1// Copyright 2011 the V8 project authors. All rights reserved.
2// Use of this source code is governed by a BSD-style license that can be
3// found in the LICENSE file.
4
5#include "src/codegen.h"
6#include "src/deoptimizer.h"
7#include "src/full-codegen/full-codegen.h"
8#include "src/register-configuration.h"
9#include "src/safepoint-table.h"
10
11namespace v8 {
12namespace internal {
13
14
15int Deoptimizer::patch_size() {
16  const int kCallInstructionSizeInWords = 6;
17  return kCallInstructionSizeInWords * Assembler::kInstrSize;
18}
19
20
21void Deoptimizer::EnsureRelocSpaceForLazyDeoptimization(Handle<Code> code) {
22  // Empty because there is no need for relocation information for the code
23  // patching in Deoptimizer::PatchCodeForDeoptimization below.
24}
25
26
27void Deoptimizer::PatchCodeForDeoptimization(Isolate* isolate, Code* code) {
28  Address code_start_address = code->instruction_start();
29  // Invalidate the relocation information, as it will become invalid by the
30  // code patching below, and is not needed any more.
31  code->InvalidateRelocation();
32
33  if (FLAG_zap_code_space) {
34    // Fail hard and early if we enter this code object again.
35    byte* pointer = code->FindCodeAgeSequence();
36    if (pointer != NULL) {
37      pointer += kNoCodeAgeSequenceLength;
38    } else {
39      pointer = code->instruction_start();
40    }
41    CodePatcher patcher(isolate, pointer, 1);
42    patcher.masm()->break_(0xCC);
43
44    DeoptimizationInputData* data =
45        DeoptimizationInputData::cast(code->deoptimization_data());
46    int osr_offset = data->OsrPcOffset()->value();
47    if (osr_offset > 0) {
48      CodePatcher osr_patcher(isolate, code->instruction_start() + osr_offset,
49                              1);
50      osr_patcher.masm()->break_(0xCC);
51    }
52  }
53
54  DeoptimizationInputData* deopt_data =
55      DeoptimizationInputData::cast(code->deoptimization_data());
56#ifdef DEBUG
57  Address prev_call_address = NULL;
58#endif
59  // For each LLazyBailout instruction insert a call to the corresponding
60  // deoptimization entry.
61  for (int i = 0; i < deopt_data->DeoptCount(); i++) {
62    if (deopt_data->Pc(i)->value() == -1) continue;
63    Address call_address = code_start_address + deopt_data->Pc(i)->value();
64    Address deopt_entry = GetDeoptimizationEntry(isolate, i, LAZY);
65    int call_size_in_bytes = MacroAssembler::CallSize(deopt_entry,
66                                                      RelocInfo::NONE32);
67    int call_size_in_words = call_size_in_bytes / Assembler::kInstrSize;
68    DCHECK(call_size_in_bytes % Assembler::kInstrSize == 0);
69    DCHECK(call_size_in_bytes <= patch_size());
70    CodePatcher patcher(isolate, call_address, call_size_in_words);
71    patcher.masm()->Call(deopt_entry, RelocInfo::NONE32);
72    DCHECK(prev_call_address == NULL ||
73           call_address >= prev_call_address + patch_size());
74    DCHECK(call_address + patch_size() <= code->instruction_end());
75
76#ifdef DEBUG
77    prev_call_address = call_address;
78#endif
79  }
80}
81
82
83void Deoptimizer::SetPlatformCompiledStubRegisters(
84    FrameDescription* output_frame, CodeStubDescriptor* descriptor) {
85  ApiFunction function(descriptor->deoptimization_handler());
86  ExternalReference xref(&function, ExternalReference::BUILTIN_CALL, isolate_);
87  intptr_t handler = reinterpret_cast<intptr_t>(xref.address());
88  int params = descriptor->GetHandlerParameterCount();
89  output_frame->SetRegister(a0.code(), params);
90  output_frame->SetRegister(a1.code(), handler);
91}
92
93
94void Deoptimizer::CopyDoubleRegisters(FrameDescription* output_frame) {
95  for (int i = 0; i < DoubleRegister::kMaxNumRegisters; ++i) {
96    double double_value = input_->GetDoubleRegister(i);
97    output_frame->SetDoubleRegister(i, double_value);
98  }
99}
100
101#define __ masm()->
102
103
104// This code tries to be close to ia32 code so that any changes can be
105// easily ported.
106void Deoptimizer::TableEntryGenerator::Generate() {
107  GeneratePrologue();
108
109  // Unlike on ARM we don't save all the registers, just the useful ones.
110  // For the rest, there are gaps on the stack, so the offsets remain the same.
111  const int kNumberOfRegisters = Register::kNumRegisters;
112
113  RegList restored_regs = kJSCallerSaved | kCalleeSaved;
114  RegList saved_regs = restored_regs | sp.bit() | ra.bit();
115
116  const int kDoubleRegsSize = kDoubleSize * DoubleRegister::kMaxNumRegisters;
117
118  // Save all FPU registers before messing with them.
119  __ Dsubu(sp, sp, Operand(kDoubleRegsSize));
120  const RegisterConfiguration* config = RegisterConfiguration::Crankshaft();
121  for (int i = 0; i < config->num_allocatable_double_registers(); ++i) {
122    int code = config->GetAllocatableDoubleCode(i);
123    const DoubleRegister fpu_reg = DoubleRegister::from_code(code);
124    int offset = code * kDoubleSize;
125    __ sdc1(fpu_reg, MemOperand(sp, offset));
126  }
127
128  // Push saved_regs (needed to populate FrameDescription::registers_).
129  // Leave gaps for other registers.
130  __ Dsubu(sp, sp, kNumberOfRegisters * kPointerSize);
131  for (int16_t i = kNumberOfRegisters - 1; i >= 0; i--) {
132    if ((saved_regs & (1 << i)) != 0) {
133      __ sd(ToRegister(i), MemOperand(sp, kPointerSize * i));
134    }
135  }
136
137  __ li(a2, Operand(ExternalReference(Isolate::kCEntryFPAddress, isolate())));
138  __ sd(fp, MemOperand(a2));
139
140  const int kSavedRegistersAreaSize =
141      (kNumberOfRegisters * kPointerSize) + kDoubleRegsSize;
142
143  // Get the bailout id from the stack.
144  __ ld(a2, MemOperand(sp, kSavedRegistersAreaSize));
145
146  // Get the address of the location in the code object (a3) (return
147  // address for lazy deoptimization) and compute the fp-to-sp delta in
148  // register a4.
149  __ mov(a3, ra);
150  // Correct one word for bailout id.
151  __ Daddu(a4, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
152
153  __ Dsubu(a4, fp, a4);
154
155  // Allocate a new deoptimizer object.
156  __ PrepareCallCFunction(6, a5);
157  // Pass six arguments, according to n64 ABI.
158  __ mov(a0, zero_reg);
159  Label context_check;
160  __ ld(a1, MemOperand(fp, CommonFrameConstants::kContextOrFrameTypeOffset));
161  __ JumpIfSmi(a1, &context_check);
162  __ ld(a0, MemOperand(fp, JavaScriptFrameConstants::kFunctionOffset));
163  __ bind(&context_check);
164  __ li(a1, Operand(type()));  // Bailout type.
165  // a2: bailout id already loaded.
166  // a3: code address or 0 already loaded.
167  // a4: already has fp-to-sp delta.
168  __ li(a5, Operand(ExternalReference::isolate_address(isolate())));
169
170  // Call Deoptimizer::New().
171  {
172    AllowExternalCallThatCantCauseGC scope(masm());
173    __ CallCFunction(ExternalReference::new_deoptimizer_function(isolate()), 6);
174  }
175
176  // Preserve "deoptimizer" object in register v0 and get the input
177  // frame descriptor pointer to a1 (deoptimizer->input_);
178  // Move deopt-obj to a0 for call to Deoptimizer::ComputeOutputFrames() below.
179  __ mov(a0, v0);
180  __ ld(a1, MemOperand(v0, Deoptimizer::input_offset()));
181
182  // Copy core registers into FrameDescription::registers_[kNumRegisters].
183  DCHECK(Register::kNumRegisters == kNumberOfRegisters);
184  for (int i = 0; i < kNumberOfRegisters; i++) {
185    int offset = (i * kPointerSize) + FrameDescription::registers_offset();
186    if ((saved_regs & (1 << i)) != 0) {
187      __ ld(a2, MemOperand(sp, i * kPointerSize));
188      __ sd(a2, MemOperand(a1, offset));
189    } else if (FLAG_debug_code) {
190      __ li(a2, kDebugZapValue);
191      __ sd(a2, MemOperand(a1, offset));
192    }
193  }
194
195  int double_regs_offset = FrameDescription::double_registers_offset();
196  // Copy FPU registers to
197  // double_registers_[DoubleRegister::kNumAllocatableRegisters]
198  for (int i = 0; i < config->num_allocatable_double_registers(); ++i) {
199    int code = config->GetAllocatableDoubleCode(i);
200    int dst_offset = code * kDoubleSize + double_regs_offset;
201    int src_offset = code * kDoubleSize + kNumberOfRegisters * kPointerSize;
202    __ ldc1(f0, MemOperand(sp, src_offset));
203    __ sdc1(f0, MemOperand(a1, dst_offset));
204  }
205
206  // Remove the bailout id and the saved registers from the stack.
207  __ Daddu(sp, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
208
209  // Compute a pointer to the unwinding limit in register a2; that is
210  // the first stack slot not part of the input frame.
211  __ ld(a2, MemOperand(a1, FrameDescription::frame_size_offset()));
212  __ Daddu(a2, a2, sp);
213
214  // Unwind the stack down to - but not including - the unwinding
215  // limit and copy the contents of the activation frame to the input
216  // frame description.
217  __ Daddu(a3, a1, Operand(FrameDescription::frame_content_offset()));
218  Label pop_loop;
219  Label pop_loop_header;
220  __ BranchShort(&pop_loop_header);
221  __ bind(&pop_loop);
222  __ pop(a4);
223  __ sd(a4, MemOperand(a3, 0));
224  __ daddiu(a3, a3, sizeof(uint64_t));
225  __ bind(&pop_loop_header);
226  __ BranchShort(&pop_loop, ne, a2, Operand(sp));
227  // Compute the output frame in the deoptimizer.
228  __ push(a0);  // Preserve deoptimizer object across call.
229  // a0: deoptimizer object; a1: scratch.
230  __ PrepareCallCFunction(1, a1);
231  // Call Deoptimizer::ComputeOutputFrames().
232  {
233    AllowExternalCallThatCantCauseGC scope(masm());
234    __ CallCFunction(
235        ExternalReference::compute_output_frames_function(isolate()), 1);
236  }
237  __ pop(a0);  // Restore deoptimizer object (class Deoptimizer).
238
239  __ ld(sp, MemOperand(a0, Deoptimizer::caller_frame_top_offset()));
240
241  // Replace the current (input) frame with the output frames.
242  Label outer_push_loop, inner_push_loop,
243      outer_loop_header, inner_loop_header;
244  // Outer loop state: a4 = current "FrameDescription** output_",
245  // a1 = one past the last FrameDescription**.
246  __ lw(a1, MemOperand(a0, Deoptimizer::output_count_offset()));
247  __ ld(a4, MemOperand(a0, Deoptimizer::output_offset()));  // a4 is output_.
248  __ Dlsa(a1, a4, a1, kPointerSizeLog2);
249  __ BranchShort(&outer_loop_header);
250  __ bind(&outer_push_loop);
251  // Inner loop state: a2 = current FrameDescription*, a3 = loop index.
252  __ ld(a2, MemOperand(a4, 0));  // output_[ix]
253  __ ld(a3, MemOperand(a2, FrameDescription::frame_size_offset()));
254  __ BranchShort(&inner_loop_header);
255  __ bind(&inner_push_loop);
256  __ Dsubu(a3, a3, Operand(sizeof(uint64_t)));
257  __ Daddu(a6, a2, Operand(a3));
258  __ ld(a7, MemOperand(a6, FrameDescription::frame_content_offset()));
259  __ push(a7);
260  __ bind(&inner_loop_header);
261  __ BranchShort(&inner_push_loop, ne, a3, Operand(zero_reg));
262
263  __ Daddu(a4, a4, Operand(kPointerSize));
264  __ bind(&outer_loop_header);
265  __ BranchShort(&outer_push_loop, lt, a4, Operand(a1));
266
267  __ ld(a1, MemOperand(a0, Deoptimizer::input_offset()));
268  for (int i = 0; i < config->num_allocatable_double_registers(); ++i) {
269    int code = config->GetAllocatableDoubleCode(i);
270    const DoubleRegister fpu_reg = DoubleRegister::from_code(code);
271    int src_offset = code * kDoubleSize + double_regs_offset;
272    __ ldc1(fpu_reg, MemOperand(a1, src_offset));
273  }
274
275  // Push state, pc, and continuation from the last output frame.
276  __ ld(a6, MemOperand(a2, FrameDescription::state_offset()));
277  __ push(a6);
278
279  __ ld(a6, MemOperand(a2, FrameDescription::pc_offset()));
280  __ push(a6);
281  __ ld(a6, MemOperand(a2, FrameDescription::continuation_offset()));
282  __ push(a6);
283
284
285  // Technically restoring 'at' should work unless zero_reg is also restored
286  // but it's safer to check for this.
287  DCHECK(!(at.bit() & restored_regs));
288  // Restore the registers from the last output frame.
289  __ mov(at, a2);
290  for (int i = kNumberOfRegisters - 1; i >= 0; i--) {
291    int offset = (i * kPointerSize) + FrameDescription::registers_offset();
292    if ((restored_regs & (1 << i)) != 0) {
293      __ ld(ToRegister(i), MemOperand(at, offset));
294    }
295  }
296
297  __ InitializeRootRegister();
298
299  __ pop(at);  // Get continuation, leave pc on stack.
300  __ pop(ra);
301  __ Jump(at);
302  __ stop("Unreachable.");
303}
304
305
306// Maximum size of a table entry generated below.
307const int Deoptimizer::table_entry_size_ = 2 * Assembler::kInstrSize;
308
309void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
310  Assembler::BlockTrampolinePoolScope block_trampoline_pool(masm());
311
312  // Create a sequence of deoptimization entries.
313  // Note that registers are still live when jumping to an entry.
314  Label table_start, done, done_special, trampoline_jump;
315  __ bind(&table_start);
316  int kMaxEntriesBranchReach =
317      (1 << (kImm16Bits - 2)) / (table_entry_size_ / Assembler::kInstrSize);
318
319  if (count() <= kMaxEntriesBranchReach) {
320    // Common case.
321    for (int i = 0; i < count(); i++) {
322      Label start;
323      __ bind(&start);
324      DCHECK(is_int16(i));
325      __ BranchShort(USE_DELAY_SLOT, &done);  // Expose delay slot.
326      __ li(at, i);                      // In the delay slot.
327
328      DCHECK_EQ(table_entry_size_, masm()->SizeOfCodeGeneratedSince(&start));
329    }
330
331    DCHECK_EQ(masm()->SizeOfCodeGeneratedSince(&table_start),
332              count() * table_entry_size_);
333    __ bind(&done);
334    __ Push(at);
335  } else {
336    // Uncommon case, the branch cannot reach.
337    // Create mini trampoline and adjust id constants to get proper value at
338    // the end of table.
339    for (int i = kMaxEntriesBranchReach; i > 1; i--) {
340      Label start;
341      __ bind(&start);
342      DCHECK(is_int16(i));
343      __ BranchShort(USE_DELAY_SLOT, &trampoline_jump);  // Expose delay slot.
344      __ li(at, -i);                                // In the delay slot.
345      DCHECK_EQ(table_entry_size_, masm()->SizeOfCodeGeneratedSince(&start));
346    }
347    // Entry with id == kMaxEntriesBranchReach - 1.
348    __ bind(&trampoline_jump);
349    __ BranchShort(USE_DELAY_SLOT, &done_special);
350    __ li(at, -1);
351
352    for (int i = kMaxEntriesBranchReach; i < count(); i++) {
353      Label start;
354      __ bind(&start);
355      DCHECK(is_int16(i));
356      __ Branch(USE_DELAY_SLOT, &done);  // Expose delay slot.
357      __ li(at, i);                      // In the delay slot.
358    }
359
360    DCHECK_EQ(masm()->SizeOfCodeGeneratedSince(&table_start),
361              count() * table_entry_size_);
362    __ bind(&done_special);
363    __ daddiu(at, at, kMaxEntriesBranchReach);
364    __ bind(&done);
365    __ Push(at);
366  }
367}
368
369
370void FrameDescription::SetCallerPc(unsigned offset, intptr_t value) {
371  SetFrameSlot(offset, value);
372}
373
374
375void FrameDescription::SetCallerFp(unsigned offset, intptr_t value) {
376  SetFrameSlot(offset, value);
377}
378
379
380void FrameDescription::SetCallerConstantPool(unsigned offset, intptr_t value) {
381  // No embedded constant pool support.
382  UNREACHABLE();
383}
384
385
386#undef __
387
388
389}  // namespace internal
390}  // namespace v8
391