1 // Copyright 2012 the V8 project authors. All rights reserved. 2 // Redistribution and use in source and binary forms, with or without 3 // modification, are permitted provided that the following conditions are 4 // met: 5 // 6 // * Redistributions of source code must retain the above copyright 7 // notice, this list of conditions and the following disclaimer. 8 // * Redistributions in binary form must reproduce the above 9 // copyright notice, this list of conditions and the following 10 // disclaimer in the documentation and/or other materials provided 11 // with the distribution. 12 // * Neither the name of Google Inc. nor the names of its 13 // contributors may be used to endorse or promote products derived 14 // from this software without specific prior written permission. 15 // 16 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 17 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 18 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 19 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 20 // OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 21 // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 22 // LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23 // DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24 // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 26 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 28 #include "v8.h" 29 30 #if defined(V8_TARGET_ARCH_IA32) 31 32 #include "codegen.h" 33 #include "deoptimizer.h" 34 #include "full-codegen.h" 35 #include "safepoint-table.h" 36 37 namespace v8 { 38 namespace internal { 39 40 const int Deoptimizer::table_entry_size_ = 10; 41 42 43 int Deoptimizer::patch_size() { 44 return Assembler::kCallInstructionLength; 45 } 46 47 48 void Deoptimizer::EnsureRelocSpaceForLazyDeoptimization(Handle<Code> code) { 49 Isolate* isolate = code->GetIsolate(); 50 HandleScope scope(isolate); 51 52 // Compute the size of relocation information needed for the code 53 // patching in Deoptimizer::DeoptimizeFunction. 54 int min_reloc_size = 0; 55 int prev_pc_offset = 0; 56 DeoptimizationInputData* deopt_data = 57 DeoptimizationInputData::cast(code->deoptimization_data()); 58 for (int i = 0; i < deopt_data->DeoptCount(); i++) { 59 int pc_offset = deopt_data->Pc(i)->value(); 60 if (pc_offset == -1) continue; 61 ASSERT_GE(pc_offset, prev_pc_offset); 62 int pc_delta = pc_offset - prev_pc_offset; 63 // We use RUNTIME_ENTRY reloc info which has a size of 2 bytes 64 // if encodable with small pc delta encoding and up to 6 bytes 65 // otherwise. 66 if (pc_delta <= RelocInfo::kMaxSmallPCDelta) { 67 min_reloc_size += 2; 68 } else { 69 min_reloc_size += 6; 70 } 71 prev_pc_offset = pc_offset; 72 } 73 74 // If the relocation information is not big enough we create a new 75 // relocation info object that is padded with comments to make it 76 // big enough for lazy doptimization. 77 int reloc_length = code->relocation_info()->length(); 78 if (min_reloc_size > reloc_length) { 79 int comment_reloc_size = RelocInfo::kMinRelocCommentSize; 80 // Padding needed. 81 int min_padding = min_reloc_size - reloc_length; 82 // Number of comments needed to take up at least that much space. 83 int additional_comments = 84 (min_padding + comment_reloc_size - 1) / comment_reloc_size; 85 // Actual padding size. 86 int padding = additional_comments * comment_reloc_size; 87 // Allocate new relocation info and copy old relocation to the end 88 // of the new relocation info array because relocation info is 89 // written and read backwards. 90 Factory* factory = isolate->factory(); 91 Handle<ByteArray> new_reloc = 92 factory->NewByteArray(reloc_length + padding, TENURED); 93 memcpy(new_reloc->GetDataStartAddress() + padding, 94 code->relocation_info()->GetDataStartAddress(), 95 reloc_length); 96 // Create a relocation writer to write the comments in the padding 97 // space. Use position 0 for everything to ensure short encoding. 98 RelocInfoWriter reloc_info_writer( 99 new_reloc->GetDataStartAddress() + padding, 0); 100 intptr_t comment_string 101 = reinterpret_cast<intptr_t>(RelocInfo::kFillerCommentString); 102 RelocInfo rinfo(0, RelocInfo::COMMENT, comment_string, NULL); 103 for (int i = 0; i < additional_comments; ++i) { 104 #ifdef DEBUG 105 byte* pos_before = reloc_info_writer.pos(); 106 #endif 107 reloc_info_writer.Write(&rinfo); 108 ASSERT(RelocInfo::kMinRelocCommentSize == 109 pos_before - reloc_info_writer.pos()); 110 } 111 // Replace relocation information on the code object. 112 code->set_relocation_info(*new_reloc); 113 } 114 } 115 116 117 void Deoptimizer::DeoptimizeFunction(JSFunction* function) { 118 if (!function->IsOptimized()) return; 119 120 Isolate* isolate = function->GetIsolate(); 121 HandleScope scope(isolate); 122 AssertNoAllocation no_allocation; 123 124 // Get the optimized code. 125 Code* code = function->code(); 126 Address code_start_address = code->instruction_start(); 127 128 // We will overwrite the code's relocation info in-place. Relocation info 129 // is written backward. The relocation info is the payload of a byte 130 // array. Later on we will slide this to the start of the byte array and 131 // create a filler object in the remaining space. 132 ByteArray* reloc_info = code->relocation_info(); 133 Address reloc_end_address = reloc_info->address() + reloc_info->Size(); 134 RelocInfoWriter reloc_info_writer(reloc_end_address, code_start_address); 135 136 // For each LLazyBailout instruction insert a call to the corresponding 137 // deoptimization entry. 138 139 // Since the call is a relative encoding, write new 140 // reloc info. We do not need any of the existing reloc info because the 141 // existing code will not be used again (we zap it in debug builds). 142 // 143 // Emit call to lazy deoptimization at all lazy deopt points. 144 DeoptimizationInputData* deopt_data = 145 DeoptimizationInputData::cast(code->deoptimization_data()); 146 #ifdef DEBUG 147 Address prev_call_address = NULL; 148 #endif 149 for (int i = 0; i < deopt_data->DeoptCount(); i++) { 150 if (deopt_data->Pc(i)->value() == -1) continue; 151 // Patch lazy deoptimization entry. 152 Address call_address = code_start_address + deopt_data->Pc(i)->value(); 153 CodePatcher patcher(call_address, patch_size()); 154 Address deopt_entry = GetDeoptimizationEntry(i, LAZY); 155 patcher.masm()->call(deopt_entry, RelocInfo::NONE); 156 // We use RUNTIME_ENTRY for deoptimization bailouts. 157 RelocInfo rinfo(call_address + 1, // 1 after the call opcode. 158 RelocInfo::RUNTIME_ENTRY, 159 reinterpret_cast<intptr_t>(deopt_entry), 160 NULL); 161 reloc_info_writer.Write(&rinfo); 162 ASSERT_GE(reloc_info_writer.pos(), 163 reloc_info->address() + ByteArray::kHeaderSize); 164 ASSERT(prev_call_address == NULL || 165 call_address >= prev_call_address + patch_size()); 166 ASSERT(call_address + patch_size() <= code->instruction_end()); 167 #ifdef DEBUG 168 prev_call_address = call_address; 169 #endif 170 } 171 172 // Move the relocation info to the beginning of the byte array. 173 int new_reloc_size = reloc_end_address - reloc_info_writer.pos(); 174 memmove(code->relocation_start(), reloc_info_writer.pos(), new_reloc_size); 175 176 // The relocation info is in place, update the size. 177 reloc_info->set_length(new_reloc_size); 178 179 // Handle the junk part after the new relocation info. We will create 180 // a non-live object in the extra space at the end of the former reloc info. 181 Address junk_address = reloc_info->address() + reloc_info->Size(); 182 ASSERT(junk_address <= reloc_end_address); 183 isolate->heap()->CreateFillerObjectAt(junk_address, 184 reloc_end_address - junk_address); 185 186 // Add the deoptimizing code to the list. 187 DeoptimizingCodeListNode* node = new DeoptimizingCodeListNode(code); 188 DeoptimizerData* data = isolate->deoptimizer_data(); 189 node->set_next(data->deoptimizing_code_list_); 190 data->deoptimizing_code_list_ = node; 191 192 // We might be in the middle of incremental marking with compaction. 193 // Tell collector to treat this code object in a special way and 194 // ignore all slots that might have been recorded on it. 195 isolate->heap()->mark_compact_collector()->InvalidateCode(code); 196 197 // Set the code for the function to non-optimized version. 198 function->ReplaceCode(function->shared()->code()); 199 200 if (FLAG_trace_deopt) { 201 PrintF("[forced deoptimization: "); 202 function->PrintName(); 203 PrintF(" / %x]\n", reinterpret_cast<uint32_t>(function)); 204 } 205 } 206 207 208 static const byte kJnsInstruction = 0x79; 209 static const byte kJnsOffset = 0x13; 210 static const byte kJaeInstruction = 0x73; 211 static const byte kJaeOffset = 0x07; 212 static const byte kCallInstruction = 0xe8; 213 static const byte kNopByteOne = 0x66; 214 static const byte kNopByteTwo = 0x90; 215 216 217 void Deoptimizer::PatchStackCheckCodeAt(Code* unoptimized_code, 218 Address pc_after, 219 Code* check_code, 220 Code* replacement_code) { 221 Address call_target_address = pc_after - kIntSize; 222 ASSERT_EQ(check_code->entry(), 223 Assembler::target_address_at(call_target_address)); 224 // The stack check code matches the pattern: 225 // 226 // cmp esp, <limit> 227 // jae ok 228 // call <stack guard> 229 // test eax, <loop nesting depth> 230 // ok: ... 231 // 232 // We will patch away the branch so the code is: 233 // 234 // cmp esp, <limit> ;; Not changed 235 // nop 236 // nop 237 // call <on-stack replacment> 238 // test eax, <loop nesting depth> 239 // ok: 240 241 if (FLAG_count_based_interrupts) { 242 ASSERT_EQ(*(call_target_address - 3), kJnsInstruction); 243 ASSERT_EQ(*(call_target_address - 2), kJnsOffset); 244 } else { 245 ASSERT_EQ(*(call_target_address - 3), kJaeInstruction); 246 ASSERT_EQ(*(call_target_address - 2), kJaeOffset); 247 } 248 ASSERT_EQ(*(call_target_address - 1), kCallInstruction); 249 *(call_target_address - 3) = kNopByteOne; 250 *(call_target_address - 2) = kNopByteTwo; 251 Assembler::set_target_address_at(call_target_address, 252 replacement_code->entry()); 253 254 unoptimized_code->GetHeap()->incremental_marking()->RecordCodeTargetPatch( 255 unoptimized_code, call_target_address, replacement_code); 256 } 257 258 259 void Deoptimizer::RevertStackCheckCodeAt(Code* unoptimized_code, 260 Address pc_after, 261 Code* check_code, 262 Code* replacement_code) { 263 Address call_target_address = pc_after - kIntSize; 264 ASSERT_EQ(replacement_code->entry(), 265 Assembler::target_address_at(call_target_address)); 266 267 // Replace the nops from patching (Deoptimizer::PatchStackCheckCode) to 268 // restore the conditional branch. 269 ASSERT_EQ(*(call_target_address - 3), kNopByteOne); 270 ASSERT_EQ(*(call_target_address - 2), kNopByteTwo); 271 ASSERT_EQ(*(call_target_address - 1), kCallInstruction); 272 if (FLAG_count_based_interrupts) { 273 *(call_target_address - 3) = kJnsInstruction; 274 *(call_target_address - 2) = kJnsOffset; 275 } else { 276 *(call_target_address - 3) = kJaeInstruction; 277 *(call_target_address - 2) = kJaeOffset; 278 } 279 Assembler::set_target_address_at(call_target_address, 280 check_code->entry()); 281 282 check_code->GetHeap()->incremental_marking()->RecordCodeTargetPatch( 283 unoptimized_code, call_target_address, check_code); 284 } 285 286 287 static int LookupBailoutId(DeoptimizationInputData* data, unsigned ast_id) { 288 ByteArray* translations = data->TranslationByteArray(); 289 int length = data->DeoptCount(); 290 for (int i = 0; i < length; i++) { 291 if (static_cast<unsigned>(data->AstId(i)->value()) == ast_id) { 292 TranslationIterator it(translations, data->TranslationIndex(i)->value()); 293 int value = it.Next(); 294 ASSERT(Translation::BEGIN == static_cast<Translation::Opcode>(value)); 295 // Read the number of frames. 296 value = it.Next(); 297 if (value == 1) return i; 298 } 299 } 300 UNREACHABLE(); 301 return -1; 302 } 303 304 305 void Deoptimizer::DoComputeOsrOutputFrame() { 306 DeoptimizationInputData* data = DeoptimizationInputData::cast( 307 optimized_code_->deoptimization_data()); 308 unsigned ast_id = data->OsrAstId()->value(); 309 // TODO(kasperl): This should not be the bailout_id_. It should be 310 // the ast id. Confusing. 311 ASSERT(bailout_id_ == ast_id); 312 313 int bailout_id = LookupBailoutId(data, ast_id); 314 unsigned translation_index = data->TranslationIndex(bailout_id)->value(); 315 ByteArray* translations = data->TranslationByteArray(); 316 317 TranslationIterator iterator(translations, translation_index); 318 Translation::Opcode opcode = 319 static_cast<Translation::Opcode>(iterator.Next()); 320 ASSERT(Translation::BEGIN == opcode); 321 USE(opcode); 322 int count = iterator.Next(); 323 iterator.Next(); // Drop JS frames count. 324 ASSERT(count == 1); 325 USE(count); 326 327 opcode = static_cast<Translation::Opcode>(iterator.Next()); 328 USE(opcode); 329 ASSERT(Translation::JS_FRAME == opcode); 330 unsigned node_id = iterator.Next(); 331 USE(node_id); 332 ASSERT(node_id == ast_id); 333 JSFunction* function = JSFunction::cast(ComputeLiteral(iterator.Next())); 334 USE(function); 335 ASSERT(function == function_); 336 unsigned height = iterator.Next(); 337 unsigned height_in_bytes = height * kPointerSize; 338 USE(height_in_bytes); 339 340 unsigned fixed_size = ComputeFixedSize(function_); 341 unsigned input_frame_size = input_->GetFrameSize(); 342 ASSERT(fixed_size + height_in_bytes == input_frame_size); 343 344 unsigned stack_slot_size = optimized_code_->stack_slots() * kPointerSize; 345 unsigned outgoing_height = data->ArgumentsStackHeight(bailout_id)->value(); 346 unsigned outgoing_size = outgoing_height * kPointerSize; 347 unsigned output_frame_size = fixed_size + stack_slot_size + outgoing_size; 348 ASSERT(outgoing_size == 0); // OSR does not happen in the middle of a call. 349 350 if (FLAG_trace_osr) { 351 PrintF("[on-stack replacement: begin 0x%08" V8PRIxPTR " ", 352 reinterpret_cast<intptr_t>(function_)); 353 function_->PrintName(); 354 PrintF(" => node=%u, frame=%d->%d]\n", 355 ast_id, 356 input_frame_size, 357 output_frame_size); 358 } 359 360 // There's only one output frame in the OSR case. 361 output_count_ = 1; 362 output_ = new FrameDescription*[1]; 363 output_[0] = new(output_frame_size) FrameDescription( 364 output_frame_size, function_); 365 output_[0]->SetFrameType(StackFrame::JAVA_SCRIPT); 366 367 // Clear the incoming parameters in the optimized frame to avoid 368 // confusing the garbage collector. 369 unsigned output_offset = output_frame_size - kPointerSize; 370 int parameter_count = function_->shared()->formal_parameter_count() + 1; 371 for (int i = 0; i < parameter_count; ++i) { 372 output_[0]->SetFrameSlot(output_offset, 0); 373 output_offset -= kPointerSize; 374 } 375 376 // Translate the incoming parameters. This may overwrite some of the 377 // incoming argument slots we've just cleared. 378 int input_offset = input_frame_size - kPointerSize; 379 bool ok = true; 380 int limit = input_offset - (parameter_count * kPointerSize); 381 while (ok && input_offset > limit) { 382 ok = DoOsrTranslateCommand(&iterator, &input_offset); 383 } 384 385 // There are no translation commands for the caller's pc and fp, the 386 // context, and the function. Set them up explicitly. 387 for (int i = StandardFrameConstants::kCallerPCOffset; 388 ok && i >= StandardFrameConstants::kMarkerOffset; 389 i -= kPointerSize) { 390 uint32_t input_value = input_->GetFrameSlot(input_offset); 391 if (FLAG_trace_osr) { 392 const char* name = "UNKNOWN"; 393 switch (i) { 394 case StandardFrameConstants::kCallerPCOffset: 395 name = "caller's pc"; 396 break; 397 case StandardFrameConstants::kCallerFPOffset: 398 name = "fp"; 399 break; 400 case StandardFrameConstants::kContextOffset: 401 name = "context"; 402 break; 403 case StandardFrameConstants::kMarkerOffset: 404 name = "function"; 405 break; 406 } 407 PrintF(" [esp + %d] <- 0x%08x ; [esp + %d] (fixed part - %s)\n", 408 output_offset, 409 input_value, 410 input_offset, 411 name); 412 } 413 output_[0]->SetFrameSlot(output_offset, input_->GetFrameSlot(input_offset)); 414 input_offset -= kPointerSize; 415 output_offset -= kPointerSize; 416 } 417 418 // Translate the rest of the frame. 419 while (ok && input_offset >= 0) { 420 ok = DoOsrTranslateCommand(&iterator, &input_offset); 421 } 422 423 // If translation of any command failed, continue using the input frame. 424 if (!ok) { 425 delete output_[0]; 426 output_[0] = input_; 427 output_[0]->SetPc(reinterpret_cast<uint32_t>(from_)); 428 } else { 429 // Set up the frame pointer and the context pointer. 430 output_[0]->SetRegister(ebp.code(), input_->GetRegister(ebp.code())); 431 output_[0]->SetRegister(esi.code(), input_->GetRegister(esi.code())); 432 433 unsigned pc_offset = data->OsrPcOffset()->value(); 434 uint32_t pc = reinterpret_cast<uint32_t>( 435 optimized_code_->entry() + pc_offset); 436 output_[0]->SetPc(pc); 437 } 438 Code* continuation = 439 function->GetIsolate()->builtins()->builtin(Builtins::kNotifyOSR); 440 output_[0]->SetContinuation( 441 reinterpret_cast<uint32_t>(continuation->entry())); 442 443 if (FLAG_trace_osr) { 444 PrintF("[on-stack replacement translation %s: 0x%08" V8PRIxPTR " ", 445 ok ? "finished" : "aborted", 446 reinterpret_cast<intptr_t>(function)); 447 function->PrintName(); 448 PrintF(" => pc=0x%0x]\n", output_[0]->GetPc()); 449 } 450 } 451 452 453 void Deoptimizer::DoComputeArgumentsAdaptorFrame(TranslationIterator* iterator, 454 int frame_index) { 455 JSFunction* function = JSFunction::cast(ComputeLiteral(iterator->Next())); 456 unsigned height = iterator->Next(); 457 unsigned height_in_bytes = height * kPointerSize; 458 if (FLAG_trace_deopt) { 459 PrintF(" translating arguments adaptor => height=%d\n", height_in_bytes); 460 } 461 462 unsigned fixed_frame_size = ArgumentsAdaptorFrameConstants::kFrameSize; 463 unsigned output_frame_size = height_in_bytes + fixed_frame_size; 464 465 // Allocate and store the output frame description. 466 FrameDescription* output_frame = 467 new(output_frame_size) FrameDescription(output_frame_size, function); 468 output_frame->SetFrameType(StackFrame::ARGUMENTS_ADAPTOR); 469 470 // Arguments adaptor can not be topmost or bottommost. 471 ASSERT(frame_index > 0 && frame_index < output_count_ - 1); 472 ASSERT(output_[frame_index] == NULL); 473 output_[frame_index] = output_frame; 474 475 // The top address of the frame is computed from the previous 476 // frame's top and this frame's size. 477 uint32_t top_address; 478 top_address = output_[frame_index - 1]->GetTop() - output_frame_size; 479 output_frame->SetTop(top_address); 480 481 // Compute the incoming parameter translation. 482 int parameter_count = height; 483 unsigned output_offset = output_frame_size; 484 for (int i = 0; i < parameter_count; ++i) { 485 output_offset -= kPointerSize; 486 DoTranslateCommand(iterator, frame_index, output_offset); 487 } 488 489 // Read caller's PC from the previous frame. 490 output_offset -= kPointerSize; 491 intptr_t callers_pc = output_[frame_index - 1]->GetPc(); 492 output_frame->SetFrameSlot(output_offset, callers_pc); 493 if (FLAG_trace_deopt) { 494 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; caller's pc\n", 495 top_address + output_offset, output_offset, callers_pc); 496 } 497 498 // Read caller's FP from the previous frame, and set this frame's FP. 499 output_offset -= kPointerSize; 500 intptr_t value = output_[frame_index - 1]->GetFp(); 501 output_frame->SetFrameSlot(output_offset, value); 502 intptr_t fp_value = top_address + output_offset; 503 output_frame->SetFp(fp_value); 504 if (FLAG_trace_deopt) { 505 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; caller's fp\n", 506 fp_value, output_offset, value); 507 } 508 509 // A marker value is used in place of the context. 510 output_offset -= kPointerSize; 511 intptr_t context = reinterpret_cast<intptr_t>( 512 Smi::FromInt(StackFrame::ARGUMENTS_ADAPTOR)); 513 output_frame->SetFrameSlot(output_offset, context); 514 if (FLAG_trace_deopt) { 515 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; context (adaptor sentinel)\n", 516 top_address + output_offset, output_offset, context); 517 } 518 519 // The function was mentioned explicitly in the ARGUMENTS_ADAPTOR_FRAME. 520 output_offset -= kPointerSize; 521 value = reinterpret_cast<intptr_t>(function); 522 output_frame->SetFrameSlot(output_offset, value); 523 if (FLAG_trace_deopt) { 524 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; function\n", 525 top_address + output_offset, output_offset, value); 526 } 527 528 // Number of incoming arguments. 529 output_offset -= kPointerSize; 530 value = reinterpret_cast<uint32_t>(Smi::FromInt(height - 1)); 531 output_frame->SetFrameSlot(output_offset, value); 532 if (FLAG_trace_deopt) { 533 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; argc (%d)\n", 534 top_address + output_offset, output_offset, value, height - 1); 535 } 536 537 ASSERT(0 == output_offset); 538 539 Builtins* builtins = isolate_->builtins(); 540 Code* adaptor_trampoline = 541 builtins->builtin(Builtins::kArgumentsAdaptorTrampoline); 542 uint32_t pc = reinterpret_cast<uint32_t>( 543 adaptor_trampoline->instruction_start() + 544 isolate_->heap()->arguments_adaptor_deopt_pc_offset()->value()); 545 output_frame->SetPc(pc); 546 } 547 548 549 void Deoptimizer::DoComputeConstructStubFrame(TranslationIterator* iterator, 550 int frame_index) { 551 Builtins* builtins = isolate_->builtins(); 552 Code* construct_stub = builtins->builtin(Builtins::kJSConstructStubGeneric); 553 JSFunction* function = JSFunction::cast(ComputeLiteral(iterator->Next())); 554 unsigned height = iterator->Next(); 555 unsigned height_in_bytes = height * kPointerSize; 556 if (FLAG_trace_deopt) { 557 PrintF(" translating construct stub => height=%d\n", height_in_bytes); 558 } 559 560 unsigned fixed_frame_size = 7 * kPointerSize; 561 unsigned output_frame_size = height_in_bytes + fixed_frame_size; 562 563 // Allocate and store the output frame description. 564 FrameDescription* output_frame = 565 new(output_frame_size) FrameDescription(output_frame_size, function); 566 output_frame->SetFrameType(StackFrame::CONSTRUCT); 567 568 // Construct stub can not be topmost or bottommost. 569 ASSERT(frame_index > 0 && frame_index < output_count_ - 1); 570 ASSERT(output_[frame_index] == NULL); 571 output_[frame_index] = output_frame; 572 573 // The top address of the frame is computed from the previous 574 // frame's top and this frame's size. 575 uint32_t top_address; 576 top_address = output_[frame_index - 1]->GetTop() - output_frame_size; 577 output_frame->SetTop(top_address); 578 579 // Compute the incoming parameter translation. 580 int parameter_count = height; 581 unsigned output_offset = output_frame_size; 582 for (int i = 0; i < parameter_count; ++i) { 583 output_offset -= kPointerSize; 584 DoTranslateCommand(iterator, frame_index, output_offset); 585 } 586 587 // Read caller's PC from the previous frame. 588 output_offset -= kPointerSize; 589 intptr_t callers_pc = output_[frame_index - 1]->GetPc(); 590 output_frame->SetFrameSlot(output_offset, callers_pc); 591 if (FLAG_trace_deopt) { 592 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; caller's pc\n", 593 top_address + output_offset, output_offset, callers_pc); 594 } 595 596 // Read caller's FP from the previous frame, and set this frame's FP. 597 output_offset -= kPointerSize; 598 intptr_t value = output_[frame_index - 1]->GetFp(); 599 output_frame->SetFrameSlot(output_offset, value); 600 intptr_t fp_value = top_address + output_offset; 601 output_frame->SetFp(fp_value); 602 if (FLAG_trace_deopt) { 603 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; caller's fp\n", 604 fp_value, output_offset, value); 605 } 606 607 // The context can be gotten from the previous frame. 608 output_offset -= kPointerSize; 609 value = output_[frame_index - 1]->GetContext(); 610 output_frame->SetFrameSlot(output_offset, value); 611 if (FLAG_trace_deopt) { 612 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; context\n", 613 top_address + output_offset, output_offset, value); 614 } 615 616 // A marker value is used in place of the function. 617 output_offset -= kPointerSize; 618 value = reinterpret_cast<intptr_t>(Smi::FromInt(StackFrame::CONSTRUCT)); 619 output_frame->SetFrameSlot(output_offset, value); 620 if (FLAG_trace_deopt) { 621 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; function (construct sentinel)\n", 622 top_address + output_offset, output_offset, value); 623 } 624 625 // The output frame reflects a JSConstructStubGeneric frame. 626 output_offset -= kPointerSize; 627 value = reinterpret_cast<intptr_t>(construct_stub); 628 output_frame->SetFrameSlot(output_offset, value); 629 if (FLAG_trace_deopt) { 630 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; code object\n", 631 top_address + output_offset, output_offset, value); 632 } 633 634 // Number of incoming arguments. 635 output_offset -= kPointerSize; 636 value = reinterpret_cast<uint32_t>(Smi::FromInt(height - 1)); 637 output_frame->SetFrameSlot(output_offset, value); 638 if (FLAG_trace_deopt) { 639 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; argc (%d)\n", 640 top_address + output_offset, output_offset, value, height - 1); 641 } 642 643 // The newly allocated object was passed as receiver in the artificial 644 // constructor stub environment created by HEnvironment::CopyForInlining(). 645 output_offset -= kPointerSize; 646 value = output_frame->GetFrameSlot(output_frame_size - kPointerSize); 647 output_frame->SetFrameSlot(output_offset, value); 648 if (FLAG_trace_deopt) { 649 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; allocated receiver\n", 650 top_address + output_offset, output_offset, value); 651 } 652 653 ASSERT(0 == output_offset); 654 655 uint32_t pc = reinterpret_cast<uint32_t>( 656 construct_stub->instruction_start() + 657 isolate_->heap()->construct_stub_deopt_pc_offset()->value()); 658 output_frame->SetPc(pc); 659 } 660 661 662 void Deoptimizer::DoComputeJSFrame(TranslationIterator* iterator, 663 int frame_index) { 664 int node_id = iterator->Next(); 665 JSFunction* function = JSFunction::cast(ComputeLiteral(iterator->Next())); 666 unsigned height = iterator->Next(); 667 unsigned height_in_bytes = height * kPointerSize; 668 if (FLAG_trace_deopt) { 669 PrintF(" translating "); 670 function->PrintName(); 671 PrintF(" => node=%d, height=%d\n", node_id, height_in_bytes); 672 } 673 674 // The 'fixed' part of the frame consists of the incoming parameters and 675 // the part described by JavaScriptFrameConstants. 676 unsigned fixed_frame_size = ComputeFixedSize(function); 677 unsigned input_frame_size = input_->GetFrameSize(); 678 unsigned output_frame_size = height_in_bytes + fixed_frame_size; 679 680 // Allocate and store the output frame description. 681 FrameDescription* output_frame = 682 new(output_frame_size) FrameDescription(output_frame_size, function); 683 output_frame->SetFrameType(StackFrame::JAVA_SCRIPT); 684 685 bool is_bottommost = (0 == frame_index); 686 bool is_topmost = (output_count_ - 1 == frame_index); 687 ASSERT(frame_index >= 0 && frame_index < output_count_); 688 ASSERT(output_[frame_index] == NULL); 689 output_[frame_index] = output_frame; 690 691 // The top address for the bottommost output frame can be computed from 692 // the input frame pointer and the output frame's height. For all 693 // subsequent output frames, it can be computed from the previous one's 694 // top address and the current frame's size. 695 uint32_t top_address; 696 if (is_bottommost) { 697 // 2 = context and function in the frame. 698 top_address = 699 input_->GetRegister(ebp.code()) - (2 * kPointerSize) - height_in_bytes; 700 } else { 701 top_address = output_[frame_index - 1]->GetTop() - output_frame_size; 702 } 703 output_frame->SetTop(top_address); 704 705 // Compute the incoming parameter translation. 706 int parameter_count = function->shared()->formal_parameter_count() + 1; 707 unsigned output_offset = output_frame_size; 708 unsigned input_offset = input_frame_size; 709 for (int i = 0; i < parameter_count; ++i) { 710 output_offset -= kPointerSize; 711 DoTranslateCommand(iterator, frame_index, output_offset); 712 } 713 input_offset -= (parameter_count * kPointerSize); 714 715 // There are no translation commands for the caller's pc and fp, the 716 // context, and the function. Synthesize their values and set them up 717 // explicitly. 718 // 719 // The caller's pc for the bottommost output frame is the same as in the 720 // input frame. For all subsequent output frames, it can be read from the 721 // previous one. This frame's pc can be computed from the non-optimized 722 // function code and AST id of the bailout. 723 output_offset -= kPointerSize; 724 input_offset -= kPointerSize; 725 intptr_t value; 726 if (is_bottommost) { 727 value = input_->GetFrameSlot(input_offset); 728 } else { 729 value = output_[frame_index - 1]->GetPc(); 730 } 731 output_frame->SetFrameSlot(output_offset, value); 732 if (FLAG_trace_deopt) { 733 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; caller's pc\n", 734 top_address + output_offset, output_offset, value); 735 } 736 737 // The caller's frame pointer for the bottommost output frame is the same 738 // as in the input frame. For all subsequent output frames, it can be 739 // read from the previous one. Also compute and set this frame's frame 740 // pointer. 741 output_offset -= kPointerSize; 742 input_offset -= kPointerSize; 743 if (is_bottommost) { 744 value = input_->GetFrameSlot(input_offset); 745 } else { 746 value = output_[frame_index - 1]->GetFp(); 747 } 748 output_frame->SetFrameSlot(output_offset, value); 749 intptr_t fp_value = top_address + output_offset; 750 ASSERT(!is_bottommost || input_->GetRegister(ebp.code()) == fp_value); 751 output_frame->SetFp(fp_value); 752 if (is_topmost) output_frame->SetRegister(ebp.code(), fp_value); 753 if (FLAG_trace_deopt) { 754 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; caller's fp\n", 755 fp_value, output_offset, value); 756 } 757 758 // For the bottommost output frame the context can be gotten from the input 759 // frame. For all subsequent output frames it can be gotten from the function 760 // so long as we don't inline functions that need local contexts. 761 output_offset -= kPointerSize; 762 input_offset -= kPointerSize; 763 if (is_bottommost) { 764 value = input_->GetFrameSlot(input_offset); 765 } else { 766 value = reinterpret_cast<uint32_t>(function->context()); 767 } 768 output_frame->SetFrameSlot(output_offset, value); 769 output_frame->SetContext(value); 770 if (is_topmost) output_frame->SetRegister(esi.code(), value); 771 if (FLAG_trace_deopt) { 772 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; context\n", 773 top_address + output_offset, output_offset, value); 774 } 775 776 // The function was mentioned explicitly in the BEGIN_FRAME. 777 output_offset -= kPointerSize; 778 input_offset -= kPointerSize; 779 value = reinterpret_cast<uint32_t>(function); 780 // The function for the bottommost output frame should also agree with the 781 // input frame. 782 ASSERT(!is_bottommost || input_->GetFrameSlot(input_offset) == value); 783 output_frame->SetFrameSlot(output_offset, value); 784 if (FLAG_trace_deopt) { 785 PrintF(" 0x%08x: [top + %d] <- 0x%08x ; function\n", 786 top_address + output_offset, output_offset, value); 787 } 788 789 // Translate the rest of the frame. 790 for (unsigned i = 0; i < height; ++i) { 791 output_offset -= kPointerSize; 792 DoTranslateCommand(iterator, frame_index, output_offset); 793 } 794 ASSERT(0 == output_offset); 795 796 // Compute this frame's PC, state, and continuation. 797 Code* non_optimized_code = function->shared()->code(); 798 FixedArray* raw_data = non_optimized_code->deoptimization_data(); 799 DeoptimizationOutputData* data = DeoptimizationOutputData::cast(raw_data); 800 Address start = non_optimized_code->instruction_start(); 801 unsigned pc_and_state = GetOutputInfo(data, node_id, function->shared()); 802 unsigned pc_offset = FullCodeGenerator::PcField::decode(pc_and_state); 803 uint32_t pc_value = reinterpret_cast<uint32_t>(start + pc_offset); 804 output_frame->SetPc(pc_value); 805 806 FullCodeGenerator::State state = 807 FullCodeGenerator::StateField::decode(pc_and_state); 808 output_frame->SetState(Smi::FromInt(state)); 809 810 // Set the continuation for the topmost frame. 811 if (is_topmost && bailout_type_ != DEBUGGER) { 812 Builtins* builtins = isolate_->builtins(); 813 Code* continuation = (bailout_type_ == EAGER) 814 ? builtins->builtin(Builtins::kNotifyDeoptimized) 815 : builtins->builtin(Builtins::kNotifyLazyDeoptimized); 816 output_frame->SetContinuation( 817 reinterpret_cast<uint32_t>(continuation->entry())); 818 } 819 } 820 821 822 void Deoptimizer::FillInputFrame(Address tos, JavaScriptFrame* frame) { 823 // Set the register values. The values are not important as there are no 824 // callee saved registers in JavaScript frames, so all registers are 825 // spilled. Registers ebp and esp are set to the correct values though. 826 827 for (int i = 0; i < Register::kNumRegisters; i++) { 828 input_->SetRegister(i, i * 4); 829 } 830 input_->SetRegister(esp.code(), reinterpret_cast<intptr_t>(frame->sp())); 831 input_->SetRegister(ebp.code(), reinterpret_cast<intptr_t>(frame->fp())); 832 for (int i = 0; i < DoubleRegister::kNumAllocatableRegisters; i++) { 833 input_->SetDoubleRegister(i, 0.0); 834 } 835 836 // Fill the frame content from the actual data on the frame. 837 for (unsigned i = 0; i < input_->GetFrameSize(); i += kPointerSize) { 838 input_->SetFrameSlot(i, Memory::uint32_at(tos + i)); 839 } 840 } 841 842 843 #define __ masm()-> 844 845 void Deoptimizer::EntryGenerator::Generate() { 846 GeneratePrologue(); 847 CpuFeatures::Scope scope(SSE2); 848 849 Isolate* isolate = masm()->isolate(); 850 851 // Save all general purpose registers before messing with them. 852 const int kNumberOfRegisters = Register::kNumRegisters; 853 854 const int kDoubleRegsSize = kDoubleSize * 855 XMMRegister::kNumAllocatableRegisters; 856 __ sub(esp, Immediate(kDoubleRegsSize)); 857 for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) { 858 XMMRegister xmm_reg = XMMRegister::FromAllocationIndex(i); 859 int offset = i * kDoubleSize; 860 __ movdbl(Operand(esp, offset), xmm_reg); 861 } 862 863 __ pushad(); 864 865 const int kSavedRegistersAreaSize = kNumberOfRegisters * kPointerSize + 866 kDoubleRegsSize; 867 868 // Get the bailout id from the stack. 869 __ mov(ebx, Operand(esp, kSavedRegistersAreaSize)); 870 871 // Get the address of the location in the code object if possible 872 // and compute the fp-to-sp delta in register edx. 873 if (type() == EAGER) { 874 __ Set(ecx, Immediate(0)); 875 __ lea(edx, Operand(esp, kSavedRegistersAreaSize + 1 * kPointerSize)); 876 } else { 877 __ mov(ecx, Operand(esp, kSavedRegistersAreaSize + 1 * kPointerSize)); 878 __ lea(edx, Operand(esp, kSavedRegistersAreaSize + 2 * kPointerSize)); 879 } 880 __ sub(edx, ebp); 881 __ neg(edx); 882 883 // Allocate a new deoptimizer object. 884 __ PrepareCallCFunction(6, eax); 885 __ mov(eax, Operand(ebp, JavaScriptFrameConstants::kFunctionOffset)); 886 __ mov(Operand(esp, 0 * kPointerSize), eax); // Function. 887 __ mov(Operand(esp, 1 * kPointerSize), Immediate(type())); // Bailout type. 888 __ mov(Operand(esp, 2 * kPointerSize), ebx); // Bailout id. 889 __ mov(Operand(esp, 3 * kPointerSize), ecx); // Code address or 0. 890 __ mov(Operand(esp, 4 * kPointerSize), edx); // Fp-to-sp delta. 891 __ mov(Operand(esp, 5 * kPointerSize), 892 Immediate(ExternalReference::isolate_address())); 893 { 894 AllowExternalCallThatCantCauseGC scope(masm()); 895 __ CallCFunction(ExternalReference::new_deoptimizer_function(isolate), 6); 896 } 897 898 // Preserve deoptimizer object in register eax and get the input 899 // frame descriptor pointer. 900 __ mov(ebx, Operand(eax, Deoptimizer::input_offset())); 901 902 // Fill in the input registers. 903 for (int i = kNumberOfRegisters - 1; i >= 0; i--) { 904 int offset = (i * kPointerSize) + FrameDescription::registers_offset(); 905 __ pop(Operand(ebx, offset)); 906 } 907 908 // Fill in the double input registers. 909 int double_regs_offset = FrameDescription::double_registers_offset(); 910 for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) { 911 int dst_offset = i * kDoubleSize + double_regs_offset; 912 int src_offset = i * kDoubleSize; 913 __ movdbl(xmm0, Operand(esp, src_offset)); 914 __ movdbl(Operand(ebx, dst_offset), xmm0); 915 } 916 917 // Remove the bailout id and the double registers from the stack. 918 if (type() == EAGER) { 919 __ add(esp, Immediate(kDoubleRegsSize + kPointerSize)); 920 } else { 921 __ add(esp, Immediate(kDoubleRegsSize + 2 * kPointerSize)); 922 } 923 924 // Compute a pointer to the unwinding limit in register ecx; that is 925 // the first stack slot not part of the input frame. 926 __ mov(ecx, Operand(ebx, FrameDescription::frame_size_offset())); 927 __ add(ecx, esp); 928 929 // Unwind the stack down to - but not including - the unwinding 930 // limit and copy the contents of the activation frame to the input 931 // frame description. 932 __ lea(edx, Operand(ebx, FrameDescription::frame_content_offset())); 933 Label pop_loop; 934 __ bind(&pop_loop); 935 __ pop(Operand(edx, 0)); 936 __ add(edx, Immediate(sizeof(uint32_t))); 937 __ cmp(ecx, esp); 938 __ j(not_equal, &pop_loop); 939 940 // Compute the output frame in the deoptimizer. 941 __ push(eax); 942 __ PrepareCallCFunction(1, ebx); 943 __ mov(Operand(esp, 0 * kPointerSize), eax); 944 { 945 AllowExternalCallThatCantCauseGC scope(masm()); 946 __ CallCFunction( 947 ExternalReference::compute_output_frames_function(isolate), 1); 948 } 949 __ pop(eax); 950 951 // Replace the current frame with the output frames. 952 Label outer_push_loop, inner_push_loop; 953 // Outer loop state: eax = current FrameDescription**, edx = one past the 954 // last FrameDescription**. 955 __ mov(edx, Operand(eax, Deoptimizer::output_count_offset())); 956 __ mov(eax, Operand(eax, Deoptimizer::output_offset())); 957 __ lea(edx, Operand(eax, edx, times_4, 0)); 958 __ bind(&outer_push_loop); 959 // Inner loop state: ebx = current FrameDescription*, ecx = loop index. 960 __ mov(ebx, Operand(eax, 0)); 961 __ mov(ecx, Operand(ebx, FrameDescription::frame_size_offset())); 962 __ bind(&inner_push_loop); 963 __ sub(ecx, Immediate(sizeof(uint32_t))); 964 __ push(Operand(ebx, ecx, times_1, FrameDescription::frame_content_offset())); 965 __ test(ecx, ecx); 966 __ j(not_zero, &inner_push_loop); 967 __ add(eax, Immediate(kPointerSize)); 968 __ cmp(eax, edx); 969 __ j(below, &outer_push_loop); 970 971 // In case of OSR, we have to restore the XMM registers. 972 if (type() == OSR) { 973 for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) { 974 XMMRegister xmm_reg = XMMRegister::FromAllocationIndex(i); 975 int src_offset = i * kDoubleSize + double_regs_offset; 976 __ movdbl(xmm_reg, Operand(ebx, src_offset)); 977 } 978 } 979 980 // Push state, pc, and continuation from the last output frame. 981 if (type() != OSR) { 982 __ push(Operand(ebx, FrameDescription::state_offset())); 983 } 984 __ push(Operand(ebx, FrameDescription::pc_offset())); 985 __ push(Operand(ebx, FrameDescription::continuation_offset())); 986 987 988 // Push the registers from the last output frame. 989 for (int i = 0; i < kNumberOfRegisters; i++) { 990 int offset = (i * kPointerSize) + FrameDescription::registers_offset(); 991 __ push(Operand(ebx, offset)); 992 } 993 994 // Restore the registers from the stack. 995 __ popad(); 996 997 // Return to the continuation point. 998 __ ret(0); 999 } 1000 1001 1002 void Deoptimizer::TableEntryGenerator::GeneratePrologue() { 1003 // Create a sequence of deoptimization entries. 1004 Label done; 1005 for (int i = 0; i < count(); i++) { 1006 int start = masm()->pc_offset(); 1007 USE(start); 1008 __ push_imm32(i); 1009 __ jmp(&done); 1010 ASSERT(masm()->pc_offset() - start == table_entry_size_); 1011 } 1012 __ bind(&done); 1013 } 1014 1015 #undef __ 1016 1017 1018 } } // namespace v8::internal 1019 1020 #endif // V8_TARGET_ARCH_IA32 1021