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.
7 #if V8_TARGET_ARCH_IA32
10 #include "deoptimizer.h"
11 #include "full-codegen.h"
12 #include "safepoint-table.h"
17 const int Deoptimizer::table_entry_size_ = 10;
20 int Deoptimizer::patch_size() {
21 return Assembler::kCallInstructionLength;
25 void Deoptimizer::EnsureRelocSpaceForLazyDeoptimization(Handle<Code> code) {
26 Isolate* isolate = code->GetIsolate();
27 HandleScope scope(isolate);
29 // Compute the size of relocation information needed for the code
30 // patching in Deoptimizer::DeoptimizeFunction.
31 int min_reloc_size = 0;
32 int prev_pc_offset = 0;
33 DeoptimizationInputData* deopt_data =
34 DeoptimizationInputData::cast(code->deoptimization_data());
35 for (int i = 0; i < deopt_data->DeoptCount(); i++) {
36 int pc_offset = deopt_data->Pc(i)->value();
37 if (pc_offset == -1) continue;
38 ASSERT_GE(pc_offset, prev_pc_offset);
39 int pc_delta = pc_offset - prev_pc_offset;
40 // We use RUNTIME_ENTRY reloc info which has a size of 2 bytes
41 // if encodable with small pc delta encoding and up to 6 bytes
43 if (pc_delta <= RelocInfo::kMaxSmallPCDelta) {
48 prev_pc_offset = pc_offset;
51 // If the relocation information is not big enough we create a new
52 // relocation info object that is padded with comments to make it
53 // big enough for lazy doptimization.
54 int reloc_length = code->relocation_info()->length();
55 if (min_reloc_size > reloc_length) {
56 int comment_reloc_size = RelocInfo::kMinRelocCommentSize;
58 int min_padding = min_reloc_size - reloc_length;
59 // Number of comments needed to take up at least that much space.
60 int additional_comments =
61 (min_padding + comment_reloc_size - 1) / comment_reloc_size;
62 // Actual padding size.
63 int padding = additional_comments * comment_reloc_size;
64 // Allocate new relocation info and copy old relocation to the end
65 // of the new relocation info array because relocation info is
66 // written and read backwards.
67 Factory* factory = isolate->factory();
68 Handle<ByteArray> new_reloc =
69 factory->NewByteArray(reloc_length + padding, TENURED);
70 OS::MemCopy(new_reloc->GetDataStartAddress() + padding,
71 code->relocation_info()->GetDataStartAddress(),
73 // Create a relocation writer to write the comments in the padding
74 // space. Use position 0 for everything to ensure short encoding.
75 RelocInfoWriter reloc_info_writer(
76 new_reloc->GetDataStartAddress() + padding, 0);
77 intptr_t comment_string
78 = reinterpret_cast<intptr_t>(RelocInfo::kFillerCommentString);
79 RelocInfo rinfo(0, RelocInfo::COMMENT, comment_string, NULL);
80 for (int i = 0; i < additional_comments; ++i) {
82 byte* pos_before = reloc_info_writer.pos();
84 reloc_info_writer.Write(&rinfo);
85 ASSERT(RelocInfo::kMinRelocCommentSize ==
86 pos_before - reloc_info_writer.pos());
88 // Replace relocation information on the code object.
89 code->set_relocation_info(*new_reloc);
94 void Deoptimizer::PatchCodeForDeoptimization(Isolate* isolate, Code* code) {
95 Address code_start_address = code->instruction_start();
97 if (FLAG_zap_code_space) {
98 // Fail hard and early if we enter this code object again.
99 byte* pointer = code->FindCodeAgeSequence();
100 if (pointer != NULL) {
101 pointer += kNoCodeAgeSequenceLength;
103 pointer = code->instruction_start();
105 CodePatcher patcher(pointer, 1);
106 patcher.masm()->int3();
108 DeoptimizationInputData* data =
109 DeoptimizationInputData::cast(code->deoptimization_data());
110 int osr_offset = data->OsrPcOffset()->value();
111 if (osr_offset > 0) {
112 CodePatcher osr_patcher(code->instruction_start() + osr_offset, 1);
113 osr_patcher.masm()->int3();
117 // We will overwrite the code's relocation info in-place. Relocation info
118 // is written backward. The relocation info is the payload of a byte
119 // array. Later on we will slide this to the start of the byte array and
120 // create a filler object in the remaining space.
121 ByteArray* reloc_info = code->relocation_info();
122 Address reloc_end_address = reloc_info->address() + reloc_info->Size();
123 RelocInfoWriter reloc_info_writer(reloc_end_address, code_start_address);
125 // Since the call is a relative encoding, write new
126 // reloc info. We do not need any of the existing reloc info because the
127 // existing code will not be used again (we zap it in debug builds).
129 // Emit call to lazy deoptimization at all lazy deopt points.
130 DeoptimizationInputData* deopt_data =
131 DeoptimizationInputData::cast(code->deoptimization_data());
132 SharedFunctionInfo* shared =
133 SharedFunctionInfo::cast(deopt_data->SharedFunctionInfo());
134 shared->EvictFromOptimizedCodeMap(code, "deoptimized code");
136 Address prev_call_address = NULL;
138 // For each LLazyBailout instruction insert a call to the corresponding
139 // deoptimization entry.
140 for (int i = 0; i < deopt_data->DeoptCount(); i++) {
141 if (deopt_data->Pc(i)->value() == -1) continue;
142 // Patch lazy deoptimization entry.
143 Address call_address = code_start_address + deopt_data->Pc(i)->value();
144 CodePatcher patcher(call_address, patch_size());
145 Address deopt_entry = GetDeoptimizationEntry(isolate, i, LAZY);
146 patcher.masm()->call(deopt_entry, RelocInfo::NONE32);
147 // We use RUNTIME_ENTRY for deoptimization bailouts.
148 RelocInfo rinfo(call_address + 1, // 1 after the call opcode.
149 RelocInfo::RUNTIME_ENTRY,
150 reinterpret_cast<intptr_t>(deopt_entry),
152 reloc_info_writer.Write(&rinfo);
153 ASSERT_GE(reloc_info_writer.pos(),
154 reloc_info->address() + ByteArray::kHeaderSize);
155 ASSERT(prev_call_address == NULL ||
156 call_address >= prev_call_address + patch_size());
157 ASSERT(call_address + patch_size() <= code->instruction_end());
159 prev_call_address = call_address;
163 // Move the relocation info to the beginning of the byte array.
164 int new_reloc_size = reloc_end_address - reloc_info_writer.pos();
166 code->relocation_start(), reloc_info_writer.pos(), new_reloc_size);
168 // The relocation info is in place, update the size.
169 reloc_info->set_length(new_reloc_size);
171 // Handle the junk part after the new relocation info. We will create
172 // a non-live object in the extra space at the end of the former reloc info.
173 Address junk_address = reloc_info->address() + reloc_info->Size();
174 ASSERT(junk_address <= reloc_end_address);
175 isolate->heap()->CreateFillerObjectAt(junk_address,
176 reloc_end_address - junk_address);
180 void Deoptimizer::FillInputFrame(Address tos, JavaScriptFrame* frame) {
181 // Set the register values. The values are not important as there are no
182 // callee saved registers in JavaScript frames, so all registers are
183 // spilled. Registers ebp and esp are set to the correct values though.
185 for (int i = 0; i < Register::kNumRegisters; i++) {
186 input_->SetRegister(i, i * 4);
188 input_->SetRegister(esp.code(), reinterpret_cast<intptr_t>(frame->sp()));
189 input_->SetRegister(ebp.code(), reinterpret_cast<intptr_t>(frame->fp()));
190 simd128_value_t zero = {{0.0, 0.0}};
191 for (int i = 0; i < DoubleRegister::NumAllocatableRegisters(); i++) {
192 input_->SetSIMD128Register(i, zero);
195 // Fill the frame content from the actual data on the frame.
196 for (unsigned i = 0; i < input_->GetFrameSize(); i += kPointerSize) {
197 input_->SetFrameSlot(i, Memory::uint32_at(tos + i));
202 void Deoptimizer::SetPlatformCompiledStubRegisters(
203 FrameDescription* output_frame, CodeStubInterfaceDescriptor* descriptor) {
205 reinterpret_cast<intptr_t>(descriptor->deoptimization_handler_);
206 int params = descriptor->GetHandlerParameterCount();
207 output_frame->SetRegister(eax.code(), params);
208 output_frame->SetRegister(ebx.code(), handler);
212 void Deoptimizer::CopySIMD128Registers(FrameDescription* output_frame) {
213 if (!CpuFeatures::IsSupported(SSE2)) return;
214 for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
215 simd128_value_t xmm_value = input_->GetSIMD128Register(i);
216 output_frame->SetSIMD128Register(i, xmm_value);
221 bool Deoptimizer::HasAlignmentPadding(JSFunction* function) {
222 int parameter_count = function->shared()->formal_parameter_count() + 1;
223 unsigned input_frame_size = input_->GetFrameSize();
224 unsigned alignment_state_offset =
225 input_frame_size - parameter_count * kPointerSize -
226 StandardFrameConstants::kFixedFrameSize -
228 ASSERT(JavaScriptFrameConstants::kDynamicAlignmentStateOffset ==
229 JavaScriptFrameConstants::kLocal0Offset);
230 int32_t alignment_state = input_->GetFrameSlot(alignment_state_offset);
231 return (alignment_state == kAlignmentPaddingPushed);
235 Code* Deoptimizer::NotifyStubFailureBuiltin() {
236 Builtins::Name name = CpuFeatures::IsSupported(SSE2) ?
237 Builtins::kNotifyStubFailureSaveDoubles : Builtins::kNotifyStubFailure;
238 return isolate_->builtins()->builtin(name);
244 void Deoptimizer::EntryGenerator::Generate() {
247 // Save all general purpose registers before messing with them.
248 const int kNumberOfRegisters = Register::kNumRegisters;
250 const int kXMMRegsSize = kSIMD128Size *
251 XMMRegister::kNumAllocatableRegisters;
252 __ sub(esp, Immediate(kXMMRegsSize));
253 if (CpuFeatures::IsSupported(SSE2)) {
254 CpuFeatureScope scope(masm(), SSE2);
255 for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
256 XMMRegister xmm_reg = XMMRegister::FromAllocationIndex(i);
257 int offset = i * kSIMD128Size;
258 __ movups(Operand(esp, offset), xmm_reg);
264 const int kSavedRegistersAreaSize = kNumberOfRegisters * kPointerSize +
267 // Get the bailout id from the stack.
268 __ mov(ebx, Operand(esp, kSavedRegistersAreaSize));
270 // Get the address of the location in the code object
271 // and compute the fp-to-sp delta in register edx.
272 __ mov(ecx, Operand(esp, kSavedRegistersAreaSize + 1 * kPointerSize));
273 __ lea(edx, Operand(esp, kSavedRegistersAreaSize + 2 * kPointerSize));
278 // Allocate a new deoptimizer object.
279 __ PrepareCallCFunction(6, eax);
280 __ mov(eax, Operand(ebp, JavaScriptFrameConstants::kFunctionOffset));
281 __ mov(Operand(esp, 0 * kPointerSize), eax); // Function.
282 __ mov(Operand(esp, 1 * kPointerSize), Immediate(type())); // Bailout type.
283 __ mov(Operand(esp, 2 * kPointerSize), ebx); // Bailout id.
284 __ mov(Operand(esp, 3 * kPointerSize), ecx); // Code address or 0.
285 __ mov(Operand(esp, 4 * kPointerSize), edx); // Fp-to-sp delta.
286 __ mov(Operand(esp, 5 * kPointerSize),
287 Immediate(ExternalReference::isolate_address(isolate())));
289 AllowExternalCallThatCantCauseGC scope(masm());
290 __ CallCFunction(ExternalReference::new_deoptimizer_function(isolate()), 6);
293 // Preserve deoptimizer object in register eax and get the input
294 // frame descriptor pointer.
295 __ mov(ebx, Operand(eax, Deoptimizer::input_offset()));
297 // Fill in the input registers.
298 for (int i = kNumberOfRegisters - 1; i >= 0; i--) {
299 int offset = (i * kPointerSize) + FrameDescription::registers_offset();
300 __ pop(Operand(ebx, offset));
303 int xmm_regs_offset = FrameDescription::simd128_registers_offset();
304 if (CpuFeatures::IsSupported(SSE2)) {
305 CpuFeatureScope scope(masm(), SSE2);
306 // Fill in the xmm input registers.
307 for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
308 int dst_offset = i * kSIMD128Size + xmm_regs_offset;
309 int src_offset = i * kSIMD128Size;
310 __ movups(xmm0, Operand(esp, src_offset));
311 __ movups(Operand(ebx, dst_offset), xmm0);
315 // Clear FPU all exceptions.
316 // TODO(ulan): Find out why the TOP register is not zero here in some cases,
317 // and check that the generated code never deoptimizes with unbalanced stack.
320 // Remove the bailout id, return address and the double registers.
321 __ add(esp, Immediate(kXMMRegsSize + 2 * kPointerSize));
323 // Compute a pointer to the unwinding limit in register ecx; that is
324 // the first stack slot not part of the input frame.
325 __ mov(ecx, Operand(ebx, FrameDescription::frame_size_offset()));
328 // Unwind the stack down to - but not including - the unwinding
329 // limit and copy the contents of the activation frame to the input
330 // frame description.
331 __ lea(edx, Operand(ebx, FrameDescription::frame_content_offset()));
332 Label pop_loop_header;
333 __ jmp(&pop_loop_header);
336 __ pop(Operand(edx, 0));
337 __ add(edx, Immediate(sizeof(uint32_t)));
338 __ bind(&pop_loop_header);
340 __ j(not_equal, &pop_loop);
342 // Compute the output frame in the deoptimizer.
344 __ PrepareCallCFunction(1, ebx);
345 __ mov(Operand(esp, 0 * kPointerSize), eax);
347 AllowExternalCallThatCantCauseGC scope(masm());
349 ExternalReference::compute_output_frames_function(isolate()), 1);
353 // If frame was dynamically aligned, pop padding.
355 __ cmp(Operand(eax, Deoptimizer::has_alignment_padding_offset()),
357 __ j(equal, &no_padding);
359 if (FLAG_debug_code) {
360 __ cmp(ecx, Immediate(kAlignmentZapValue));
361 __ Assert(equal, kAlignmentMarkerExpected);
363 __ bind(&no_padding);
365 // Replace the current frame with the output frames.
366 Label outer_push_loop, inner_push_loop,
367 outer_loop_header, inner_loop_header;
368 // Outer loop state: eax = current FrameDescription**, edx = one past the
369 // last FrameDescription**.
370 __ mov(edx, Operand(eax, Deoptimizer::output_count_offset()));
371 __ mov(eax, Operand(eax, Deoptimizer::output_offset()));
372 __ lea(edx, Operand(eax, edx, times_4, 0));
373 __ jmp(&outer_loop_header);
374 __ bind(&outer_push_loop);
375 // Inner loop state: ebx = current FrameDescription*, ecx = loop index.
376 __ mov(ebx, Operand(eax, 0));
377 __ mov(ecx, Operand(ebx, FrameDescription::frame_size_offset()));
378 __ jmp(&inner_loop_header);
379 __ bind(&inner_push_loop);
380 __ sub(ecx, Immediate(sizeof(uint32_t)));
381 __ push(Operand(ebx, ecx, times_1, FrameDescription::frame_content_offset()));
382 __ bind(&inner_loop_header);
384 __ j(not_zero, &inner_push_loop);
385 __ add(eax, Immediate(kPointerSize));
386 __ bind(&outer_loop_header);
388 __ j(below, &outer_push_loop);
390 // In case of a failed STUB, we have to restore the XMM registers.
391 if (CpuFeatures::IsSupported(SSE2)) {
392 CpuFeatureScope scope(masm(), SSE2);
393 for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
394 XMMRegister xmm_reg = XMMRegister::FromAllocationIndex(i);
395 int src_offset = i * kSIMD128Size + xmm_regs_offset;
396 __ movups(xmm_reg, Operand(ebx, src_offset));
400 // Push state, pc, and continuation from the last output frame.
401 __ push(Operand(ebx, FrameDescription::state_offset()));
402 __ push(Operand(ebx, FrameDescription::pc_offset()));
403 __ push(Operand(ebx, FrameDescription::continuation_offset()));
406 // Push the registers from the last output frame.
407 for (int i = 0; i < kNumberOfRegisters; i++) {
408 int offset = (i * kPointerSize) + FrameDescription::registers_offset();
409 __ push(Operand(ebx, offset));
412 // Restore the registers from the stack.
415 // Return to the continuation point.
420 void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
421 // Create a sequence of deoptimization entries.
423 for (int i = 0; i < count(); i++) {
424 int start = masm()->pc_offset();
428 ASSERT(masm()->pc_offset() - start == table_entry_size_);
434 void FrameDescription::SetCallerPc(unsigned offset, intptr_t value) {
435 SetFrameSlot(offset, value);
439 void FrameDescription::SetCallerFp(unsigned offset, intptr_t value) {
440 SetFrameSlot(offset, value);
444 void FrameDescription::SetCallerConstantPool(unsigned offset, intptr_t value) {
445 // No out-of-line constant pool support.
450 double FrameDescription::GetDoubleRegister(unsigned n) const {
451 ASSERT(n < ARRAY_SIZE(simd128_registers_));
452 return simd128_registers_[n].d[0];
456 void FrameDescription::SetDoubleRegister(unsigned n, double value) {
457 ASSERT(n < ARRAY_SIZE(simd128_registers_));
458 simd128_registers_[n].d[0] = value;
465 } } // namespace v8::internal
467 #endif // V8_TARGET_ARCH_IA32