quick_trampoline_entrypoints.cc revision 36fea8dd490ab6439f391b8cd7f366c59f026fd2
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    Primitive::Type type = GetParamPrimitiveType();
439    switch (type) {
440      case Primitive::kPrimLong:  // Fall-through.
441      case Primitive::kPrimDouble:
442        if (IsSplitLongOrDouble()) {
443          sf_->SetVRegLong(cur_reg_, ReadSplitLongParam());
444        } else {
445          sf_->SetVRegLong(cur_reg_, *reinterpret_cast<jlong*>(GetParamAddress()));
446        }
447        ++cur_reg_;
448        break;
449      case Primitive::kPrimNot: {
450          StackReference<mirror::Object>* stack_ref =
451              reinterpret_cast<StackReference<mirror::Object>*>(GetParamAddress());
452          sf_->SetVRegReference(cur_reg_, stack_ref->AsMirrorPtr());
453        }
454        break;
455      case Primitive::kPrimBoolean:  // Fall-through.
456      case Primitive::kPrimByte:     // Fall-through.
457      case Primitive::kPrimChar:     // Fall-through.
458      case Primitive::kPrimShort:    // Fall-through.
459      case Primitive::kPrimInt:      // Fall-through.
460      case Primitive::kPrimFloat:
461        sf_->SetVReg(cur_reg_, *reinterpret_cast<jint*>(GetParamAddress()));
462        break;
463      case Primitive::kPrimVoid:
464        LOG(FATAL) << "UNREACHABLE";
465        break;
466    }
467    ++cur_reg_;
468  }
469
470 private:
471  ShadowFrame* const sf_;
472  uint32_t cur_reg_;
473
474  DISALLOW_COPY_AND_ASSIGN(BuildQuickShadowFrameVisitor);
475};
476
477extern "C" uint64_t artQuickToInterpreterBridge(mirror::ArtMethod* method, Thread* self,
478                                                mirror::ArtMethod** sp)
479    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
480  // Ensure we don't get thread suspension until the object arguments are safely in the shadow
481  // frame.
482  FinishCalleeSaveFrameSetup(self, sp, Runtime::kRefsAndArgs);
483
484  if (method->IsAbstract()) {
485    ThrowAbstractMethodError(method);
486    return 0;
487  } else {
488    DCHECK(!method->IsNative()) << PrettyMethod(method);
489    const char* old_cause = self->StartAssertNoThreadSuspension("Building interpreter shadow frame");
490    MethodHelper mh(method);
491    const DexFile::CodeItem* code_item = mh.GetCodeItem();
492    DCHECK(code_item != nullptr) << PrettyMethod(method);
493    uint16_t num_regs = code_item->registers_size_;
494    void* memory = alloca(ShadowFrame::ComputeSize(num_regs));
495    ShadowFrame* shadow_frame(ShadowFrame::Create(num_regs, NULL,  // No last shadow coming from quick.
496                                                  method, 0, memory));
497    size_t first_arg_reg = code_item->registers_size_ - code_item->ins_size_;
498    BuildQuickShadowFrameVisitor shadow_frame_builder(sp, mh.IsStatic(), mh.GetShorty(),
499                                                      mh.GetShortyLength(),
500                                                      shadow_frame, first_arg_reg);
501    shadow_frame_builder.VisitArguments();
502    // Push a transition back into managed code onto the linked list in thread.
503    ManagedStack fragment;
504    self->PushManagedStackFragment(&fragment);
505    self->PushShadowFrame(shadow_frame);
506    self->EndAssertNoThreadSuspension(old_cause);
507
508    if (method->IsStatic() && !method->GetDeclaringClass()->IsInitializing()) {
509      // Ensure static method's class is initialized.
510      SirtRef<mirror::Class> sirt_c(self, method->GetDeclaringClass());
511      if (!Runtime::Current()->GetClassLinker()->EnsureInitialized(sirt_c, true, true)) {
512        DCHECK(Thread::Current()->IsExceptionPending()) << PrettyMethod(method);
513        self->PopManagedStackFragment(fragment);
514        return 0;
515      }
516    }
517
518    JValue result = interpreter::EnterInterpreterFromStub(self, mh, code_item, *shadow_frame);
519    // Pop transition.
520    self->PopManagedStackFragment(fragment);
521    // No need to restore the args since the method has already been run by the interpreter.
522    return result.GetJ();
523  }
524}
525
526// Visits arguments on the stack placing them into the args vector, Object* arguments are converted
527// to jobjects.
528class BuildQuickArgumentVisitor FINAL : public QuickArgumentVisitor {
529 public:
530  BuildQuickArgumentVisitor(mirror::ArtMethod** sp, bool is_static, const char* shorty,
531                            uint32_t shorty_len, ScopedObjectAccessUnchecked* soa,
532                            std::vector<jvalue>* args) :
533    QuickArgumentVisitor(sp, is_static, shorty, shorty_len), soa_(soa), args_(args) {}
534
535  void Visit() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) OVERRIDE {
536    jvalue val;
537    Primitive::Type type = GetParamPrimitiveType();
538    switch (type) {
539      case Primitive::kPrimNot: {
540        StackReference<mirror::Object>* stack_ref =
541            reinterpret_cast<StackReference<mirror::Object>*>(GetParamAddress());
542        val.l = soa_->AddLocalReference<jobject>(stack_ref->AsMirrorPtr());
543        references_.push_back(std::make_pair(val.l, stack_ref));
544        break;
545      }
546      case Primitive::kPrimLong:  // Fall-through.
547      case Primitive::kPrimDouble:
548        if (IsSplitLongOrDouble()) {
549          val.j = ReadSplitLongParam();
550        } else {
551          val.j = *reinterpret_cast<jlong*>(GetParamAddress());
552        }
553        break;
554      case Primitive::kPrimBoolean:  // Fall-through.
555      case Primitive::kPrimByte:     // Fall-through.
556      case Primitive::kPrimChar:     // Fall-through.
557      case Primitive::kPrimShort:    // Fall-through.
558      case Primitive::kPrimInt:      // Fall-through.
559      case Primitive::kPrimFloat:
560        val.i = *reinterpret_cast<jint*>(GetParamAddress());
561        break;
562      case Primitive::kPrimVoid:
563        LOG(FATAL) << "UNREACHABLE";
564        val.j = 0;
565        break;
566    }
567    args_->push_back(val);
568  }
569
570  void FixupReferences() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
571    // Fixup any references which may have changed.
572    for (const auto& pair : references_) {
573      pair.second->Assign(soa_->Decode<mirror::Object*>(pair.first));
574    }
575  }
576
577 private:
578  ScopedObjectAccessUnchecked* soa_;
579  std::vector<jvalue>* args_;
580  // References which we must update when exiting in case the GC moved the objects.
581  std::vector<std::pair<jobject, StackReference<mirror::Object>*> > references_;
582  DISALLOW_COPY_AND_ASSIGN(BuildQuickArgumentVisitor);
583};
584
585// Handler for invocation on proxy methods. On entry a frame will exist for the proxy object method
586// which is responsible for recording callee save registers. We explicitly place into jobjects the
587// incoming reference arguments (so they survive GC). We invoke the invocation handler, which is a
588// field within the proxy object, which will box the primitive arguments and deal with error cases.
589extern "C" uint64_t artQuickProxyInvokeHandler(mirror::ArtMethod* proxy_method,
590                                               mirror::Object* receiver,
591                                               Thread* self, mirror::ArtMethod** sp)
592    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
593  DCHECK(proxy_method->IsProxyMethod()) << PrettyMethod(proxy_method);
594  DCHECK(receiver->GetClass()->IsProxyClass()) << PrettyMethod(proxy_method);
595  // Ensure we don't get thread suspension until the object arguments are safely in jobjects.
596  const char* old_cause =
597      self->StartAssertNoThreadSuspension("Adding to IRT proxy object arguments");
598  // Register the top of the managed stack, making stack crawlable.
599  DCHECK_EQ(*sp, proxy_method) << PrettyMethod(proxy_method);
600  self->SetTopOfStack(sp, 0);
601  DCHECK_EQ(proxy_method->GetFrameSizeInBytes(),
602            Runtime::Current()->GetCalleeSaveMethod(Runtime::kRefsAndArgs)->GetFrameSizeInBytes())
603      << PrettyMethod(proxy_method);
604  self->VerifyStack();
605  // Start new JNI local reference state.
606  JNIEnvExt* env = self->GetJniEnv();
607  ScopedObjectAccessUnchecked soa(env);
608  ScopedJniEnvLocalRefState env_state(env);
609  // Create local ref. copies of proxy method and the receiver.
610  jobject rcvr_jobj = soa.AddLocalReference<jobject>(receiver);
611
612  // Placing arguments into args vector and remove the receiver.
613  MethodHelper proxy_mh(proxy_method);
614  DCHECK(!proxy_mh.IsStatic()) << PrettyMethod(proxy_method);
615  std::vector<jvalue> args;
616  BuildQuickArgumentVisitor local_ref_visitor(sp, proxy_mh.IsStatic(), proxy_mh.GetShorty(),
617                                              proxy_mh.GetShortyLength(), &soa, &args);
618
619  local_ref_visitor.VisitArguments();
620  DCHECK_GT(args.size(), 0U) << PrettyMethod(proxy_method);
621  args.erase(args.begin());
622
623  // Convert proxy method into expected interface method.
624  mirror::ArtMethod* interface_method = proxy_method->FindOverriddenMethod();
625  DCHECK(interface_method != NULL) << PrettyMethod(proxy_method);
626  DCHECK(!interface_method->IsProxyMethod()) << PrettyMethod(interface_method);
627  jobject interface_method_jobj = soa.AddLocalReference<jobject>(interface_method);
628
629  // All naked Object*s should now be in jobjects, so its safe to go into the main invoke code
630  // that performs allocations.
631  self->EndAssertNoThreadSuspension(old_cause);
632  JValue result = InvokeProxyInvocationHandler(soa, proxy_mh.GetShorty(),
633                                               rcvr_jobj, interface_method_jobj, args);
634  // Restore references which might have moved.
635  local_ref_visitor.FixupReferences();
636  return result.GetJ();
637}
638
639// Read object references held in arguments from quick frames and place in a JNI local references,
640// so they don't get garbage collected.
641class RememberForGcArgumentVisitor FINAL : public QuickArgumentVisitor {
642 public:
643  RememberForGcArgumentVisitor(mirror::ArtMethod** sp, bool is_static, const char* shorty,
644                               uint32_t shorty_len, ScopedObjectAccessUnchecked* soa) :
645    QuickArgumentVisitor(sp, is_static, shorty, shorty_len), soa_(soa) {}
646
647  void Visit() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) OVERRIDE {
648    if (IsParamAReference()) {
649      StackReference<mirror::Object>* stack_ref =
650          reinterpret_cast<StackReference<mirror::Object>*>(GetParamAddress());
651      jobject reference =
652          soa_->AddLocalReference<jobject>(stack_ref->AsMirrorPtr());
653      references_.push_back(std::make_pair(reference, stack_ref));
654    }
655  }
656
657  void FixupReferences() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
658    // Fixup any references which may have changed.
659    for (const auto& pair : references_) {
660      pair.second->Assign(soa_->Decode<mirror::Object*>(pair.first));
661    }
662  }
663
664 private:
665  ScopedObjectAccessUnchecked* soa_;
666  // References which we must update when exiting in case the GC moved the objects.
667  std::vector<std::pair<jobject, StackReference<mirror::Object>*> > references_;
668  DISALLOW_COPY_AND_ASSIGN(RememberForGcArgumentVisitor);
669};
670
671// Lazily resolve a method for quick. Called by stub code.
672extern "C" const void* artQuickResolutionTrampoline(mirror::ArtMethod* called,
673                                                    mirror::Object* receiver,
674                                                    Thread* self, mirror::ArtMethod** sp)
675    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
676  FinishCalleeSaveFrameSetup(self, sp, Runtime::kRefsAndArgs);
677  // Start new JNI local reference state
678  JNIEnvExt* env = self->GetJniEnv();
679  ScopedObjectAccessUnchecked soa(env);
680  ScopedJniEnvLocalRefState env_state(env);
681  const char* old_cause = self->StartAssertNoThreadSuspension("Quick method resolution set up");
682
683  // Compute details about the called method (avoid GCs)
684  ClassLinker* linker = Runtime::Current()->GetClassLinker();
685  mirror::ArtMethod* caller = QuickArgumentVisitor::GetCallingMethod(sp);
686  InvokeType invoke_type;
687  const DexFile* dex_file;
688  uint32_t dex_method_idx;
689  if (called->IsRuntimeMethod()) {
690    uint32_t dex_pc = caller->ToDexPc(QuickArgumentVisitor::GetCallingPc(sp));
691    const DexFile::CodeItem* code;
692    {
693      MethodHelper mh(caller);
694      dex_file = &mh.GetDexFile();
695      code = mh.GetCodeItem();
696    }
697    CHECK_LT(dex_pc, code->insns_size_in_code_units_);
698    const Instruction* instr = Instruction::At(&code->insns_[dex_pc]);
699    Instruction::Code instr_code = instr->Opcode();
700    bool is_range;
701    switch (instr_code) {
702      case Instruction::INVOKE_DIRECT:
703        invoke_type = kDirect;
704        is_range = false;
705        break;
706      case Instruction::INVOKE_DIRECT_RANGE:
707        invoke_type = kDirect;
708        is_range = true;
709        break;
710      case Instruction::INVOKE_STATIC:
711        invoke_type = kStatic;
712        is_range = false;
713        break;
714      case Instruction::INVOKE_STATIC_RANGE:
715        invoke_type = kStatic;
716        is_range = true;
717        break;
718      case Instruction::INVOKE_SUPER:
719        invoke_type = kSuper;
720        is_range = false;
721        break;
722      case Instruction::INVOKE_SUPER_RANGE:
723        invoke_type = kSuper;
724        is_range = true;
725        break;
726      case Instruction::INVOKE_VIRTUAL:
727        invoke_type = kVirtual;
728        is_range = false;
729        break;
730      case Instruction::INVOKE_VIRTUAL_RANGE:
731        invoke_type = kVirtual;
732        is_range = true;
733        break;
734      case Instruction::INVOKE_INTERFACE:
735        invoke_type = kInterface;
736        is_range = false;
737        break;
738      case Instruction::INVOKE_INTERFACE_RANGE:
739        invoke_type = kInterface;
740        is_range = true;
741        break;
742      default:
743        LOG(FATAL) << "Unexpected call into trampoline: " << instr->DumpString(NULL);
744        // Avoid used uninitialized warnings.
745        invoke_type = kDirect;
746        is_range = false;
747    }
748    dex_method_idx = (is_range) ? instr->VRegB_3rc() : instr->VRegB_35c();
749
750  } else {
751    invoke_type = kStatic;
752    dex_file = &MethodHelper(called).GetDexFile();
753    dex_method_idx = called->GetDexMethodIndex();
754  }
755  uint32_t shorty_len;
756  const char* shorty =
757      dex_file->GetMethodShorty(dex_file->GetMethodId(dex_method_idx), &shorty_len);
758  RememberForGcArgumentVisitor visitor(sp, invoke_type == kStatic, shorty, shorty_len, &soa);
759  visitor.VisitArguments();
760  self->EndAssertNoThreadSuspension(old_cause);
761  bool virtual_or_interface = invoke_type == kVirtual || invoke_type == kInterface;
762  // Resolve method filling in dex cache.
763  if (called->IsRuntimeMethod()) {
764    SirtRef<mirror::Object> sirt_receiver(soa.Self(), virtual_or_interface ? receiver : nullptr);
765    called = linker->ResolveMethod(dex_method_idx, caller, invoke_type);
766    receiver = sirt_receiver.get();
767  }
768  const void* code = NULL;
769  if (LIKELY(!self->IsExceptionPending())) {
770    // Incompatible class change should have been handled in resolve method.
771    CHECK(!called->CheckIncompatibleClassChange(invoke_type))
772        << PrettyMethod(called) << " " << invoke_type;
773    if (virtual_or_interface) {
774      // Refine called method based on receiver.
775      CHECK(receiver != nullptr) << invoke_type;
776      if (invoke_type == kVirtual) {
777        called = receiver->GetClass()->FindVirtualMethodForVirtual(called);
778      } else {
779        called = receiver->GetClass()->FindVirtualMethodForInterface(called);
780      }
781      // We came here because of sharpening. Ensure the dex cache is up-to-date on the method index
782      // of the sharpened method.
783      if (called->GetDexCacheResolvedMethods() == caller->GetDexCacheResolvedMethods()) {
784        caller->GetDexCacheResolvedMethods()->Set<false>(called->GetDexMethodIndex(), called);
785      } else {
786        // Calling from one dex file to another, need to compute the method index appropriate to
787        // the caller's dex file. Since we get here only if the original called was a runtime
788        // method, we've got the correct dex_file and a dex_method_idx from above.
789        DCHECK(&MethodHelper(caller).GetDexFile() == dex_file);
790        uint32_t method_index =
791            MethodHelper(called).FindDexMethodIndexInOtherDexFile(*dex_file, dex_method_idx);
792        if (method_index != DexFile::kDexNoIndex) {
793          caller->GetDexCacheResolvedMethods()->Set<false>(method_index, called);
794        }
795      }
796    }
797    // Ensure that the called method's class is initialized.
798    SirtRef<mirror::Class> called_class(soa.Self(), called->GetDeclaringClass());
799    linker->EnsureInitialized(called_class, true, true);
800    if (LIKELY(called_class->IsInitialized())) {
801      code = called->GetEntryPointFromQuickCompiledCode();
802    } else if (called_class->IsInitializing()) {
803      if (invoke_type == kStatic) {
804        // Class is still initializing, go to oat and grab code (trampoline must be left in place
805        // until class is initialized to stop races between threads).
806        code = linker->GetQuickOatCodeFor(called);
807      } else {
808        // No trampoline for non-static methods.
809        code = called->GetEntryPointFromQuickCompiledCode();
810      }
811    } else {
812      DCHECK(called_class->IsErroneous());
813    }
814  }
815  CHECK_EQ(code == NULL, self->IsExceptionPending());
816  // Fixup any locally saved objects may have moved during a GC.
817  visitor.FixupReferences();
818  // Place called method in callee-save frame to be placed as first argument to quick method.
819  *sp = called;
820  return code;
821}
822
823
824
825/*
826 * This class uses a couple of observations to unite the different calling conventions through
827 * a few constants.
828 *
829 * 1) Number of registers used for passing is normally even, so counting down has no penalty for
830 *    possible alignment.
831 * 2) Known 64b architectures store 8B units on the stack, both for integral and floating point
832 *    types, so using uintptr_t is OK. Also means that we can use kRegistersNeededX to denote
833 *    when we have to split things
834 * 3) The only soft-float, Arm, is 32b, so no widening needs to be taken into account for floats
835 *    and we can use Int handling directly.
836 * 4) Only 64b architectures widen, and their stack is aligned 8B anyways, so no padding code
837 *    necessary when widening. Also, widening of Ints will take place implicitly, and the
838 *    extension should be compatible with Aarch64, which mandates copying the available bits
839 *    into LSB and leaving the rest unspecified.
840 * 5) Aligning longs and doubles is necessary on arm only, and it's the same in registers and on
841 *    the stack.
842 * 6) There is only little endian.
843 *
844 *
845 * Actual work is supposed to be done in a delegate of the template type. The interface is as
846 * follows:
847 *
848 * void PushGpr(uintptr_t):   Add a value for the next GPR
849 *
850 * void PushFpr4(float):      Add a value for the next FPR of size 32b. Is only called if we need
851 *                            padding, that is, think the architecture is 32b and aligns 64b.
852 *
853 * void PushFpr8(uint64_t):   Push a double. We _will_ call this on 32b, it's the callee's job to
854 *                            split this if necessary. The current state will have aligned, if
855 *                            necessary.
856 *
857 * void PushStack(uintptr_t): Push a value to the stack.
858 *
859 * uintptr_t PushSirt(mirror::Object* ref): Add a reference to the Sirt. This _will_ have nullptr,
860 *                                          as this might be important for null initialization.
861 *                                          Must return the jobject, that is, the reference to the
862 *                                          entry in the Sirt (nullptr if necessary).
863 *
864 */
865template <class T> class BuildGenericJniFrameStateMachine {
866 public:
867#if defined(__arm__)
868  // TODO: These are all dummy values!
869  static constexpr bool kNativeSoftFloatAbi = true;
870  static constexpr size_t kNumNativeGprArgs = 4;  // 4 arguments passed in GPRs, r0-r3
871  static constexpr size_t kNumNativeFprArgs = 0;  // 0 arguments passed in FPRs.
872
873  static constexpr size_t kRegistersNeededForLong = 2;
874  static constexpr size_t kRegistersNeededForDouble = 2;
875  static constexpr bool kMultiRegistersAligned = true;
876  static constexpr bool kMultiRegistersWidened = false;
877  static constexpr bool kAlignLongOnStack = true;
878  static constexpr bool kAlignDoubleOnStack = true;
879#elif defined(__mips__)
880  // TODO: These are all dummy values!
881  static constexpr bool kNativeSoftFloatAbi = true;  // This is a hard float ABI.
882  static constexpr size_t kNumNativeGprArgs = 0;  // 6 arguments passed in GPRs.
883  static constexpr size_t kNumNativeFprArgs = 0;  // 8 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 = true;
889  static constexpr bool kAlignLongOnStack = false;
890  static constexpr bool kAlignDoubleOnStack = false;
891#elif defined(__i386__)
892  // TODO: Check these!
893  static constexpr bool kNativeSoftFloatAbi = false;  // Not using int registers for fp
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 = false;       // x86 not using regs, anyways
900  static constexpr bool kMultiRegistersWidened = false;
901  static constexpr bool kAlignLongOnStack = false;
902  static constexpr bool kAlignDoubleOnStack = false;
903#elif defined(__x86_64__)
904  static constexpr bool kNativeSoftFloatAbi = false;  // This is a hard float ABI.
905  static constexpr size_t kNumNativeGprArgs = 6;  // 6 arguments passed in GPRs.
906  static constexpr size_t kNumNativeFprArgs = 8;  // 8 arguments passed in FPRs.
907
908  static constexpr size_t kRegistersNeededForLong = 1;
909  static constexpr size_t kRegistersNeededForDouble = 1;
910  static constexpr bool kMultiRegistersAligned = false;
911  static constexpr bool kMultiRegistersWidened = false;
912  static constexpr bool kAlignLongOnStack = false;
913  static constexpr bool kAlignDoubleOnStack = false;
914#else
915#error "Unsupported architecture"
916#endif
917
918 public:
919  explicit BuildGenericJniFrameStateMachine(T* delegate) : gpr_index_(kNumNativeGprArgs),
920                                                           fpr_index_(kNumNativeFprArgs),
921                                                           stack_entries_(0),
922                                                           delegate_(delegate) {
923    // For register alignment, we want to assume that counters (gpr_index_, fpr_index_) are even iff
924    // the next register is even; counting down is just to make the compiler happy...
925    CHECK_EQ(kNumNativeGprArgs % 2, 0U);
926    CHECK_EQ(kNumNativeFprArgs % 2, 0U);
927  }
928
929  virtual ~BuildGenericJniFrameStateMachine() {}
930
931  bool HavePointerGpr() {
932    return gpr_index_ > 0;
933  }
934
935  void AdvancePointer(void* val) {
936    if (HavePointerGpr()) {
937      gpr_index_--;
938      PushGpr(reinterpret_cast<uintptr_t>(val));
939    } else {
940      stack_entries_++;         // TODO: have a field for pointer length as multiple of 32b
941      PushStack(reinterpret_cast<uintptr_t>(val));
942      gpr_index_ = 0;
943    }
944  }
945
946
947  bool HaveSirtGpr() {
948    return gpr_index_ > 0;
949  }
950
951  void AdvanceSirt(mirror::Object* ptr) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
952    uintptr_t sirtRef = PushSirt(ptr);
953    if (HaveSirtGpr()) {
954      gpr_index_--;
955      PushGpr(sirtRef);
956    } else {
957      stack_entries_++;
958      PushStack(sirtRef);
959      gpr_index_ = 0;
960    }
961  }
962
963
964  bool HaveIntGpr() {
965    return gpr_index_ > 0;
966  }
967
968  void AdvanceInt(uint32_t val) {
969    if (HaveIntGpr()) {
970      gpr_index_--;
971      PushGpr(val);
972    } else {
973      stack_entries_++;
974      PushStack(val);
975      gpr_index_ = 0;
976    }
977  }
978
979
980  bool HaveLongGpr() {
981    return gpr_index_ >= kRegistersNeededForLong + (LongGprNeedsPadding() ? 1 : 0);
982  }
983
984  bool LongGprNeedsPadding() {
985    return kRegistersNeededForLong > 1 &&     // only pad when using multiple registers
986        kAlignLongOnStack &&                  // and when it needs alignment
987        (gpr_index_ & 1) == 1;                // counter is odd, see constructor
988  }
989
990  bool LongStackNeedsPadding() {
991    return kRegistersNeededForLong > 1 &&     // only pad when using multiple registers
992        kAlignLongOnStack &&                  // and when it needs 8B alignment
993        (stack_entries_ & 1) == 1;            // counter is odd
994  }
995
996  void AdvanceLong(uint64_t val) {
997    if (HaveLongGpr()) {
998      if (LongGprNeedsPadding()) {
999        PushGpr(0);
1000        gpr_index_--;
1001      }
1002      if (kRegistersNeededForLong == 1) {
1003        PushGpr(static_cast<uintptr_t>(val));
1004      } else {
1005        PushGpr(static_cast<uintptr_t>(val & 0xFFFFFFFF));
1006        PushGpr(static_cast<uintptr_t>((val >> 32) & 0xFFFFFFFF));
1007      }
1008      gpr_index_ -= kRegistersNeededForLong;
1009    } else {
1010      if (LongStackNeedsPadding()) {
1011        PushStack(0);
1012        stack_entries_++;
1013      }
1014      if (kRegistersNeededForLong == 1) {
1015        PushStack(static_cast<uintptr_t>(val));
1016        stack_entries_++;
1017      } else {
1018        PushStack(static_cast<uintptr_t>(val & 0xFFFFFFFF));
1019        PushStack(static_cast<uintptr_t>((val >> 32) & 0xFFFFFFFF));
1020        stack_entries_ += 2;
1021      }
1022      gpr_index_ = 0;
1023    }
1024  }
1025
1026
1027  bool HaveFloatFpr() {
1028    return fpr_index_ > 0;
1029  }
1030
1031  // TODO: please review this bit representation retrieving.
1032  template <typename U, typename V> V convert(U in) {
1033    CHECK_LE(sizeof(U), sizeof(V));
1034    union { U u; V v; } tmp;
1035    tmp.u = in;
1036    return tmp.v;
1037  }
1038
1039  void AdvanceFloat(float val) {
1040    if (kNativeSoftFloatAbi) {
1041      AdvanceInt(convert<float, uint32_t>(val));
1042    } else {
1043      if (HaveFloatFpr()) {
1044        fpr_index_--;
1045        if (kRegistersNeededForDouble == 1) {
1046          if (kMultiRegistersWidened) {
1047            PushFpr8(convert<double, uint64_t>(val));
1048          } else {
1049            // No widening, just use the bits.
1050            PushFpr8(convert<float, uint64_t>(val));
1051          }
1052        } else {
1053          PushFpr4(val);
1054        }
1055      } else {
1056        stack_entries_++;
1057        if (kRegistersNeededForDouble == 1 && kMultiRegistersWidened) {
1058          // Need to widen before storing: Note the "double" in the template instantiation.
1059          PushStack(convert<double, uintptr_t>(val));
1060        } else {
1061          PushStack(convert<float, uintptr_t>(val));
1062        }
1063        fpr_index_ = 0;
1064      }
1065    }
1066  }
1067
1068
1069  bool HaveDoubleFpr() {
1070    return fpr_index_ >= kRegistersNeededForDouble + (DoubleFprNeedsPadding() ? 1 : 0);
1071  }
1072
1073  bool DoubleFprNeedsPadding() {
1074    return kRegistersNeededForDouble > 1 &&     // only pad when using multiple registers
1075        kAlignDoubleOnStack &&                  // and when it needs alignment
1076        (fpr_index_ & 1) == 1;                  // counter is odd, see constructor
1077  }
1078
1079  bool DoubleStackNeedsPadding() {
1080    return kRegistersNeededForDouble > 1 &&     // only pad when using multiple registers
1081        kAlignDoubleOnStack &&                  // and when it needs 8B alignment
1082        (stack_entries_ & 1) == 1;              // counter is odd
1083  }
1084
1085  void AdvanceDouble(uint64_t val) {
1086    if (kNativeSoftFloatAbi) {
1087      AdvanceLong(val);
1088    } else {
1089      if (HaveDoubleFpr()) {
1090        if (DoubleFprNeedsPadding()) {
1091          PushFpr4(0);
1092          fpr_index_--;
1093        }
1094        PushFpr8(val);
1095        fpr_index_ -= kRegistersNeededForDouble;
1096      } else {
1097        if (DoubleStackNeedsPadding()) {
1098          PushStack(0);
1099          stack_entries_++;
1100        }
1101        if (kRegistersNeededForDouble == 1) {
1102          PushStack(static_cast<uintptr_t>(val));
1103          stack_entries_++;
1104        } else {
1105          PushStack(static_cast<uintptr_t>(val & 0xFFFFFFFF));
1106          PushStack(static_cast<uintptr_t>((val >> 32) & 0xFFFFFFFF));
1107          stack_entries_ += 2;
1108        }
1109        fpr_index_ = 0;
1110      }
1111    }
1112  }
1113
1114  uint32_t getStackEntries() {
1115    return stack_entries_;
1116  }
1117
1118  uint32_t getNumberOfUsedGprs() {
1119    return kNumNativeGprArgs - gpr_index_;
1120  }
1121
1122  uint32_t getNumberOfUsedFprs() {
1123    return kNumNativeFprArgs - fpr_index_;
1124  }
1125
1126 private:
1127  void PushGpr(uintptr_t val) {
1128    delegate_->PushGpr(val);
1129  }
1130  void PushFpr4(float val) {
1131    delegate_->PushFpr4(val);
1132  }
1133  void PushFpr8(uint64_t val) {
1134    delegate_->PushFpr8(val);
1135  }
1136  void PushStack(uintptr_t val) {
1137    delegate_->PushStack(val);
1138  }
1139  uintptr_t PushSirt(mirror::Object* ref) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1140    return delegate_->PushSirt(ref);
1141  }
1142
1143  uint32_t gpr_index_;      // Number of free GPRs
1144  uint32_t fpr_index_;      // Number of free FPRs
1145  uint32_t stack_entries_;  // Stack entries are in multiples of 32b, as floats are usually not
1146                            // extended
1147  T* delegate_;             // What Push implementation gets called
1148};
1149
1150class ComputeGenericJniFrameSize FINAL {
1151 public:
1152  ComputeGenericJniFrameSize() : num_sirt_references_(0), num_stack_entries_(0) {}
1153
1154  uint32_t GetStackSize() {
1155    return num_stack_entries_ * sizeof(uintptr_t);
1156  }
1157
1158  // WARNING: After this, *sp won't be pointing to the method anymore!
1159  void ComputeLayout(mirror::ArtMethod*** m, bool is_static, const char* shorty, uint32_t shorty_len,
1160                     void* sp, StackIndirectReferenceTable** table, uint32_t* sirt_entries,
1161                     uintptr_t** start_stack, uintptr_t** start_gpr, uint32_t** start_fpr,
1162                     void** code_return, size_t* overall_size)
1163      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1164    ComputeAll(is_static, shorty, shorty_len);
1165
1166    mirror::ArtMethod* method = **m;
1167
1168    uint8_t* sp8 = reinterpret_cast<uint8_t*>(sp);
1169
1170    // First, fix up the layout of the callee-save frame.
1171    // We have to squeeze in the Sirt, and relocate the method pointer.
1172
1173    // "Free" the slot for the method.
1174    sp8 += kPointerSize;
1175
1176    // Add the Sirt.
1177    *sirt_entries = num_sirt_references_;
1178    size_t sirt_size = StackIndirectReferenceTable::GetAlignedSirtSize(num_sirt_references_);
1179    sp8 -= sirt_size;
1180    *table = reinterpret_cast<StackIndirectReferenceTable*>(sp8);
1181    (*table)->SetNumberOfReferences(num_sirt_references_);
1182
1183    // Add a slot for the method pointer, and fill it. Fix the pointer-pointer given to us.
1184    sp8 -= kPointerSize;
1185    uint8_t* method_pointer = sp8;
1186    *(reinterpret_cast<mirror::ArtMethod**>(method_pointer)) = method;
1187    *m = reinterpret_cast<mirror::ArtMethod**>(method_pointer);
1188
1189    // Reference cookie and padding
1190    sp8 -= 8;
1191    // Store Sirt size
1192    *reinterpret_cast<uint32_t*>(sp8) = static_cast<uint32_t>(sirt_size & 0xFFFFFFFF);
1193
1194    // Next comes the native call stack.
1195    sp8 -= GetStackSize();
1196    // Now align the call stack below. This aligns by 16, as AArch64 seems to require.
1197    uintptr_t mask = ~0x0F;
1198    sp8 = reinterpret_cast<uint8_t*>(reinterpret_cast<uintptr_t>(sp8) & mask);
1199    *start_stack = reinterpret_cast<uintptr_t*>(sp8);
1200
1201    // put fprs and gprs below
1202    // Assumption is OK right now, as we have soft-float arm
1203    size_t fregs = BuildGenericJniFrameStateMachine<ComputeGenericJniFrameSize>::kNumNativeFprArgs;
1204    sp8 -= fregs * sizeof(uintptr_t);
1205    *start_fpr = reinterpret_cast<uint32_t*>(sp8);
1206    size_t iregs = BuildGenericJniFrameStateMachine<ComputeGenericJniFrameSize>::kNumNativeGprArgs;
1207    sp8 -= iregs * sizeof(uintptr_t);
1208    *start_gpr = reinterpret_cast<uintptr_t*>(sp8);
1209
1210    // reserve space for the code pointer
1211    sp8 -= kPointerSize;
1212    *code_return = reinterpret_cast<void*>(sp8);
1213
1214    *overall_size = reinterpret_cast<uint8_t*>(sp) - sp8;
1215
1216    // The new SP is stored at the end of the alloca, so it can be immediately popped
1217    sp8 = reinterpret_cast<uint8_t*>(sp) - 5 * KB;
1218    *(reinterpret_cast<uint8_t**>(sp8)) = method_pointer;
1219  }
1220
1221  void ComputeSirtOffset() { }  // nothing to do, static right now
1222
1223  void ComputeAll(bool is_static, const char* shorty, uint32_t shorty_len)
1224      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1225    BuildGenericJniFrameStateMachine<ComputeGenericJniFrameSize> sm(this);
1226
1227    // JNIEnv
1228    sm.AdvancePointer(nullptr);
1229
1230    // Class object or this as first argument
1231    sm.AdvanceSirt(reinterpret_cast<mirror::Object*>(0x12345678));
1232
1233    for (uint32_t i = 1; i < shorty_len; ++i) {
1234      Primitive::Type cur_type_ = Primitive::GetType(shorty[i]);
1235      switch (cur_type_) {
1236        case Primitive::kPrimNot:
1237          sm.AdvanceSirt(reinterpret_cast<mirror::Object*>(0x12345678));
1238          break;
1239
1240        case Primitive::kPrimBoolean:
1241        case Primitive::kPrimByte:
1242        case Primitive::kPrimChar:
1243        case Primitive::kPrimShort:
1244        case Primitive::kPrimInt:
1245          sm.AdvanceInt(0);
1246          break;
1247        case Primitive::kPrimFloat:
1248          sm.AdvanceFloat(0);
1249          break;
1250        case Primitive::kPrimDouble:
1251          sm.AdvanceDouble(0);
1252          break;
1253        case Primitive::kPrimLong:
1254          sm.AdvanceLong(0);
1255          break;
1256        default:
1257          LOG(FATAL) << "Unexpected type: " << cur_type_ << " in " << shorty;
1258      }
1259    }
1260
1261    num_stack_entries_ = sm.getStackEntries();
1262  }
1263
1264  void PushGpr(uintptr_t /* val */) {
1265    // not optimizing registers, yet
1266  }
1267
1268  void PushFpr4(float /* val */) {
1269    // not optimizing registers, yet
1270  }
1271
1272  void PushFpr8(uint64_t /* val */) {
1273    // not optimizing registers, yet
1274  }
1275
1276  void PushStack(uintptr_t /* val */) {
1277    // counting is already done in the superclass
1278  }
1279
1280  uintptr_t PushSirt(mirror::Object* /* ptr */) {
1281    num_sirt_references_++;
1282    return reinterpret_cast<uintptr_t>(nullptr);
1283  }
1284
1285 private:
1286  uint32_t num_sirt_references_;
1287  uint32_t num_stack_entries_;
1288};
1289
1290// Visits arguments on the stack placing them into a region lower down the stack for the benefit
1291// of transitioning into native code.
1292class BuildGenericJniFrameVisitor FINAL : public QuickArgumentVisitor {
1293 public:
1294  BuildGenericJniFrameVisitor(mirror::ArtMethod*** sp, bool is_static, const char* shorty,
1295                              uint32_t shorty_len, Thread* self) :
1296      QuickArgumentVisitor(*sp, is_static, shorty, shorty_len), sm_(this) {
1297    ComputeGenericJniFrameSize fsc;
1298    fsc.ComputeLayout(sp, is_static, shorty, shorty_len, *sp, &sirt_, &sirt_expected_refs_,
1299                      &cur_stack_arg_, &cur_gpr_reg_, &cur_fpr_reg_, &code_return_,
1300                      &alloca_used_size_);
1301    sirt_number_of_references_ = 0;
1302    cur_sirt_entry_ = reinterpret_cast<StackReference<mirror::Object>*>(GetFirstSirtEntry());
1303
1304    // jni environment is always first argument
1305    sm_.AdvancePointer(self->GetJniEnv());
1306
1307    if (is_static) {
1308      sm_.AdvanceSirt((**sp)->GetDeclaringClass());
1309    }
1310  }
1311
1312  void Visit() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) OVERRIDE {
1313    Primitive::Type type = GetParamPrimitiveType();
1314    switch (type) {
1315      case Primitive::kPrimLong: {
1316        jlong long_arg;
1317        if (IsSplitLongOrDouble()) {
1318          long_arg = ReadSplitLongParam();
1319        } else {
1320          long_arg = *reinterpret_cast<jlong*>(GetParamAddress());
1321        }
1322        sm_.AdvanceLong(long_arg);
1323        break;
1324      }
1325      case Primitive::kPrimDouble: {
1326        uint64_t double_arg;
1327        if (IsSplitLongOrDouble()) {
1328          // Read into union so that we don't case to a double.
1329          double_arg = ReadSplitLongParam();
1330        } else {
1331          double_arg = *reinterpret_cast<uint64_t*>(GetParamAddress());
1332        }
1333        sm_.AdvanceDouble(double_arg);
1334        break;
1335      }
1336      case Primitive::kPrimNot: {
1337        StackReference<mirror::Object>* stack_ref =
1338            reinterpret_cast<StackReference<mirror::Object>*>(GetParamAddress());
1339        sm_.AdvanceSirt(stack_ref->AsMirrorPtr());
1340        break;
1341      }
1342      case Primitive::kPrimFloat:
1343        sm_.AdvanceFloat(*reinterpret_cast<float*>(GetParamAddress()));
1344        break;
1345      case Primitive::kPrimBoolean:  // Fall-through.
1346      case Primitive::kPrimByte:     // Fall-through.
1347      case Primitive::kPrimChar:     // Fall-through.
1348      case Primitive::kPrimShort:    // Fall-through.
1349      case Primitive::kPrimInt:      // Fall-through.
1350        sm_.AdvanceInt(*reinterpret_cast<jint*>(GetParamAddress()));
1351        break;
1352      case Primitive::kPrimVoid:
1353        LOG(FATAL) << "UNREACHABLE";
1354        break;
1355    }
1356  }
1357
1358  void FinalizeSirt(Thread* self) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1359    // Initialize padding entries.
1360    while (sirt_number_of_references_ < sirt_expected_refs_) {
1361      *cur_sirt_entry_ = StackReference<mirror::Object>();
1362      cur_sirt_entry_++;
1363      sirt_number_of_references_++;
1364    }
1365    sirt_->SetNumberOfReferences(sirt_expected_refs_);
1366    DCHECK_NE(sirt_expected_refs_, 0U);
1367    // Install Sirt.
1368    self->PushSirt(sirt_);
1369  }
1370
1371  jobject GetFirstSirtEntry() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1372    return reinterpret_cast<jobject>(sirt_->GetStackReference(0));
1373  }
1374
1375  void PushGpr(uintptr_t val) {
1376    *cur_gpr_reg_ = val;
1377    cur_gpr_reg_++;
1378  }
1379
1380  void PushFpr4(float val) {
1381    *cur_fpr_reg_ = val;
1382    cur_fpr_reg_++;
1383  }
1384
1385  void PushFpr8(uint64_t val) {
1386    uint64_t* tmp = reinterpret_cast<uint64_t*>(cur_fpr_reg_);
1387    *tmp = val;
1388    cur_fpr_reg_ += 2;
1389  }
1390
1391  void PushStack(uintptr_t val) {
1392    *cur_stack_arg_ = val;
1393    cur_stack_arg_++;
1394  }
1395
1396  uintptr_t PushSirt(mirror::Object* ref) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1397    uintptr_t tmp;
1398    if (ref == nullptr) {
1399      *cur_sirt_entry_ = StackReference<mirror::Object>();
1400      tmp = reinterpret_cast<uintptr_t>(nullptr);
1401    } else {
1402      *cur_sirt_entry_ = StackReference<mirror::Object>::FromMirrorPtr(ref);
1403      tmp = reinterpret_cast<uintptr_t>(cur_sirt_entry_);
1404    }
1405    cur_sirt_entry_++;
1406    sirt_number_of_references_++;
1407    return tmp;
1408  }
1409
1410  // Size of the part of the alloca that we actually need.
1411  size_t GetAllocaUsedSize() {
1412    return alloca_used_size_;
1413  }
1414
1415  void* GetCodeReturn() {
1416    return code_return_;
1417  }
1418
1419 private:
1420  uint32_t sirt_number_of_references_;
1421  StackReference<mirror::Object>* cur_sirt_entry_;
1422  StackIndirectReferenceTable* sirt_;
1423  uint32_t sirt_expected_refs_;
1424  uintptr_t* cur_gpr_reg_;
1425  uint32_t* cur_fpr_reg_;
1426  uintptr_t* cur_stack_arg_;
1427  // StackReference<mirror::Object>* top_of_sirt_;
1428  void* code_return_;
1429  size_t alloca_used_size_;
1430
1431  BuildGenericJniFrameStateMachine<BuildGenericJniFrameVisitor> sm_;
1432
1433  DISALLOW_COPY_AND_ASSIGN(BuildGenericJniFrameVisitor);
1434};
1435
1436/*
1437 * Initializes an alloca region assumed to be directly below sp for a native call:
1438 * Create a Sirt and call stack and fill a mini stack with values to be pushed to registers.
1439 * The final element on the stack is a pointer to the native code.
1440 *
1441 * On entry, the stack has a standard callee-save frame above sp, and an alloca below it.
1442 * We need to fix this, as the Sirt needs to go into the callee-save frame.
1443 *
1444 * The return of this function denotes:
1445 * 1) How many bytes of the alloca can be released, if the value is non-negative.
1446 * 2) An error, if the value is negative.
1447 */
1448extern "C" ssize_t artQuickGenericJniTrampoline(Thread* self, mirror::ArtMethod** sp)
1449    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1450  mirror::ArtMethod* called = *sp;
1451  DCHECK(called->IsNative()) << PrettyMethod(called, true);
1452
1453  // run the visitor
1454  MethodHelper mh(called);
1455
1456  BuildGenericJniFrameVisitor visitor(&sp, called->IsStatic(), mh.GetShorty(), mh.GetShortyLength(),
1457                                      self);
1458  visitor.VisitArguments();
1459  visitor.FinalizeSirt(self);
1460
1461  // fix up managed-stack things in Thread
1462  self->SetTopOfStack(sp, 0);
1463
1464  self->VerifyStack();
1465
1466  // start JNI, save the cookie
1467  uint32_t cookie;
1468  if (called->IsSynchronized()) {
1469    cookie = JniMethodStartSynchronized(visitor.GetFirstSirtEntry(), self);
1470    if (self->IsExceptionPending()) {
1471      self->PopSirt();
1472      // A negative value denotes an error.
1473      // TODO: Do we still need to fix the stack pointer? I think so. Then it's necessary to push
1474      //       that value!
1475      return -1;
1476    }
1477  } else {
1478    cookie = JniMethodStart(self);
1479  }
1480  uint32_t* sp32 = reinterpret_cast<uint32_t*>(sp);
1481  *(sp32 - 1) = cookie;
1482
1483  // retrieve native code
1484  const void* nativeCode = called->GetNativeMethod();
1485  if (nativeCode == nullptr) {
1486    // TODO: is this really an error, or do we need to try to find native code?
1487    LOG(FATAL) << "Finding native code not implemented yet.";
1488  }
1489
1490  uintptr_t* code_pointer = reinterpret_cast<uintptr_t*>(visitor.GetCodeReturn());
1491  size_t window_size = visitor.GetAllocaUsedSize();
1492  *code_pointer = reinterpret_cast<uintptr_t>(nativeCode);
1493
1494  // 5K reserved, window_size + frame pointer used.
1495  return (5 * KB) - window_size - kPointerSize;
1496}
1497
1498/*
1499 * Is called after the native JNI code. Responsible for cleanup (SIRT, saved state) and
1500 * unlocking.
1501 */
1502extern "C" uint64_t artQuickGenericJniEndTrampoline(Thread* self, mirror::ArtMethod** sp,
1503                                                    jvalue result, uint64_t result_f)
1504    SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
1505  uint32_t* sp32 = reinterpret_cast<uint32_t*>(sp);
1506  mirror::ArtMethod* called = *sp;
1507  uint32_t cookie = *(sp32 - 1);
1508
1509  MethodHelper mh(called);
1510  char return_shorty_char = mh.GetShorty()[0];
1511
1512  if (return_shorty_char == 'L') {
1513    // the only special ending call
1514    if (called->IsSynchronized()) {
1515      StackIndirectReferenceTable* table =
1516          reinterpret_cast<StackIndirectReferenceTable*>(
1517              reinterpret_cast<uint8_t*>(sp) + kPointerSize);
1518      jobject tmp = reinterpret_cast<jobject>(table->GetStackReference(0));
1519
1520      return reinterpret_cast<uint64_t>(JniMethodEndWithReferenceSynchronized(result.l, cookie, tmp,
1521                                                                              self));
1522    } else {
1523      return reinterpret_cast<uint64_t>(JniMethodEndWithReference(result.l, cookie, self));
1524    }
1525  } else {
1526    if (called->IsSynchronized()) {
1527      StackIndirectReferenceTable* table =
1528          reinterpret_cast<StackIndirectReferenceTable*>(
1529              reinterpret_cast<uint8_t*>(sp) + kPointerSize);
1530      jobject tmp = reinterpret_cast<jobject>(table->GetStackReference(0));
1531
1532      JniMethodEndSynchronized(cookie, tmp, self);
1533    } else {
1534      JniMethodEnd(cookie, self);
1535    }
1536
1537    switch (return_shorty_char) {
1538      case 'F':  // Fall-through.
1539      case 'D':
1540        return result_f;
1541      case 'Z':
1542        return result.z;
1543      case 'B':
1544        return result.b;
1545      case 'C':
1546        return result.c;
1547      case 'S':
1548        return result.s;
1549      case 'I':
1550        return result.i;
1551      case 'J':
1552        return result.j;
1553      case 'V':
1554        return 0;
1555      default:
1556        LOG(FATAL) << "Unexpected return shorty character " << return_shorty_char;
1557        return 0;
1558    }
1559  }
1560}
1561
1562}  // namespace art
1563