1 // Copyright 2011 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/codegen.h"
6 #include "src/deoptimizer.h"
7 #include "src/full-codegen/full-codegen.h"
8 #include "src/register-configuration.h"
9 #include "src/safepoint-table.h"
10
11 namespace v8 {
12 namespace internal {
13
14
patch_size()15 int Deoptimizer::patch_size() {
16 const int kCallInstructionSizeInWords = 4;
17 return kCallInstructionSizeInWords * Assembler::kInstrSize;
18 }
19
20
EnsureRelocSpaceForLazyDeoptimization(Handle<Code> code)21 void Deoptimizer::EnsureRelocSpaceForLazyDeoptimization(Handle<Code> code) {
22 // Empty because there is no need for relocation information for the code
23 // patching in Deoptimizer::PatchCodeForDeoptimization below.
24 }
25
26
PatchCodeForDeoptimization(Isolate * isolate,Code * code)27 void Deoptimizer::PatchCodeForDeoptimization(Isolate* isolate, Code* code) {
28 Address code_start_address = code->instruction_start();
29 // Invalidate the relocation information, as it will become invalid by the
30 // code patching below, and is not needed any more.
31 code->InvalidateRelocation();
32
33 if (FLAG_zap_code_space) {
34 // Fail hard and early if we enter this code object again.
35 byte* pointer = code->FindCodeAgeSequence();
36 if (pointer != NULL) {
37 pointer += kNoCodeAgeSequenceLength;
38 } else {
39 pointer = code->instruction_start();
40 }
41 CodePatcher patcher(isolate, pointer, 1);
42 patcher.masm()->break_(0xCC);
43
44 DeoptimizationInputData* data =
45 DeoptimizationInputData::cast(code->deoptimization_data());
46 int osr_offset = data->OsrPcOffset()->value();
47 if (osr_offset > 0) {
48 CodePatcher osr_patcher(isolate, code->instruction_start() + osr_offset,
49 1);
50 osr_patcher.masm()->break_(0xCC);
51 }
52 }
53
54 DeoptimizationInputData* deopt_data =
55 DeoptimizationInputData::cast(code->deoptimization_data());
56 #ifdef DEBUG
57 Address prev_call_address = NULL;
58 #endif
59 // For each LLazyBailout instruction insert a call to the corresponding
60 // deoptimization entry.
61 for (int i = 0; i < deopt_data->DeoptCount(); i++) {
62 if (deopt_data->Pc(i)->value() == -1) continue;
63 Address call_address = code_start_address + deopt_data->Pc(i)->value();
64 Address deopt_entry = GetDeoptimizationEntry(isolate, i, LAZY);
65 int call_size_in_bytes = MacroAssembler::CallSize(deopt_entry,
66 RelocInfo::NONE32);
67 int call_size_in_words = call_size_in_bytes / Assembler::kInstrSize;
68 DCHECK(call_size_in_bytes % Assembler::kInstrSize == 0);
69 DCHECK(call_size_in_bytes <= patch_size());
70 CodePatcher patcher(isolate, call_address, call_size_in_words);
71 patcher.masm()->Call(deopt_entry, RelocInfo::NONE32);
72 DCHECK(prev_call_address == NULL ||
73 call_address >= prev_call_address + patch_size());
74 DCHECK(call_address + patch_size() <= code->instruction_end());
75
76 #ifdef DEBUG
77 prev_call_address = call_address;
78 #endif
79 }
80 }
81
82
SetPlatformCompiledStubRegisters(FrameDescription * output_frame,CodeStubDescriptor * descriptor)83 void Deoptimizer::SetPlatformCompiledStubRegisters(
84 FrameDescription* output_frame, CodeStubDescriptor* descriptor) {
85 ApiFunction function(descriptor->deoptimization_handler());
86 ExternalReference xref(&function, ExternalReference::BUILTIN_CALL, isolate_);
87 intptr_t handler = reinterpret_cast<intptr_t>(xref.address());
88 int params = descriptor->GetHandlerParameterCount();
89 output_frame->SetRegister(a0.code(), params);
90 output_frame->SetRegister(a1.code(), handler);
91 }
92
93
CopyDoubleRegisters(FrameDescription * output_frame)94 void Deoptimizer::CopyDoubleRegisters(FrameDescription* output_frame) {
95 for (int i = 0; i < DoubleRegister::kMaxNumRegisters; ++i) {
96 double double_value = input_->GetDoubleRegister(i);
97 output_frame->SetDoubleRegister(i, double_value);
98 }
99 }
100
101 #define __ masm()->
102
103
104 // This code tries to be close to ia32 code so that any changes can be
105 // easily ported.
Generate()106 void Deoptimizer::TableEntryGenerator::Generate() {
107 GeneratePrologue();
108
109 // Unlike on ARM we don't save all the registers, just the useful ones.
110 // For the rest, there are gaps on the stack, so the offsets remain the same.
111 const int kNumberOfRegisters = Register::kNumRegisters;
112
113 RegList restored_regs = kJSCallerSaved | kCalleeSaved;
114 RegList saved_regs = restored_regs | sp.bit() | ra.bit();
115
116 const int kDoubleRegsSize = kDoubleSize * DoubleRegister::kMaxNumRegisters;
117
118 // Save all FPU registers before messing with them.
119 __ Subu(sp, sp, Operand(kDoubleRegsSize));
120 const RegisterConfiguration* config = RegisterConfiguration::Crankshaft();
121 for (int i = 0; i < config->num_allocatable_double_registers(); ++i) {
122 int code = config->GetAllocatableDoubleCode(i);
123 const DoubleRegister fpu_reg = DoubleRegister::from_code(code);
124 int offset = code * kDoubleSize;
125 __ sdc1(fpu_reg, MemOperand(sp, offset));
126 }
127
128 // Push saved_regs (needed to populate FrameDescription::registers_).
129 // Leave gaps for other registers.
130 __ Subu(sp, sp, kNumberOfRegisters * kPointerSize);
131 for (int16_t i = kNumberOfRegisters - 1; i >= 0; i--) {
132 if ((saved_regs & (1 << i)) != 0) {
133 __ sw(ToRegister(i), MemOperand(sp, kPointerSize * i));
134 }
135 }
136
137 __ li(a2, Operand(ExternalReference(Isolate::kCEntryFPAddress, isolate())));
138 __ sw(fp, MemOperand(a2));
139
140 const int kSavedRegistersAreaSize =
141 (kNumberOfRegisters * kPointerSize) + kDoubleRegsSize;
142
143 // Get the bailout id from the stack.
144 __ lw(a2, MemOperand(sp, kSavedRegistersAreaSize));
145
146 // Get the address of the location in the code object (a3) (return
147 // address for lazy deoptimization) and compute the fp-to-sp delta in
148 // register t0.
149 __ mov(a3, ra);
150 // Correct one word for bailout id.
151 __ Addu(t0, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
152
153 __ Subu(t0, fp, t0);
154
155 // Allocate a new deoptimizer object.
156 __ PrepareCallCFunction(6, t1);
157 // Pass four arguments in a0 to a3 and fifth & sixth arguments on stack.
158 __ mov(a0, zero_reg);
159 Label context_check;
160 __ lw(a1, MemOperand(fp, CommonFrameConstants::kContextOrFrameTypeOffset));
161 __ JumpIfSmi(a1, &context_check);
162 __ lw(a0, MemOperand(fp, JavaScriptFrameConstants::kFunctionOffset));
163 __ bind(&context_check);
164 __ li(a1, Operand(type())); // Bailout type.
165 // a2: bailout id already loaded.
166 // a3: code address or 0 already loaded.
167 __ sw(t0, CFunctionArgumentOperand(5)); // Fp-to-sp delta.
168 __ li(t1, Operand(ExternalReference::isolate_address(isolate())));
169 __ sw(t1, CFunctionArgumentOperand(6)); // Isolate.
170 // Call Deoptimizer::New().
171 {
172 AllowExternalCallThatCantCauseGC scope(masm());
173 __ CallCFunction(ExternalReference::new_deoptimizer_function(isolate()), 6);
174 }
175
176 // Preserve "deoptimizer" object in register v0 and get the input
177 // frame descriptor pointer to a1 (deoptimizer->input_);
178 // Move deopt-obj to a0 for call to Deoptimizer::ComputeOutputFrames() below.
179 __ mov(a0, v0);
180 __ lw(a1, MemOperand(v0, Deoptimizer::input_offset()));
181
182 // Copy core registers into FrameDescription::registers_[kNumRegisters].
183 DCHECK(Register::kNumRegisters == kNumberOfRegisters);
184 for (int i = 0; i < kNumberOfRegisters; i++) {
185 int offset = (i * kPointerSize) + FrameDescription::registers_offset();
186 if ((saved_regs & (1 << i)) != 0) {
187 __ lw(a2, MemOperand(sp, i * kPointerSize));
188 __ sw(a2, MemOperand(a1, offset));
189 } else if (FLAG_debug_code) {
190 __ li(a2, kDebugZapValue);
191 __ sw(a2, MemOperand(a1, offset));
192 }
193 }
194
195 int double_regs_offset = FrameDescription::double_registers_offset();
196 // Copy FPU registers to
197 // double_registers_[DoubleRegister::kNumAllocatableRegisters]
198 for (int i = 0; i < config->num_allocatable_double_registers(); ++i) {
199 int code = config->GetAllocatableDoubleCode(i);
200 int dst_offset = code * kDoubleSize + double_regs_offset;
201 int src_offset = code * kDoubleSize + kNumberOfRegisters * kPointerSize;
202 __ ldc1(f0, MemOperand(sp, src_offset));
203 __ sdc1(f0, MemOperand(a1, dst_offset));
204 }
205
206 // Remove the bailout id and the saved registers from the stack.
207 __ Addu(sp, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
208
209 // Compute a pointer to the unwinding limit in register a2; that is
210 // the first stack slot not part of the input frame.
211 __ lw(a2, MemOperand(a1, FrameDescription::frame_size_offset()));
212 __ Addu(a2, a2, sp);
213
214 // Unwind the stack down to - but not including - the unwinding
215 // limit and copy the contents of the activation frame to the input
216 // frame description.
217 __ Addu(a3, a1, Operand(FrameDescription::frame_content_offset()));
218 Label pop_loop;
219 Label pop_loop_header;
220 __ BranchShort(&pop_loop_header);
221 __ bind(&pop_loop);
222 __ pop(t0);
223 __ sw(t0, MemOperand(a3, 0));
224 __ addiu(a3, a3, sizeof(uint32_t));
225 __ bind(&pop_loop_header);
226 __ BranchShort(&pop_loop, ne, a2, Operand(sp));
227
228 // Compute the output frame in the deoptimizer.
229 __ push(a0); // Preserve deoptimizer object across call.
230 // a0: deoptimizer object; a1: scratch.
231 __ PrepareCallCFunction(1, a1);
232 // Call Deoptimizer::ComputeOutputFrames().
233 {
234 AllowExternalCallThatCantCauseGC scope(masm());
235 __ CallCFunction(
236 ExternalReference::compute_output_frames_function(isolate()), 1);
237 }
238 __ pop(a0); // Restore deoptimizer object (class Deoptimizer).
239
240 __ lw(sp, MemOperand(a0, Deoptimizer::caller_frame_top_offset()));
241
242 // Replace the current (input) frame with the output frames.
243 Label outer_push_loop, inner_push_loop,
244 outer_loop_header, inner_loop_header;
245 // Outer loop state: t0 = current "FrameDescription** output_",
246 // a1 = one past the last FrameDescription**.
247 __ lw(a1, MemOperand(a0, Deoptimizer::output_count_offset()));
248 __ lw(t0, MemOperand(a0, Deoptimizer::output_offset())); // t0 is output_.
249 __ Lsa(a1, t0, a1, kPointerSizeLog2);
250 __ BranchShort(&outer_loop_header);
251 __ bind(&outer_push_loop);
252 // Inner loop state: a2 = current FrameDescription*, a3 = loop index.
253 __ lw(a2, MemOperand(t0, 0)); // output_[ix]
254 __ lw(a3, MemOperand(a2, FrameDescription::frame_size_offset()));
255 __ BranchShort(&inner_loop_header);
256 __ bind(&inner_push_loop);
257 __ Subu(a3, a3, Operand(sizeof(uint32_t)));
258 __ Addu(t2, a2, Operand(a3));
259 __ lw(t3, MemOperand(t2, FrameDescription::frame_content_offset()));
260 __ push(t3);
261 __ bind(&inner_loop_header);
262 __ BranchShort(&inner_push_loop, ne, a3, Operand(zero_reg));
263
264 __ Addu(t0, t0, Operand(kPointerSize));
265 __ bind(&outer_loop_header);
266 __ BranchShort(&outer_push_loop, lt, t0, Operand(a1));
267
268 __ lw(a1, MemOperand(a0, Deoptimizer::input_offset()));
269 for (int i = 0; i < config->num_allocatable_double_registers(); ++i) {
270 int code = config->GetAllocatableDoubleCode(i);
271 const DoubleRegister fpu_reg = DoubleRegister::from_code(code);
272 int src_offset = code * kDoubleSize + double_regs_offset;
273 __ ldc1(fpu_reg, MemOperand(a1, src_offset));
274 }
275
276 // Push state, pc, and continuation from the last output frame.
277 __ lw(t2, MemOperand(a2, FrameDescription::state_offset()));
278 __ push(t2);
279
280 __ lw(t2, MemOperand(a2, FrameDescription::pc_offset()));
281 __ push(t2);
282 __ lw(t2, MemOperand(a2, FrameDescription::continuation_offset()));
283 __ push(t2);
284
285
286 // Technically restoring 'at' should work unless zero_reg is also restored
287 // but it's safer to check for this.
288 DCHECK(!(at.bit() & restored_regs));
289 // Restore the registers from the last output frame.
290 __ mov(at, a2);
291 for (int i = kNumberOfRegisters - 1; i >= 0; i--) {
292 int offset = (i * kPointerSize) + FrameDescription::registers_offset();
293 if ((restored_regs & (1 << i)) != 0) {
294 __ lw(ToRegister(i), MemOperand(at, offset));
295 }
296 }
297
298 __ InitializeRootRegister();
299
300 __ pop(at); // Get continuation, leave pc on stack.
301 __ pop(ra);
302 __ Jump(at);
303 __ stop("Unreachable.");
304 }
305
306
307 // Maximum size of a table entry generated below.
308 const int Deoptimizer::table_entry_size_ = 2 * Assembler::kInstrSize;
309
GeneratePrologue()310 void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
311 Assembler::BlockTrampolinePoolScope block_trampoline_pool(masm());
312
313 // Create a sequence of deoptimization entries.
314 // Note that registers are still live when jumping to an entry.
315 Label table_start, done, done_special, trampoline_jump;
316 __ bind(&table_start);
317 int kMaxEntriesBranchReach = (1 << (kImm16Bits - 2))/
318 (table_entry_size_ / Assembler::kInstrSize);
319
320 if (count() <= kMaxEntriesBranchReach) {
321 // Common case.
322 for (int i = 0; i < count(); i++) {
323 Label start;
324 __ bind(&start);
325 DCHECK(is_int16(i));
326 __ BranchShort(USE_DELAY_SLOT, &done); // Expose delay slot.
327 __ li(at, i); // In the delay slot.
328
329 DCHECK_EQ(table_entry_size_, masm()->SizeOfCodeGeneratedSince(&start));
330 }
331
332 DCHECK_EQ(masm()->SizeOfCodeGeneratedSince(&table_start),
333 count() * table_entry_size_);
334 __ bind(&done);
335 __ Push(at);
336 } else {
337 // Uncommon case, the branch cannot reach.
338 // Create mini trampoline and adjust id constants to get proper value at
339 // the end of table.
340 for (int i = kMaxEntriesBranchReach; i > 1; i--) {
341 Label start;
342 __ bind(&start);
343 DCHECK(is_int16(i));
344 __ BranchShort(USE_DELAY_SLOT, &trampoline_jump); // Expose delay slot.
345 __ li(at, - i); // In the delay slot.
346 DCHECK_EQ(table_entry_size_, masm()->SizeOfCodeGeneratedSince(&start));
347 }
348 // Entry with id == kMaxEntriesBranchReach - 1.
349 __ bind(&trampoline_jump);
350 __ BranchShort(USE_DELAY_SLOT, &done_special);
351 __ li(at, -1);
352
353 for (int i = kMaxEntriesBranchReach ; i < count(); i++) {
354 Label start;
355 __ bind(&start);
356 DCHECK(is_int16(i));
357 __ BranchShort(USE_DELAY_SLOT, &done); // Expose delay slot.
358 __ li(at, i); // In the delay slot.
359 }
360
361 DCHECK_EQ(masm()->SizeOfCodeGeneratedSince(&table_start),
362 count() * table_entry_size_);
363 __ bind(&done_special);
364 __ addiu(at, at, kMaxEntriesBranchReach);
365 __ bind(&done);
366 __ Push(at);
367 }
368 }
369
370
SetCallerPc(unsigned offset,intptr_t value)371 void FrameDescription::SetCallerPc(unsigned offset, intptr_t value) {
372 SetFrameSlot(offset, value);
373 }
374
375
SetCallerFp(unsigned offset,intptr_t value)376 void FrameDescription::SetCallerFp(unsigned offset, intptr_t value) {
377 SetFrameSlot(offset, value);
378 }
379
380
SetCallerConstantPool(unsigned offset,intptr_t value)381 void FrameDescription::SetCallerConstantPool(unsigned offset, intptr_t value) {
382 // No embedded constant pool support.
383 UNREACHABLE();
384 }
385
386
387 #undef __
388
389
390 } // namespace internal
391 } // namespace v8
392