quick_trampoline_entrypoints.cc revision 9a6a99aaac2e4c973e0bc71075f196b8b084100f
1/*
2 * Copyright (C) 2012 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 *      http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17#include "callee_save_frame.h"
18#include "common_throws.h"
19#include "dex_file-inl.h"
20#include "dex_instruction-inl.h"
21#include "entrypoints/entrypoint_utils.h"
22#include "gc/accounting/card_table-inl.h"
23#include "interpreter/interpreter.h"
24#include "invoke_arg_array_builder.h"
25#include "mirror/art_method-inl.h"
26#include "mirror/class-inl.h"
27#include "mirror/object-inl.h"
28#include "mirror/object_array-inl.h"
29#include "object_utils.h"
30#include "runtime.h"
31
32namespace art {
33
34// Visits the arguments as saved to the stack by a Runtime::kRefAndArgs callee save frame.
35class QuickArgumentVisitor {
36  // Size of each spilled GPR.
37#ifdef __LP64__
38  static constexpr size_t kBytesPerGprSpillLocation = 8;
39#else
40  static constexpr size_t kBytesPerGprSpillLocation = 4;
41#endif
42  // Number of bytes for each out register in the caller method's frame.
43  static constexpr size_t kBytesStackArgLocation = 4;
44#if defined(__arm__)
45  // The callee save frame is pointed to by SP.
46  // | argN       |  |
47  // | ...        |  |
48  // | arg4       |  |
49  // | arg3 spill |  |  Caller's frame
50  // | arg2 spill |  |
51  // | arg1 spill |  |
52  // | Method*    | ---
53  // | LR         |
54  // | ...        |    callee saves
55  // | R3         |    arg3
56  // | R2         |    arg2
57  // | R1         |    arg1
58  // | R0         |    padding
59  // | Method*    |  <- sp
60  static constexpr bool kQuickSoftFloatAbi = true;  // This is a soft float ABI.
61  static constexpr size_t kNumQuickGprArgs = 3;  // 3 arguments passed in GPRs.
62  static constexpr size_t kNumQuickFprArgs = 0;  // 0 arguments passed in FPRs.
63  static constexpr size_t kBytesPerFprSpillLocation = 4;  // FPR spill size is 4 bytes.
64  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Fpr1Offset = 0;  // Offset of first FPR arg.
65  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Gpr1Offset = 8;  // Offset of first GPR arg.
66  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_LrOffset = 44;  // Offset of return address.
67  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_FrameSize = 48;  // Frame size.
68  static size_t GprIndexToGprOffset(uint32_t gpr_index) {
69    return gpr_index * kBytesPerGprSpillLocation;
70  }
71#elif defined(__mips__)
72  // The callee save frame is pointed to by SP.
73  // | argN       |  |
74  // | ...        |  |
75  // | arg4       |  |
76  // | arg3 spill |  |  Caller's frame
77  // | arg2 spill |  |
78  // | arg1 spill |  |
79  // | Method*    | ---
80  // | RA         |
81  // | ...        |    callee saves
82  // | A3         |    arg3
83  // | A2         |    arg2
84  // | A1         |    arg1
85  // | A0/Method* |  <- sp
86  static constexpr bool kQuickSoftFloatAbi = true;  // This is a soft float ABI.
87  static constexpr size_t kNumQuickGprArgs = 3;  // 3 arguments passed in GPRs.
88  static constexpr size_t kNumQuickFprArgs = 0;  // 0 arguments passed in FPRs.
89  static constexpr size_t kBytesPerFprSpillLocation = 4;  // FPR spill size is 4 bytes.
90  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Fpr1Offset = 0;  // Offset of first FPR arg.
91  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Gpr1Offset = 4;  // Offset of first GPR arg.
92  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_LrOffset = 60;  // Offset of return address.
93  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_FrameSize = 64;  // Frame size.
94  static size_t GprIndexToGprOffset(uint32_t gpr_index) {
95    return gpr_index * kBytesPerGprSpillLocation;
96  }
97#elif defined(__i386__)
98  // The callee save frame is pointed to by SP.
99  // | argN        |  |
100  // | ...         |  |
101  // | arg4        |  |
102  // | arg3 spill  |  |  Caller's frame
103  // | arg2 spill  |  |
104  // | arg1 spill  |  |
105  // | Method*     | ---
106  // | Return      |
107  // | EBP,ESI,EDI |    callee saves
108  // | EBX         |    arg3
109  // | EDX         |    arg2
110  // | ECX         |    arg1
111  // | EAX/Method* |  <- sp
112  static constexpr bool kQuickSoftFloatAbi = true;  // This is a soft float ABI.
113  static constexpr size_t kNumQuickGprArgs = 3;  // 3 arguments passed in GPRs.
114  static constexpr size_t kNumQuickFprArgs = 0;  // 0 arguments passed in FPRs.
115  static constexpr size_t kBytesPerFprSpillLocation = 8;  // FPR spill size is 8 bytes.
116  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Fpr1Offset = 0;  // Offset of first FPR arg.
117  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Gpr1Offset = 4;  // Offset of first GPR arg.
118  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_LrOffset = 28;  // Offset of return address.
119  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_FrameSize = 32;  // Frame size.
120  static size_t GprIndexToGprOffset(uint32_t gpr_index) {
121    return gpr_index * kBytesPerGprSpillLocation;
122  }
123#elif defined(__x86_64__)
124  // The callee save frame is pointed to by SP.
125  // | argN            |  |
126  // | ...             |  |
127  // | reg. arg spills |  |  Caller's frame
128  // | Method*         | ---
129  // | Return          |
130  // | R15             |    callee save
131  // | R14             |    callee save
132  // | R13             |    callee save
133  // | R12             |    callee save
134  // | R9              |    arg5
135  // | R8              |    arg4
136  // | RSI/R6          |    arg1
137  // | RBP/R5          |    callee save
138  // | RBX/R3          |    callee save
139  // | RDX/R2          |    arg2
140  // | RCX/R1          |    arg3
141  // | XMM7            |    float arg 8
142  // | XMM6            |    float arg 7
143  // | XMM5            |    float arg 6
144  // | XMM4            |    float arg 5
145  // | XMM3            |    float arg 4
146  // | XMM2            |    float arg 3
147  // | XMM1            |    float arg 2
148  // | XMM0            |    float arg 1
149  // | Padding         |
150  // | RDI/Method*     |  <- sp
151  static constexpr bool kQuickSoftFloatAbi = false;  // This is a hard float ABI.
152  static constexpr size_t kNumQuickGprArgs = 5;  // 3 arguments passed in GPRs.
153  static constexpr size_t kNumQuickFprArgs = 8;  // 0 arguments passed in FPRs.
154  static constexpr size_t kBytesPerFprSpillLocation = 8;  // FPR spill size is 8 bytes.
155  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Fpr1Offset = 16;  // Offset of first FPR arg.
156  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_Gpr1Offset = 80;  // Offset of first GPR arg.
157  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_LrOffset = 168;  // Offset of return address.
158  static constexpr size_t kQuickCalleeSaveFrame_RefAndArgs_FrameSize = 176;  // Frame size.
159  static size_t GprIndexToGprOffset(uint32_t gpr_index) {
160    switch (gpr_index) {
161      case 0: return (4 * kBytesPerGprSpillLocation);
162      case 1: return (1 * kBytesPerGprSpillLocation);
163      case 2: return (0 * kBytesPerGprSpillLocation);
164      case 3: return (5 * kBytesPerGprSpillLocation);
165      case 4: return (6 * kBytesPerGprSpillLocation);
166      default:
167        LOG(FATAL) << "Unexpected GPR index: " << gpr_index;
168        return 0;
169    }
170  }
171#else
172#error "Unsupported architecture"
173#endif
174
175 public:
176  static mirror::ArtMethod* GetCallingMethod(mirror::ArtMethod** sp)
177      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
178    DCHECK((*sp)->IsCalleeSaveMethod());
179    byte* previous_sp = reinterpret_cast<byte*>(sp) + kQuickCalleeSaveFrame_RefAndArgs_FrameSize;
180    return *reinterpret_cast<mirror::ArtMethod**>(previous_sp);
181  }
182
183  // For the given quick ref and args quick frame, return the caller's PC.
184  static uintptr_t GetCallingPc(mirror::ArtMethod** sp)
185      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
186    DCHECK((*sp)->IsCalleeSaveMethod());
187    byte* lr = reinterpret_cast<byte*>(sp) + kQuickCalleeSaveFrame_RefAndArgs_LrOffset;
188    return *reinterpret_cast<uintptr_t*>(lr);
189  }
190
191  QuickArgumentVisitor(mirror::ArtMethod** sp, bool is_static,
192                       const char* shorty, uint32_t shorty_len)
193      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) :
194      is_static_(is_static), shorty_(shorty), shorty_len_(shorty_len),
195      gpr_args_(reinterpret_cast<byte*>(sp) + kQuickCalleeSaveFrame_RefAndArgs_Gpr1Offset),
196      fpr_args_(reinterpret_cast<byte*>(sp) + kQuickCalleeSaveFrame_RefAndArgs_Fpr1Offset),
197      stack_args_(reinterpret_cast<byte*>(sp) + kQuickCalleeSaveFrame_RefAndArgs_FrameSize
198                  + StackArgumentStartFromShorty(is_static, shorty, shorty_len)),
199      gpr_index_(0), fpr_index_(0), stack_index_(0), cur_type_(Primitive::kPrimVoid),
200      is_split_long_or_double_(false) {
201    DCHECK_EQ(kQuickCalleeSaveFrame_RefAndArgs_FrameSize,
202              Runtime::Current()->GetCalleeSaveMethod(Runtime::kRefsAndArgs)->GetFrameSizeInBytes());
203  }
204
205  virtual ~QuickArgumentVisitor() {}
206
207  virtual void Visit() = 0;
208
209  Primitive::Type GetParamPrimitiveType() const {
210    return cur_type_;
211  }
212
213  byte* GetParamAddress() const {
214    if (!kQuickSoftFloatAbi) {
215      Primitive::Type type = GetParamPrimitiveType();
216      if (UNLIKELY((type == Primitive::kPrimDouble) || (type == Primitive::kPrimFloat))) {
217        if ((kNumQuickFprArgs != 0) && (fpr_index_ + 1 < kNumQuickFprArgs + 1)) {
218          return fpr_args_ + (fpr_index_ * kBytesPerFprSpillLocation);
219        }
220      }
221    }
222    if (gpr_index_ < kNumQuickGprArgs) {
223      return gpr_args_ + GprIndexToGprOffset(gpr_index_);
224    }
225    return stack_args_ + (stack_index_ * kBytesStackArgLocation);
226  }
227
228  bool IsSplitLongOrDouble() const {
229    if ((kBytesPerGprSpillLocation == 4) || (kBytesPerFprSpillLocation == 4)) {
230      return is_split_long_or_double_;
231    } else {
232      return false;  // An optimization for when GPR and FPRs are 64bit.
233    }
234  }
235
236  bool IsParamAReference() const {
237    return GetParamPrimitiveType() == Primitive::kPrimNot;
238  }
239
240  bool IsParamALongOrDouble() const {
241    Primitive::Type type = GetParamPrimitiveType();
242    return type == Primitive::kPrimLong || type == Primitive::kPrimDouble;
243  }
244
245  uint64_t ReadSplitLongParam() const {
246    DCHECK(IsSplitLongOrDouble());
247    uint64_t low_half = *reinterpret_cast<uint32_t*>(GetParamAddress());
248    uint64_t high_half = *reinterpret_cast<uint32_t*>(stack_args_);
249    return (low_half & 0xffffffffULL) | (high_half << 32);
250  }
251
252  void VisitArguments() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
253    gpr_index_ = 0;
254    fpr_index_ = 0;
255    stack_index_ = 0;
256    if (!is_static_) {  // Handle this.
257      cur_type_ = Primitive::kPrimNot;
258      is_split_long_or_double_ = false;
259      Visit();
260      if (kNumQuickGprArgs > 0) {
261        gpr_index_++;
262      } else {
263        stack_index_++;
264      }
265    }
266    for (uint32_t shorty_index = 1; shorty_index < shorty_len_; ++shorty_index) {
267      cur_type_ = Primitive::GetType(shorty_[shorty_index]);
268      switch (cur_type_) {
269        case Primitive::kPrimNot:
270        case Primitive::kPrimBoolean:
271        case Primitive::kPrimByte:
272        case Primitive::kPrimChar:
273        case Primitive::kPrimShort:
274        case Primitive::kPrimInt:
275          is_split_long_or_double_ = false;
276          Visit();
277          if (gpr_index_ < kNumQuickGprArgs) {
278            gpr_index_++;
279          } else {
280            stack_index_++;
281          }
282          break;
283        case Primitive::kPrimFloat:
284          is_split_long_or_double_ = false;
285          Visit();
286          if (kQuickSoftFloatAbi) {
287            if (gpr_index_ < kNumQuickGprArgs) {
288              gpr_index_++;
289            } else {
290              stack_index_++;
291            }
292          } else {
293            if ((kNumQuickFprArgs != 0) && (fpr_index_ + 1 < kNumQuickFprArgs + 1)) {
294              fpr_index_++;
295            } else {
296              stack_index_++;
297            }
298          }
299          break;
300        case Primitive::kPrimDouble:
301        case Primitive::kPrimLong:
302          if (kQuickSoftFloatAbi || (cur_type_ == Primitive::kPrimLong)) {
303            is_split_long_or_double_ = (kBytesPerGprSpillLocation == 4) &&
304                ((gpr_index_ + 1) == kNumQuickGprArgs);
305            Visit();
306            if (gpr_index_ < kNumQuickGprArgs) {
307              gpr_index_++;
308              if (kBytesPerGprSpillLocation == 4) {
309                if (gpr_index_ < kNumQuickGprArgs) {
310                  gpr_index_++;
311                } else {
312                  stack_index_++;
313                }
314              }
315            } else {
316              if (kBytesStackArgLocation == 4) {
317                stack_index_+= 2;
318              } else {
319                CHECK_EQ(kBytesStackArgLocation, 8U);
320                stack_index_++;
321              }
322            }
323          } else {
324            is_split_long_or_double_ = (kBytesPerFprSpillLocation == 4) &&
325                ((fpr_index_ + 1) == kNumQuickFprArgs);
326            Visit();
327            if ((kNumQuickFprArgs != 0) && (fpr_index_ + 1 < kNumQuickFprArgs + 1)) {
328              fpr_index_++;
329              if (kBytesPerFprSpillLocation == 4) {
330                if ((kNumQuickFprArgs != 0) && (fpr_index_ + 1 < kNumQuickFprArgs + 1)) {
331                  fpr_index_++;
332                } else {
333                  stack_index_++;
334                }
335              }
336            } else {
337              if (kBytesStackArgLocation == 4) {
338                stack_index_+= 2;
339              } else {
340                CHECK_EQ(kBytesStackArgLocation, 8U);
341                stack_index_++;
342              }
343            }
344          }
345          break;
346        default:
347          LOG(FATAL) << "Unexpected type: " << cur_type_ << " in " << shorty_;
348      }
349    }
350  }
351
352 private:
353  static size_t StackArgumentStartFromShorty(bool is_static, const char* shorty,
354                                             uint32_t shorty_len) {
355    if (kQuickSoftFloatAbi) {
356      CHECK_EQ(kNumQuickFprArgs, 0U);
357      return (kNumQuickGprArgs * kBytesPerGprSpillLocation) + kBytesPerGprSpillLocation /* ArtMethod* */;
358    } else {
359      size_t offset = kBytesPerGprSpillLocation;  // Skip Method*.
360      size_t gprs_seen = 0;
361      size_t fprs_seen = 0;
362      if (!is_static && (gprs_seen < kNumQuickGprArgs)) {
363        gprs_seen++;
364        offset += kBytesStackArgLocation;
365      }
366      for (uint32_t i = 1; i < shorty_len; ++i) {
367        switch (shorty[i]) {
368          case 'Z':
369          case 'B':
370          case 'C':
371          case 'S':
372          case 'I':
373          case 'L':
374            if (gprs_seen < kNumQuickGprArgs) {
375              gprs_seen++;
376              offset += kBytesStackArgLocation;
377            }
378            break;
379          case 'J':
380            if (gprs_seen < kNumQuickGprArgs) {
381              gprs_seen++;
382              offset += 2 * kBytesStackArgLocation;
383              if (kBytesPerGprSpillLocation == 4) {
384                if (gprs_seen < kNumQuickGprArgs) {
385                  gprs_seen++;
386                }
387              }
388            }
389            break;
390          case 'F':
391            if ((kNumQuickFprArgs != 0) && (fprs_seen + 1 < kNumQuickFprArgs + 1)) {
392              fprs_seen++;
393              offset += kBytesStackArgLocation;
394            }
395            break;
396          case 'D':
397            if ((kNumQuickFprArgs != 0) && (fprs_seen + 1 < kNumQuickFprArgs + 1)) {
398              fprs_seen++;
399              offset += 2 * kBytesStackArgLocation;
400              if (kBytesPerFprSpillLocation == 4) {
401                if ((kNumQuickFprArgs != 0) && (fprs_seen + 1 < kNumQuickFprArgs + 1)) {
402                  fprs_seen++;
403                }
404              }
405            }
406            break;
407          default:
408            LOG(FATAL) << "Unexpected shorty character: " << shorty[i] << " in " << shorty;
409        }
410      }
411      return offset;
412    }
413  }
414
415  const bool is_static_;
416  const char* const shorty_;
417  const uint32_t shorty_len_;
418  byte* const gpr_args_;  // Address of GPR arguments in callee save frame.
419  byte* const fpr_args_;  // Address of FPR arguments in callee save frame.
420  byte* const stack_args_;  // Address of stack arguments in caller's frame.
421  uint32_t gpr_index_;  // Index into spilled GPRs.
422  uint32_t fpr_index_;  // Index into spilled FPRs.
423  uint32_t stack_index_;  // Index into arguments on the stack.
424  // The current type of argument during VisitArguments.
425  Primitive::Type cur_type_;
426  // Does a 64bit parameter straddle the register and stack arguments?
427  bool is_split_long_or_double_;
428};
429
430// Visits arguments on the stack placing them into the shadow frame.
431class BuildQuickShadowFrameVisitor FINAL : public QuickArgumentVisitor {
432 public:
433  BuildQuickShadowFrameVisitor(mirror::ArtMethod** sp, bool is_static, const char* shorty,
434                               uint32_t shorty_len, ShadowFrame* sf, size_t first_arg_reg) :
435    QuickArgumentVisitor(sp, is_static, shorty, shorty_len), sf_(sf), cur_reg_(first_arg_reg) {}
436
437  void Visit() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) OVERRIDE;
438
439 private:
440  ShadowFrame* const sf_;
441  uint32_t cur_reg_;
442
443  DISALLOW_COPY_AND_ASSIGN(BuildQuickShadowFrameVisitor);
444};
445
446void BuildQuickShadowFrameVisitor::Visit()  {
447  Primitive::Type type = GetParamPrimitiveType();
448  switch (type) {
449    case Primitive::kPrimLong:  // Fall-through.
450    case Primitive::kPrimDouble:
451      if (IsSplitLongOrDouble()) {
452        sf_->SetVRegLong(cur_reg_, ReadSplitLongParam());
453      } else {
454        sf_->SetVRegLong(cur_reg_, *reinterpret_cast<jlong*>(GetParamAddress()));
455      }
456      ++cur_reg_;
457      break;
458    case Primitive::kPrimNot: {
459        StackReference<mirror::Object>* stack_ref =
460            reinterpret_cast<StackReference<mirror::Object>*>(GetParamAddress());
461        sf_->SetVRegReference(cur_reg_, stack_ref->AsMirrorPtr());
462      }
463      break;
464    case Primitive::kPrimBoolean:  // Fall-through.
465    case Primitive::kPrimByte:     // Fall-through.
466    case Primitive::kPrimChar:     // Fall-through.
467    case Primitive::kPrimShort:    // Fall-through.
468    case Primitive::kPrimInt:      // Fall-through.
469    case Primitive::kPrimFloat:
470      sf_->SetVReg(cur_reg_, *reinterpret_cast<jint*>(GetParamAddress()));
471      break;
472    case Primitive::kPrimVoid:
473      LOG(FATAL) << "UNREACHABLE";
474      break;
475  }
476  ++cur_reg_;
477}
478
479extern "C" uint64_t artQuickToInterpreterBridge(mirror::ArtMethod* method, Thread* self,
480                                                mirror::ArtMethod** sp)
481    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
482  // Ensure we don't get thread suspension until the object arguments are safely in the shadow
483  // frame.
484  FinishCalleeSaveFrameSetup(self, sp, Runtime::kRefsAndArgs);
485
486  if (method->IsAbstract()) {
487    ThrowAbstractMethodError(method);
488    return 0;
489  } else {
490    DCHECK(!method->IsNative()) << PrettyMethod(method);
491    const char* old_cause = self->StartAssertNoThreadSuspension("Building interpreter shadow frame");
492    MethodHelper mh(method);
493    const DexFile::CodeItem* code_item = mh.GetCodeItem();
494    DCHECK(code_item != nullptr) << PrettyMethod(method);
495    uint16_t num_regs = code_item->registers_size_;
496    void* memory = alloca(ShadowFrame::ComputeSize(num_regs));
497    ShadowFrame* shadow_frame(ShadowFrame::Create(num_regs, NULL,  // No last shadow coming from quick.
498                                                  method, 0, memory));
499    size_t first_arg_reg = code_item->registers_size_ - code_item->ins_size_;
500    BuildQuickShadowFrameVisitor shadow_frame_builder(sp, mh.IsStatic(), mh.GetShorty(),
501                                                      mh.GetShortyLength(),
502                                                      shadow_frame, first_arg_reg);
503    shadow_frame_builder.VisitArguments();
504    // Push a transition back into managed code onto the linked list in thread.
505    ManagedStack fragment;
506    self->PushManagedStackFragment(&fragment);
507    self->PushShadowFrame(shadow_frame);
508    self->EndAssertNoThreadSuspension(old_cause);
509
510    if (method->IsStatic() && !method->GetDeclaringClass()->IsInitializing()) {
511      // Ensure static method's class is initialized.
512      SirtRef<mirror::Class> sirt_c(self, method->GetDeclaringClass());
513      if (!Runtime::Current()->GetClassLinker()->EnsureInitialized(sirt_c, true, true)) {
514        DCHECK(Thread::Current()->IsExceptionPending()) << PrettyMethod(method);
515        self->PopManagedStackFragment(fragment);
516        return 0;
517      }
518    }
519
520    JValue result = interpreter::EnterInterpreterFromStub(self, mh, code_item, *shadow_frame);
521    // Pop transition.
522    self->PopManagedStackFragment(fragment);
523    // No need to restore the args since the method has already been run by the interpreter.
524    return result.GetJ();
525  }
526}
527
528// Visits arguments on the stack placing them into the args vector, Object* arguments are converted
529// to jobjects.
530class BuildQuickArgumentVisitor FINAL : public QuickArgumentVisitor {
531 public:
532  BuildQuickArgumentVisitor(mirror::ArtMethod** sp, bool is_static, const char* shorty,
533                            uint32_t shorty_len, ScopedObjectAccessUnchecked* soa,
534                            std::vector<jvalue>* args) :
535    QuickArgumentVisitor(sp, is_static, shorty, shorty_len), soa_(soa), args_(args) {}
536
537  void Visit() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) OVERRIDE;
538
539  void FixupReferences() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_);
540
541 private:
542  ScopedObjectAccessUnchecked* const soa_;
543  std::vector<jvalue>* const args_;
544  // References which we must update when exiting in case the GC moved the objects.
545  std::vector<std::pair<jobject, StackReference<mirror::Object>*> > references_;
546
547  DISALLOW_COPY_AND_ASSIGN(BuildQuickArgumentVisitor);
548};
549
550void BuildQuickArgumentVisitor::Visit() {
551  jvalue val;
552  Primitive::Type type = GetParamPrimitiveType();
553  switch (type) {
554    case Primitive::kPrimNot: {
555      StackReference<mirror::Object>* stack_ref =
556          reinterpret_cast<StackReference<mirror::Object>*>(GetParamAddress());
557      val.l = soa_->AddLocalReference<jobject>(stack_ref->AsMirrorPtr());
558      references_.push_back(std::make_pair(val.l, stack_ref));
559      break;
560    }
561    case Primitive::kPrimLong:  // Fall-through.
562    case Primitive::kPrimDouble:
563      if (IsSplitLongOrDouble()) {
564        val.j = ReadSplitLongParam();
565      } else {
566        val.j = *reinterpret_cast<jlong*>(GetParamAddress());
567      }
568      break;
569    case Primitive::kPrimBoolean:  // Fall-through.
570    case Primitive::kPrimByte:     // Fall-through.
571    case Primitive::kPrimChar:     // Fall-through.
572    case Primitive::kPrimShort:    // Fall-through.
573    case Primitive::kPrimInt:      // Fall-through.
574    case Primitive::kPrimFloat:
575      val.i = *reinterpret_cast<jint*>(GetParamAddress());
576      break;
577    case Primitive::kPrimVoid:
578      LOG(FATAL) << "UNREACHABLE";
579      val.j = 0;
580      break;
581  }
582  args_->push_back(val);
583}
584
585void BuildQuickArgumentVisitor::FixupReferences() {
586  // Fixup any references which may have changed.
587  for (const auto& pair : references_) {
588    pair.second->Assign(soa_->Decode<mirror::Object*>(pair.first));
589  }
590}
591
592// Handler for invocation on proxy methods. On entry a frame will exist for the proxy object method
593// which is responsible for recording callee save registers. We explicitly place into jobjects the
594// incoming reference arguments (so they survive GC). We invoke the invocation handler, which is a
595// field within the proxy object, which will box the primitive arguments and deal with error cases.
596extern "C" uint64_t artQuickProxyInvokeHandler(mirror::ArtMethod* proxy_method,
597                                               mirror::Object* receiver,
598                                               Thread* self, mirror::ArtMethod** sp)
599    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
600  DCHECK(proxy_method->IsProxyMethod()) << PrettyMethod(proxy_method);
601  DCHECK(receiver->GetClass()->IsProxyClass()) << PrettyMethod(proxy_method);
602  // Ensure we don't get thread suspension until the object arguments are safely in jobjects.
603  const char* old_cause =
604      self->StartAssertNoThreadSuspension("Adding to IRT proxy object arguments");
605  // Register the top of the managed stack, making stack crawlable.
606  DCHECK_EQ(*sp, proxy_method) << PrettyMethod(proxy_method);
607  self->SetTopOfStack(sp, 0);
608  DCHECK_EQ(proxy_method->GetFrameSizeInBytes(),
609            Runtime::Current()->GetCalleeSaveMethod(Runtime::kRefsAndArgs)->GetFrameSizeInBytes())
610      << PrettyMethod(proxy_method);
611  self->VerifyStack();
612  // Start new JNI local reference state.
613  JNIEnvExt* env = self->GetJniEnv();
614  ScopedObjectAccessUnchecked soa(env);
615  ScopedJniEnvLocalRefState env_state(env);
616  // Create local ref. copies of proxy method and the receiver.
617  jobject rcvr_jobj = soa.AddLocalReference<jobject>(receiver);
618
619  // Placing arguments into args vector and remove the receiver.
620  MethodHelper proxy_mh(proxy_method);
621  DCHECK(!proxy_mh.IsStatic()) << PrettyMethod(proxy_method);
622  std::vector<jvalue> args;
623  BuildQuickArgumentVisitor local_ref_visitor(sp, proxy_mh.IsStatic(), proxy_mh.GetShorty(),
624                                              proxy_mh.GetShortyLength(), &soa, &args);
625
626  local_ref_visitor.VisitArguments();
627  DCHECK_GT(args.size(), 0U) << PrettyMethod(proxy_method);
628  args.erase(args.begin());
629
630  // Convert proxy method into expected interface method.
631  mirror::ArtMethod* interface_method = proxy_method->FindOverriddenMethod();
632  DCHECK(interface_method != NULL) << PrettyMethod(proxy_method);
633  DCHECK(!interface_method->IsProxyMethod()) << PrettyMethod(interface_method);
634  jobject interface_method_jobj = soa.AddLocalReference<jobject>(interface_method);
635
636  // All naked Object*s should now be in jobjects, so its safe to go into the main invoke code
637  // that performs allocations.
638  self->EndAssertNoThreadSuspension(old_cause);
639  JValue result = InvokeProxyInvocationHandler(soa, proxy_mh.GetShorty(),
640                                               rcvr_jobj, interface_method_jobj, args);
641  // Restore references which might have moved.
642  local_ref_visitor.FixupReferences();
643  return result.GetJ();
644}
645
646// Read object references held in arguments from quick frames and place in a JNI local references,
647// so they don't get garbage collected.
648class RememberForGcArgumentVisitor FINAL : public QuickArgumentVisitor {
649 public:
650  RememberForGcArgumentVisitor(mirror::ArtMethod** sp, bool is_static, const char* shorty,
651                               uint32_t shorty_len, ScopedObjectAccessUnchecked* soa) :
652    QuickArgumentVisitor(sp, is_static, shorty, shorty_len), soa_(soa) {}
653
654  void Visit() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) OVERRIDE;
655
656  void FixupReferences() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_);
657
658 private:
659  ScopedObjectAccessUnchecked* const soa_;
660  // References which we must update when exiting in case the GC moved the objects.
661  std::vector<std::pair<jobject, StackReference<mirror::Object>*> > references_;
662  DISALLOW_COPY_AND_ASSIGN(RememberForGcArgumentVisitor);
663};
664
665void RememberForGcArgumentVisitor::Visit() {
666  if (IsParamAReference()) {
667    StackReference<mirror::Object>* stack_ref =
668        reinterpret_cast<StackReference<mirror::Object>*>(GetParamAddress());
669    jobject reference =
670        soa_->AddLocalReference<jobject>(stack_ref->AsMirrorPtr());
671    references_.push_back(std::make_pair(reference, stack_ref));
672  }
673}
674
675void RememberForGcArgumentVisitor::FixupReferences() {
676  // Fixup any references which may have changed.
677  for (const auto& pair : references_) {
678    pair.second->Assign(soa_->Decode<mirror::Object*>(pair.first));
679  }
680}
681
682
683// Lazily resolve a method for quick. Called by stub code.
684extern "C" const void* artQuickResolutionTrampoline(mirror::ArtMethod* called,
685                                                    mirror::Object* receiver,
686                                                    Thread* self, mirror::ArtMethod** sp)
687    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
688  FinishCalleeSaveFrameSetup(self, sp, Runtime::kRefsAndArgs);
689  // Start new JNI local reference state
690  JNIEnvExt* env = self->GetJniEnv();
691  ScopedObjectAccessUnchecked soa(env);
692  ScopedJniEnvLocalRefState env_state(env);
693  const char* old_cause = self->StartAssertNoThreadSuspension("Quick method resolution set up");
694
695  // Compute details about the called method (avoid GCs)
696  ClassLinker* linker = Runtime::Current()->GetClassLinker();
697  mirror::ArtMethod* caller = QuickArgumentVisitor::GetCallingMethod(sp);
698  InvokeType invoke_type;
699  const DexFile* dex_file;
700  uint32_t dex_method_idx;
701  if (called->IsRuntimeMethod()) {
702    uint32_t dex_pc = caller->ToDexPc(QuickArgumentVisitor::GetCallingPc(sp));
703    const DexFile::CodeItem* code;
704    {
705      MethodHelper mh(caller);
706      dex_file = &mh.GetDexFile();
707      code = mh.GetCodeItem();
708    }
709    CHECK_LT(dex_pc, code->insns_size_in_code_units_);
710    const Instruction* instr = Instruction::At(&code->insns_[dex_pc]);
711    Instruction::Code instr_code = instr->Opcode();
712    bool is_range;
713    switch (instr_code) {
714      case Instruction::INVOKE_DIRECT:
715        invoke_type = kDirect;
716        is_range = false;
717        break;
718      case Instruction::INVOKE_DIRECT_RANGE:
719        invoke_type = kDirect;
720        is_range = true;
721        break;
722      case Instruction::INVOKE_STATIC:
723        invoke_type = kStatic;
724        is_range = false;
725        break;
726      case Instruction::INVOKE_STATIC_RANGE:
727        invoke_type = kStatic;
728        is_range = true;
729        break;
730      case Instruction::INVOKE_SUPER:
731        invoke_type = kSuper;
732        is_range = false;
733        break;
734      case Instruction::INVOKE_SUPER_RANGE:
735        invoke_type = kSuper;
736        is_range = true;
737        break;
738      case Instruction::INVOKE_VIRTUAL:
739        invoke_type = kVirtual;
740        is_range = false;
741        break;
742      case Instruction::INVOKE_VIRTUAL_RANGE:
743        invoke_type = kVirtual;
744        is_range = true;
745        break;
746      case Instruction::INVOKE_INTERFACE:
747        invoke_type = kInterface;
748        is_range = false;
749        break;
750      case Instruction::INVOKE_INTERFACE_RANGE:
751        invoke_type = kInterface;
752        is_range = true;
753        break;
754      default:
755        LOG(FATAL) << "Unexpected call into trampoline: " << instr->DumpString(NULL);
756        // Avoid used uninitialized warnings.
757        invoke_type = kDirect;
758        is_range = false;
759    }
760    dex_method_idx = (is_range) ? instr->VRegB_3rc() : instr->VRegB_35c();
761
762  } else {
763    invoke_type = kStatic;
764    dex_file = &MethodHelper(called).GetDexFile();
765    dex_method_idx = called->GetDexMethodIndex();
766  }
767  uint32_t shorty_len;
768  const char* shorty =
769      dex_file->GetMethodShorty(dex_file->GetMethodId(dex_method_idx), &shorty_len);
770  RememberForGcArgumentVisitor visitor(sp, invoke_type == kStatic, shorty, shorty_len, &soa);
771  visitor.VisitArguments();
772  self->EndAssertNoThreadSuspension(old_cause);
773  bool virtual_or_interface = invoke_type == kVirtual || invoke_type == kInterface;
774  // Resolve method filling in dex cache.
775  if (called->IsRuntimeMethod()) {
776    SirtRef<mirror::Object> sirt_receiver(soa.Self(), virtual_or_interface ? receiver : nullptr);
777    called = linker->ResolveMethod(dex_method_idx, caller, invoke_type);
778    receiver = sirt_receiver.get();
779  }
780  const void* code = NULL;
781  if (LIKELY(!self->IsExceptionPending())) {
782    // Incompatible class change should have been handled in resolve method.
783    CHECK(!called->CheckIncompatibleClassChange(invoke_type))
784        << PrettyMethod(called) << " " << invoke_type;
785    if (virtual_or_interface) {
786      // Refine called method based on receiver.
787      CHECK(receiver != nullptr) << invoke_type;
788      if (invoke_type == kVirtual) {
789        called = receiver->GetClass()->FindVirtualMethodForVirtual(called);
790      } else {
791        called = receiver->GetClass()->FindVirtualMethodForInterface(called);
792      }
793      // We came here because of sharpening. Ensure the dex cache is up-to-date on the method index
794      // of the sharpened method.
795      if (called->GetDexCacheResolvedMethods() == caller->GetDexCacheResolvedMethods()) {
796        caller->GetDexCacheResolvedMethods()->Set<false>(called->GetDexMethodIndex(), called);
797      } else {
798        // Calling from one dex file to another, need to compute the method index appropriate to
799        // the caller's dex file. Since we get here only if the original called was a runtime
800        // method, we've got the correct dex_file and a dex_method_idx from above.
801        DCHECK(&MethodHelper(caller).GetDexFile() == dex_file);
802        uint32_t method_index =
803            MethodHelper(called).FindDexMethodIndexInOtherDexFile(*dex_file, dex_method_idx);
804        if (method_index != DexFile::kDexNoIndex) {
805          caller->GetDexCacheResolvedMethods()->Set<false>(method_index, called);
806        }
807      }
808    }
809    // Ensure that the called method's class is initialized.
810    SirtRef<mirror::Class> called_class(soa.Self(), called->GetDeclaringClass());
811    linker->EnsureInitialized(called_class, true, true);
812    if (LIKELY(called_class->IsInitialized())) {
813      code = called->GetEntryPointFromQuickCompiledCode();
814    } else if (called_class->IsInitializing()) {
815      if (invoke_type == kStatic) {
816        // Class is still initializing, go to oat and grab code (trampoline must be left in place
817        // until class is initialized to stop races between threads).
818        code = linker->GetQuickOatCodeFor(called);
819      } else {
820        // No trampoline for non-static methods.
821        code = called->GetEntryPointFromQuickCompiledCode();
822      }
823    } else {
824      DCHECK(called_class->IsErroneous());
825    }
826  }
827  CHECK_EQ(code == NULL, self->IsExceptionPending());
828  // Fixup any locally saved objects may have moved during a GC.
829  visitor.FixupReferences();
830  // Place called method in callee-save frame to be placed as first argument to quick method.
831  *sp = called;
832  return code;
833}
834
835
836
837/*
838 * This class uses a couple of observations to unite the different calling conventions through
839 * a few constants.
840 *
841 * 1) Number of registers used for passing is normally even, so counting down has no penalty for
842 *    possible alignment.
843 * 2) Known 64b architectures store 8B units on the stack, both for integral and floating point
844 *    types, so using uintptr_t is OK. Also means that we can use kRegistersNeededX to denote
845 *    when we have to split things
846 * 3) The only soft-float, Arm, is 32b, so no widening needs to be taken into account for floats
847 *    and we can use Int handling directly.
848 * 4) Only 64b architectures widen, and their stack is aligned 8B anyways, so no padding code
849 *    necessary when widening. Also, widening of Ints will take place implicitly, and the
850 *    extension should be compatible with Aarch64, which mandates copying the available bits
851 *    into LSB and leaving the rest unspecified.
852 * 5) Aligning longs and doubles is necessary on arm only, and it's the same in registers and on
853 *    the stack.
854 * 6) There is only little endian.
855 *
856 *
857 * Actual work is supposed to be done in a delegate of the template type. The interface is as
858 * follows:
859 *
860 * void PushGpr(uintptr_t):   Add a value for the next GPR
861 *
862 * void PushFpr4(float):      Add a value for the next FPR of size 32b. Is only called if we need
863 *                            padding, that is, think the architecture is 32b and aligns 64b.
864 *
865 * void PushFpr8(uint64_t):   Push a double. We _will_ call this on 32b, it's the callee's job to
866 *                            split this if necessary. The current state will have aligned, if
867 *                            necessary.
868 *
869 * void PushStack(uintptr_t): Push a value to the stack.
870 *
871 * uintptr_t PushSirt(mirror::Object* ref): Add a reference to the Sirt. This _will_ have nullptr,
872 *                                          as this might be important for null initialization.
873 *                                          Must return the jobject, that is, the reference to the
874 *                                          entry in the Sirt (nullptr if necessary).
875 *
876 */
877template <class T> class BuildGenericJniFrameStateMachine {
878 public:
879#if defined(__arm__)
880  // TODO: These are all dummy values!
881  static constexpr bool kNativeSoftFloatAbi = true;
882  static constexpr size_t kNumNativeGprArgs = 4;  // 4 arguments passed in GPRs, r0-r3
883  static constexpr size_t kNumNativeFprArgs = 0;  // 0 arguments passed in FPRs.
884
885  static constexpr size_t kRegistersNeededForLong = 2;
886  static constexpr size_t kRegistersNeededForDouble = 2;
887  static constexpr bool kMultiRegistersAligned = true;
888  static constexpr bool kMultiRegistersWidened = false;
889  static constexpr bool kAlignLongOnStack = true;
890  static constexpr bool kAlignDoubleOnStack = true;
891#elif defined(__mips__)
892  // TODO: These are all dummy values!
893  static constexpr bool kNativeSoftFloatAbi = true;  // This is a hard float ABI.
894  static constexpr size_t kNumNativeGprArgs = 0;  // 6 arguments passed in GPRs.
895  static constexpr size_t kNumNativeFprArgs = 0;  // 8 arguments passed in FPRs.
896
897  static constexpr size_t kRegistersNeededForLong = 2;
898  static constexpr size_t kRegistersNeededForDouble = 2;
899  static constexpr bool kMultiRegistersAligned = true;
900  static constexpr bool kMultiRegistersWidened = true;
901  static constexpr bool kAlignLongOnStack = false;
902  static constexpr bool kAlignDoubleOnStack = false;
903#elif defined(__i386__)
904  // TODO: Check these!
905  static constexpr bool kNativeSoftFloatAbi = false;  // Not using int registers for fp
906  static constexpr size_t kNumNativeGprArgs = 0;  // 6 arguments passed in GPRs.
907  static constexpr size_t kNumNativeFprArgs = 0;  // 8 arguments passed in FPRs.
908
909  static constexpr size_t kRegistersNeededForLong = 2;
910  static constexpr size_t kRegistersNeededForDouble = 2;
911  static constexpr bool kMultiRegistersAligned = false;       // x86 not using regs, anyways
912  static constexpr bool kMultiRegistersWidened = false;
913  static constexpr bool kAlignLongOnStack = false;
914  static constexpr bool kAlignDoubleOnStack = false;
915#elif defined(__x86_64__)
916  static constexpr bool kNativeSoftFloatAbi = false;  // This is a hard float ABI.
917  static constexpr size_t kNumNativeGprArgs = 6;  // 6 arguments passed in GPRs.
918  static constexpr size_t kNumNativeFprArgs = 8;  // 8 arguments passed in FPRs.
919
920  static constexpr size_t kRegistersNeededForLong = 1;
921  static constexpr size_t kRegistersNeededForDouble = 1;
922  static constexpr bool kMultiRegistersAligned = false;
923  static constexpr bool kMultiRegistersWidened = false;
924  static constexpr bool kAlignLongOnStack = false;
925  static constexpr bool kAlignDoubleOnStack = false;
926#else
927#error "Unsupported architecture"
928#endif
929
930 public:
931  explicit BuildGenericJniFrameStateMachine(T* delegate) : gpr_index_(kNumNativeGprArgs),
932                                                           fpr_index_(kNumNativeFprArgs),
933                                                           stack_entries_(0),
934                                                           delegate_(delegate) {
935    // For register alignment, we want to assume that counters (gpr_index_, fpr_index_) are even iff
936    // the next register is even; counting down is just to make the compiler happy...
937    CHECK_EQ(kNumNativeGprArgs % 2, 0U);
938    CHECK_EQ(kNumNativeFprArgs % 2, 0U);
939  }
940
941  virtual ~BuildGenericJniFrameStateMachine() {}
942
943  bool HavePointerGpr() {
944    return gpr_index_ > 0;
945  }
946
947  void AdvancePointer(void* val) {
948    if (HavePointerGpr()) {
949      gpr_index_--;
950      PushGpr(reinterpret_cast<uintptr_t>(val));
951    } else {
952      stack_entries_++;         // TODO: have a field for pointer length as multiple of 32b
953      PushStack(reinterpret_cast<uintptr_t>(val));
954      gpr_index_ = 0;
955    }
956  }
957
958
959  bool HaveSirtGpr() {
960    return gpr_index_ > 0;
961  }
962
963  void AdvanceSirt(mirror::Object* ptr) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
964    uintptr_t sirtRef = PushSirt(ptr);
965    if (HaveSirtGpr()) {
966      gpr_index_--;
967      PushGpr(sirtRef);
968    } else {
969      stack_entries_++;
970      PushStack(sirtRef);
971      gpr_index_ = 0;
972    }
973  }
974
975
976  bool HaveIntGpr() {
977    return gpr_index_ > 0;
978  }
979
980  void AdvanceInt(uint32_t val) {
981    if (HaveIntGpr()) {
982      gpr_index_--;
983      PushGpr(val);
984    } else {
985      stack_entries_++;
986      PushStack(val);
987      gpr_index_ = 0;
988    }
989  }
990
991
992  bool HaveLongGpr() {
993    return gpr_index_ >= kRegistersNeededForLong + (LongGprNeedsPadding() ? 1 : 0);
994  }
995
996  bool LongGprNeedsPadding() {
997    return kRegistersNeededForLong > 1 &&     // only pad when using multiple registers
998        kAlignLongOnStack &&                  // and when it needs alignment
999        (gpr_index_ & 1) == 1;                // counter is odd, see constructor
1000  }
1001
1002  bool LongStackNeedsPadding() {
1003    return kRegistersNeededForLong > 1 &&     // only pad when using multiple registers
1004        kAlignLongOnStack &&                  // and when it needs 8B alignment
1005        (stack_entries_ & 1) == 1;            // counter is odd
1006  }
1007
1008  void AdvanceLong(uint64_t val) {
1009    if (HaveLongGpr()) {
1010      if (LongGprNeedsPadding()) {
1011        PushGpr(0);
1012        gpr_index_--;
1013      }
1014      if (kRegistersNeededForLong == 1) {
1015        PushGpr(static_cast<uintptr_t>(val));
1016      } else {
1017        PushGpr(static_cast<uintptr_t>(val & 0xFFFFFFFF));
1018        PushGpr(static_cast<uintptr_t>((val >> 32) & 0xFFFFFFFF));
1019      }
1020      gpr_index_ -= kRegistersNeededForLong;
1021    } else {
1022      if (LongStackNeedsPadding()) {
1023        PushStack(0);
1024        stack_entries_++;
1025      }
1026      if (kRegistersNeededForLong == 1) {
1027        PushStack(static_cast<uintptr_t>(val));
1028        stack_entries_++;
1029      } else {
1030        PushStack(static_cast<uintptr_t>(val & 0xFFFFFFFF));
1031        PushStack(static_cast<uintptr_t>((val >> 32) & 0xFFFFFFFF));
1032        stack_entries_ += 2;
1033      }
1034      gpr_index_ = 0;
1035    }
1036  }
1037
1038
1039  bool HaveFloatFpr() {
1040    return fpr_index_ > 0;
1041  }
1042
1043  template <typename U, typename V> V convert(U in) {
1044    CHECK_LE(sizeof(U), sizeof(V));
1045    union { U u; V v; } tmp;
1046    tmp.u = in;
1047    return tmp.v;
1048  }
1049
1050  void AdvanceFloat(float val) {
1051    if (kNativeSoftFloatAbi) {
1052      AdvanceInt(convert<float, uint32_t>(val));
1053    } else {
1054      if (HaveFloatFpr()) {
1055        fpr_index_--;
1056        if (kRegistersNeededForDouble == 1) {
1057          if (kMultiRegistersWidened) {
1058            PushFpr8(convert<double, uint64_t>(val));
1059          } else {
1060            // No widening, just use the bits.
1061            PushFpr8(convert<float, uint64_t>(val));
1062          }
1063        } else {
1064          PushFpr4(val);
1065        }
1066      } else {
1067        stack_entries_++;
1068        if (kRegistersNeededForDouble == 1 && kMultiRegistersWidened) {
1069          // Need to widen before storing: Note the "double" in the template instantiation.
1070          PushStack(convert<double, uintptr_t>(val));
1071        } else {
1072          PushStack(convert<float, uintptr_t>(val));
1073        }
1074        fpr_index_ = 0;
1075      }
1076    }
1077  }
1078
1079
1080  bool HaveDoubleFpr() {
1081    return fpr_index_ >= kRegistersNeededForDouble + (DoubleFprNeedsPadding() ? 1 : 0);
1082  }
1083
1084  bool DoubleFprNeedsPadding() {
1085    return kRegistersNeededForDouble > 1 &&     // only pad when using multiple registers
1086        kAlignDoubleOnStack &&                  // and when it needs alignment
1087        (fpr_index_ & 1) == 1;                  // counter is odd, see constructor
1088  }
1089
1090  bool DoubleStackNeedsPadding() {
1091    return kRegistersNeededForDouble > 1 &&     // only pad when using multiple registers
1092        kAlignDoubleOnStack &&                  // and when it needs 8B alignment
1093        (stack_entries_ & 1) == 1;              // counter is odd
1094  }
1095
1096  void AdvanceDouble(uint64_t val) {
1097    if (kNativeSoftFloatAbi) {
1098      AdvanceLong(val);
1099    } else {
1100      if (HaveDoubleFpr()) {
1101        if (DoubleFprNeedsPadding()) {
1102          PushFpr4(0);
1103          fpr_index_--;
1104        }
1105        PushFpr8(val);
1106        fpr_index_ -= kRegistersNeededForDouble;
1107      } else {
1108        if (DoubleStackNeedsPadding()) {
1109          PushStack(0);
1110          stack_entries_++;
1111        }
1112        if (kRegistersNeededForDouble == 1) {
1113          PushStack(static_cast<uintptr_t>(val));
1114          stack_entries_++;
1115        } else {
1116          PushStack(static_cast<uintptr_t>(val & 0xFFFFFFFF));
1117          PushStack(static_cast<uintptr_t>((val >> 32) & 0xFFFFFFFF));
1118          stack_entries_ += 2;
1119        }
1120        fpr_index_ = 0;
1121      }
1122    }
1123  }
1124
1125  uint32_t getStackEntries() {
1126    return stack_entries_;
1127  }
1128
1129  uint32_t getNumberOfUsedGprs() {
1130    return kNumNativeGprArgs - gpr_index_;
1131  }
1132
1133  uint32_t getNumberOfUsedFprs() {
1134    return kNumNativeFprArgs - fpr_index_;
1135  }
1136
1137 private:
1138  void PushGpr(uintptr_t val) {
1139    delegate_->PushGpr(val);
1140  }
1141  void PushFpr4(float val) {
1142    delegate_->PushFpr4(val);
1143  }
1144  void PushFpr8(uint64_t val) {
1145    delegate_->PushFpr8(val);
1146  }
1147  void PushStack(uintptr_t val) {
1148    delegate_->PushStack(val);
1149  }
1150  uintptr_t PushSirt(mirror::Object* ref) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1151    return delegate_->PushSirt(ref);
1152  }
1153
1154  uint32_t gpr_index_;      // Number of free GPRs
1155  uint32_t fpr_index_;      // Number of free FPRs
1156  uint32_t stack_entries_;  // Stack entries are in multiples of 32b, as floats are usually not
1157                            // extended
1158  T* delegate_;             // What Push implementation gets called
1159};
1160
1161class ComputeGenericJniFrameSize FINAL {
1162 public:
1163  ComputeGenericJniFrameSize() : num_sirt_references_(0), num_stack_entries_(0) {}
1164
1165  uint32_t GetStackSize() {
1166    return num_stack_entries_ * sizeof(uintptr_t);
1167  }
1168
1169  // WARNING: After this, *sp won't be pointing to the method anymore!
1170  void ComputeLayout(mirror::ArtMethod*** m, bool is_static, const char* shorty, uint32_t shorty_len,
1171                     void* sp, StackIndirectReferenceTable** table, uint32_t* sirt_entries,
1172                     uintptr_t** start_stack, uintptr_t** start_gpr, uint32_t** start_fpr,
1173                     void** code_return, size_t* overall_size)
1174      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1175    ComputeAll(is_static, shorty, shorty_len);
1176
1177    mirror::ArtMethod* method = **m;
1178
1179    uint8_t* sp8 = reinterpret_cast<uint8_t*>(sp);
1180
1181    // First, fix up the layout of the callee-save frame.
1182    // We have to squeeze in the Sirt, and relocate the method pointer.
1183
1184    // "Free" the slot for the method.
1185    sp8 += kPointerSize;
1186
1187    // Add the Sirt.
1188    *sirt_entries = num_sirt_references_;
1189    size_t sirt_size = StackIndirectReferenceTable::GetAlignedSirtSize(num_sirt_references_);
1190    sp8 -= sirt_size;
1191    *table = reinterpret_cast<StackIndirectReferenceTable*>(sp8);
1192    (*table)->SetNumberOfReferences(num_sirt_references_);
1193
1194    // Add a slot for the method pointer, and fill it. Fix the pointer-pointer given to us.
1195    sp8 -= kPointerSize;
1196    uint8_t* method_pointer = sp8;
1197    *(reinterpret_cast<mirror::ArtMethod**>(method_pointer)) = method;
1198    *m = reinterpret_cast<mirror::ArtMethod**>(method_pointer);
1199
1200    // Reference cookie and padding
1201    sp8 -= 8;
1202    // Store Sirt size
1203    *reinterpret_cast<uint32_t*>(sp8) = static_cast<uint32_t>(sirt_size & 0xFFFFFFFF);
1204
1205    // Next comes the native call stack.
1206    sp8 -= GetStackSize();
1207    // Now align the call stack below. This aligns by 16, as AArch64 seems to require.
1208    uintptr_t mask = ~0x0F;
1209    sp8 = reinterpret_cast<uint8_t*>(reinterpret_cast<uintptr_t>(sp8) & mask);
1210    *start_stack = reinterpret_cast<uintptr_t*>(sp8);
1211
1212    // put fprs and gprs below
1213    // Assumption is OK right now, as we have soft-float arm
1214    size_t fregs = BuildGenericJniFrameStateMachine<ComputeGenericJniFrameSize>::kNumNativeFprArgs;
1215    sp8 -= fregs * sizeof(uintptr_t);
1216    *start_fpr = reinterpret_cast<uint32_t*>(sp8);
1217    size_t iregs = BuildGenericJniFrameStateMachine<ComputeGenericJniFrameSize>::kNumNativeGprArgs;
1218    sp8 -= iregs * sizeof(uintptr_t);
1219    *start_gpr = reinterpret_cast<uintptr_t*>(sp8);
1220
1221    // reserve space for the code pointer
1222    sp8 -= kPointerSize;
1223    *code_return = reinterpret_cast<void*>(sp8);
1224
1225    *overall_size = reinterpret_cast<uint8_t*>(sp) - sp8;
1226
1227    // The new SP is stored at the end of the alloca, so it can be immediately popped
1228    sp8 = reinterpret_cast<uint8_t*>(sp) - 5 * KB;
1229    *(reinterpret_cast<uint8_t**>(sp8)) = method_pointer;
1230  }
1231
1232  void ComputeSirtOffset() { }  // nothing to do, static right now
1233
1234  void ComputeAll(bool is_static, const char* shorty, uint32_t shorty_len)
1235      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1236    BuildGenericJniFrameStateMachine<ComputeGenericJniFrameSize> sm(this);
1237
1238    // JNIEnv
1239    sm.AdvancePointer(nullptr);
1240
1241    // Class object or this as first argument
1242    sm.AdvanceSirt(reinterpret_cast<mirror::Object*>(0x12345678));
1243
1244    for (uint32_t i = 1; i < shorty_len; ++i) {
1245      Primitive::Type cur_type_ = Primitive::GetType(shorty[i]);
1246      switch (cur_type_) {
1247        case Primitive::kPrimNot:
1248          sm.AdvanceSirt(reinterpret_cast<mirror::Object*>(0x12345678));
1249          break;
1250
1251        case Primitive::kPrimBoolean:
1252        case Primitive::kPrimByte:
1253        case Primitive::kPrimChar:
1254        case Primitive::kPrimShort:
1255        case Primitive::kPrimInt:
1256          sm.AdvanceInt(0);
1257          break;
1258        case Primitive::kPrimFloat:
1259          sm.AdvanceFloat(0);
1260          break;
1261        case Primitive::kPrimDouble:
1262          sm.AdvanceDouble(0);
1263          break;
1264        case Primitive::kPrimLong:
1265          sm.AdvanceLong(0);
1266          break;
1267        default:
1268          LOG(FATAL) << "Unexpected type: " << cur_type_ << " in " << shorty;
1269      }
1270    }
1271
1272    num_stack_entries_ = sm.getStackEntries();
1273  }
1274
1275  void PushGpr(uintptr_t /* val */) {
1276    // not optimizing registers, yet
1277  }
1278
1279  void PushFpr4(float /* val */) {
1280    // not optimizing registers, yet
1281  }
1282
1283  void PushFpr8(uint64_t /* val */) {
1284    // not optimizing registers, yet
1285  }
1286
1287  void PushStack(uintptr_t /* val */) {
1288    // counting is already done in the superclass
1289  }
1290
1291  uintptr_t PushSirt(mirror::Object* /* ptr */) {
1292    num_sirt_references_++;
1293    return reinterpret_cast<uintptr_t>(nullptr);
1294  }
1295
1296 private:
1297  uint32_t num_sirt_references_;
1298  uint32_t num_stack_entries_;
1299};
1300
1301// Visits arguments on the stack placing them into a region lower down the stack for the benefit
1302// of transitioning into native code.
1303class BuildGenericJniFrameVisitor FINAL : public QuickArgumentVisitor {
1304 public:
1305  BuildGenericJniFrameVisitor(mirror::ArtMethod*** sp, bool is_static, const char* shorty,
1306                              uint32_t shorty_len, Thread* self) :
1307      QuickArgumentVisitor(*sp, is_static, shorty, shorty_len), sm_(this) {
1308    ComputeGenericJniFrameSize fsc;
1309    fsc.ComputeLayout(sp, is_static, shorty, shorty_len, *sp, &sirt_, &sirt_expected_refs_,
1310                      &cur_stack_arg_, &cur_gpr_reg_, &cur_fpr_reg_, &code_return_,
1311                      &alloca_used_size_);
1312    sirt_number_of_references_ = 0;
1313    cur_sirt_entry_ = reinterpret_cast<StackReference<mirror::Object>*>(GetFirstSirtEntry());
1314
1315    // jni environment is always first argument
1316    sm_.AdvancePointer(self->GetJniEnv());
1317
1318    if (is_static) {
1319      sm_.AdvanceSirt((**sp)->GetDeclaringClass());
1320    }
1321  }
1322
1323  void Visit() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) OVERRIDE;
1324
1325  void FinalizeSirt(Thread* self) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_);
1326
1327  jobject GetFirstSirtEntry() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1328    return reinterpret_cast<jobject>(sirt_->GetStackReference(0));
1329  }
1330
1331  void PushGpr(uintptr_t val) {
1332    *cur_gpr_reg_ = val;
1333    cur_gpr_reg_++;
1334  }
1335
1336  void PushFpr4(float val) {
1337    *cur_fpr_reg_ = val;
1338    cur_fpr_reg_++;
1339  }
1340
1341  void PushFpr8(uint64_t val) {
1342    uint64_t* tmp = reinterpret_cast<uint64_t*>(cur_fpr_reg_);
1343    *tmp = val;
1344    cur_fpr_reg_ += 2;
1345  }
1346
1347  void PushStack(uintptr_t val) {
1348    *cur_stack_arg_ = val;
1349    cur_stack_arg_++;
1350  }
1351
1352  uintptr_t PushSirt(mirror::Object* ref) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1353    uintptr_t tmp;
1354    if (ref == nullptr) {
1355      *cur_sirt_entry_ = StackReference<mirror::Object>();
1356      tmp = reinterpret_cast<uintptr_t>(nullptr);
1357    } else {
1358      *cur_sirt_entry_ = StackReference<mirror::Object>::FromMirrorPtr(ref);
1359      tmp = reinterpret_cast<uintptr_t>(cur_sirt_entry_);
1360    }
1361    cur_sirt_entry_++;
1362    sirt_number_of_references_++;
1363    return tmp;
1364  }
1365
1366  // Size of the part of the alloca that we actually need.
1367  size_t GetAllocaUsedSize() {
1368    return alloca_used_size_;
1369  }
1370
1371  void* GetCodeReturn() {
1372    return code_return_;
1373  }
1374
1375 private:
1376  uint32_t sirt_number_of_references_;
1377  StackReference<mirror::Object>* cur_sirt_entry_;
1378  StackIndirectReferenceTable* sirt_;
1379  uint32_t sirt_expected_refs_;
1380  uintptr_t* cur_gpr_reg_;
1381  uint32_t* cur_fpr_reg_;
1382  uintptr_t* cur_stack_arg_;
1383  // StackReference<mirror::Object>* top_of_sirt_;
1384  void* code_return_;
1385  size_t alloca_used_size_;
1386
1387  BuildGenericJniFrameStateMachine<BuildGenericJniFrameVisitor> sm_;
1388
1389  DISALLOW_COPY_AND_ASSIGN(BuildGenericJniFrameVisitor);
1390};
1391
1392void BuildGenericJniFrameVisitor::Visit() {
1393  Primitive::Type type = GetParamPrimitiveType();
1394  switch (type) {
1395    case Primitive::kPrimLong: {
1396      jlong long_arg;
1397      if (IsSplitLongOrDouble()) {
1398        long_arg = ReadSplitLongParam();
1399      } else {
1400        long_arg = *reinterpret_cast<jlong*>(GetParamAddress());
1401      }
1402      sm_.AdvanceLong(long_arg);
1403      break;
1404    }
1405    case Primitive::kPrimDouble: {
1406      uint64_t double_arg;
1407      if (IsSplitLongOrDouble()) {
1408        // Read into union so that we don't case to a double.
1409        double_arg = ReadSplitLongParam();
1410      } else {
1411        double_arg = *reinterpret_cast<uint64_t*>(GetParamAddress());
1412      }
1413      sm_.AdvanceDouble(double_arg);
1414      break;
1415    }
1416    case Primitive::kPrimNot: {
1417      StackReference<mirror::Object>* stack_ref =
1418          reinterpret_cast<StackReference<mirror::Object>*>(GetParamAddress());
1419      sm_.AdvanceSirt(stack_ref->AsMirrorPtr());
1420      break;
1421    }
1422    case Primitive::kPrimFloat:
1423      sm_.AdvanceFloat(*reinterpret_cast<float*>(GetParamAddress()));
1424      break;
1425    case Primitive::kPrimBoolean:  // Fall-through.
1426    case Primitive::kPrimByte:     // Fall-through.
1427    case Primitive::kPrimChar:     // Fall-through.
1428    case Primitive::kPrimShort:    // Fall-through.
1429    case Primitive::kPrimInt:      // Fall-through.
1430      sm_.AdvanceInt(*reinterpret_cast<jint*>(GetParamAddress()));
1431      break;
1432    case Primitive::kPrimVoid:
1433      LOG(FATAL) << "UNREACHABLE";
1434      break;
1435  }
1436}
1437
1438void BuildGenericJniFrameVisitor::FinalizeSirt(Thread* self) {
1439  // Initialize padding entries.
1440  while (sirt_number_of_references_ < sirt_expected_refs_) {
1441    *cur_sirt_entry_ = StackReference<mirror::Object>();
1442    cur_sirt_entry_++;
1443    sirt_number_of_references_++;
1444  }
1445  sirt_->SetNumberOfReferences(sirt_expected_refs_);
1446  DCHECK_NE(sirt_expected_refs_, 0U);
1447  // Install Sirt.
1448  self->PushSirt(sirt_);
1449}
1450
1451extern "C" void* artFindNativeMethod();
1452
1453/*
1454 * Initializes an alloca region assumed to be directly below sp for a native call:
1455 * Create a Sirt and call stack and fill a mini stack with values to be pushed to registers.
1456 * The final element on the stack is a pointer to the native code.
1457 *
1458 * On entry, the stack has a standard callee-save frame above sp, and an alloca below it.
1459 * We need to fix this, as the Sirt needs to go into the callee-save frame.
1460 *
1461 * The return of this function denotes:
1462 * 1) How many bytes of the alloca can be released, if the value is non-negative.
1463 * 2) An error, if the value is negative.
1464 */
1465extern "C" ssize_t artQuickGenericJniTrampoline(Thread* self, mirror::ArtMethod** sp)
1466    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1467  mirror::ArtMethod* called = *sp;
1468  DCHECK(called->IsNative()) << PrettyMethod(called, true);
1469
1470  // run the visitor
1471  MethodHelper mh(called);
1472
1473  BuildGenericJniFrameVisitor visitor(&sp, called->IsStatic(), mh.GetShorty(), mh.GetShortyLength(),
1474                                      self);
1475  visitor.VisitArguments();
1476  visitor.FinalizeSirt(self);
1477
1478  // fix up managed-stack things in Thread
1479  self->SetTopOfStack(sp, 0);
1480
1481  self->VerifyStack();
1482
1483  // Start JNI, save the cookie.
1484  uint32_t cookie;
1485  if (called->IsSynchronized()) {
1486    cookie = JniMethodStartSynchronized(visitor.GetFirstSirtEntry(), self);
1487    if (self->IsExceptionPending()) {
1488      self->PopSirt();
1489      // A negative value denotes an error.
1490      return -1;
1491    }
1492  } else {
1493    cookie = JniMethodStart(self);
1494  }
1495  uint32_t* sp32 = reinterpret_cast<uint32_t*>(sp);
1496  *(sp32 - 1) = cookie;
1497
1498  // Retrieve the stored native code.
1499  const void* nativeCode = called->GetNativeMethod();
1500
1501  // There are two cases for the content of nativeCode:
1502  // 1) Pointer to the native function.
1503  // 2) Pointer to the trampoline for native code binding.
1504  // In the second case, we need to execute the binding and continue with the actual native function
1505  // pointer.
1506  DCHECK(nativeCode != nullptr);
1507  if (nativeCode == GetJniDlsymLookupStub()) {
1508    nativeCode = artFindNativeMethod();
1509
1510    if (nativeCode == nullptr) {
1511      DCHECK(self->IsExceptionPending());    // There should be an exception pending now.
1512      return -1;
1513    }
1514    // Note that the native code pointer will be automatically set by artFindNativeMethod().
1515  }
1516
1517  // Store the native code pointer in the stack at the right location.
1518  uintptr_t* code_pointer = reinterpret_cast<uintptr_t*>(visitor.GetCodeReturn());
1519  *code_pointer = reinterpret_cast<uintptr_t>(nativeCode);
1520
1521  // 5K reserved, window_size + frame pointer used.
1522  size_t window_size = visitor.GetAllocaUsedSize();
1523  return (5 * KB) - window_size - kPointerSize;
1524}
1525
1526/*
1527 * Is called after the native JNI code. Responsible for cleanup (SIRT, saved state) and
1528 * unlocking.
1529 */
1530extern "C" uint64_t artQuickGenericJniEndTrampoline(Thread* self, mirror::ArtMethod** sp,
1531                                                    jvalue result, uint64_t result_f)
1532    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1533  uint32_t* sp32 = reinterpret_cast<uint32_t*>(sp);
1534  mirror::ArtMethod* called = *sp;
1535  uint32_t cookie = *(sp32 - 1);
1536
1537  MethodHelper mh(called);
1538  char return_shorty_char = mh.GetShorty()[0];
1539
1540  if (return_shorty_char == 'L') {
1541    // the only special ending call
1542    if (called->IsSynchronized()) {
1543      StackIndirectReferenceTable* table =
1544          reinterpret_cast<StackIndirectReferenceTable*>(
1545              reinterpret_cast<uint8_t*>(sp) + kPointerSize);
1546      jobject tmp = reinterpret_cast<jobject>(table->GetStackReference(0));
1547
1548      return reinterpret_cast<uint64_t>(JniMethodEndWithReferenceSynchronized(result.l, cookie, tmp,
1549                                                                              self));
1550    } else {
1551      return reinterpret_cast<uint64_t>(JniMethodEndWithReference(result.l, cookie, self));
1552    }
1553  } else {
1554    if (called->IsSynchronized()) {
1555      StackIndirectReferenceTable* table =
1556          reinterpret_cast<StackIndirectReferenceTable*>(
1557              reinterpret_cast<uint8_t*>(sp) + kPointerSize);
1558      jobject tmp = reinterpret_cast<jobject>(table->GetStackReference(0));
1559
1560      JniMethodEndSynchronized(cookie, tmp, self);
1561    } else {
1562      JniMethodEnd(cookie, self);
1563    }
1564
1565    switch (return_shorty_char) {
1566      case 'F':  // Fall-through.
1567      case 'D':
1568        return result_f;
1569      case 'Z':
1570        return result.z;
1571      case 'B':
1572        return result.b;
1573      case 'C':
1574        return result.c;
1575      case 'S':
1576        return result.s;
1577      case 'I':
1578        return result.i;
1579      case 'J':
1580        return result.j;
1581      case 'V':
1582        return 0;
1583      default:
1584        LOG(FATAL) << "Unexpected return shorty character " << return_shorty_char;
1585        return 0;
1586    }
1587  }
1588}
1589
1590}  // namespace art
1591