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