method_in), rl_result_(rl_result_in) {
1023        }
1024
1025        void Compile() {
1026          GenerateTargetLabel();
1027
1028          m2l_->CallRuntimeHelperImmReg(kQuickInitializeType, type_idx_, rl_method_.reg, true);
1029          m2l_->OpRegCopy(rl_result_.reg,  m2l_->TargetReg(kRet0, kRef));
1030          m2l_->OpUnconditionalBranch(cont_);
1031        }
1032
1033       private:
1034        const int type_idx_;
1035        const RegLocation rl_method_;
1036        const RegLocation rl_result_;
1037      };
1038
1039      // Add to list for future.
1040      AddSlowPath(new (arena_) SlowPath(this, branch, cont, type_idx, rl_method, rl_result));
1041
1042      StoreValue(rl_dest, rl_result);
1043     } else {
1044      // Fast path, we're done - just store result
1045      StoreValue(rl_dest, rl_result);
1046    }
1047  }
1048}
1049
1050void Mir2Lir::GenConstString(uint32_t string_idx, RegLocation rl_dest) {
1051  /* NOTE: Most strings should be available at compile time */
1052  int32_t offset_of_string = mirror::ObjectArray<mirror::String>::OffsetOfElement(string_idx).
1053                                                                                      Int32Value();
1054  if (!cu_->compiler_driver->CanAssumeStringIsPresentInDexCache(
1055      *cu_->dex_file, string_idx) || SLOW_STRING_PATH) {
1056    // slow path, resolve string if not in dex cache
1057    FlushAllRegs();
1058    LockCallTemps();  // Using explicit registers
1059
1060    // If the Method* is already in a register, we can save a copy.
1061    RegLocation rl_method = mir_graph_->GetMethodLoc();
1062    RegStorage r_method;
1063    if (rl_method.location == kLocPhysReg) {
1064      // A temp would conflict with register use below.
1065      DCHECK(!IsTemp(rl_method.reg));
1066      r_method = rl_method.reg;
1067    } else {
1068      r_method = TargetReg(kArg2, kRef);
1069      LoadCurrMethodDirect(r_method);
1070    }
1071    LoadRefDisp(r_method, mirror::ArtMethod::DexCacheStringsOffset().Int32Value(),
1072                TargetReg(kArg0, kRef), kNotVolatile);
1073
1074    // Might call out to helper, which will return resolved string in kRet0
1075    LoadRefDisp(TargetReg(kArg0, kRef), offset_of_string, TargetReg(kRet0, kRef), kNotVolatile);
1076    LIR* fromfast = OpCmpImmBranch(kCondEq, TargetReg(kRet0, kRef), 0, NULL);
1077    LIR* cont = NewLIR0(kPseudoTargetLabel);
1078
1079    {
1080      // Object to generate the slow path for string resolution.
1081      class SlowPath : public LIRSlowPath {
1082       public:
1083        SlowPath(Mir2Lir* m2l, LIR* fromfast_in, LIR* cont_in, RegStorage r_method_in,
1084                 int32_t string_idx_in) :
1085            LIRSlowPath(m2l, m2l->GetCurrentDexPc(), fromfast_in, cont_in),
1086            r_method_(r_method_in), string_idx_(string_idx_in) {
1087        }
1088
1089        void Compile() {
1090          GenerateTargetLabel();
1091          m2l_->CallRuntimeHelperRegImm(kQuickResolveString, r_method_, string_idx_, true);
1092          m2l_->OpUnconditionalBranch(cont_);
1093        }
1094
1095       private:
1096         const RegStorage r_method_;
1097         const int32_t string_idx_;
1098      };
1099
1100      AddSlowPath(new (arena_) SlowPath(this, fromfast, cont, r_method, string_idx));
1101    }
1102
1103    GenBarrier();
1104    StoreValue(rl_dest, GetReturn(kRefReg));
1105  } else {
1106    RegLocation rl_method = LoadCurrMethod();
1107    RegStorage res_reg = AllocTempRef();
1108    RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true);
1109    LoadRefDisp(rl_method.reg, mirror::ArtMethod::DexCacheStringsOffset().Int32Value(), res_reg,
1110                kNotVolatile);
1111    LoadRefDisp(res_reg, offset_of_string, rl_result.reg, kNotVolatile);
1112    StoreValue(rl_dest, rl_result);
1113  }
1114}
1115
1116/*
1117 * Let helper function take care of everything.  Will
1118 * call Class::NewInstanceFromCode(type_idx, method);
1119 */
1120void Mir2Lir::GenNewInstance(uint32_t type_idx, RegLocation rl_dest) {
1121  FlushAllRegs();  /* Everything to home location */
1122  // alloc will always check for resolution, do we also need to verify
1123  // access because the verifier was unable to?
1124  const DexFile* dex_file = cu_->dex_file;
1125  CompilerDriver* driver = cu_->compiler_driver;
1126  if (driver->CanAccessInstantiableTypeWithoutChecks(cu_->method_idx, *dex_file, type_idx)) {
1127    bool is_type_initialized;
1128    bool use_direct_type_ptr;
1129    uintptr_t direct_type_ptr;
1130    bool is_finalizable;
1131    if (kEmbedClassInCode &&
1132        driver->CanEmbedTypeInCode(*dex_file, type_idx, &is_type_initialized, &use_direct_type_ptr,
1133                                   &direct_type_ptr, &is_finalizable) &&
1134                                   !is_finalizable) {
1135      // The fast path.
1136      if (!use_direct_type_ptr) {
1137        LoadClassType(*dex_file, type_idx, kArg0);
1138        if (!is_type_initialized) {
1139          CallRuntimeHelperRegMethod(kQuickAllocObjectResolved, TargetReg(kArg0, kRef), true);
1140        } else {
1141          CallRuntimeHelperRegMethod(kQuickAllocObjectInitialized, TargetReg(kArg0, kRef), true);
1142        }
1143      } else {
1144        // Use the direct pointer.
1145        if (!is_type_initialized) {
1146          CallRuntimeHelperImmMethod(kQuickAllocObjectResolved, direct_type_ptr, true);
1147        } else {
1148          CallRuntimeHelperImmMethod(kQuickAllocObjectInitialized, direct_type_ptr, true);
1149        }
1150      }
1151    } else {
1152      // The slow path.
1153      CallRuntimeHelperImmMethod(kQuickAllocObject, type_idx, true);
1154    }
1155  } else {
1156    CallRuntimeHelperImmMethod(kQuickAllocObjectWithAccessCheck, type_idx, true);
1157  }
1158  StoreValue(rl_dest, GetReturn(kRefReg));
1159}
1160
1161void Mir2Lir::GenThrow(RegLocation rl_src) {
1162  FlushAllRegs();
1163  CallRuntimeHelperRegLocation(kQuickDeliverException, rl_src, true);
1164}
1165
1166// For final classes there are no sub-classes to check and so we can answer the instance-of
1167// question with simple comparisons.
1168void Mir2Lir::GenInstanceofFinal(bool use_declaring_class, uint32_t type_idx, RegLocation rl_dest,
1169                                 RegLocation rl_src) {
1170  // X86 has its own implementation.
1171  DCHECK(cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64);
1172
1173  RegLocation object = LoadValue(rl_src, kRefReg);
1174  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1175  RegStorage result_reg = rl_result.reg;
1176  if (IsSameReg(result_reg, object.reg)) {
1177    result_reg = AllocTypedTemp(false, kCoreReg);
1178    DCHECK(!IsSameReg(result_reg, object.reg));
1179  }
1180  LoadConstant(result_reg, 0);     // assume false
1181  LIR* null_branchover = OpCmpImmBranch(kCondEq, object.reg, 0, NULL);
1182
1183  RegStorage check_class = AllocTypedTemp(false, kRefReg);
1184  RegStorage object_class = AllocTypedTemp(false, kRefReg);
1185
1186  LoadCurrMethodDirect(check_class);
1187  if (use_declaring_class) {
1188    LoadRefDisp(check_class, mirror::ArtMethod::DeclaringClassOffset().Int32Value(), check_class,
1189                kNotVolatile);
1190    LoadRefDisp(object.reg,  mirror::Object::ClassOffset().Int32Value(), object_class,
1191                kNotVolatile);
1192  } else {
1193    LoadRefDisp(check_class, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(),
1194                check_class, kNotVolatile);
1195    LoadRefDisp(object.reg,  mirror::Object::ClassOffset().Int32Value(), object_class,
1196                kNotVolatile);
1197    int32_t offset_of_type = ClassArray::OffsetOfElement(type_idx).Int32Value();
1198    LoadRefDisp(check_class, offset_of_type, check_class, kNotVolatile);
1199  }
1200
1201  // FIXME: what should we be comparing here? compressed or decompressed references?
1202  if (cu_->instruction_set == kThumb2) {
1203    OpRegReg(kOpCmp, check_class, object_class);  // Same?
1204    LIR* it = OpIT(kCondEq, "");   // if-convert the test
1205    LoadConstant(result_reg, 1);     // .eq case - load true
1206    OpEndIT(it);
1207  } else {
1208    GenSelectConst32(check_class, object_class, kCondEq, 1, 0, result_reg, kCoreReg);
1209  }
1210  LIR* target = NewLIR0(kPseudoTargetLabel);
1211  null_branchover->target = target;
1212  FreeTemp(object_class);
1213  FreeTemp(check_class);
1214  if (IsTemp(result_reg)) {
1215    OpRegCopy(rl_result.reg, result_reg);
1216    FreeTemp(result_reg);
1217  }
1218  StoreValue(rl_dest, rl_result);
1219}
1220
1221void Mir2Lir::GenInstanceofCallingHelper(bool needs_access_check, bool type_known_final,
1222                                         bool type_known_abstract, bool use_declaring_class,
1223                                         bool can_assume_type_is_in_dex_cache,
1224                                         uint32_t type_idx, RegLocation rl_dest,
1225                                         RegLocation rl_src) {
1226  FlushAllRegs();
1227  // May generate a call - use explicit registers
1228  LockCallTemps();
1229  RegStorage method_reg = TargetReg(kArg1, kRef);
1230  LoadCurrMethodDirect(method_reg);   // kArg1 <= current Method*
1231  RegStorage class_reg = TargetReg(kArg2, kRef);  // kArg2 will hold the Class*
1232  RegStorage ref_reg = TargetReg(kArg0, kRef);  // kArg0 will hold the ref.
1233  RegStorage ret_reg = GetReturn(kRefReg).reg;
1234  if (needs_access_check) {
1235    // Check we have access to type_idx and if not throw IllegalAccessError,
1236    // returns Class* in kArg0
1237    CallRuntimeHelperImm(kQuickInitializeTypeAndVerifyAccess, type_idx, true);
1238    OpRegCopy(class_reg, ret_reg);  // Align usage with fast path
1239    LoadValueDirectFixed(rl_src, ref_reg);  // kArg0 <= ref
1240  } else if (use_declaring_class) {
1241    LoadValueDirectFixed(rl_src, ref_reg);  // kArg0 <= ref
1242    LoadRefDisp(method_reg, mirror::ArtMethod::DeclaringClassOffset().Int32Value(),
1243                class_reg, kNotVolatile);
1244  } else {
1245    if (can_assume_type_is_in_dex_cache) {
1246      // Conditionally, as in the other case we will also load it.
1247      LoadValueDirectFixed(rl_src, ref_reg);  // kArg0 <= ref
1248    }
1249
1250    // Load dex cache entry into class_reg (kArg2)
1251    LoadRefDisp(method_reg, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(),
1252                class_reg, kNotVolatile);
1253    int32_t offset_of_type = ClassArray::OffsetOfElement(type_idx).Int32Value();
1254    LoadRefDisp(class_reg, offset_of_type, class_reg, kNotVolatile);
1255    if (!can_assume_type_is_in_dex_cache) {
1256      LIR* slow_path_branch = OpCmpImmBranch(kCondEq, class_reg, 0, NULL);
1257      LIR* slow_path_target = NewLIR0(kPseudoTargetLabel);
1258
1259      // Should load value here.
1260      LoadValueDirectFixed(rl_src, ref_reg);  // kArg0 <= ref
1261
1262      class InitTypeSlowPath : public Mir2Lir::LIRSlowPath {
1263       public:
1264        InitTypeSlowPath(Mir2Lir* m2l, LIR* branch, LIR* cont, uint32_t type_idx_in,
1265                         RegLocation rl_src_in)
1266            : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), branch, cont), type_idx_(type_idx_in),
1267              rl_src_(rl_src_in) {
1268        }
1269
1270        void Compile() OVERRIDE {
1271          GenerateTargetLabel();
1272
1273          m2l_->CallRuntimeHelperImm(kQuickInitializeType, type_idx_, true);
1274          m2l_->OpRegCopy(m2l_->TargetReg(kArg2, kRef),
1275                          m2l_->TargetReg(kRet0, kRef));  // Align usage with fast path
1276          m2l_->OpUnconditionalBranch(cont_);
1277        }
1278
1279       private:
1280        uint32_t type_idx_;
1281        RegLocation rl_src_;
1282      };
1283
1284      AddSlowPath(new (arena_) InitTypeSlowPath(this, slow_path_branch, slow_path_target,
1285                                                type_idx, rl_src));
1286    }
1287  }
1288  /* kArg0 is ref, kArg2 is class. If ref==null, use directly as bool result */
1289  RegLocation rl_result = GetReturn(kCoreReg);
1290  if (!IsSameReg(rl_result.reg, ref_reg)) {
1291    // On MIPS and x86_64 rArg0 != rl_result, place false in result if branch is taken.
1292    LoadConstant(rl_result.reg, 0);
1293  }
1294  LIR* branch1 = OpCmpImmBranch(kCondEq, ref_reg, 0, NULL);
1295
1296  /* load object->klass_ */
1297  RegStorage ref_class_reg = TargetReg(kArg1, kRef);  // kArg1 will hold the Class* of ref.
1298  DCHECK_EQ(mirror::Object::ClassOffset().Int32Value(), 0);
1299  LoadRefDisp(ref_reg, mirror::Object::ClassOffset().Int32Value(),
1300              ref_class_reg, kNotVolatile);
1301  /* kArg0 is ref, kArg1 is ref->klass_, kArg2 is class */
1302  LIR* branchover = NULL;
1303  if (type_known_final) {
1304    // rl_result == ref == class.
1305    GenSelectConst32(ref_class_reg, class_reg, kCondEq, 1, 0, rl_result.reg,
1306                     kCoreReg);
1307  } else {
1308    if (cu_->instruction_set == kThumb2) {
1309      RegStorage r_tgt = LoadHelper(kQuickInstanceofNonTrivial);
1310      LIR* it = nullptr;
1311      if (!type_known_abstract) {
1312      /* Uses conditional nullification */
1313        OpRegReg(kOpCmp, ref_class_reg, class_reg);  // Same?
1314        it = OpIT(kCondEq, "EE");   // if-convert the test
1315        LoadConstant(rl_result.reg, 1);     // .eq case - load true
1316      }
1317      OpRegCopy(ref_reg, class_reg);    // .ne case - arg0 <= class
1318      OpReg(kOpBlx, r_tgt);    // .ne case: helper(class, ref->class)
1319      if (it != nullptr) {
1320        OpEndIT(it);
1321      }
1322      FreeTemp(r_tgt);
1323    } else {
1324      if (!type_known_abstract) {
1325        /* Uses branchovers */
1326        LoadConstant(rl_result.reg, 1);     // assume true
1327        branchover = OpCmpBranch(kCondEq, TargetReg(kArg1, kRef), TargetReg(kArg2, kRef), NULL);
1328      }
1329
1330      OpRegCopy(TargetReg(kArg0, kRef), class_reg);    // .ne case - arg0 <= class
1331      CallRuntimeHelper(kQuickInstanceofNonTrivial, false);
1332    }
1333  }
1334  // TODO: only clobber when type isn't final?
1335  ClobberCallerSave();
1336  /* branch targets here */
1337  LIR* target = NewLIR0(kPseudoTargetLabel);
1338  StoreValue(rl_dest, rl_result);
1339  branch1->target = target;
1340  if (branchover != nullptr) {
1341    branchover->target = target;
1342  }
1343}
1344
1345void Mir2Lir::GenInstanceof(uint32_t type_idx, RegLocation rl_dest, RegLocation rl_src) {
1346  bool type_known_final, type_known_abstract, use_declaring_class;
1347  bool needs_access_check = !cu_->compiler_driver->CanAccessTypeWithoutChecks(cu_->method_idx,
1348                                                                              *cu_->dex_file,
1349                                                                              type_idx,
1350                                                                              &type_known_final,
1351                                                                              &type_known_abstract,
1352                                                                              &use_declaring_class);
1353  bool can_assume_type_is_in_dex_cache = !needs_access_check &&
1354      cu_->compiler_driver->CanAssumeTypeIsPresentInDexCache(*cu_->dex_file, type_idx);
1355
1356  if ((use_declaring_class || can_assume_type_is_in_dex_cache) && type_known_final) {
1357    GenInstanceofFinal(use_declaring_class, type_idx, rl_dest, rl_src);
1358  } else {
1359    GenInstanceofCallingHelper(needs_access_check, type_known_final, type_known_abstract,
1360                               use_declaring_class, can_assume_type_is_in_dex_cache,
1361                               type_idx, rl_dest, rl_src);
1362  }
1363}
1364
1365void Mir2Lir::GenCheckCast(uint32_t insn_idx, uint32_t type_idx, RegLocation rl_src) {
1366  bool type_known_final, type_known_abstract, use_declaring_class;
1367  bool needs_access_check = !cu_->compiler_driver->CanAccessTypeWithoutChecks(cu_->method_idx,
1368                                                                              *cu_->dex_file,
1369                                                                              type_idx,
1370                                                                              &type_known_final,
1371                                                                              &type_known_abstract,
1372                                                                              &use_declaring_class);
1373  // Note: currently type_known_final is unused, as optimizing will only improve the performance
1374  // of the exception throw path.
1375  DexCompilationUnit* cu = mir_graph_->GetCurrentDexCompilationUnit();
1376  if (!needs_access_check && cu_->compiler_driver->IsSafeCast(cu, insn_idx)) {
1377    // Verifier type analysis proved this check cast would never cause an exception.
1378    return;
1379  }
1380  FlushAllRegs();
1381  // May generate a call - use explicit registers
1382  LockCallTemps();
1383  RegStorage method_reg = TargetReg(kArg1, kRef);
1384  LoadCurrMethodDirect(method_reg);  // kArg1 <= current Method*
1385  RegStorage class_reg = TargetReg(kArg2, kRef);  // kArg2 will hold the Class*
1386  if (needs_access_check) {
1387    // Check we have access to type_idx and if not throw IllegalAccessError,
1388    // returns Class* in kRet0
1389    // InitializeTypeAndVerifyAccess(idx, method)
1390    CallRuntimeHelperImm(kQuickInitializeTypeAndVerifyAccess, type_idx, true);
1391    OpRegCopy(class_reg, TargetReg(kRet0, kRef));  // Align usage with fast path
1392  } else if (use_declaring_class) {
1393    LoadRefDisp(method_reg, mirror::ArtMethod::DeclaringClassOffset().Int32Value(),
1394                class_reg, kNotVolatile);
1395  } else {
1396    // Load dex cache entry into class_reg (kArg2)
1397    LoadRefDisp(method_reg, mirror::ArtMethod::DexCacheResolvedTypesOffset().Int32Value(),
1398                class_reg, kNotVolatile);
1399    int32_t offset_of_type = ClassArray::OffsetOfElement(type_idx).Int32Value();
1400    LoadRefDisp(class_reg, offset_of_type, class_reg, kNotVolatile);
1401    if (!cu_->compiler_driver->CanAssumeTypeIsPresentInDexCache(*cu_->dex_file, type_idx)) {
1402      // Need to test presence of type in dex cache at runtime
1403      LIR* hop_branch = OpCmpImmBranch(kCondEq, class_reg, 0, NULL);
1404      LIR* cont = NewLIR0(kPseudoTargetLabel);
1405
1406      // Slow path to initialize the type.  Executed if the type is NULL.
1407      class SlowPath : public LIRSlowPath {
1408       public:
1409        SlowPath(Mir2Lir* m2l, LIR* fromfast, LIR* cont_in, const int type_idx_in,
1410                 const RegStorage class_reg_in) :
1411                   LIRSlowPath(m2l, m2l->GetCurrentDexPc(), fromfast, cont_in),
1412                   type_idx_(type_idx_in), class_reg_(class_reg_in) {
1413        }
1414
1415        void Compile() {
1416          GenerateTargetLabel();
1417
1418          // Call out to helper, which will return resolved type in kArg0
1419          // InitializeTypeFromCode(idx, method)
1420          m2l_->CallRuntimeHelperImmReg(kQuickInitializeType, type_idx_,
1421                                        m2l_->TargetReg(kArg1, kRef), true);
1422          m2l_->OpRegCopy(class_reg_, m2l_->TargetReg(kRet0, kRef));  // Align usage with fast path
1423          m2l_->OpUnconditionalBranch(cont_);
1424        }
1425
1426       public:
1427        const int type_idx_;
1428        const RegStorage class_reg_;
1429      };
1430
1431      AddSlowPath(new (arena_) SlowPath(this, hop_branch, cont, type_idx, class_reg));
1432    }
1433  }
1434  // At this point, class_reg (kArg2) has class
1435  LoadValueDirectFixed(rl_src, TargetReg(kArg0, kRef));  // kArg0 <= ref
1436
1437  // Slow path for the case where the classes are not equal.  In this case we need
1438  // to call a helper function to do the check.
1439  class SlowPath : public LIRSlowPath {
1440   public:
1441    SlowPath(Mir2Lir* m2l, LIR* fromfast, LIR* cont, bool load):
1442               LIRSlowPath(m2l, m2l->GetCurrentDexPc(), fromfast, cont), load_(load) {
1443    }
1444
1445    void Compile() {
1446      GenerateTargetLabel();
1447
1448      if (load_) {
1449        m2l_->LoadRefDisp(m2l_->TargetReg(kArg0, kRef), mirror::Object::ClassOffset().Int32Value(),
1450                          m2l_->TargetReg(kArg1, kRef), kNotVolatile);
1451      }
1452      m2l_->CallRuntimeHelperRegReg(kQuickCheckCast, m2l_->TargetReg(kArg2, kRef),
1453                                    m2l_->TargetReg(kArg1, kRef), true);
1454      m2l_->OpUnconditionalBranch(cont_);
1455    }
1456
1457   private:
1458    const bool load_;
1459  };
1460
1461  if (type_known_abstract) {
1462    // Easier case, run slow path if target is non-null (slow path will load from target)
1463    LIR* branch = OpCmpImmBranch(kCondNe, TargetReg(kArg0, kRef), 0, nullptr);
1464    LIR* cont = NewLIR0(kPseudoTargetLabel);
1465    AddSlowPath(new (arena_) SlowPath(this, branch, cont, true));
1466  } else {
1467    // Harder, more common case.  We need to generate a forward branch over the load
1468    // if the target is null.  If it's non-null we perform the load and branch to the
1469    // slow path if the classes are not equal.
1470
1471    /* Null is OK - continue */
1472    LIR* branch1 = OpCmpImmBranch(kCondEq, TargetReg(kArg0, kRef), 0, nullptr);
1473    /* load object->klass_ */
1474    DCHECK_EQ(mirror::Object::ClassOffset().Int32Value(), 0);
1475    LoadRefDisp(TargetReg(kArg0, kRef), mirror::Object::ClassOffset().Int32Value(),
1476                TargetReg(kArg1, kRef), kNotVolatile);
1477
1478    LIR* branch2 = OpCmpBranch(kCondNe, TargetReg(kArg1, kRef), class_reg, nullptr);
1479    LIR* cont = NewLIR0(kPseudoTargetLabel);
1480
1481    // Add the slow path that will not perform load since this is already done.
1482    AddSlowPath(new (arena_) SlowPath(this, branch2, cont, false));
1483
1484    // Set the null check to branch to the continuation.
1485    branch1->target = cont;
1486  }
1487}
1488
1489void Mir2Lir::GenLong3Addr(OpKind first_op, OpKind second_op, RegLocation rl_dest,
1490                           RegLocation rl_src1, RegLocation rl_src2) {
1491  RegLocation rl_result;
1492  if (cu_->instruction_set == kThumb2) {
1493    /*
1494     * NOTE:  This is the one place in the code in which we might have
1495     * as many as six live temporary registers.  There are 5 in the normal
1496     * set for Arm.  Until we have spill capabilities, temporarily add
1497     * lr to the temp set.  It is safe to do this locally, but note that
1498     * lr is used explicitly elsewhere in the code generator and cannot
1499     * normally be used as a general temp register.
1500     */
1501    MarkTemp(TargetReg(kLr, kNotWide));   // Add lr to the temp pool
1502    FreeTemp(TargetReg(kLr, kNotWide));   // and make it available
1503  }
1504  rl_src1 = LoadValueWide(rl_src1, kCoreReg);
1505  rl_src2 = LoadValueWide(rl_src2, kCoreReg);
1506  rl_result = EvalLoc(rl_dest, kCoreReg, true);
1507  // The longs may overlap - use intermediate temp if so
1508  if ((rl_result.reg.GetLowReg() == rl_src1.reg.GetHighReg()) || (rl_result.reg.GetLowReg() == rl_src2.reg.GetHighReg())) {
1509    RegStorage t_reg = AllocTemp();
1510    OpRegRegReg(first_op, t_reg, rl_src1.reg.GetLow(), rl_src2.reg.GetLow());
1511    OpRegRegReg(second_op, rl_result.reg.GetHigh(), rl_src1.reg.GetHigh(), rl_src2.reg.GetHigh());
1512    OpRegCopy(rl_result.reg.GetLow(), t_reg);
1513    FreeTemp(t_reg);
1514  } else {
1515    OpRegRegReg(first_op, rl_result.reg.GetLow(), rl_src1.reg.GetLow(), rl_src2.reg.GetLow());
1516    OpRegRegReg(second_op, rl_result.reg.GetHigh(), rl_src1.reg.GetHigh(), rl_src2.reg.GetHigh());
1517  }
1518  /*
1519   * NOTE: If rl_dest refers to a frame variable in a large frame, the
1520   * following StoreValueWide might need to allocate a temp register.
1521   * To further work around the lack of a spill capability, explicitly
1522   * free any temps from rl_src1 & rl_src2 that aren't still live in rl_result.
1523   * Remove when spill is functional.
1524   */
1525  FreeRegLocTemps(rl_result, rl_src1);
1526  FreeRegLocTemps(rl_result, rl_src2);
1527  StoreValueWide(rl_dest, rl_result);
1528  if (cu_->instruction_set == kThumb2) {
1529    Clobber(TargetReg(kLr, kNotWide));
1530    UnmarkTemp(TargetReg(kLr, kNotWide));  // Remove lr from the temp pool
1531  }
1532}
1533
1534void Mir2Lir::GenShiftOpLong(Instruction::Code opcode, RegLocation rl_dest,
1535                             RegLocation rl_src1, RegLocation rl_shift) {
1536  QuickEntrypointEnum target;
1537  switch (opcode) {
1538    case Instruction::SHL_LONG:
1539    case Instruction::SHL_LONG_2ADDR:
1540      target = kQuickShlLong;
1541      break;
1542    case Instruction::SHR_LONG:
1543    case Instruction::SHR_LONG_2ADDR:
1544      target = kQuickShrLong;
1545      break;
1546    case Instruction::USHR_LONG:
1547    case Instruction::USHR_LONG_2ADDR:
1548      target = kQuickUshrLong;
1549      break;
1550    default:
1551      LOG(FATAL) << "Unexpected case";
1552      target = kQuickShlLong;
1553  }
1554  FlushAllRegs();   /* Send everything to home location */
1555  CallRuntimeHelperRegLocationRegLocation(target, rl_src1, rl_shift, false);
1556  RegLocation rl_result = GetReturnWide(kCoreReg);
1557  StoreValueWide(rl_dest, rl_result);
1558}
1559
1560
1561void Mir2Lir::GenArithOpInt(Instruction::Code opcode, RegLocation rl_dest,
1562                            RegLocation rl_src1, RegLocation rl_src2, int flags) {
1563  DCHECK(cu_->instruction_set != kX86 && cu_->instruction_set != kX86_64);
1564  OpKind op = kOpBkpt;
1565  bool is_div_rem = false;
1566  bool check_zero = false;
1567  bool unary = false;
1568  RegLocation rl_result;
1569  bool shift_op = false;
1570  switch (opcode) {
1571    case Instruction::NEG_INT:
1572      op = kOpNeg;
1573      unary = true;
1574      break;
1575    case Instruction::NOT_INT:
1576      op = kOpMvn;
1577      unary = true;
1578      break;
1579    case Instruction::ADD_INT:
1580    case Instruction::ADD_INT_2ADDR:
1581      op = kOpAdd;
1582      break;
1583    case Instruction::SUB_INT:
1584    case Instruction::SUB_INT_2ADDR:
1585      op = kOpSub;
1586      break;
1587    case Instruction::MUL_INT:
1588    case Instruction::MUL_INT_2ADDR:
1589      op = kOpMul;
1590      break;
1591    case Instruction::DIV_INT:
1592    case Instruction::DIV_INT_2ADDR:
1593      check_zero = true;
1594      op = kOpDiv;
1595      is_div_rem = true;
1596      break;
1597    /* NOTE: returns in kArg1 */
1598    case Instruction::REM_INT:
1599    case Instruction::REM_INT_2ADDR:
1600      check_zero = true;
1601      op = kOpRem;
1602      is_div_rem = true;
1603      break;
1604    case Instruction::AND_INT:
1605    case Instruction::AND_INT_2ADDR:
1606      op = kOpAnd;
1607      break;
1608    case Instruction::OR_INT:
1609    case Instruction::OR_INT_2ADDR:
1610      op = kOpOr;
1611      break;
1612    case Instruction::XOR_INT:
1613    case Instruction::XOR_INT_2ADDR:
1614      op = kOpXor;
1615      break;
1616    case Instruction::SHL_INT:
1617    case Instruction::SHL_INT_2ADDR:
1618      shift_op = true;
1619      op = kOpLsl;
1620      break;
1621    case Instruction::SHR_INT:
1622    case Instruction::SHR_INT_2ADDR:
1623      shift_op = true;
1624      op = kOpAsr;
1625      break;
1626    case Instruction::USHR_INT:
1627    case Instruction::USHR_INT_2ADDR:
1628      shift_op = true;
1629      op = kOpLsr;
1630      break;
1631    default:
1632      LOG(FATAL) << "Invalid word arith op: " << opcode;
1633  }
1634  if (!is_div_rem) {
1635    if (unary) {
1636      rl_src1 = LoadValue(rl_src1, kCoreReg);
1637      rl_result = EvalLoc(rl_dest, kCoreReg, true);
1638      OpRegReg(op, rl_result.reg, rl_src1.reg);
1639    } else {
1640      if ((shift_op) && (cu_->instruction_set != kArm64)) {
1641        rl_src2 = LoadValue(rl_src2, kCoreReg);
1642        RegStorage t_reg = AllocTemp();
1643        OpRegRegImm(kOpAnd, t_reg, rl_src2.reg, 31);
1644        rl_src1 = LoadValue(rl_src1, kCoreReg);
1645        rl_result = EvalLoc(rl_dest, kCoreReg, true);
1646        OpRegRegReg(op, rl_result.reg, rl_src1.reg, t_reg);
1647        FreeTemp(t_reg);
1648      } else {
1649        rl_src1 = LoadValue(rl_src1, kCoreReg);
1650        rl_src2 = LoadValue(rl_src2, kCoreReg);
1651        rl_result = EvalLoc(rl_dest, kCoreReg, true);
1652        OpRegRegReg(op, rl_result.reg, rl_src1.reg, rl_src2.reg);
1653      }
1654    }
1655    StoreValue(rl_dest, rl_result);
1656  } else {
1657    bool done = false;      // Set to true if we happen to find a way to use a real instruction.
1658    if (cu_->instruction_set == kMips || cu_->instruction_set == kArm64) {
1659      rl_src1 = LoadValue(rl_src1, kCoreReg);
1660      rl_src2 = LoadValue(rl_src2, kCoreReg);
1661      if (check_zero && (flags & MIR_IGNORE_DIV_ZERO_CHECK) == 0) {
1662        GenDivZeroCheck(rl_src2.reg);
1663      }
1664      rl_result = GenDivRem(rl_dest, rl_src1.reg, rl_src2.reg, op == kOpDiv);
1665      done = true;
1666    } else if (cu_->instruction_set == kThumb2) {
1667      if (cu_->GetInstructionSetFeatures()->AsArmInstructionSetFeatures()->
1668              HasDivideInstruction()) {
1669        // Use ARM SDIV instruction for division.  For remainder we also need to
1670        // calculate using a MUL and subtract.
1671        rl_src1 = LoadValue(rl_src1, kCoreReg);
1672        rl_src2 = LoadValue(rl_src2, kCoreReg);
1673        if (check_zero && (flags & MIR_IGNORE_DIV_ZERO_CHECK) == 0) {
1674          GenDivZeroCheck(rl_src2.reg);
1675        }
1676        rl_result = GenDivRem(rl_dest, rl_src1.reg, rl_src2.reg, op == kOpDiv);
1677        done = true;
1678      }
1679    }
1680
1681    // If we haven't already generated the code use the callout function.
1682    if (!done) {
1683      FlushAllRegs();   /* Send everything to home location */
1684      LoadValueDirectFixed(rl_src2, TargetReg(kArg1, kNotWide));
1685      RegStorage r_tgt = CallHelperSetup(kQuickIdivmod);
1686      LoadValueDirectFixed(rl_src1, TargetReg(kArg0, kNotWide));
1687      if (check_zero && (flags & MIR_IGNORE_DIV_ZERO_CHECK) == 0) {
1688        GenDivZeroCheck(TargetReg(kArg1, kNotWide));
1689      }
1690      // NOTE: callout here is not a safepoint.
1691      CallHelper(r_tgt, kQuickIdivmod, false /* not a safepoint */);
1692      if (op == kOpDiv)
1693        rl_result = GetReturn(kCoreReg);
1694      else
1695        rl_result = GetReturnAlt();
1696    }
1697    StoreValue(rl_dest, rl_result);
1698  }
1699}
1700
1701/*
1702 * The following are the first-level codegen routines that analyze the format
1703 * of each bytecode then either dispatch special purpose codegen routines
1704 * or produce corresponding Thumb instructions directly.
1705 */
1706
1707// Returns true if no more than two bits are set in 'x'.
1708static bool IsPopCountLE2(unsigned int x) {
1709  x &= x - 1;
1710  return (x & (x - 1)) == 0;
1711}
1712
1713// Returns true if it added instructions to 'cu' to divide 'rl_src' by 'lit'
1714// and store the result in 'rl_dest'.
1715bool Mir2Lir::HandleEasyDivRem(Instruction::Code dalvik_opcode, bool is_div,
1716                               RegLocation rl_src, RegLocation rl_dest, int lit) {
1717  if ((lit < 2) || ((cu_->instruction_set != kThumb2) && !IsPowerOfTwo(lit))) {
1718    return false;
1719  }
1720  // No divide instruction for Arm, so check for more special cases
1721  if ((cu_->instruction_set == kThumb2) && !IsPowerOfTwo(lit)) {
1722    return SmallLiteralDivRem(dalvik_opcode, is_div, rl_src, rl_dest, lit);
1723  }
1724  int k = LowestSetBit(lit);
1725  if (k >= 30) {
1726    // Avoid special cases.
1727    return false;
1728  }
1729  rl_src = LoadValue(rl_src, kCoreReg);
1730  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1731  if (is_div) {
1732    RegStorage t_reg = AllocTemp();
1733    if (lit == 2) {
1734      // Division by 2 is by far the most common division by constant.
1735      OpRegRegImm(kOpLsr, t_reg, rl_src.reg, 32 - k);
1736      OpRegRegReg(kOpAdd, t_reg, t_reg, rl_src.reg);
1737      OpRegRegImm(kOpAsr, rl_result.reg, t_reg, k);
1738    } else {
1739      OpRegRegImm(kOpAsr, t_reg, rl_src.reg, 31);
1740      OpRegRegImm(kOpLsr, t_reg, t_reg, 32 - k);
1741      OpRegRegReg(kOpAdd, t_reg, t_reg, rl_src.reg);
1742      OpRegRegImm(kOpAsr, rl_result.reg, t_reg, k);
1743    }
1744  } else {
1745    RegStorage t_reg1 = AllocTemp();
1746    RegStorage t_reg2 = AllocTemp();
1747    if (lit == 2) {
1748      OpRegRegImm(kOpLsr, t_reg1, rl_src.reg, 32 - k);
1749      OpRegRegReg(kOpAdd, t_reg2, t_reg1, rl_src.reg);
1750      OpRegRegImm(kOpAnd, t_reg2, t_reg2, lit -1);
1751      OpRegRegReg(kOpSub, rl_result.reg, t_reg2, t_reg1);
1752    } else {
1753      OpRegRegImm(kOpAsr, t_reg1, rl_src.reg, 31);
1754      OpRegRegImm(kOpLsr, t_reg1, t_reg1, 32 - k);
1755      OpRegRegReg(kOpAdd, t_reg2, t_reg1, rl_src.reg);
1756      OpRegRegImm(kOpAnd, t_reg2, t_reg2, lit - 1);
1757      OpRegRegReg(kOpSub, rl_result.reg, t_reg2, t_reg1);
1758    }
1759  }
1760  StoreValue(rl_dest, rl_result);
1761  return true;
1762}
1763
1764// Returns true if it added instructions to 'cu' to multiply 'rl_src' by 'lit'
1765// and store the result in 'rl_dest'.
1766bool Mir2Lir::HandleEasyMultiply(RegLocation rl_src, RegLocation rl_dest, int lit) {
1767  if (lit < 0) {
1768    return false;
1769  }
1770  if (lit == 0) {
1771    RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1772    LoadConstant(rl_result.reg, 0);
1773    StoreValue(rl_dest, rl_result);
1774    return true;
1775  }
1776  if (lit == 1) {
1777    rl_src = LoadValue(rl_src, kCoreReg);
1778    RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1779    OpRegCopy(rl_result.reg, rl_src.reg);
1780    StoreValue(rl_dest, rl_result);
1781    return true;
1782  }
1783  // There is RegRegRegShift on Arm, so check for more special cases
1784  if (cu_->instruction_set == kThumb2) {
1785    return EasyMultiply(rl_src, rl_dest, lit);
1786  }
1787  // Can we simplify this multiplication?
1788  bool power_of_two = false;
1789  bool pop_count_le2 = false;
1790  bool power_of_two_minus_one = false;
1791  if (IsPowerOfTwo(lit)) {
1792    power_of_two = true;
1793  } else if (IsPopCountLE2(lit)) {
1794    pop_count_le2 = true;
1795  } else if (IsPowerOfTwo(lit + 1)) {
1796    power_of_two_minus_one = true;
1797  } else {
1798    return false;
1799  }
1800  rl_src = LoadValue(rl_src, kCoreReg);
1801  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
1802  if (power_of_two) {
1803    // Shift.
1804    OpRegRegImm(kOpLsl, rl_result.reg, rl_src.reg, LowestSetBit(lit));
1805  } else if (pop_count_le2) {
1806    // Shift and add and shift.
1807    int first_bit = LowestSetBit(lit);
1808    int second_bit = LowestSetBit(lit ^ (1 << first_bit));
1809    GenMultiplyByTwoBitMultiplier(rl_src, rl_result, lit, first_bit, second_bit);
1810  } else {
1811    // Reverse subtract: (src << (shift + 1)) - src.
1812    DCHECK(power_of_two_minus_one);
1813    // TUNING: rsb dst, src, src lsl#LowestSetBit(lit + 1)
1814    RegStorage t_reg = AllocTemp();
1815    OpRegRegImm(kOpLsl, t_reg, rl_src.reg, LowestSetBit(lit + 1));
1816    OpRegRegReg(kOpSub, rl_result.reg, t_reg, rl_src.reg);
1817  }
1818  StoreValue(rl_dest, rl_result);
1819  return true;
1820}
1821
1822// Returns true if it generates instructions.
1823bool Mir2Lir::HandleEasyFloatingPointDiv(RegLocation rl_dest, RegLocation rl_src1,
1824                                         RegLocation rl_src2) {
1825  if (!rl_src2.is_const ||
1826      ((cu_->instruction_set != kThumb2) && (cu_->instruction_set != kArm64))) {
1827    return false;
1828  }
1829
1830  if (!rl_src2.wide) {
1831    int32_t divisor = mir_graph_->ConstantValue(rl_src2);
1832    if (CanDivideByReciprocalMultiplyFloat(divisor)) {
1833      // Generate multiply by reciprocal instead of div.
1834      float recip = 1.0f/bit_cast<int32_t, float>(divisor);
1835      GenMultiplyByConstantFloat(rl_dest, rl_src1, bit_cast<float, int32_t>(recip));
1836      return true;
1837    }
1838  } else {
1839    int64_t divisor = mir_graph_->ConstantValueWide(rl_src2);
1840    if (CanDivideByReciprocalMultiplyDouble(divisor)) {
1841      // Generate multiply by reciprocal instead of div.
1842      double recip = 1.0/bit_cast<double, int64_t>(divisor);
1843      GenMultiplyByConstantDouble(rl_dest, rl_src1, bit_cast<double, int64_t>(recip));
1844      return true;
1845    }
1846  }
1847  return false;
1848}
1849
1850void Mir2Lir::GenArithOpIntLit(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src,
1851                               int lit) {
1852  RegLocation rl_result;
1853  OpKind op = static_cast<OpKind>(0);    /* Make gcc happy */
1854  int shift_op = false;
1855  bool is_div = false;
1856
1857  switch (opcode) {
1858    case Instruction::RSUB_INT_LIT8:
1859    case Instruction::RSUB_INT: {
1860      rl_src = LoadValue(rl_src, kCoreReg);
1861      rl_result = EvalLoc(rl_dest, kCoreReg, true);
1862      if (cu_->instruction_set == kThumb2) {
1863        OpRegRegImm(kOpRsub, rl_result.reg, rl_src.reg, lit);
1864      } else {
1865        OpRegReg(kOpNeg, rl_result.reg, rl_src.reg);
1866        OpRegImm(kOpAdd, rl_result.reg, lit);
1867      }
1868      StoreValue(rl_dest, rl_result);
1869      return;
1870    }
1871
1872    case Instruction::SUB_INT:
1873    case Instruction::SUB_INT_2ADDR:
1874      lit = -lit;
1875      FALLTHROUGH_INTENDED;
1876    case Instruction::ADD_INT:
1877    case Instruction::ADD_INT_2ADDR:
1878    case Instruction::ADD_INT_LIT8:
1879    case Instruction::ADD_INT_LIT16:
1880      op = kOpAdd;
1881      break;
1882    case Instruction::MUL_INT:
1883    case Instruction::MUL_INT_2ADDR:
1884    case Instruction::MUL_INT_LIT8:
1885    case Instruction::MUL_INT_LIT16: {
1886      if (HandleEasyMultiply(rl_src, rl_dest, lit)) {
1887        return;
1888      }
1889      op = kOpMul;
1890      break;
1891    }
1892    case Instruction::AND_INT:
1893    case Instruction::AND_INT_2ADDR:
1894    case Instruction::AND_INT_LIT8:
1895    case Instruction::AND_INT_LIT16:
1896      op = kOpAnd;
1897      break;
1898    case Instruction::OR_INT:
1899    case Instruction::OR_INT_2ADDR:
1900    case Instruction::OR_INT_LIT8:
1901    case Instruction::OR_INT_LIT16:
1902      op = kOpOr;
1903      break;
1904    case Instruction::XOR_INT:
1905    case Instruction::XOR_INT_2ADDR:
1906    case Instruction::XOR_INT_LIT8:
1907    case Instruction::XOR_INT_LIT16:
1908      op = kOpXor;
1909      break;
1910    case Instruction::SHL_INT_LIT8:
1911    case Instruction::SHL_INT:
1912    case Instruction::SHL_INT_2ADDR:
1913      lit &= 31;
1914      shift_op = true;
1915      op = kOpLsl;
1916      break;
1917    case Instruction::SHR_INT_LIT8:
1918    case Instruction::SHR_INT:
1919    case Instruction::SHR_INT_2ADDR:
1920      lit &= 31;
1921      shift_op = true;
1922      op = kOpAsr;
1923      break;
1924    case Instruction::USHR_INT_LIT8:
1925    case Instruction::USHR_INT:
1926    case Instruction::USHR_INT_2ADDR:
1927      lit &= 31;
1928      shift_op = true;
1929      op = kOpLsr;
1930      break;
1931
1932    case Instruction::DIV_INT:
1933    case Instruction::DIV_INT_2ADDR:
1934    case Instruction::DIV_INT_LIT8:
1935    case Instruction::DIV_INT_LIT16:
1936    case Instruction::REM_INT:
1937    case Instruction::REM_INT_2ADDR:
1938    case Instruction::REM_INT_LIT8:
1939    case Instruction::REM_INT_LIT16: {
1940      if (lit == 0) {
1941        GenDivZeroException();
1942        return;
1943      }
1944      if ((opcode == Instruction::DIV_INT) ||
1945          (opcode == Instruction::DIV_INT_2ADDR) ||
1946          (opcode == Instruction::DIV_INT_LIT8) ||
1947          (opcode == Instruction::DIV_INT_LIT16)) {
1948        is_div = true;
1949      } else {
1950        is_div = false;
1951      }
1952      if (HandleEasyDivRem(opcode, is_div, rl_src, rl_dest, lit)) {
1953        return;
1954      }
1955
1956      bool done = false;
1957      if (cu_->instruction_set == kMips || cu_->instruction_set == kArm64) {
1958        rl_src = LoadValue(rl_src, kCoreReg);
1959        rl_result = GenDivRemLit(rl_dest, rl_src.reg, lit, is_div);
1960        done = true;
1961      } else if (cu_->instruction_set == kX86 || cu_->instruction_set == kX86_64) {
1962        rl_result = GenDivRemLit(rl_dest, rl_src, lit, is_div);
1963        done = true;
1964      } else if (cu_->instruction_set == kThumb2) {
1965        if (cu_->GetInstructionSetFeatures()->AsArmInstructionSetFeatures()->
1966                HasDivideInstruction()) {
1967          // Use ARM SDIV instruction for division.  For remainder we also need to
1968          // calculate using a MUL and subtract.
1969          rl_src = LoadValue(rl_src, kCoreReg);
1970          rl_result = GenDivRemLit(rl_dest, rl_src.reg, lit, is_div);
1971          done = true;
1972        }
1973      }
1974
1975      if (!done) {
1976        FlushAllRegs();   /* Everything to home location. */
1977        LoadValueDirectFixed(rl_src, TargetReg(kArg0, kNotWide));
1978        Clobber(TargetReg(kArg0, kNotWide));
1979        CallRuntimeHelperRegImm(kQuickIdivmod, TargetReg(kArg0, kNotWide), lit, false);
1980        if (is_div)
1981          rl_result = GetReturn(kCoreReg);
1982        else
1983          rl_result = GetReturnAlt();
1984      }
1985      StoreValue(rl_dest, rl_result);
1986      return;
1987    }
1988    default:
1989      LOG(FATAL) << "Unexpected opcode " << opcode;
1990  }
1991  rl_src = LoadValue(rl_src, kCoreReg);
1992  rl_result = EvalLoc(rl_dest, kCoreReg, true);
1993  // Avoid shifts by literal 0 - no support in Thumb.  Change to copy.
1994  if (shift_op && (lit == 0)) {
1995    OpRegCopy(rl_result.reg, rl_src.reg);
1996  } else {
1997    OpRegRegImm(op, rl_result.reg, rl_src.reg, lit);
1998  }
1999  StoreValue(rl_dest, rl_result);
2000}
2001
2002void Mir2Lir::GenArithOpLong(Instruction::Code opcode, RegLocation rl_dest,
2003                             RegLocation rl_src1, RegLocation rl_src2, int flags) {
2004  RegLocation rl_result;
2005  OpKind first_op = kOpBkpt;
2006  OpKind second_op = kOpBkpt;
2007  bool call_out = false;
2008  bool check_zero = false;
2009  int ret_reg = TargetReg(kRet0, kNotWide).GetReg();
2010  QuickEntrypointEnum target;
2011
2012  switch (opcode) {
2013    case Instruction::NOT_LONG:
2014      rl_src2 = LoadValueWide(rl_src2, kCoreReg);
2015      rl_result = EvalLoc(rl_dest, kCoreReg, true);
2016      // Check for destructive overlap
2017      if (rl_result.reg.GetLowReg() == rl_src2.reg.GetHighReg()) {
2018        RegStorage t_reg = AllocTemp();
2019        OpRegCopy(t_reg, rl_src2.reg.GetHigh());
2020        OpRegReg(kOpMvn, rl_result.reg.GetLow(), rl_src2.reg.GetLow());
2021        OpRegReg(kOpMvn, rl_result.reg.GetHigh(), t_reg);
2022        FreeTemp(t_reg);
2023      } else {
2024        OpRegReg(kOpMvn, rl_result.reg.GetLow(), rl_src2.reg.GetLow());
2025        OpRegReg(kOpMvn, rl_result.reg.GetHigh(), rl_src2.reg.GetHigh());
2026      }
2027      StoreValueWide(rl_dest, rl_result);
2028      return;
2029    case Instruction::ADD_LONG:
2030    case Instruction::ADD_LONG_2ADDR:
2031      first_op = kOpAdd;
2032      second_op = kOpAdc;
2033      break;
2034    case Instruction::SUB_LONG:
2035    case Instruction::SUB_LONG_2ADDR:
2036      first_op = kOpSub;
2037      second_op = kOpSbc;
2038      break;
2039    case Instruction::MUL_LONG:
2040    case Instruction::MUL_LONG_2ADDR:
2041      call_out = true;
2042      ret_reg = TargetReg(kRet0, kNotWide).GetReg();
2043      target = kQuickLmul;
2044      break;
2045    case Instruction::DIV_LONG:
2046    case Instruction::DIV_LONG_2ADDR:
2047      call_out = true;
2048      check_zero = true;
2049      ret_reg = TargetReg(kRet0, kNotWide).GetReg();
2050      target = kQuickLdiv;
2051      break;
2052    case Instruction::REM_LONG:
2053    case Instruction::REM_LONG_2ADDR:
2054      call_out = true;
2055      check_zero = true;
2056      target = kQuickLmod;
2057      /* NOTE - for Arm, result is in kArg2/kArg3 instead of kRet0/kRet1 */
2058      ret_reg = (cu_->instruction_set == kThumb2) ? TargetReg(kArg2, kNotWide).GetReg() :
2059          TargetReg(kRet0, kNotWide).GetReg();
2060      break;
2061    case Instruction::AND_LONG_2ADDR:
2062    case Instruction::AND_LONG:
2063      first_op = kOpAnd;
2064      second_op = kOpAnd;
2065      break;
2066    case Instruction::OR_LONG:
2067    case Instruction::OR_LONG_2ADDR:
2068      first_op = kOpOr;
2069      second_op = kOpOr;
2070      break;
2071    case Instruction::XOR_LONG:
2072    case Instruction::XOR_LONG_2ADDR:
2073      first_op = kOpXor;
2074      second_op = kOpXor;
2075      break;
2076    default:
2077      LOG(FATAL) << "Invalid long arith op";
2078  }
2079  if (!call_out) {
2080    GenLong3Addr(first_op, second_op, rl_dest, rl_src1, rl_src2);
2081  } else {
2082    FlushAllRegs();   /* Send everything to home location */
2083    if (check_zero) {
2084      RegStorage r_tmp1 = TargetReg(kArg0, kWide);
2085      RegStorage r_tmp2 = TargetReg(kArg2, kWide);
2086      LoadValueDirectWideFixed(rl_src2, r_tmp2);
2087      RegStorage r_tgt = CallHelperSetup(target);
2088      if ((flags & MIR_IGNORE_DIV_ZERO_CHECK) == 0) {
2089        GenDivZeroCheckWide(r_tmp2);
2090      }
2091      LoadValueDirectWideFixed(rl_src1, r_tmp1);
2092      // NOTE: callout here is not a safepoint
2093      CallHelper(r_tgt, target, false /* not safepoint */);
2094    } else {
2095      CallRuntimeHelperRegLocationRegLocation(target, rl_src1, rl_src2, false);
2096    }
2097    // Adjust return regs in to handle case of rem returning kArg2/kArg3
2098    if (ret_reg == TargetReg(kRet0, kNotWide).GetReg())
2099      rl_result = GetReturnWide(kCoreReg);
2100    else
2101      rl_result = GetReturnWideAlt();
2102    StoreValueWide(rl_dest, rl_result);
2103  }
2104}
2105
2106void Mir2Lir::GenConst(RegLocation rl_dest, int value) {
2107  RegLocation rl_result = EvalLoc(rl_dest, kAnyReg, true);
2108  LoadConstantNoClobber(rl_result.reg, value);
2109  StoreValue(rl_dest, rl_result);
2110  if (value == 0) {
2111    Workaround7250540(rl_dest, rl_result.reg);
2112  }
2113}
2114
2115void Mir2Lir::GenConversionCall(QuickEntrypointEnum trampoline, RegLocation rl_dest,
2116                                RegLocation rl_src) {
2117  /*
2118   * Don't optimize the register usage since it calls out to support
2119   * functions
2120   */
2121
2122  FlushAllRegs();   /* Send everything to home location */
2123  CallRuntimeHelperRegLocation(trampoline, rl_src, false);
2124  if (rl_dest.wide) {
2125    RegLocation rl_result;
2126    rl_result = GetReturnWide(LocToRegClass(rl_dest));
2127    StoreValueWide(rl_dest, rl_result);
2128  } else {
2129    RegLocation rl_result;
2130    rl_result = GetReturn(LocToRegClass(rl_dest));
2131    StoreValue(rl_dest, rl_result);
2132  }
2133}
2134
2135class SuspendCheckSlowPath : public Mir2Lir::LIRSlowPath {
2136 public:
2137  SuspendCheckSlowPath(Mir2Lir* m2l, LIR* branch, LIR* cont)
2138      : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), branch, cont) {
2139  }
2140
2141  void Compile() OVERRIDE {
2142    m2l_->ResetRegPool();
2143    m2l_->ResetDefTracking();
2144    GenerateTargetLabel(kPseudoSuspendTarget);
2145    m2l_->CallRuntimeHelper(kQuickTestSuspend, true);
2146    if (cont_ != nullptr) {
2147      m2l_->OpUnconditionalBranch(cont_);
2148    }
2149  }
2150};
2151
2152/* Check if we need to check for pending suspend request */
2153void Mir2Lir::GenSuspendTest(int opt_flags) {
2154  if (!cu_->compiler_driver->GetCompilerOptions().GetImplicitSuspendChecks()) {
2155    if (NO_SUSPEND || (opt_flags & MIR_IGNORE_SUSPEND_CHECK)) {
2156      return;
2157    }
2158    FlushAllRegs();
2159    LIR* branch = OpTestSuspend(NULL);
2160    LIR* cont = NewLIR0(kPseudoTargetLabel);
2161    AddSlowPath(new (arena_) SuspendCheckSlowPath(this, branch, cont));
2162  } else {
2163    if (NO_SUSPEND || (opt_flags & MIR_IGNORE_SUSPEND_CHECK)) {
2164      return;
2165    }
2166    FlushAllRegs();     // TODO: needed?
2167    LIR* inst = CheckSuspendUsingLoad();
2168    MarkSafepointPC(inst);
2169  }
2170}
2171
2172/* Check if we need to check for pending suspend request */
2173void Mir2Lir::GenSuspendTestAndBranch(int opt_flags, LIR* target) {
2174  if (!cu_->compiler_driver->GetCompilerOptions().GetImplicitSuspendChecks()) {
2175    if (NO_SUSPEND || (opt_flags & MIR_IGNORE_SUSPEND_CHECK)) {
2176      OpUnconditionalBranch(target);
2177      return;
2178    }
2179    OpTestSuspend(target);
2180    FlushAllRegs();
2181    LIR* branch = OpUnconditionalBranch(nullptr);
2182    AddSlowPath(new (arena_) SuspendCheckSlowPath(this, branch, target));
2183  } else {
2184    // For the implicit suspend check, just perform the trigger
2185    // load and branch to the target.
2186    if (NO_SUSPEND || (opt_flags & MIR_IGNORE_SUSPEND_CHECK)) {
2187      OpUnconditionalBranch(target);
2188      return;
2189    }
2190    FlushAllRegs();
2191    LIR* inst = CheckSuspendUsingLoad();
2192    MarkSafepointPC(inst);
2193    OpUnconditionalBranch(target);
2194  }
2195}
2196
2197/* Call out to helper assembly routine that will null check obj and then lock it. */
2198void Mir2Lir::GenMonitorEnter(int opt_flags, RegLocation rl_src) {
2199  UNUSED(opt_flags);  // TODO: avoid null check with specialized non-null helper.
2200  FlushAllRegs();
2201  CallRuntimeHelperRegLocation(kQuickLockObject, rl_src, true);
2202}
2203
2204/* Call out to helper assembly routine that will null check obj and then unlock it. */
2205void Mir2Lir::GenMonitorExit(int opt_flags, RegLocation rl_src) {
2206  UNUSED(opt_flags);  // TODO: avoid null check with specialized non-null helper.
2207  FlushAllRegs();
2208  CallRuntimeHelperRegLocation(kQuickUnlockObject, rl_src, true);
2209}
2210
2211/* Generic code for generating a wide constant into a VR. */
2212void Mir2Lir::GenConstWide(RegLocation rl_dest, int64_t value) {
2213  RegLocation rl_result = EvalLoc(rl_dest, kAnyReg, true);
2214  LoadConstantWide(rl_result.reg, value);
2215  StoreValueWide(rl_dest, rl_result);
2216}
2217
2218void Mir2Lir::GenSmallPackedSwitch(MIR* mir, DexOffset table_offset, RegLocation rl_src) {
2219  const uint16_t* table = mir_graph_->GetTable(mir, table_offset);
2220  const uint16_t entries = table[1];
2221  // Chained cmp-and-branch.
2222  const int32_t* as_int32 = reinterpret_cast<const int32_t*>(&table[2]);
2223  int32_t starting_key = as_int32[0];
2224  const int32_t* targets = &as_int32[1];
2225  rl_src = LoadValue(rl_src, kCoreReg);
2226  int i = 0;
2227  for (; i < entries; i++) {
2228    if (!InexpensiveConstantInt(starting_key + i, Instruction::Code::IF_EQ)) {
2229      // Switch to using a temp and add.
2230      break;
2231    }
2232    BasicBlock* case_block =
2233        mir_graph_->FindBlock(current_dalvik_offset_ + targets[i]);
2234    OpCmpImmBranch(kCondEq, rl_src.reg, starting_key + i, &block_label_list_[case_block->id]);
2235  }
2236  if (i < entries) {
2237    // The rest do not seem to be inexpensive. Try to allocate a temp and use add.
2238    RegStorage key_temp = AllocTypedTemp(false, kCoreReg, false);
2239    if (key_temp.Valid()) {
2240      LoadConstantNoClobber(key_temp, starting_key + i);
2241      for (; i < entries - 1; i++) {
2242        BasicBlock* case_block =
2243            mir_graph_->FindBlock(current_dalvik_offset_ + targets[i]);
2244        OpCmpBranch(kCondEq, rl_src.reg, key_temp, &block_label_list_[case_block->id]);
2245        OpRegImm(kOpAdd, key_temp, 1);  // Increment key.
2246      }
2247      BasicBlock* case_block =
2248          mir_graph_->FindBlock(current_dalvik_offset_ + targets[i]);
2249      OpCmpBranch(kCondEq, rl_src.reg, key_temp, &block_label_list_[case_block->id]);
2250    } else {
2251      // No free temp, just finish the old loop.
2252      for (; i < entries; i++) {
2253        BasicBlock* case_block =
2254            mir_graph_->FindBlock(current_dalvik_offset_ + targets[i]);
2255        OpCmpImmBranch(kCondEq, rl_src.reg, starting_key + i, &block_label_list_[case_block->id]);
2256      }
2257    }
2258  }
2259}
2260
2261void Mir2Lir::GenPackedSwitch(MIR* mir, DexOffset table_offset, RegLocation rl_src) {
2262  const uint16_t* table = mir_graph_->GetTable(mir, table_offset);
2263  if (cu_->verbose) {
2264    DumpSparseSwitchTable(table);
2265  }
2266
2267  const uint16_t entries = table[1];
2268  if (entries <= kSmallSwitchThreshold) {
2269    GenSmallPackedSwitch(mir, table_offset, rl_src);
2270  } else {
2271    // Use the backend-specific implementation.
2272    GenLargePackedSwitch(mir, table_offset, rl_src);
2273  }
2274}
2275
2276void Mir2Lir::GenSmallSparseSwitch(MIR* mir, DexOffset table_offset, RegLocation rl_src) {
2277  const uint16_t* table = mir_graph_->GetTable(mir, table_offset);
2278  const uint16_t entries = table[1];
2279  // Chained cmp-and-branch.
2280  const int32_t* keys = reinterpret_cast<const int32_t*>(&table[2]);
2281  const int32_t* targets = &keys[entries];
2282  rl_src = LoadValue(rl_src, kCoreReg);
2283  for (int i = 0; i < entries; i++) {
2284    int key = keys[i];
2285    BasicBlock* case_block =
2286        mir_graph_->FindBlock(current_dalvik_offset_ + targets[i]);
2287    OpCmpImmBranch(kCondEq, rl_src.reg, key, &block_label_list_[case_block->id]);
2288  }
2289}
2290
2291void Mir2Lir::GenSparseSwitch(MIR* mir, DexOffset table_offset, RegLocation rl_src) {
2292  const uint16_t* table = mir_graph_->GetTable(mir, table_offset);
2293  if (cu_->verbose) {
2294    DumpSparseSwitchTable(table);
2295  }
2296
2297  const uint16_t entries = table[1];
2298  if (entries <= kSmallSwitchThreshold) {
2299    GenSmallSparseSwitch(mir, table_offset, rl_src);
2300  } else {
2301    // Use the backend-specific implementation.
2302    GenLargeSparseSwitch(mir, table_offset, rl_src);
2303  }
2304}
2305
2306bool Mir2Lir::SizeMatchesTypeForEntrypoint(OpSize size, Primitive::Type type) {
2307  switch (size) {
2308    case kReference:
2309      return type == Primitive::kPrimNot;
2310    case k64:
2311    case kDouble:
2312      return type == Primitive::kPrimLong || type == Primitive::kPrimDouble;
2313    case k32:
2314    case kSingle:
2315      return type == Primitive::kPrimInt || type == Primitive::kPrimFloat;
2316    case kSignedHalf:
2317      return type == Primitive::kPrimShort;
2318    case kUnsignedHalf:
2319      return type == Primitive::kPrimChar;
2320    case kSignedByte:
2321      return type == Primitive::kPrimByte;
2322    case kUnsignedByte:
2323      return type == Primitive::kPrimBoolean;
2324    case kWord:  // Intentional fallthrough.
2325    default:
2326      return false;  // There are no sane types with this op size.
2327  }
2328}
2329
2330}  // namespace art
2331