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#if V8_TARGET_ARCH_X64
6
7#include "src/codegen.h"
8#include "src/deoptimizer.h"
9#include "src/full-codegen/full-codegen.h"
10#include "src/register-configuration.h"
11#include "src/safepoint-table.h"
12
13namespace v8 {
14namespace internal {
15
16
17const int Deoptimizer::table_entry_size_ = 10;
18
19
20int Deoptimizer::patch_size() {
21  return Assembler::kCallSequenceLength;
22}
23
24
25void Deoptimizer::EnsureRelocSpaceForLazyDeoptimization(Handle<Code> code) {
26  // Empty because there is no need for relocation information for the code
27  // patching in Deoptimizer::PatchCodeForDeoptimization below.
28}
29
30
31void Deoptimizer::PatchCodeForDeoptimization(Isolate* isolate, Code* code) {
32  // Invalidate the relocation information, as it will become invalid by the
33  // code patching below, and is not needed any more.
34  code->InvalidateRelocation();
35
36  if (FLAG_zap_code_space) {
37    // Fail hard and early if we enter this code object again.
38    byte* pointer = code->FindCodeAgeSequence();
39    if (pointer != NULL) {
40      pointer += kNoCodeAgeSequenceLength;
41    } else {
42      pointer = code->instruction_start();
43    }
44    CodePatcher patcher(isolate, pointer, 1);
45    patcher.masm()->int3();
46
47    DeoptimizationInputData* data =
48        DeoptimizationInputData::cast(code->deoptimization_data());
49    int osr_offset = data->OsrPcOffset()->value();
50    if (osr_offset > 0) {
51      CodePatcher osr_patcher(isolate, code->instruction_start() + osr_offset,
52                              1);
53      osr_patcher.masm()->int3();
54    }
55  }
56
57  // For each LLazyBailout instruction insert a absolute call to the
58  // corresponding deoptimization entry, or a short call to an absolute
59  // jump if space is short. The absolute jumps are put in a table just
60  // before the safepoint table (space was allocated there when the Code
61  // object was created, if necessary).
62
63  Address instruction_start = code->instruction_start();
64#ifdef DEBUG
65  Address prev_call_address = NULL;
66#endif
67  DeoptimizationInputData* deopt_data =
68      DeoptimizationInputData::cast(code->deoptimization_data());
69  deopt_data->SetSharedFunctionInfo(Smi::FromInt(0));
70  // For each LLazyBailout instruction insert a call to the corresponding
71  // deoptimization entry.
72  for (int i = 0; i < deopt_data->DeoptCount(); i++) {
73    if (deopt_data->Pc(i)->value() == -1) continue;
74    // Position where Call will be patched in.
75    Address call_address = instruction_start + deopt_data->Pc(i)->value();
76    // There is room enough to write a long call instruction because we pad
77    // LLazyBailout instructions with nops if necessary.
78    CodePatcher patcher(isolate, call_address, Assembler::kCallSequenceLength);
79    patcher.masm()->Call(GetDeoptimizationEntry(isolate, i, LAZY),
80                         Assembler::RelocInfoNone());
81    DCHECK(prev_call_address == NULL ||
82           call_address >= prev_call_address + patch_size());
83    DCHECK(call_address + patch_size() <= code->instruction_end());
84#ifdef DEBUG
85    prev_call_address = call_address;
86#endif
87  }
88}
89
90
91void Deoptimizer::FillInputFrame(Address tos, JavaScriptFrame* frame) {
92  // Set the register values. The values are not important as there are no
93  // callee saved registers in JavaScript frames, so all registers are
94  // spilled. Registers rbp and rsp are set to the correct values though.
95  for (int i = 0; i < Register::kNumRegisters; i++) {
96    input_->SetRegister(i, i * 4);
97  }
98  input_->SetRegister(rsp.code(), reinterpret_cast<intptr_t>(frame->sp()));
99  input_->SetRegister(rbp.code(), reinterpret_cast<intptr_t>(frame->fp()));
100  for (int i = 0; i < DoubleRegister::kMaxNumRegisters; i++) {
101    input_->SetDoubleRegister(i, 0.0);
102  }
103
104  // Fill the frame content from the actual data on the frame.
105  for (unsigned i = 0; i < input_->GetFrameSize(); i += kPointerSize) {
106    input_->SetFrameSlot(i, Memory::uintptr_at(tos + i));
107  }
108}
109
110
111void Deoptimizer::SetPlatformCompiledStubRegisters(
112    FrameDescription* output_frame, CodeStubDescriptor* descriptor) {
113  intptr_t handler =
114      reinterpret_cast<intptr_t>(descriptor->deoptimization_handler());
115  int params = descriptor->GetHandlerParameterCount();
116  output_frame->SetRegister(rax.code(), params);
117  output_frame->SetRegister(rbx.code(), handler);
118}
119
120
121void Deoptimizer::CopyDoubleRegisters(FrameDescription* output_frame) {
122  for (int i = 0; i < XMMRegister::kMaxNumRegisters; ++i) {
123    double double_value = input_->GetDoubleRegister(i);
124    output_frame->SetDoubleRegister(i, double_value);
125  }
126}
127
128
129bool Deoptimizer::HasAlignmentPadding(JSFunction* function) {
130  // There is no dynamic alignment padding on x64 in the input frame.
131  return false;
132}
133
134
135#define __ masm()->
136
137void Deoptimizer::TableEntryGenerator::Generate() {
138  GeneratePrologue();
139
140  // Save all general purpose registers before messing with them.
141  const int kNumberOfRegisters = Register::kNumRegisters;
142
143  const int kDoubleRegsSize = kDoubleSize * XMMRegister::kMaxNumRegisters;
144  __ subp(rsp, Immediate(kDoubleRegsSize));
145
146  const RegisterConfiguration* config =
147      RegisterConfiguration::ArchDefault(RegisterConfiguration::CRANKSHAFT);
148  for (int i = 0; i < config->num_allocatable_double_registers(); ++i) {
149    int code = config->GetAllocatableDoubleCode(i);
150    XMMRegister xmm_reg = XMMRegister::from_code(code);
151    int offset = code * kDoubleSize;
152    __ Movsd(Operand(rsp, offset), xmm_reg);
153  }
154
155  // We push all registers onto the stack, even though we do not need
156  // to restore all later.
157  for (int i = 0; i < kNumberOfRegisters; i++) {
158    Register r = Register::from_code(i);
159    __ pushq(r);
160  }
161
162  const int kSavedRegistersAreaSize = kNumberOfRegisters * kRegisterSize +
163                                      kDoubleRegsSize;
164
165  __ Store(ExternalReference(Isolate::kCEntryFPAddress, isolate()), rbp);
166
167  // We use this to keep the value of the fifth argument temporarily.
168  // Unfortunately we can't store it directly in r8 (used for passing
169  // this on linux), since it is another parameter passing register on windows.
170  Register arg5 = r11;
171
172  // Get the bailout id from the stack.
173  __ movp(arg_reg_3, Operand(rsp, kSavedRegistersAreaSize));
174
175  // Get the address of the location in the code object
176  // and compute the fp-to-sp delta in register arg5.
177  __ movp(arg_reg_4, Operand(rsp, kSavedRegistersAreaSize + 1 * kRegisterSize));
178  __ leap(arg5, Operand(rsp, kSavedRegistersAreaSize + 1 * kRegisterSize +
179                            kPCOnStackSize));
180
181  __ subp(arg5, rbp);
182  __ negp(arg5);
183
184  // Allocate a new deoptimizer object.
185  __ PrepareCallCFunction(6);
186  __ movp(rax, Operand(rbp, JavaScriptFrameConstants::kFunctionOffset));
187  __ movp(arg_reg_1, rax);
188  __ Set(arg_reg_2, type());
189  // Args 3 and 4 are already in the right registers.
190
191  // On windows put the arguments on the stack (PrepareCallCFunction
192  // has created space for this). On linux pass the arguments in r8 and r9.
193#ifdef _WIN64
194  __ movq(Operand(rsp, 4 * kRegisterSize), arg5);
195  __ LoadAddress(arg5, ExternalReference::isolate_address(isolate()));
196  __ movq(Operand(rsp, 5 * kRegisterSize), arg5);
197#else
198  __ movp(r8, arg5);
199  __ LoadAddress(r9, ExternalReference::isolate_address(isolate()));
200#endif
201
202  { AllowExternalCallThatCantCauseGC scope(masm());
203    __ CallCFunction(ExternalReference::new_deoptimizer_function(isolate()), 6);
204  }
205  // Preserve deoptimizer object in register rax and get the input
206  // frame descriptor pointer.
207  __ movp(rbx, Operand(rax, Deoptimizer::input_offset()));
208
209  // Fill in the input registers.
210  for (int i = kNumberOfRegisters -1; i >= 0; i--) {
211    int offset = (i * kPointerSize) + FrameDescription::registers_offset();
212    __ PopQuad(Operand(rbx, offset));
213  }
214
215  // Fill in the double input registers.
216  int double_regs_offset = FrameDescription::double_registers_offset();
217  for (int i = 0; i < XMMRegister::kMaxNumRegisters; i++) {
218    int dst_offset = i * kDoubleSize + double_regs_offset;
219    __ popq(Operand(rbx, dst_offset));
220  }
221
222  // Remove the bailout id and return address from the stack.
223  __ addp(rsp, Immediate(1 * kRegisterSize + kPCOnStackSize));
224
225  // Compute a pointer to the unwinding limit in register rcx; that is
226  // the first stack slot not part of the input frame.
227  __ movp(rcx, Operand(rbx, FrameDescription::frame_size_offset()));
228  __ addp(rcx, rsp);
229
230  // Unwind the stack down to - but not including - the unwinding
231  // limit and copy the contents of the activation frame to the input
232  // frame description.
233  __ leap(rdx, Operand(rbx, FrameDescription::frame_content_offset()));
234  Label pop_loop_header;
235  __ jmp(&pop_loop_header);
236  Label pop_loop;
237  __ bind(&pop_loop);
238  __ Pop(Operand(rdx, 0));
239  __ addp(rdx, Immediate(sizeof(intptr_t)));
240  __ bind(&pop_loop_header);
241  __ cmpp(rcx, rsp);
242  __ j(not_equal, &pop_loop);
243
244  // Compute the output frame in the deoptimizer.
245  __ pushq(rax);
246  __ PrepareCallCFunction(2);
247  __ movp(arg_reg_1, rax);
248  __ LoadAddress(arg_reg_2, ExternalReference::isolate_address(isolate()));
249  {
250    AllowExternalCallThatCantCauseGC scope(masm());
251    __ CallCFunction(
252        ExternalReference::compute_output_frames_function(isolate()), 2);
253  }
254  __ popq(rax);
255
256  // Replace the current frame with the output frames.
257  Label outer_push_loop, inner_push_loop,
258      outer_loop_header, inner_loop_header;
259  // Outer loop state: rax = current FrameDescription**, rdx = one past the
260  // last FrameDescription**.
261  __ movl(rdx, Operand(rax, Deoptimizer::output_count_offset()));
262  __ movp(rax, Operand(rax, Deoptimizer::output_offset()));
263  __ leap(rdx, Operand(rax, rdx, times_pointer_size, 0));
264  __ jmp(&outer_loop_header);
265  __ bind(&outer_push_loop);
266  // Inner loop state: rbx = current FrameDescription*, rcx = loop index.
267  __ movp(rbx, Operand(rax, 0));
268  __ movp(rcx, Operand(rbx, FrameDescription::frame_size_offset()));
269  __ jmp(&inner_loop_header);
270  __ bind(&inner_push_loop);
271  __ subp(rcx, Immediate(sizeof(intptr_t)));
272  __ Push(Operand(rbx, rcx, times_1, FrameDescription::frame_content_offset()));
273  __ bind(&inner_loop_header);
274  __ testp(rcx, rcx);
275  __ j(not_zero, &inner_push_loop);
276  __ addp(rax, Immediate(kPointerSize));
277  __ bind(&outer_loop_header);
278  __ cmpp(rax, rdx);
279  __ j(below, &outer_push_loop);
280
281  for (int i = 0; i < config->num_allocatable_double_registers(); ++i) {
282    int code = config->GetAllocatableDoubleCode(i);
283    XMMRegister xmm_reg = XMMRegister::from_code(code);
284    int src_offset = code * kDoubleSize + double_regs_offset;
285    __ Movsd(xmm_reg, Operand(rbx, src_offset));
286  }
287
288  // Push state, pc, and continuation from the last output frame.
289  __ Push(Operand(rbx, FrameDescription::state_offset()));
290  __ PushQuad(Operand(rbx, FrameDescription::pc_offset()));
291  __ PushQuad(Operand(rbx, FrameDescription::continuation_offset()));
292
293  // Push the registers from the last output frame.
294  for (int i = 0; i < kNumberOfRegisters; i++) {
295    int offset = (i * kPointerSize) + FrameDescription::registers_offset();
296    __ PushQuad(Operand(rbx, offset));
297  }
298
299  // Restore the registers from the stack.
300  for (int i = kNumberOfRegisters - 1; i >= 0 ; i--) {
301    Register r = Register::from_code(i);
302    // Do not restore rsp, simply pop the value into the next register
303    // and overwrite this afterwards.
304    if (r.is(rsp)) {
305      DCHECK(i > 0);
306      r = Register::from_code(i - 1);
307    }
308    __ popq(r);
309  }
310
311  // Set up the roots register.
312  __ InitializeRootRegister();
313
314  // Return to the continuation point.
315  __ ret(0);
316}
317
318
319void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
320  // Create a sequence of deoptimization entries.
321  Label done;
322  for (int i = 0; i < count(); i++) {
323    int start = masm()->pc_offset();
324    USE(start);
325    __ pushq_imm32(i);
326    __ jmp(&done);
327    DCHECK(masm()->pc_offset() - start == table_entry_size_);
328  }
329  __ bind(&done);
330}
331
332
333void FrameDescription::SetCallerPc(unsigned offset, intptr_t value) {
334  if (kPCOnStackSize == 2 * kPointerSize) {
335    // Zero out the high-32 bit of PC for x32 port.
336    SetFrameSlot(offset + kPointerSize, 0);
337  }
338  SetFrameSlot(offset, value);
339}
340
341
342void FrameDescription::SetCallerFp(unsigned offset, intptr_t value) {
343  if (kFPOnStackSize == 2 * kPointerSize) {
344    // Zero out the high-32 bit of FP for x32 port.
345    SetFrameSlot(offset + kPointerSize, 0);
346  }
347  SetFrameSlot(offset, value);
348}
349
350
351void FrameDescription::SetCallerConstantPool(unsigned offset, intptr_t value) {
352  // No embedded constant pool support.
353  UNREACHABLE();
354}
355
356
357#undef __
358
359
360}  // namespace internal
361}  // namespace v8
362
363#endif  // V8_TARGET_ARCH_X64
364