1// Copyright 2012 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
14const int Deoptimizer::table_entry_size_ = 8;
15
16
17int Deoptimizer::patch_size() {
18  const int kCallInstructionSizeInWords = 3;
19  return kCallInstructionSizeInWords * Assembler::kInstrSize;
20}
21
22
23void Deoptimizer::EnsureRelocSpaceForLazyDeoptimization(Handle<Code> code) {
24  // Empty because there is no need for relocation information for the code
25  // patching in Deoptimizer::PatchCodeForDeoptimization below.
26}
27
28
29void Deoptimizer::PatchCodeForDeoptimization(Isolate* isolate, Code* code) {
30  Address code_start_address = code->instruction_start();
31  // Invalidate the relocation information, as it will become invalid by the
32  // code patching below, and is not needed any more.
33  code->InvalidateRelocation();
34
35  if (FLAG_zap_code_space) {
36    // Fail hard and early if we enter this code object again.
37    byte* pointer = code->FindCodeAgeSequence();
38    if (pointer != NULL) {
39      pointer += kNoCodeAgeSequenceLength;
40    } else {
41      pointer = code->instruction_start();
42    }
43    CodePatcher patcher(isolate, pointer, 1);
44    patcher.masm()->bkpt(0);
45
46    DeoptimizationInputData* data =
47        DeoptimizationInputData::cast(code->deoptimization_data());
48    int osr_offset = data->OsrPcOffset()->value();
49    if (osr_offset > 0) {
50      CodePatcher osr_patcher(isolate, code->instruction_start() + osr_offset,
51                              1);
52      osr_patcher.masm()->bkpt(0);
53    }
54  }
55
56  DeoptimizationInputData* deopt_data =
57      DeoptimizationInputData::cast(code->deoptimization_data());
58#ifdef DEBUG
59  Address prev_call_address = NULL;
60#endif
61  // For each LLazyBailout instruction insert a call to the corresponding
62  // deoptimization entry.
63  for (int i = 0; i < deopt_data->DeoptCount(); i++) {
64    if (deopt_data->Pc(i)->value() == -1) continue;
65    Address call_address = code_start_address + deopt_data->Pc(i)->value();
66    Address deopt_entry = GetDeoptimizationEntry(isolate, i, LAZY);
67    // We need calls to have a predictable size in the unoptimized code, but
68    // this is optimized code, so we don't have to have a predictable size.
69    int call_size_in_bytes = MacroAssembler::CallDeoptimizerSize();
70    int call_size_in_words = call_size_in_bytes / Assembler::kInstrSize;
71    DCHECK(call_size_in_bytes % Assembler::kInstrSize == 0);
72    DCHECK(call_size_in_bytes <= patch_size());
73    CodePatcher patcher(isolate, call_address, call_size_in_words);
74    patcher.masm()->CallDeoptimizer(deopt_entry);
75    DCHECK(prev_call_address == NULL ||
76           call_address >= prev_call_address + patch_size());
77    DCHECK(call_address + patch_size() <= code->instruction_end());
78#ifdef DEBUG
79    prev_call_address = call_address;
80#endif
81  }
82}
83
84
85void Deoptimizer::SetPlatformCompiledStubRegisters(
86    FrameDescription* output_frame, CodeStubDescriptor* descriptor) {
87  ApiFunction function(descriptor->deoptimization_handler());
88  ExternalReference xref(&function, ExternalReference::BUILTIN_CALL, isolate_);
89  intptr_t handler = reinterpret_cast<intptr_t>(xref.address());
90  int params = descriptor->GetHandlerParameterCount();
91  output_frame->SetRegister(r0.code(), params);
92  output_frame->SetRegister(r1.code(), handler);
93}
94
95
96void Deoptimizer::CopyDoubleRegisters(FrameDescription* output_frame) {
97  for (int i = 0; i < DwVfpRegister::kMaxNumRegisters; ++i) {
98    Float64 double_value = input_->GetDoubleRegister(i);
99    output_frame->SetDoubleRegister(i, double_value);
100  }
101}
102
103#define __ masm()->
104
105// This code tries to be close to ia32 code so that any changes can be
106// easily ported.
107void Deoptimizer::TableEntryGenerator::Generate() {
108  GeneratePrologue();
109
110  // Save all general purpose registers before messing with them.
111  const int kNumberOfRegisters = Register::kNumRegisters;
112
113  // Everything but pc, lr and ip which will be saved but not restored.
114  RegList restored_regs = kJSCallerSaved | kCalleeSaved | ip.bit();
115
116  const int kDoubleRegsSize = kDoubleSize * DwVfpRegister::kMaxNumRegisters;
117
118  // Save all allocatable VFP registers before messing with them.
119  DCHECK(kDoubleRegZero.code() == 14);
120  DCHECK(kScratchDoubleReg.code() == 15);
121
122  {
123    // We use a run-time check for VFP32DREGS.
124    CpuFeatureScope scope(masm(), VFP32DREGS,
125                          CpuFeatureScope::kDontCheckSupported);
126
127    // Check CPU flags for number of registers, setting the Z condition flag.
128    __ CheckFor32DRegs(ip);
129
130    // Push registers d0-d15, and possibly d16-d31, on the stack.
131    // If d16-d31 are not pushed, decrease the stack pointer instead.
132    __ vstm(db_w, sp, d16, d31, ne);
133    __ sub(sp, sp, Operand(16 * kDoubleSize), LeaveCC, eq);
134    __ vstm(db_w, sp, d0, d15);
135  }
136
137  // Push all 16 registers (needed to populate FrameDescription::registers_).
138  // TODO(1588) Note that using pc with stm is deprecated, so we should perhaps
139  // handle this a bit differently.
140  __ stm(db_w, sp, restored_regs  | sp.bit() | lr.bit() | pc.bit());
141
142  __ mov(ip, Operand(ExternalReference(Isolate::kCEntryFPAddress, isolate())));
143  __ str(fp, MemOperand(ip));
144
145  const int kSavedRegistersAreaSize =
146      (kNumberOfRegisters * kPointerSize) + kDoubleRegsSize;
147
148  // Get the bailout id from the stack.
149  __ ldr(r2, MemOperand(sp, kSavedRegistersAreaSize));
150
151  // Get the address of the location in the code object (r3) (return
152  // address for lazy deoptimization) and compute the fp-to-sp delta in
153  // register r4.
154  __ mov(r3, lr);
155  // Correct one word for bailout id.
156  __ add(r4, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
157  __ sub(r4, fp, r4);
158
159  // Allocate a new deoptimizer object.
160  // Pass four arguments in r0 to r3 and fifth argument on stack.
161  __ PrepareCallCFunction(6, r5);
162  __ mov(r0, Operand(0));
163  Label context_check;
164  __ ldr(r1, MemOperand(fp, CommonFrameConstants::kContextOrFrameTypeOffset));
165  __ JumpIfSmi(r1, &context_check);
166  __ ldr(r0, MemOperand(fp, JavaScriptFrameConstants::kFunctionOffset));
167  __ bind(&context_check);
168  __ mov(r1, Operand(type()));  // bailout type,
169  // r2: bailout id already loaded.
170  // r3: code address or 0 already loaded.
171  __ str(r4, MemOperand(sp, 0 * kPointerSize));  // Fp-to-sp delta.
172  __ mov(r5, Operand(ExternalReference::isolate_address(isolate())));
173  __ str(r5, MemOperand(sp, 1 * kPointerSize));  // Isolate.
174  // Call Deoptimizer::New().
175  {
176    AllowExternalCallThatCantCauseGC scope(masm());
177    __ CallCFunction(ExternalReference::new_deoptimizer_function(isolate()), 6);
178  }
179
180  // Preserve "deoptimizer" object in register r0 and get the input
181  // frame descriptor pointer to r1 (deoptimizer->input_);
182  __ ldr(r1, MemOperand(r0, Deoptimizer::input_offset()));
183
184  // Copy core registers into FrameDescription::registers_[kNumRegisters].
185  DCHECK(Register::kNumRegisters == kNumberOfRegisters);
186  for (int i = 0; i < kNumberOfRegisters; i++) {
187    int offset = (i * kPointerSize) + FrameDescription::registers_offset();
188    __ ldr(r2, MemOperand(sp, i * kPointerSize));
189    __ str(r2, MemOperand(r1, offset));
190  }
191
192  // Copy VFP registers to
193  // double_registers_[DoubleRegister::kMaxNumAllocatableRegisters]
194  int double_regs_offset = FrameDescription::double_registers_offset();
195  const RegisterConfiguration* config = RegisterConfiguration::Crankshaft();
196  for (int i = 0; i < config->num_allocatable_double_registers(); ++i) {
197    int code = config->GetAllocatableDoubleCode(i);
198    int dst_offset = code * kDoubleSize + double_regs_offset;
199    int src_offset = code * kDoubleSize + kNumberOfRegisters * kPointerSize;
200    __ vldr(d0, sp, src_offset);
201    __ vstr(d0, r1, dst_offset);
202  }
203
204  // Remove the bailout id and the saved registers from the stack.
205  __ add(sp, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
206
207  // Compute a pointer to the unwinding limit in register r2; that is
208  // the first stack slot not part of the input frame.
209  __ ldr(r2, MemOperand(r1, FrameDescription::frame_size_offset()));
210  __ add(r2, r2, sp);
211
212  // Unwind the stack down to - but not including - the unwinding
213  // limit and copy the contents of the activation frame to the input
214  // frame description.
215  __ add(r3,  r1, Operand(FrameDescription::frame_content_offset()));
216  Label pop_loop;
217  Label pop_loop_header;
218  __ b(&pop_loop_header);
219  __ bind(&pop_loop);
220  __ pop(r4);
221  __ str(r4, MemOperand(r3, 0));
222  __ add(r3, r3, Operand(sizeof(uint32_t)));
223  __ bind(&pop_loop_header);
224  __ cmp(r2, sp);
225  __ b(ne, &pop_loop);
226
227  // Compute the output frame in the deoptimizer.
228  __ push(r0);  // Preserve deoptimizer object across call.
229  // r0: deoptimizer object; r1: scratch.
230  __ PrepareCallCFunction(1, r1);
231  // Call Deoptimizer::ComputeOutputFrames().
232  {
233    AllowExternalCallThatCantCauseGC scope(masm());
234    __ CallCFunction(
235        ExternalReference::compute_output_frames_function(isolate()), 1);
236  }
237  __ pop(r0);  // Restore deoptimizer object (class Deoptimizer).
238
239  __ ldr(sp, MemOperand(r0, 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: r4 = current "FrameDescription** output_",
245  // r1 = one past the last FrameDescription**.
246  __ ldr(r1, MemOperand(r0, Deoptimizer::output_count_offset()));
247  __ ldr(r4, MemOperand(r0, Deoptimizer::output_offset()));  // r4 is output_.
248  __ add(r1, r4, Operand(r1, LSL, 2));
249  __ jmp(&outer_loop_header);
250  __ bind(&outer_push_loop);
251  // Inner loop state: r2 = current FrameDescription*, r3 = loop index.
252  __ ldr(r2, MemOperand(r4, 0));  // output_[ix]
253  __ ldr(r3, MemOperand(r2, FrameDescription::frame_size_offset()));
254  __ jmp(&inner_loop_header);
255  __ bind(&inner_push_loop);
256  __ sub(r3, r3, Operand(sizeof(uint32_t)));
257  __ add(r6, r2, Operand(r3));
258  __ ldr(r6, MemOperand(r6, FrameDescription::frame_content_offset()));
259  __ push(r6);
260  __ bind(&inner_loop_header);
261  __ cmp(r3, Operand::Zero());
262  __ b(ne, &inner_push_loop);  // test for gt?
263  __ add(r4, r4, Operand(kPointerSize));
264  __ bind(&outer_loop_header);
265  __ cmp(r4, r1);
266  __ b(lt, &outer_push_loop);
267
268  __ ldr(r1, MemOperand(r0, Deoptimizer::input_offset()));
269  for (int i = 0; i < config->num_allocatable_double_registers(); ++i) {
270    int code = config->GetAllocatableDoubleCode(i);
271    DwVfpRegister reg = DwVfpRegister::from_code(code);
272    int src_offset = code * kDoubleSize + double_regs_offset;
273    __ vldr(reg, r1, src_offset);
274  }
275
276  // Push state, pc, and continuation from the last output frame.
277  __ ldr(r6, MemOperand(r2, FrameDescription::state_offset()));
278  __ push(r6);
279  __ ldr(r6, MemOperand(r2, FrameDescription::pc_offset()));
280  __ push(r6);
281  __ ldr(r6, MemOperand(r2, FrameDescription::continuation_offset()));
282  __ push(r6);
283
284  // Push the registers from the last output frame.
285  for (int i = kNumberOfRegisters - 1; i >= 0; i--) {
286    int offset = (i * kPointerSize) + FrameDescription::registers_offset();
287    __ ldr(r6, MemOperand(r2, offset));
288    __ push(r6);
289  }
290
291  // Restore the registers from the stack.
292  __ ldm(ia_w, sp, restored_regs);  // all but pc registers.
293  __ pop(ip);  // remove sp
294  __ pop(ip);  // remove lr
295
296  __ InitializeRootRegister();
297
298  __ pop(ip);  // remove pc
299  __ pop(ip);  // get continuation, leave pc on stack
300  __ pop(lr);
301  __ Jump(ip);
302  __ stop("Unreachable.");
303}
304
305
306void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
307  // Create a sequence of deoptimization entries.
308  // Note that registers are still live when jumping to an entry.
309
310  // We need to be able to generate immediates up to kMaxNumberOfEntries. On
311  // ARMv7, we can use movw (with a maximum immediate of 0xffff). On ARMv6, we
312  // need two instructions.
313  STATIC_ASSERT((kMaxNumberOfEntries - 1) <= 0xffff);
314  if (CpuFeatures::IsSupported(ARMv7)) {
315    CpuFeatureScope scope(masm(), ARMv7);
316    Label done;
317    for (int i = 0; i < count(); i++) {
318      int start = masm()->pc_offset();
319      USE(start);
320      __ movw(ip, i);
321      __ b(&done);
322      DCHECK_EQ(table_entry_size_, masm()->pc_offset() - start);
323    }
324    __ bind(&done);
325  } else {
326    // We want to keep table_entry_size_ == 8 (since this is the common case),
327    // but we need two instructions to load most immediates over 0xff. To handle
328    // this, we set the low byte in the main table, and then set the high byte
329    // in a separate table if necessary.
330    Label high_fixes[256];
331    int high_fix_max = (count() - 1) >> 8;
332    DCHECK_GT(arraysize(high_fixes), static_cast<size_t>(high_fix_max));
333    for (int i = 0; i < count(); i++) {
334      int start = masm()->pc_offset();
335      USE(start);
336      __ mov(ip, Operand(i & 0xff));  // Set the low byte.
337      __ b(&high_fixes[i >> 8]);      // Jump to the secondary table.
338      DCHECK_EQ(table_entry_size_, masm()->pc_offset() - start);
339    }
340    // Generate the secondary table, to set the high byte.
341    for (int high = 1; high <= high_fix_max; high++) {
342      __ bind(&high_fixes[high]);
343      __ orr(ip, ip, Operand(high << 8));
344      // If this isn't the last entry, emit a branch to the end of the table.
345      // The last entry can just fall through.
346      if (high < high_fix_max) __ b(&high_fixes[0]);
347    }
348    // Bind high_fixes[0] last, for indices like 0x00**. This case requires no
349    // fix-up, so for (common) small tables we can jump here, then just fall
350    // through with no additional branch.
351    __ bind(&high_fixes[0]);
352  }
353  __ push(ip);
354}
355
356
357void FrameDescription::SetCallerPc(unsigned offset, intptr_t value) {
358  SetFrameSlot(offset, value);
359}
360
361
362void FrameDescription::SetCallerFp(unsigned offset, intptr_t value) {
363  SetFrameSlot(offset, value);
364}
365
366
367void FrameDescription::SetCallerConstantPool(unsigned offset, intptr_t value) {
368  DCHECK(FLAG_enable_embedded_constant_pool);
369  SetFrameSlot(offset, value);
370}
371
372
373#undef __
374
375}  // namespace internal
376}  // namespace v8
377