2 // Copyright 2011 the V8 project authors. All rights reserved.
3 // Use of this source code is governed by a BSD-style license that can be
4 // found in the LICENSE file.
8 #include "src/codegen.h"
9 #include "src/deoptimizer.h"
10 #include "src/full-codegen.h"
11 #include "src/safepoint-table.h"
17 int Deoptimizer::patch_size() {
18 const int kCallInstructionSizeInWords = 4;
19 return kCallInstructionSizeInWords * Assembler::kInstrSize;
23 void Deoptimizer::PatchCodeForDeoptimization(Isolate* isolate, Code* code) {
24 Address code_start_address = code->instruction_start();
25 // Invalidate the relocation information, as it will become invalid by the
26 // code patching below, and is not needed any more.
27 code->InvalidateRelocation();
29 if (FLAG_zap_code_space) {
30 // Fail hard and early if we enter this code object again.
31 byte* pointer = code->FindCodeAgeSequence();
32 if (pointer != NULL) {
33 pointer += kNoCodeAgeSequenceLength;
35 pointer = code->instruction_start();
37 CodePatcher patcher(pointer, 1);
38 patcher.masm()->break_(0xCC);
40 DeoptimizationInputData* data =
41 DeoptimizationInputData::cast(code->deoptimization_data());
42 int osr_offset = data->OsrPcOffset()->value();
44 CodePatcher osr_patcher(code->instruction_start() + osr_offset, 1);
45 osr_patcher.masm()->break_(0xCC);
49 DeoptimizationInputData* deopt_data =
50 DeoptimizationInputData::cast(code->deoptimization_data());
52 Address prev_call_address = NULL;
54 // For each LLazyBailout instruction insert a call to the corresponding
55 // deoptimization entry.
56 for (int i = 0; i < deopt_data->DeoptCount(); i++) {
57 if (deopt_data->Pc(i)->value() == -1) continue;
58 Address call_address = code_start_address + deopt_data->Pc(i)->value();
59 Address deopt_entry = GetDeoptimizationEntry(isolate, i, LAZY);
60 int call_size_in_bytes = MacroAssembler::CallSize(deopt_entry,
62 int call_size_in_words = call_size_in_bytes / Assembler::kInstrSize;
63 DCHECK(call_size_in_bytes % Assembler::kInstrSize == 0);
64 DCHECK(call_size_in_bytes <= patch_size());
65 CodePatcher patcher(call_address, call_size_in_words);
66 patcher.masm()->Call(deopt_entry, RelocInfo::NONE32);
67 DCHECK(prev_call_address == NULL ||
68 call_address >= prev_call_address + patch_size());
69 DCHECK(call_address + patch_size() <= code->instruction_end());
72 prev_call_address = call_address;
78 void Deoptimizer::FillInputFrame(Address tos, JavaScriptFrame* frame) {
79 // Set the register values. The values are not important as there are no
80 // callee saved registers in JavaScript frames, so all registers are
81 // spilled. Registers fp and sp are set to the correct values though.
83 for (int i = 0; i < Register::kNumRegisters; i++) {
84 input_->SetRegister(i, i * 4);
86 input_->SetRegister(sp.code(), reinterpret_cast<intptr_t>(frame->sp()));
87 input_->SetRegister(fp.code(), reinterpret_cast<intptr_t>(frame->fp()));
88 for (int i = 0; i < DoubleRegister::NumAllocatableRegisters(); i++) {
89 input_->SetDoubleRegister(i, 0.0);
92 // Fill the frame content from the actual data on the frame.
93 for (unsigned i = 0; i < input_->GetFrameSize(); i += kPointerSize) {
94 input_->SetFrameSlot(i, Memory::uint32_at(tos + i));
99 void Deoptimizer::SetPlatformCompiledStubRegisters(
100 FrameDescription* output_frame, CodeStubDescriptor* descriptor) {
101 ApiFunction function(descriptor->deoptimization_handler());
102 ExternalReference xref(&function, ExternalReference::BUILTIN_CALL, isolate_);
103 intptr_t handler = reinterpret_cast<intptr_t>(xref.address());
104 int params = descriptor->GetHandlerParameterCount();
105 output_frame->SetRegister(a0.code(), params);
106 output_frame->SetRegister(a1.code(), handler);
110 void Deoptimizer::CopyDoubleRegisters(FrameDescription* output_frame) {
111 for (int i = 0; i < DoubleRegister::kMaxNumRegisters; ++i) {
112 double double_value = input_->GetDoubleRegister(i);
113 output_frame->SetDoubleRegister(i, double_value);
118 bool Deoptimizer::HasAlignmentPadding(JSFunction* function) {
119 // There is no dynamic alignment padding on MIPS in the input frame.
127 // This code tries to be close to ia32 code so that any changes can be
129 void Deoptimizer::EntryGenerator::Generate() {
132 // Unlike on ARM we don't save all the registers, just the useful ones.
133 // For the rest, there are gaps on the stack, so the offsets remain the same.
134 const int kNumberOfRegisters = Register::kNumRegisters;
136 RegList restored_regs = kJSCallerSaved | kCalleeSaved;
137 RegList saved_regs = restored_regs | sp.bit() | ra.bit();
139 const int kDoubleRegsSize =
140 kDoubleSize * FPURegister::kMaxNumAllocatableRegisters;
142 // Save all FPU registers before messing with them.
143 __ Subu(sp, sp, Operand(kDoubleRegsSize));
144 for (int i = 0; i < FPURegister::kMaxNumAllocatableRegisters; ++i) {
145 FPURegister fpu_reg = FPURegister::FromAllocationIndex(i);
146 int offset = i * kDoubleSize;
147 __ sdc1(fpu_reg, MemOperand(sp, offset));
150 // Push saved_regs (needed to populate FrameDescription::registers_).
151 // Leave gaps for other registers.
152 __ Subu(sp, sp, kNumberOfRegisters * kPointerSize);
153 for (int16_t i = kNumberOfRegisters - 1; i >= 0; i--) {
154 if ((saved_regs & (1 << i)) != 0) {
155 __ sw(ToRegister(i), MemOperand(sp, kPointerSize * i));
159 const int kSavedRegistersAreaSize =
160 (kNumberOfRegisters * kPointerSize) + kDoubleRegsSize;
162 // Get the bailout id from the stack.
163 __ lw(a2, MemOperand(sp, kSavedRegistersAreaSize));
165 // Get the address of the location in the code object (a3) (return
166 // address for lazy deoptimization) and compute the fp-to-sp delta in
169 // Correct one word for bailout id.
170 __ Addu(t0, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
174 // Allocate a new deoptimizer object.
175 // Pass four arguments in a0 to a3 and fifth & sixth arguments on stack.
176 __ PrepareCallCFunction(6, t1);
177 __ lw(a0, MemOperand(fp, JavaScriptFrameConstants::kFunctionOffset));
178 __ li(a1, Operand(type())); // bailout type,
179 // a2: bailout id already loaded.
180 // a3: code address or 0 already loaded.
181 __ sw(t0, CFunctionArgumentOperand(5)); // Fp-to-sp delta.
182 __ li(t1, Operand(ExternalReference::isolate_address(isolate())));
183 __ sw(t1, CFunctionArgumentOperand(6)); // Isolate.
184 // Call Deoptimizer::New().
186 AllowExternalCallThatCantCauseGC scope(masm());
187 __ CallCFunction(ExternalReference::new_deoptimizer_function(isolate()), 6);
190 // Preserve "deoptimizer" object in register v0 and get the input
191 // frame descriptor pointer to a1 (deoptimizer->input_);
192 // Move deopt-obj to a0 for call to Deoptimizer::ComputeOutputFrames() below.
194 __ lw(a1, MemOperand(v0, Deoptimizer::input_offset()));
196 // Copy core registers into FrameDescription::registers_[kNumRegisters].
197 DCHECK(Register::kNumRegisters == kNumberOfRegisters);
198 for (int i = 0; i < kNumberOfRegisters; i++) {
199 int offset = (i * kPointerSize) + FrameDescription::registers_offset();
200 if ((saved_regs & (1 << i)) != 0) {
201 __ lw(a2, MemOperand(sp, i * kPointerSize));
202 __ sw(a2, MemOperand(a1, offset));
203 } else if (FLAG_debug_code) {
204 __ li(a2, kDebugZapValue);
205 __ sw(a2, MemOperand(a1, offset));
209 int double_regs_offset = FrameDescription::double_registers_offset();
210 // Copy FPU registers to
211 // double_registers_[DoubleRegister::kNumAllocatableRegisters]
212 for (int i = 0; i < FPURegister::NumAllocatableRegisters(); ++i) {
213 int dst_offset = i * kDoubleSize + double_regs_offset;
214 int src_offset = i * kDoubleSize + kNumberOfRegisters * kPointerSize;
215 __ ldc1(f0, MemOperand(sp, src_offset));
216 __ sdc1(f0, MemOperand(a1, dst_offset));
219 // Remove the bailout id and the saved registers from the stack.
220 __ Addu(sp, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
222 // Compute a pointer to the unwinding limit in register a2; that is
223 // the first stack slot not part of the input frame.
224 __ lw(a2, MemOperand(a1, FrameDescription::frame_size_offset()));
227 // Unwind the stack down to - but not including - the unwinding
228 // limit and copy the contents of the activation frame to the input
229 // frame description.
230 __ Addu(a3, a1, Operand(FrameDescription::frame_content_offset()));
232 Label pop_loop_header;
233 __ BranchShort(&pop_loop_header);
236 __ sw(t0, MemOperand(a3, 0));
237 __ addiu(a3, a3, sizeof(uint32_t));
238 __ bind(&pop_loop_header);
239 __ BranchShort(&pop_loop, ne, a2, Operand(sp));
241 // Compute the output frame in the deoptimizer.
242 __ push(a0); // Preserve deoptimizer object across call.
243 // a0: deoptimizer object; a1: scratch.
244 __ PrepareCallCFunction(1, a1);
245 // Call Deoptimizer::ComputeOutputFrames().
247 AllowExternalCallThatCantCauseGC scope(masm());
249 ExternalReference::compute_output_frames_function(isolate()), 1);
251 __ pop(a0); // Restore deoptimizer object (class Deoptimizer).
253 // Replace the current (input) frame with the output frames.
254 Label outer_push_loop, inner_push_loop,
255 outer_loop_header, inner_loop_header;
256 // Outer loop state: t0 = current "FrameDescription** output_",
257 // a1 = one past the last FrameDescription**.
258 __ lw(a1, MemOperand(a0, Deoptimizer::output_count_offset()));
259 __ lw(t0, MemOperand(a0, Deoptimizer::output_offset())); // t0 is output_.
260 __ sll(a1, a1, kPointerSizeLog2); // Count to offset.
261 __ addu(a1, t0, a1); // a1 = one past the last FrameDescription**.
262 __ jmp(&outer_loop_header);
263 __ bind(&outer_push_loop);
264 // Inner loop state: a2 = current FrameDescription*, a3 = loop index.
265 __ lw(a2, MemOperand(t0, 0)); // output_[ix]
266 __ lw(a3, MemOperand(a2, FrameDescription::frame_size_offset()));
267 __ jmp(&inner_loop_header);
268 __ bind(&inner_push_loop);
269 __ Subu(a3, a3, Operand(sizeof(uint32_t)));
270 __ Addu(t2, a2, Operand(a3));
271 __ lw(t3, MemOperand(t2, FrameDescription::frame_content_offset()));
273 __ bind(&inner_loop_header);
274 __ BranchShort(&inner_push_loop, ne, a3, Operand(zero_reg));
276 __ Addu(t0, t0, Operand(kPointerSize));
277 __ bind(&outer_loop_header);
278 __ BranchShort(&outer_push_loop, lt, t0, Operand(a1));
280 __ lw(a1, MemOperand(a0, Deoptimizer::input_offset()));
281 for (int i = 0; i < FPURegister::kMaxNumAllocatableRegisters; ++i) {
282 const FPURegister fpu_reg = FPURegister::FromAllocationIndex(i);
283 int src_offset = i * kDoubleSize + double_regs_offset;
284 __ ldc1(fpu_reg, MemOperand(a1, src_offset));
287 // Push state, pc, and continuation from the last output frame.
288 __ lw(t2, MemOperand(a2, FrameDescription::state_offset()));
291 __ lw(t2, MemOperand(a2, FrameDescription::pc_offset()));
293 __ lw(t2, MemOperand(a2, FrameDescription::continuation_offset()));
297 // Technically restoring 'at' should work unless zero_reg is also restored
298 // but it's safer to check for this.
299 DCHECK(!(at.bit() & restored_regs));
300 // Restore the registers from the last output frame.
302 for (int i = kNumberOfRegisters - 1; i >= 0; i--) {
303 int offset = (i * kPointerSize) + FrameDescription::registers_offset();
304 if ((restored_regs & (1 << i)) != 0) {
305 __ lw(ToRegister(i), MemOperand(at, offset));
309 __ InitializeRootRegister();
311 __ pop(at); // Get continuation, leave pc on stack.
314 __ stop("Unreachable.");
318 // Maximum size of a table entry generated below.
319 const int Deoptimizer::table_entry_size_ = 2 * Assembler::kInstrSize;
321 void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
322 Assembler::BlockTrampolinePoolScope block_trampoline_pool(masm());
324 // Create a sequence of deoptimization entries.
325 // Note that registers are still live when jumping to an entry.
326 Label table_start, done, done_special, trampoline_jump;
327 __ bind(&table_start);
328 int kMaxEntriesBranchReach = (1 << (kImm16Bits - 2))/
329 (table_entry_size_ / Assembler::kInstrSize);
331 if (count() <= kMaxEntriesBranchReach) {
333 for (int i = 0; i < count(); i++) {
337 __ Branch(USE_DELAY_SLOT, &done); // Expose delay slot.
338 __ li(at, i); // In the delay slot.
340 DCHECK_EQ(table_entry_size_, masm()->SizeOfCodeGeneratedSince(&start));
343 DCHECK_EQ(masm()->SizeOfCodeGeneratedSince(&table_start),
344 count() * table_entry_size_);
348 // Uncommon case, the branch cannot reach.
349 // Create mini trampoline and adjust id constants to get proper value at
351 for (int i = kMaxEntriesBranchReach; i > 1; i--) {
355 __ Branch(USE_DELAY_SLOT, &trampoline_jump); // Expose delay slot.
356 __ li(at, - i); // In the delay slot.
357 DCHECK_EQ(table_entry_size_, masm()->SizeOfCodeGeneratedSince(&start));
359 // Entry with id == kMaxEntriesBranchReach - 1.
360 __ bind(&trampoline_jump);
361 __ Branch(USE_DELAY_SLOT, &done_special);
364 for (int i = kMaxEntriesBranchReach ; i < count(); i++) {
368 __ Branch(USE_DELAY_SLOT, &done); // Expose delay slot.
369 __ li(at, i); // In the delay slot.
372 DCHECK_EQ(masm()->SizeOfCodeGeneratedSince(&table_start),
373 count() * table_entry_size_);
374 __ bind(&done_special);
375 __ addiu(at, at, kMaxEntriesBranchReach);
382 void FrameDescription::SetCallerPc(unsigned offset, intptr_t value) {
383 SetFrameSlot(offset, value);
387 void FrameDescription::SetCallerFp(unsigned offset, intptr_t value) {
388 SetFrameSlot(offset, value);
392 void FrameDescription::SetCallerConstantPool(unsigned offset, intptr_t value) {
393 // No out-of-line constant pool support.
401 } } // namespace v8::internal