1 // Copyright 2013 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.
8 #include "deoptimizer.h"
9 #include "full-codegen.h"
10 #include "safepoint-table.h"
17 int Deoptimizer::patch_size() {
18 // Size of the code used to patch lazy bailout points.
19 // Patching is done by Deoptimizer::DeoptimizeFunction.
20 return 4 * kInstructionSize;
25 void Deoptimizer::PatchCodeForDeoptimization(Isolate* isolate, Code* code) {
26 // Invalidate the relocation information, as it will become invalid by the
27 // code patching below, and is not needed any more.
28 code->InvalidateRelocation();
30 // TODO(jkummerow): if (FLAG_zap_code_space), make the code object's
31 // entry sequence unusable (see other architectures).
33 DeoptimizationInputData* deopt_data =
34 DeoptimizationInputData::cast(code->deoptimization_data());
35 SharedFunctionInfo* shared =
36 SharedFunctionInfo::cast(deopt_data->SharedFunctionInfo());
37 shared->EvictFromOptimizedCodeMap(code, "deoptimized code");
38 Address code_start_address = code->instruction_start();
40 Address prev_call_address = NULL;
42 // For each LLazyBailout instruction insert a call to the corresponding
43 // deoptimization entry.
44 for (int i = 0; i < deopt_data->DeoptCount(); i++) {
45 if (deopt_data->Pc(i)->value() == -1) continue;
47 Address call_address = code_start_address + deopt_data->Pc(i)->value();
48 Address deopt_entry = GetDeoptimizationEntry(isolate, i, LAZY);
50 PatchingAssembler patcher(call_address, patch_size() / kInstructionSize);
51 patcher.LoadLiteral(ip0, 2 * kInstructionSize);
53 patcher.dc64(reinterpret_cast<intptr_t>(deopt_entry));
55 ASSERT((prev_call_address == NULL) ||
56 (call_address >= prev_call_address + patch_size()));
57 ASSERT(call_address + patch_size() <= code->instruction_end());
59 prev_call_address = call_address;
65 void Deoptimizer::FillInputFrame(Address tos, JavaScriptFrame* frame) {
66 // Set the register values. The values are not important as there are no
67 // callee saved registers in JavaScript frames, so all registers are
68 // spilled. Registers fp and sp are set to the correct values though.
69 for (int i = 0; i < Register::NumRegisters(); i++) {
70 input_->SetRegister(i, 0);
73 // TODO(all): Do we also need to set a value to csp?
74 input_->SetRegister(jssp.code(), reinterpret_cast<intptr_t>(frame->sp()));
75 input_->SetRegister(fp.code(), reinterpret_cast<intptr_t>(frame->fp()));
77 for (int i = 0; i < DoubleRegister::NumAllocatableRegisters(); i++) {
78 input_->SetDoubleRegister(i, 0.0);
81 // Fill the frame content from the actual data on the frame.
82 for (unsigned i = 0; i < input_->GetFrameSize(); i += kPointerSize) {
83 input_->SetFrameSlot(i, Memory::uint64_at(tos + i));
88 bool Deoptimizer::HasAlignmentPadding(JSFunction* function) {
89 // There is no dynamic alignment padding on ARM64 in the input frame.
94 void Deoptimizer::SetPlatformCompiledStubRegisters(
95 FrameDescription* output_frame, CodeStubInterfaceDescriptor* descriptor) {
96 ApiFunction function(descriptor->deoptimization_handler_);
97 ExternalReference xref(&function, ExternalReference::BUILTIN_CALL, isolate_);
98 intptr_t handler = reinterpret_cast<intptr_t>(xref.address());
99 int params = descriptor->GetHandlerParameterCount();
100 output_frame->SetRegister(x0.code(), params);
101 output_frame->SetRegister(x1.code(), handler);
105 void Deoptimizer::CopyDoubleRegisters(FrameDescription* output_frame) {
106 for (int i = 0; i < DoubleRegister::kMaxNumRegisters; ++i) {
107 double double_value = input_->GetDoubleRegister(i);
108 output_frame->SetDoubleRegister(i, double_value);
113 Code* Deoptimizer::NotifyStubFailureBuiltin() {
114 return isolate_->builtins()->builtin(Builtins::kNotifyStubFailureSaveDoubles);
120 static void CopyRegisterDumpToFrame(MacroAssembler* masm,
127 int offset0, offset1;
128 CPURegList copy_to_input = reg_list;
129 int reg_count = reg_list.Count();
130 int reg_size = reg_list.RegisterSizeInBytes();
131 for (int i = 0; i < (reg_count / 2); i++) {
132 __ PeekPair(scratch1, scratch2, src_offset + (i * reg_size * 2));
134 offset0 = (copy_to_input.PopLowestIndex().code() * reg_size) + dst_offset;
135 offset1 = (copy_to_input.PopLowestIndex().code() * reg_size) + dst_offset;
137 if ((offset0 + reg_size) == offset1) {
138 // Registers are adjacent: store in pairs.
139 __ Stp(scratch1, scratch2, MemOperand(frame, offset0));
141 // Registers are not adjacent: store individually.
142 __ Str(scratch1, MemOperand(frame, offset0));
143 __ Str(scratch2, MemOperand(frame, offset1));
146 if ((reg_count & 1) != 0) {
147 __ Peek(scratch1, src_offset + (reg_count - 1) * reg_size);
148 offset0 = (copy_to_input.PopLowestIndex().code() * reg_size) + dst_offset;
149 __ Str(scratch1, MemOperand(frame, offset0));
157 void Deoptimizer::EntryGenerator::Generate() {
160 // TODO(all): This code needs to be revisited. We probably only need to save
161 // caller-saved registers here. Callee-saved registers can be stored directly
162 // in the input frame.
164 // Save all allocatable floating point registers.
165 CPURegList saved_fp_registers(CPURegister::kFPRegister, kDRegSizeInBits,
166 FPRegister::kAllocatableFPRegisters);
167 __ PushCPURegList(saved_fp_registers);
169 // We save all the registers expcept jssp, sp and lr.
170 CPURegList saved_registers(CPURegister::kRegister, kXRegSizeInBits, 0, 27);
171 saved_registers.Combine(fp);
172 __ PushCPURegList(saved_registers);
174 const int kSavedRegistersAreaSize =
175 (saved_registers.Count() * kXRegSize) +
176 (saved_fp_registers.Count() * kDRegSize);
178 // Floating point registers are saved on the stack above core registers.
179 const int kFPRegistersOffset = saved_registers.Count() * kXRegSize;
181 // Get the bailout id from the stack.
182 Register bailout_id = x2;
183 __ Peek(bailout_id, kSavedRegistersAreaSize);
185 Register code_object = x3;
186 Register fp_to_sp = x4;
187 // Get the address of the location in the code object. This is the return
188 // address for lazy deoptimization.
189 __ Mov(code_object, lr);
190 // Compute the fp-to-sp delta, and correct one word for bailout id.
191 __ Add(fp_to_sp, masm()->StackPointer(),
192 kSavedRegistersAreaSize + (1 * kPointerSize));
193 __ Sub(fp_to_sp, fp, fp_to_sp);
195 // Allocate a new deoptimizer object.
196 __ Ldr(x0, MemOperand(fp, JavaScriptFrameConstants::kFunctionOffset));
198 // Following arguments are already loaded:
200 // - x3: code object address
201 // - x4: fp-to-sp delta
202 __ Mov(x5, ExternalReference::isolate_address(isolate()));
205 // Call Deoptimizer::New().
206 AllowExternalCallThatCantCauseGC scope(masm());
207 __ CallCFunction(ExternalReference::new_deoptimizer_function(isolate()), 6);
210 // Preserve "deoptimizer" object in register x0.
211 Register deoptimizer = x0;
213 // Get the input frame descriptor pointer.
214 __ Ldr(x1, MemOperand(deoptimizer, Deoptimizer::input_offset()));
216 // Copy core registers into the input frame.
217 CopyRegisterDumpToFrame(masm(), x1, saved_registers, x2, x4, 0,
218 FrameDescription::registers_offset());
220 // Copy FP registers to the input frame.
221 CopyRegisterDumpToFrame(masm(), x1, saved_fp_registers, x2, x4,
223 FrameDescription::double_registers_offset());
225 // Remove the bailout id and the saved registers from the stack.
226 __ Drop(1 + (kSavedRegistersAreaSize / kXRegSize));
228 // Compute a pointer to the unwinding limit in register x2; that is
229 // the first stack slot not part of the input frame.
230 Register unwind_limit = x2;
231 __ Ldr(unwind_limit, MemOperand(x1, FrameDescription::frame_size_offset()));
232 __ Add(unwind_limit, unwind_limit, __ StackPointer());
234 // Unwind the stack down to - but not including - the unwinding
235 // limit and copy the contents of the activation frame to the input
236 // frame description.
237 __ Add(x3, x1, FrameDescription::frame_content_offset());
239 Label pop_loop_header;
240 __ B(&pop_loop_header);
243 __ Str(x4, MemOperand(x3, kPointerSize, PostIndex));
244 __ Bind(&pop_loop_header);
245 __ Cmp(unwind_limit, __ StackPointer());
248 // Compute the output frame in the deoptimizer.
249 __ Push(x0); // Preserve deoptimizer object across call.
252 // Call Deoptimizer::ComputeOutputFrames().
253 AllowExternalCallThatCantCauseGC scope(masm());
255 ExternalReference::compute_output_frames_function(isolate()), 1);
257 __ Pop(x4); // Restore deoptimizer object (class Deoptimizer).
259 // Replace the current (input) frame with the output frames.
260 Label outer_push_loop, inner_push_loop,
261 outer_loop_header, inner_loop_header;
262 __ Ldrsw(x1, MemOperand(x4, Deoptimizer::output_count_offset()));
263 __ Ldr(x0, MemOperand(x4, Deoptimizer::output_offset()));
264 __ Add(x1, x0, Operand(x1, LSL, kPointerSizeLog2));
265 __ B(&outer_loop_header);
267 __ Bind(&outer_push_loop);
268 Register current_frame = x2;
269 __ Ldr(current_frame, MemOperand(x0, 0));
270 __ Ldr(x3, MemOperand(current_frame, FrameDescription::frame_size_offset()));
271 __ B(&inner_loop_header);
273 __ Bind(&inner_push_loop);
274 __ Sub(x3, x3, kPointerSize);
275 __ Add(x6, current_frame, x3);
276 __ Ldr(x7, MemOperand(x6, FrameDescription::frame_content_offset()));
278 __ Bind(&inner_loop_header);
279 __ Cbnz(x3, &inner_push_loop);
281 __ Add(x0, x0, kPointerSize);
282 __ Bind(&outer_loop_header);
284 __ B(lt, &outer_push_loop);
286 __ Ldr(x1, MemOperand(x4, Deoptimizer::input_offset()));
287 ASSERT(!saved_fp_registers.IncludesAliasOf(crankshaft_fp_scratch) &&
288 !saved_fp_registers.IncludesAliasOf(fp_zero) &&
289 !saved_fp_registers.IncludesAliasOf(fp_scratch));
290 int src_offset = FrameDescription::double_registers_offset();
291 while (!saved_fp_registers.IsEmpty()) {
292 const CPURegister reg = saved_fp_registers.PopLowestIndex();
293 __ Ldr(reg, MemOperand(x1, src_offset));
294 src_offset += kDoubleSize;
297 // Push state from the last output frame.
298 __ Ldr(x6, MemOperand(current_frame, FrameDescription::state_offset()));
301 // TODO(all): ARM copies a lot (if not all) of the last output frame onto the
302 // stack, then pops it all into registers. Here, we try to load it directly
303 // into the relevant registers. Is this correct? If so, we should improve the
306 // TODO(all): This code needs to be revisited, We probably don't need to
307 // restore all the registers as fullcodegen does not keep live values in
308 // registers (note that at least fp must be restored though).
310 // Restore registers from the last output frame.
311 // Note that lr is not in the list of saved_registers and will be restored
312 // later. We can use it to hold the address of last output frame while
313 // reloading the other registers.
314 ASSERT(!saved_registers.IncludesAliasOf(lr));
315 Register last_output_frame = lr;
316 __ Mov(last_output_frame, current_frame);
318 // We don't need to restore x7 as it will be clobbered later to hold the
319 // continuation address.
320 Register continuation = x7;
321 saved_registers.Remove(continuation);
323 while (!saved_registers.IsEmpty()) {
324 // TODO(all): Look for opportunities to optimize this by using ldp.
325 CPURegister current_reg = saved_registers.PopLowestIndex();
326 int offset = (current_reg.code() * kPointerSize) +
327 FrameDescription::registers_offset();
328 __ Ldr(current_reg, MemOperand(last_output_frame, offset));
331 __ Ldr(continuation, MemOperand(last_output_frame,
332 FrameDescription::continuation_offset()));
333 __ Ldr(lr, MemOperand(last_output_frame, FrameDescription::pc_offset()));
334 __ InitializeRootRegister();
339 // Size of an entry of the second level deopt table.
340 // This is the code size generated by GeneratePrologue for one entry.
341 const int Deoptimizer::table_entry_size_ = 2 * kInstructionSize;
344 void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
345 UseScratchRegisterScope temps(masm());
346 Register entry_id = temps.AcquireX();
348 // Create a sequence of deoptimization entries.
349 // Note that registers are still live when jumping to an entry.
352 InstructionAccurateScope scope(masm());
354 // The number of entry will never exceed kMaxNumberOfEntries.
355 // As long as kMaxNumberOfEntries is a valid 16 bits immediate you can use
356 // a movz instruction to load the entry id.
357 ASSERT(is_uint16(Deoptimizer::kMaxNumberOfEntries));
359 for (int i = 0; i < count(); i++) {
360 int start = masm()->pc_offset();
362 __ movz(entry_id, i);
364 ASSERT(masm()->pc_offset() - start == table_entry_size_);
372 void FrameDescription::SetCallerPc(unsigned offset, intptr_t value) {
373 SetFrameSlot(offset, value);
377 void FrameDescription::SetCallerFp(unsigned offset, intptr_t value) {
378 SetFrameSlot(offset, value);
382 void FrameDescription::SetCallerConstantPool(unsigned offset, intptr_t value) {
383 // No out-of-line constant pool support.
390 } } // namespace v8::internal