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. 4 5 #include "src/compiler/scheduler.h" 6 7 #include <iomanip> 8 9 #include "src/base/adapters.h" 10 #include "src/bit-vector.h" 11 #include "src/compiler/common-operator.h" 12 #include "src/compiler/control-equivalence.h" 13 #include "src/compiler/graph.h" 14 #include "src/compiler/node-marker.h" 15 #include "src/compiler/node-properties.h" 16 #include "src/compiler/node.h" 17 #include "src/zone/zone-containers.h" 18 19 namespace v8 { 20 namespace internal { 21 namespace compiler { 22 23 #define TRACE(...) \ 24 do { \ 25 if (FLAG_trace_turbo_scheduler) PrintF(__VA_ARGS__); \ 26 } while (false) 27 28 Scheduler::Scheduler(Zone* zone, Graph* graph, Schedule* schedule, Flags flags) 29 : zone_(zone), 30 graph_(graph), 31 schedule_(schedule), 32 flags_(flags), 33 scheduled_nodes_(zone), 34 schedule_root_nodes_(zone), 35 schedule_queue_(zone), 36 node_data_(graph_->NodeCount(), DefaultSchedulerData(), zone) {} 37 38 39 Schedule* Scheduler::ComputeSchedule(Zone* zone, Graph* graph, Flags flags) { 40 Schedule* schedule = new (graph->zone()) 41 Schedule(graph->zone(), static_cast<size_t>(graph->NodeCount())); 42 Scheduler scheduler(zone, graph, schedule, flags); 43 44 scheduler.BuildCFG(); 45 scheduler.ComputeSpecialRPONumbering(); 46 scheduler.GenerateImmediateDominatorTree(); 47 48 scheduler.PrepareUses(); 49 scheduler.ScheduleEarly(); 50 scheduler.ScheduleLate(); 51 52 scheduler.SealFinalSchedule(); 53 54 return schedule; 55 } 56 57 58 Scheduler::SchedulerData Scheduler::DefaultSchedulerData() { 59 SchedulerData def = {schedule_->start(), 0, kUnknown}; 60 return def; 61 } 62 63 64 Scheduler::SchedulerData* Scheduler::GetData(Node* node) { 65 return &node_data_[node->id()]; 66 } 67 68 69 Scheduler::Placement Scheduler::GetPlacement(Node* node) { 70 SchedulerData* data = GetData(node); 71 if (data->placement_ == kUnknown) { // Compute placement, once, on demand. 72 switch (node->opcode()) { 73 case IrOpcode::kParameter: 74 case IrOpcode::kOsrValue: 75 // Parameters and OSR values are always fixed to the start block. 76 data->placement_ = kFixed; 77 break; 78 case IrOpcode::kPhi: 79 case IrOpcode::kEffectPhi: { 80 // Phis and effect phis are fixed if their control inputs are, whereas 81 // otherwise they are coupled to a floating control node. 82 Placement p = GetPlacement(NodeProperties::GetControlInput(node)); 83 data->placement_ = (p == kFixed ? kFixed : kCoupled); 84 break; 85 } 86 #define DEFINE_CONTROL_CASE(V) case IrOpcode::k##V: 87 CONTROL_OP_LIST(DEFINE_CONTROL_CASE) 88 #undef DEFINE_CONTROL_CASE 89 { 90 // Control nodes that were not control-reachable from end may float. 91 data->placement_ = kSchedulable; 92 break; 93 } 94 default: 95 data->placement_ = kSchedulable; 96 break; 97 } 98 } 99 return data->placement_; 100 } 101 102 103 void Scheduler::UpdatePlacement(Node* node, Placement placement) { 104 SchedulerData* data = GetData(node); 105 if (data->placement_ != kUnknown) { // Trap on mutation, not initialization. 106 switch (node->opcode()) { 107 case IrOpcode::kParameter: 108 // Parameters are fixed once and for all. 109 UNREACHABLE(); 110 break; 111 case IrOpcode::kPhi: 112 case IrOpcode::kEffectPhi: { 113 // Phis and effect phis are coupled to their respective blocks. 114 DCHECK_EQ(Scheduler::kCoupled, data->placement_); 115 DCHECK_EQ(Scheduler::kFixed, placement); 116 Node* control = NodeProperties::GetControlInput(node); 117 BasicBlock* block = schedule_->block(control); 118 schedule_->AddNode(block, node); 119 break; 120 } 121 #define DEFINE_CONTROL_CASE(V) case IrOpcode::k##V: 122 CONTROL_OP_LIST(DEFINE_CONTROL_CASE) 123 #undef DEFINE_CONTROL_CASE 124 { 125 // Control nodes force coupled uses to be placed. 126 for (auto use : node->uses()) { 127 if (GetPlacement(use) == Scheduler::kCoupled) { 128 DCHECK_EQ(node, NodeProperties::GetControlInput(use)); 129 UpdatePlacement(use, placement); 130 } 131 } 132 break; 133 } 134 default: 135 DCHECK_EQ(Scheduler::kSchedulable, data->placement_); 136 DCHECK_EQ(Scheduler::kScheduled, placement); 137 break; 138 } 139 // Reduce the use count of the node's inputs to potentially make them 140 // schedulable. If all the uses of a node have been scheduled, then the node 141 // itself can be scheduled. 142 for (Edge const edge : node->input_edges()) { 143 DecrementUnscheduledUseCount(edge.to(), edge.index(), edge.from()); 144 } 145 } 146 data->placement_ = placement; 147 } 148 149 150 bool Scheduler::IsCoupledControlEdge(Node* node, int index) { 151 return GetPlacement(node) == kCoupled && 152 NodeProperties::FirstControlIndex(node) == index; 153 } 154 155 156 void Scheduler::IncrementUnscheduledUseCount(Node* node, int index, 157 Node* from) { 158 // Make sure that control edges from coupled nodes are not counted. 159 if (IsCoupledControlEdge(from, index)) return; 160 161 // Tracking use counts for fixed nodes is useless. 162 if (GetPlacement(node) == kFixed) return; 163 164 // Use count for coupled nodes is summed up on their control. 165 if (GetPlacement(node) == kCoupled) { 166 Node* control = NodeProperties::GetControlInput(node); 167 return IncrementUnscheduledUseCount(control, index, from); 168 } 169 170 ++(GetData(node)->unscheduled_count_); 171 if (FLAG_trace_turbo_scheduler) { 172 TRACE(" Use count of #%d:%s (used by #%d:%s)++ = %d\n", node->id(), 173 node->op()->mnemonic(), from->id(), from->op()->mnemonic(), 174 GetData(node)->unscheduled_count_); 175 } 176 } 177 178 179 void Scheduler::DecrementUnscheduledUseCount(Node* node, int index, 180 Node* from) { 181 // Make sure that control edges from coupled nodes are not counted. 182 if (IsCoupledControlEdge(from, index)) return; 183 184 // Tracking use counts for fixed nodes is useless. 185 if (GetPlacement(node) == kFixed) return; 186 187 // Use count for coupled nodes is summed up on their control. 188 if (GetPlacement(node) == kCoupled) { 189 Node* control = NodeProperties::GetControlInput(node); 190 return DecrementUnscheduledUseCount(control, index, from); 191 } 192 193 DCHECK(GetData(node)->unscheduled_count_ > 0); 194 --(GetData(node)->unscheduled_count_); 195 if (FLAG_trace_turbo_scheduler) { 196 TRACE(" Use count of #%d:%s (used by #%d:%s)-- = %d\n", node->id(), 197 node->op()->mnemonic(), from->id(), from->op()->mnemonic(), 198 GetData(node)->unscheduled_count_); 199 } 200 if (GetData(node)->unscheduled_count_ == 0) { 201 TRACE(" newly eligible #%d:%s\n", node->id(), node->op()->mnemonic()); 202 schedule_queue_.push(node); 203 } 204 } 205 206 207 // ----------------------------------------------------------------------------- 208 // Phase 1: Build control-flow graph. 209 210 211 // Internal class to build a control flow graph (i.e the basic blocks and edges 212 // between them within a Schedule) from the node graph. Visits control edges of 213 // the graph backwards from an end node in order to find the connected control 214 // subgraph, needed for scheduling. 215 class CFGBuilder : public ZoneObject { 216 public: 217 CFGBuilder(Zone* zone, Scheduler* scheduler) 218 : zone_(zone), 219 scheduler_(scheduler), 220 schedule_(scheduler->schedule_), 221 queued_(scheduler->graph_, 2), 222 queue_(zone), 223 control_(zone), 224 component_entry_(nullptr), 225 component_start_(nullptr), 226 component_end_(nullptr) {} 227 228 // Run the control flow graph construction algorithm by walking the graph 229 // backwards from end through control edges, building and connecting the 230 // basic blocks for control nodes. 231 void Run() { 232 ResetDataStructures(); 233 Queue(scheduler_->graph_->end()); 234 235 while (!queue_.empty()) { // Breadth-first backwards traversal. 236 Node* node = queue_.front(); 237 queue_.pop(); 238 int max = NodeProperties::PastControlIndex(node); 239 for (int i = NodeProperties::FirstControlIndex(node); i < max; i++) { 240 Queue(node->InputAt(i)); 241 } 242 } 243 244 for (NodeVector::iterator i = control_.begin(); i != control_.end(); ++i) { 245 ConnectBlocks(*i); // Connect block to its predecessor/successors. 246 } 247 } 248 249 // Run the control flow graph construction for a minimal control-connected 250 // component ending in {exit} and merge that component into an existing 251 // control flow graph at the bottom of {block}. 252 void Run(BasicBlock* block, Node* exit) { 253 ResetDataStructures(); 254 Queue(exit); 255 256 component_entry_ = nullptr; 257 component_start_ = block; 258 component_end_ = schedule_->block(exit); 259 scheduler_->equivalence_->Run(exit); 260 while (!queue_.empty()) { // Breadth-first backwards traversal. 261 Node* node = queue_.front(); 262 queue_.pop(); 263 264 // Use control dependence equivalence to find a canonical single-entry 265 // single-exit region that makes up a minimal component to be scheduled. 266 if (IsSingleEntrySingleExitRegion(node, exit)) { 267 TRACE("Found SESE at #%d:%s\n", node->id(), node->op()->mnemonic()); 268 DCHECK(!component_entry_); 269 component_entry_ = node; 270 continue; 271 } 272 273 int max = NodeProperties::PastControlIndex(node); 274 for (int i = NodeProperties::FirstControlIndex(node); i < max; i++) { 275 Queue(node->InputAt(i)); 276 } 277 } 278 DCHECK(component_entry_); 279 280 for (NodeVector::iterator i = control_.begin(); i != control_.end(); ++i) { 281 ConnectBlocks(*i); // Connect block to its predecessor/successors. 282 } 283 } 284 285 private: 286 friend class ScheduleLateNodeVisitor; 287 friend class Scheduler; 288 289 void FixNode(BasicBlock* block, Node* node) { 290 schedule_->AddNode(block, node); 291 scheduler_->UpdatePlacement(node, Scheduler::kFixed); 292 } 293 294 void Queue(Node* node) { 295 // Mark the connected control nodes as they are queued. 296 if (!queued_.Get(node)) { 297 BuildBlocks(node); 298 queue_.push(node); 299 queued_.Set(node, true); 300 control_.push_back(node); 301 } 302 } 303 304 void BuildBlocks(Node* node) { 305 switch (node->opcode()) { 306 case IrOpcode::kEnd: 307 FixNode(schedule_->end(), node); 308 break; 309 case IrOpcode::kStart: 310 FixNode(schedule_->start(), node); 311 break; 312 case IrOpcode::kLoop: 313 case IrOpcode::kMerge: 314 BuildBlockForNode(node); 315 break; 316 case IrOpcode::kTerminate: { 317 // Put Terminate in the loop to which it refers. 318 Node* loop = NodeProperties::GetControlInput(node); 319 BasicBlock* block = BuildBlockForNode(loop); 320 FixNode(block, node); 321 break; 322 } 323 case IrOpcode::kBranch: 324 case IrOpcode::kSwitch: 325 BuildBlocksForSuccessors(node); 326 break; 327 #define BUILD_BLOCK_JS_CASE(Name) case IrOpcode::k##Name: 328 JS_OP_LIST(BUILD_BLOCK_JS_CASE) 329 // JS opcodes are just like calls => fall through. 330 #undef BUILD_BLOCK_JS_CASE 331 case IrOpcode::kCall: 332 if (NodeProperties::IsExceptionalCall(node)) { 333 BuildBlocksForSuccessors(node); 334 } 335 break; 336 default: 337 break; 338 } 339 } 340 341 void ConnectBlocks(Node* node) { 342 switch (node->opcode()) { 343 case IrOpcode::kLoop: 344 case IrOpcode::kMerge: 345 ConnectMerge(node); 346 break; 347 case IrOpcode::kBranch: 348 scheduler_->UpdatePlacement(node, Scheduler::kFixed); 349 ConnectBranch(node); 350 break; 351 case IrOpcode::kSwitch: 352 scheduler_->UpdatePlacement(node, Scheduler::kFixed); 353 ConnectSwitch(node); 354 break; 355 case IrOpcode::kDeoptimize: 356 scheduler_->UpdatePlacement(node, Scheduler::kFixed); 357 ConnectDeoptimize(node); 358 break; 359 case IrOpcode::kTailCall: 360 scheduler_->UpdatePlacement(node, Scheduler::kFixed); 361 ConnectTailCall(node); 362 break; 363 case IrOpcode::kReturn: 364 scheduler_->UpdatePlacement(node, Scheduler::kFixed); 365 ConnectReturn(node); 366 break; 367 case IrOpcode::kThrow: 368 scheduler_->UpdatePlacement(node, Scheduler::kFixed); 369 ConnectThrow(node); 370 break; 371 #define CONNECT_BLOCK_JS_CASE(Name) case IrOpcode::k##Name: 372 JS_OP_LIST(CONNECT_BLOCK_JS_CASE) 373 // JS opcodes are just like calls => fall through. 374 #undef CONNECT_BLOCK_JS_CASE 375 case IrOpcode::kCall: 376 if (NodeProperties::IsExceptionalCall(node)) { 377 scheduler_->UpdatePlacement(node, Scheduler::kFixed); 378 ConnectCall(node); 379 } 380 break; 381 default: 382 break; 383 } 384 } 385 386 BasicBlock* BuildBlockForNode(Node* node) { 387 BasicBlock* block = schedule_->block(node); 388 if (block == nullptr) { 389 block = schedule_->NewBasicBlock(); 390 TRACE("Create block id:%d for #%d:%s\n", block->id().ToInt(), node->id(), 391 node->op()->mnemonic()); 392 FixNode(block, node); 393 } 394 return block; 395 } 396 397 void BuildBlocksForSuccessors(Node* node) { 398 size_t const successor_cnt = node->op()->ControlOutputCount(); 399 Node** successors = zone_->NewArray<Node*>(successor_cnt); 400 NodeProperties::CollectControlProjections(node, successors, successor_cnt); 401 for (size_t index = 0; index < successor_cnt; ++index) { 402 BuildBlockForNode(successors[index]); 403 } 404 } 405 406 void CollectSuccessorBlocks(Node* node, BasicBlock** successor_blocks, 407 size_t successor_cnt) { 408 Node** successors = reinterpret_cast<Node**>(successor_blocks); 409 NodeProperties::CollectControlProjections(node, successors, successor_cnt); 410 for (size_t index = 0; index < successor_cnt; ++index) { 411 successor_blocks[index] = schedule_->block(successors[index]); 412 } 413 } 414 415 BasicBlock* FindPredecessorBlock(Node* node) { 416 BasicBlock* predecessor_block = nullptr; 417 while (true) { 418 predecessor_block = schedule_->block(node); 419 if (predecessor_block != nullptr) break; 420 node = NodeProperties::GetControlInput(node); 421 } 422 return predecessor_block; 423 } 424 425 void ConnectCall(Node* call) { 426 BasicBlock* successor_blocks[2]; 427 CollectSuccessorBlocks(call, successor_blocks, arraysize(successor_blocks)); 428 429 // Consider the exception continuation to be deferred. 430 successor_blocks[1]->set_deferred(true); 431 432 Node* call_control = NodeProperties::GetControlInput(call); 433 BasicBlock* call_block = FindPredecessorBlock(call_control); 434 TraceConnect(call, call_block, successor_blocks[0]); 435 TraceConnect(call, call_block, successor_blocks[1]); 436 schedule_->AddCall(call_block, call, successor_blocks[0], 437 successor_blocks[1]); 438 } 439 440 void ConnectBranch(Node* branch) { 441 BasicBlock* successor_blocks[2]; 442 CollectSuccessorBlocks(branch, successor_blocks, 443 arraysize(successor_blocks)); 444 445 // Consider branch hints. 446 switch (BranchHintOf(branch->op())) { 447 case BranchHint::kNone: 448 break; 449 case BranchHint::kTrue: 450 successor_blocks[1]->set_deferred(true); 451 break; 452 case BranchHint::kFalse: 453 successor_blocks[0]->set_deferred(true); 454 break; 455 } 456 457 if (branch == component_entry_) { 458 TraceConnect(branch, component_start_, successor_blocks[0]); 459 TraceConnect(branch, component_start_, successor_blocks[1]); 460 schedule_->InsertBranch(component_start_, component_end_, branch, 461 successor_blocks[0], successor_blocks[1]); 462 } else { 463 Node* branch_control = NodeProperties::GetControlInput(branch); 464 BasicBlock* branch_block = FindPredecessorBlock(branch_control); 465 TraceConnect(branch, branch_block, successor_blocks[0]); 466 TraceConnect(branch, branch_block, successor_blocks[1]); 467 schedule_->AddBranch(branch_block, branch, successor_blocks[0], 468 successor_blocks[1]); 469 } 470 } 471 472 void ConnectSwitch(Node* sw) { 473 size_t const successor_count = sw->op()->ControlOutputCount(); 474 BasicBlock** successor_blocks = 475 zone_->NewArray<BasicBlock*>(successor_count); 476 CollectSuccessorBlocks(sw, successor_blocks, successor_count); 477 478 if (sw == component_entry_) { 479 for (size_t index = 0; index < successor_count; ++index) { 480 TraceConnect(sw, component_start_, successor_blocks[index]); 481 } 482 schedule_->InsertSwitch(component_start_, component_end_, sw, 483 successor_blocks, successor_count); 484 } else { 485 Node* switch_control = NodeProperties::GetControlInput(sw); 486 BasicBlock* switch_block = FindPredecessorBlock(switch_control); 487 for (size_t index = 0; index < successor_count; ++index) { 488 TraceConnect(sw, switch_block, successor_blocks[index]); 489 } 490 schedule_->AddSwitch(switch_block, sw, successor_blocks, successor_count); 491 } 492 } 493 494 void ConnectMerge(Node* merge) { 495 // Don't connect the special merge at the end to its predecessors. 496 if (IsFinalMerge(merge)) return; 497 498 BasicBlock* block = schedule_->block(merge); 499 DCHECK_NOT_NULL(block); 500 // For all of the merge's control inputs, add a goto at the end to the 501 // merge's basic block. 502 for (Node* const input : merge->inputs()) { 503 BasicBlock* predecessor_block = FindPredecessorBlock(input); 504 TraceConnect(merge, predecessor_block, block); 505 schedule_->AddGoto(predecessor_block, block); 506 } 507 } 508 509 void ConnectTailCall(Node* call) { 510 Node* call_control = NodeProperties::GetControlInput(call); 511 BasicBlock* call_block = FindPredecessorBlock(call_control); 512 TraceConnect(call, call_block, nullptr); 513 schedule_->AddTailCall(call_block, call); 514 } 515 516 void ConnectReturn(Node* ret) { 517 Node* return_control = NodeProperties::GetControlInput(ret); 518 BasicBlock* return_block = FindPredecessorBlock(return_control); 519 TraceConnect(ret, return_block, nullptr); 520 schedule_->AddReturn(return_block, ret); 521 } 522 523 void ConnectDeoptimize(Node* deopt) { 524 Node* deoptimize_control = NodeProperties::GetControlInput(deopt); 525 BasicBlock* deoptimize_block = FindPredecessorBlock(deoptimize_control); 526 TraceConnect(deopt, deoptimize_block, nullptr); 527 schedule_->AddDeoptimize(deoptimize_block, deopt); 528 } 529 530 void ConnectThrow(Node* thr) { 531 Node* throw_control = NodeProperties::GetControlInput(thr); 532 BasicBlock* throw_block = FindPredecessorBlock(throw_control); 533 TraceConnect(thr, throw_block, nullptr); 534 schedule_->AddThrow(throw_block, thr); 535 } 536 537 void TraceConnect(Node* node, BasicBlock* block, BasicBlock* succ) { 538 DCHECK_NOT_NULL(block); 539 if (succ == nullptr) { 540 TRACE("Connect #%d:%s, id:%d -> end\n", node->id(), 541 node->op()->mnemonic(), block->id().ToInt()); 542 } else { 543 TRACE("Connect #%d:%s, id:%d -> id:%d\n", node->id(), 544 node->op()->mnemonic(), block->id().ToInt(), succ->id().ToInt()); 545 } 546 } 547 548 bool IsFinalMerge(Node* node) { 549 return (node->opcode() == IrOpcode::kMerge && 550 node == scheduler_->graph_->end()->InputAt(0)); 551 } 552 553 bool IsSingleEntrySingleExitRegion(Node* entry, Node* exit) const { 554 size_t entry_class = scheduler_->equivalence_->ClassOf(entry); 555 size_t exit_class = scheduler_->equivalence_->ClassOf(exit); 556 return entry != exit && entry_class == exit_class; 557 } 558 559 void ResetDataStructures() { 560 control_.clear(); 561 DCHECK(queue_.empty()); 562 DCHECK(control_.empty()); 563 } 564 565 Zone* zone_; 566 Scheduler* scheduler_; 567 Schedule* schedule_; 568 NodeMarker<bool> queued_; // Mark indicating whether node is queued. 569 ZoneQueue<Node*> queue_; // Queue used for breadth-first traversal. 570 NodeVector control_; // List of encountered control nodes. 571 Node* component_entry_; // Component single-entry node. 572 BasicBlock* component_start_; // Component single-entry block. 573 BasicBlock* component_end_; // Component single-exit block. 574 }; 575 576 577 void Scheduler::BuildCFG() { 578 TRACE("--- CREATING CFG -------------------------------------------\n"); 579 580 // Instantiate a new control equivalence algorithm for the graph. 581 equivalence_ = new (zone_) ControlEquivalence(zone_, graph_); 582 583 // Build a control-flow graph for the main control-connected component that 584 // is being spanned by the graph's start and end nodes. 585 control_flow_builder_ = new (zone_) CFGBuilder(zone_, this); 586 control_flow_builder_->Run(); 587 588 // Initialize per-block data. 589 scheduled_nodes_.resize(schedule_->BasicBlockCount(), NodeVector(zone_)); 590 } 591 592 593 // ----------------------------------------------------------------------------- 594 // Phase 2: Compute special RPO and dominator tree. 595 596 597 // Compute the special reverse-post-order block ordering, which is essentially 598 // a RPO of the graph where loop bodies are contiguous. Properties: 599 // 1. If block A is a predecessor of B, then A appears before B in the order, 600 // unless B is a loop header and A is in the loop headed at B 601 // (i.e. A -> B is a backedge). 602 // => If block A dominates block B, then A appears before B in the order. 603 // => If block A is a loop header, A appears before all blocks in the loop 604 // headed at A. 605 // 2. All loops are contiguous in the order (i.e. no intervening blocks that 606 // do not belong to the loop.) 607 // Note a simple RPO traversal satisfies (1) but not (2). 608 class SpecialRPONumberer : public ZoneObject { 609 public: 610 SpecialRPONumberer(Zone* zone, Schedule* schedule) 611 : zone_(zone), 612 schedule_(schedule), 613 order_(nullptr), 614 beyond_end_(nullptr), 615 loops_(zone), 616 backedges_(zone), 617 stack_(zone), 618 previous_block_count_(0), 619 empty_(0, zone) {} 620 621 // Computes the special reverse-post-order for the main control flow graph, 622 // that is for the graph spanned between the schedule's start and end blocks. 623 void ComputeSpecialRPO() { 624 DCHECK(schedule_->end()->SuccessorCount() == 0); 625 DCHECK(!order_); // Main order does not exist yet. 626 ComputeAndInsertSpecialRPO(schedule_->start(), schedule_->end()); 627 } 628 629 // Computes the special reverse-post-order for a partial control flow graph, 630 // that is for the graph spanned between the given {entry} and {end} blocks, 631 // then updates the existing ordering with this new information. 632 void UpdateSpecialRPO(BasicBlock* entry, BasicBlock* end) { 633 DCHECK(order_); // Main order to be updated is present. 634 ComputeAndInsertSpecialRPO(entry, end); 635 } 636 637 // Serialize the previously computed order as a special reverse-post-order 638 // numbering for basic blocks into the final schedule. 639 void SerializeRPOIntoSchedule() { 640 int32_t number = 0; 641 for (BasicBlock* b = order_; b != nullptr; b = b->rpo_next()) { 642 b->set_rpo_number(number++); 643 schedule_->rpo_order()->push_back(b); 644 } 645 BeyondEndSentinel()->set_rpo_number(number); 646 } 647 648 // Print and verify the special reverse-post-order. 649 void PrintAndVerifySpecialRPO() { 650 #if DEBUG 651 if (FLAG_trace_turbo_scheduler) PrintRPO(); 652 VerifySpecialRPO(); 653 #endif 654 } 655 656 const ZoneVector<BasicBlock*>& GetOutgoingBlocks(BasicBlock* block) { 657 if (HasLoopNumber(block)) { 658 LoopInfo const& loop = loops_[GetLoopNumber(block)]; 659 if (loop.outgoing) return *loop.outgoing; 660 } 661 return empty_; 662 } 663 664 private: 665 typedef std::pair<BasicBlock*, size_t> Backedge; 666 667 // Numbering for BasicBlock::rpo_number for this block traversal: 668 static const int kBlockOnStack = -2; 669 static const int kBlockVisited1 = -3; 670 static const int kBlockVisited2 = -4; 671 static const int kBlockUnvisited1 = -1; 672 static const int kBlockUnvisited2 = kBlockVisited1; 673 674 struct SpecialRPOStackFrame { 675 BasicBlock* block; 676 size_t index; 677 }; 678 679 struct LoopInfo { 680 BasicBlock* header; 681 ZoneVector<BasicBlock*>* outgoing; 682 BitVector* members; 683 LoopInfo* prev; 684 BasicBlock* end; 685 BasicBlock* start; 686 687 void AddOutgoing(Zone* zone, BasicBlock* block) { 688 if (outgoing == nullptr) { 689 outgoing = new (zone->New(sizeof(ZoneVector<BasicBlock*>))) 690 ZoneVector<BasicBlock*>(zone); 691 } 692 outgoing->push_back(block); 693 } 694 }; 695 696 int Push(ZoneVector<SpecialRPOStackFrame>& stack, int depth, 697 BasicBlock* child, int unvisited) { 698 if (child->rpo_number() == unvisited) { 699 stack[depth].block = child; 700 stack[depth].index = 0; 701 child->set_rpo_number(kBlockOnStack); 702 return depth + 1; 703 } 704 return depth; 705 } 706 707 BasicBlock* PushFront(BasicBlock* head, BasicBlock* block) { 708 block->set_rpo_next(head); 709 return block; 710 } 711 712 static int GetLoopNumber(BasicBlock* block) { return block->loop_number(); } 713 static void SetLoopNumber(BasicBlock* block, int loop_number) { 714 return block->set_loop_number(loop_number); 715 } 716 static bool HasLoopNumber(BasicBlock* block) { 717 return block->loop_number() >= 0; 718 } 719 720 // TODO(mstarzinger): We only need this special sentinel because some tests 721 // use the schedule's end block in actual control flow (e.g. with end having 722 // successors). Once this has been cleaned up we can use the end block here. 723 BasicBlock* BeyondEndSentinel() { 724 if (beyond_end_ == nullptr) { 725 BasicBlock::Id id = BasicBlock::Id::FromInt(-1); 726 beyond_end_ = new (schedule_->zone()) BasicBlock(schedule_->zone(), id); 727 } 728 return beyond_end_; 729 } 730 731 // Compute special RPO for the control flow graph between {entry} and {end}, 732 // mutating any existing order so that the result is still valid. 733 void ComputeAndInsertSpecialRPO(BasicBlock* entry, BasicBlock* end) { 734 // RPO should not have been serialized for this schedule yet. 735 CHECK_EQ(kBlockUnvisited1, schedule_->start()->loop_number()); 736 CHECK_EQ(kBlockUnvisited1, schedule_->start()->rpo_number()); 737 CHECK_EQ(0, static_cast<int>(schedule_->rpo_order()->size())); 738 739 // Find correct insertion point within existing order. 740 BasicBlock* insertion_point = entry->rpo_next(); 741 BasicBlock* order = insertion_point; 742 743 // Perform an iterative RPO traversal using an explicit stack, 744 // recording backedges that form cycles. O(|B|). 745 DCHECK_LT(previous_block_count_, schedule_->BasicBlockCount()); 746 stack_.resize(schedule_->BasicBlockCount() - previous_block_count_); 747 previous_block_count_ = schedule_->BasicBlockCount(); 748 int stack_depth = Push(stack_, 0, entry, kBlockUnvisited1); 749 int num_loops = static_cast<int>(loops_.size()); 750 751 while (stack_depth > 0) { 752 int current = stack_depth - 1; 753 SpecialRPOStackFrame* frame = &stack_[current]; 754 755 if (frame->block != end && 756 frame->index < frame->block->SuccessorCount()) { 757 // Process the next successor. 758 BasicBlock* succ = frame->block->SuccessorAt(frame->index++); 759 if (succ->rpo_number() == kBlockVisited1) continue; 760 if (succ->rpo_number() == kBlockOnStack) { 761 // The successor is on the stack, so this is a backedge (cycle). 762 backedges_.push_back(Backedge(frame->block, frame->index - 1)); 763 if (!HasLoopNumber(succ)) { 764 // Assign a new loop number to the header if it doesn't have one. 765 SetLoopNumber(succ, num_loops++); 766 } 767 } else { 768 // Push the successor onto the stack. 769 DCHECK(succ->rpo_number() == kBlockUnvisited1); 770 stack_depth = Push(stack_, stack_depth, succ, kBlockUnvisited1); 771 } 772 } else { 773 // Finished with all successors; pop the stack and add the block. 774 order = PushFront(order, frame->block); 775 frame->block->set_rpo_number(kBlockVisited1); 776 stack_depth--; 777 } 778 } 779 780 // If no loops were encountered, then the order we computed was correct. 781 if (num_loops > static_cast<int>(loops_.size())) { 782 // Otherwise, compute the loop information from the backedges in order 783 // to perform a traversal that groups loop bodies together. 784 ComputeLoopInfo(stack_, num_loops, &backedges_); 785 786 // Initialize the "loop stack". Note the entry could be a loop header. 787 LoopInfo* loop = 788 HasLoopNumber(entry) ? &loops_[GetLoopNumber(entry)] : nullptr; 789 order = insertion_point; 790 791 // Perform an iterative post-order traversal, visiting loop bodies before 792 // edges that lead out of loops. Visits each block once, but linking loop 793 // sections together is linear in the loop size, so overall is 794 // O(|B| + max(loop_depth) * max(|loop|)) 795 stack_depth = Push(stack_, 0, entry, kBlockUnvisited2); 796 while (stack_depth > 0) { 797 SpecialRPOStackFrame* frame = &stack_[stack_depth - 1]; 798 BasicBlock* block = frame->block; 799 BasicBlock* succ = nullptr; 800 801 if (block != end && frame->index < block->SuccessorCount()) { 802 // Process the next normal successor. 803 succ = block->SuccessorAt(frame->index++); 804 } else if (HasLoopNumber(block)) { 805 // Process additional outgoing edges from the loop header. 806 if (block->rpo_number() == kBlockOnStack) { 807 // Finish the loop body the first time the header is left on the 808 // stack. 809 DCHECK(loop != nullptr && loop->header == block); 810 loop->start = PushFront(order, block); 811 order = loop->end; 812 block->set_rpo_number(kBlockVisited2); 813 // Pop the loop stack and continue visiting outgoing edges within 814 // the context of the outer loop, if any. 815 loop = loop->prev; 816 // We leave the loop header on the stack; the rest of this iteration 817 // and later iterations will go through its outgoing edges list. 818 } 819 820 // Use the next outgoing edge if there are any. 821 size_t outgoing_index = frame->index - block->SuccessorCount(); 822 LoopInfo* info = &loops_[GetLoopNumber(block)]; 823 DCHECK(loop != info); 824 if (block != entry && info->outgoing != nullptr && 825 outgoing_index < info->outgoing->size()) { 826 succ = info->outgoing->at(outgoing_index); 827 frame->index++; 828 } 829 } 830 831 if (succ != nullptr) { 832 // Process the next successor. 833 if (succ->rpo_number() == kBlockOnStack) continue; 834 if (succ->rpo_number() == kBlockVisited2) continue; 835 DCHECK(succ->rpo_number() == kBlockUnvisited2); 836 if (loop != nullptr && !loop->members->Contains(succ->id().ToInt())) { 837 // The successor is not in the current loop or any nested loop. 838 // Add it to the outgoing edges of this loop and visit it later. 839 loop->AddOutgoing(zone_, succ); 840 } else { 841 // Push the successor onto the stack. 842 stack_depth = Push(stack_, stack_depth, succ, kBlockUnvisited2); 843 if (HasLoopNumber(succ)) { 844 // Push the inner loop onto the loop stack. 845 DCHECK(GetLoopNumber(succ) < num_loops); 846 LoopInfo* next = &loops_[GetLoopNumber(succ)]; 847 next->end = order; 848 next->prev = loop; 849 loop = next; 850 } 851 } 852 } else { 853 // Finished with all successors of the current block. 854 if (HasLoopNumber(block)) { 855 // If we are going to pop a loop header, then add its entire body. 856 LoopInfo* info = &loops_[GetLoopNumber(block)]; 857 for (BasicBlock* b = info->start; true; b = b->rpo_next()) { 858 if (b->rpo_next() == info->end) { 859 b->set_rpo_next(order); 860 info->end = order; 861 break; 862 } 863 } 864 order = info->start; 865 } else { 866 // Pop a single node off the stack and add it to the order. 867 order = PushFront(order, block); 868 block->set_rpo_number(kBlockVisited2); 869 } 870 stack_depth--; 871 } 872 } 873 } 874 875 // Publish new order the first time. 876 if (order_ == nullptr) order_ = order; 877 878 // Compute the correct loop headers and set the correct loop ends. 879 LoopInfo* current_loop = nullptr; 880 BasicBlock* current_header = entry->loop_header(); 881 int32_t loop_depth = entry->loop_depth(); 882 if (entry->IsLoopHeader()) --loop_depth; // Entry might be a loop header. 883 for (BasicBlock* b = order; b != insertion_point; b = b->rpo_next()) { 884 BasicBlock* current = b; 885 886 // Reset BasicBlock::rpo_number again. 887 current->set_rpo_number(kBlockUnvisited1); 888 889 // Finish the previous loop(s) if we just exited them. 890 while (current_header != nullptr && 891 current == current_header->loop_end()) { 892 DCHECK(current_header->IsLoopHeader()); 893 DCHECK_NOT_NULL(current_loop); 894 current_loop = current_loop->prev; 895 current_header = 896 current_loop == nullptr ? nullptr : current_loop->header; 897 --loop_depth; 898 } 899 current->set_loop_header(current_header); 900 901 // Push a new loop onto the stack if this loop is a loop header. 902 if (HasLoopNumber(current)) { 903 ++loop_depth; 904 current_loop = &loops_[GetLoopNumber(current)]; 905 BasicBlock* end = current_loop->end; 906 current->set_loop_end(end == nullptr ? BeyondEndSentinel() : end); 907 current_header = current_loop->header; 908 TRACE("id:%d is a loop header, increment loop depth to %d\n", 909 current->id().ToInt(), loop_depth); 910 } 911 912 current->set_loop_depth(loop_depth); 913 914 if (current->loop_header() == nullptr) { 915 TRACE("id:%d is not in a loop (depth == %d)\n", current->id().ToInt(), 916 current->loop_depth()); 917 } else { 918 TRACE("id:%d has loop header id:%d, (depth == %d)\n", 919 current->id().ToInt(), current->loop_header()->id().ToInt(), 920 current->loop_depth()); 921 } 922 } 923 } 924 925 // Computes loop membership from the backedges of the control flow graph. 926 void ComputeLoopInfo(ZoneVector<SpecialRPOStackFrame>& queue, 927 size_t num_loops, ZoneVector<Backedge>* backedges) { 928 // Extend existing loop membership vectors. 929 for (LoopInfo& loop : loops_) { 930 BitVector* new_members = new (zone_) 931 BitVector(static_cast<int>(schedule_->BasicBlockCount()), zone_); 932 new_members->CopyFrom(*loop.members); 933 loop.members = new_members; 934 } 935 936 // Extend loop information vector. 937 loops_.resize(num_loops, LoopInfo()); 938 939 // Compute loop membership starting from backedges. 940 // O(max(loop_depth) * max(|loop|) 941 for (size_t i = 0; i < backedges->size(); i++) { 942 BasicBlock* member = backedges->at(i).first; 943 BasicBlock* header = member->SuccessorAt(backedges->at(i).second); 944 size_t loop_num = GetLoopNumber(header); 945 if (loops_[loop_num].header == nullptr) { 946 loops_[loop_num].header = header; 947 loops_[loop_num].members = new (zone_) 948 BitVector(static_cast<int>(schedule_->BasicBlockCount()), zone_); 949 } 950 951 int queue_length = 0; 952 if (member != header) { 953 // As long as the header doesn't have a backedge to itself, 954 // Push the member onto the queue and process its predecessors. 955 if (!loops_[loop_num].members->Contains(member->id().ToInt())) { 956 loops_[loop_num].members->Add(member->id().ToInt()); 957 } 958 queue[queue_length++].block = member; 959 } 960 961 // Propagate loop membership backwards. All predecessors of M up to the 962 // loop header H are members of the loop too. O(|blocks between M and H|). 963 while (queue_length > 0) { 964 BasicBlock* block = queue[--queue_length].block; 965 for (size_t i = 0; i < block->PredecessorCount(); i++) { 966 BasicBlock* pred = block->PredecessorAt(i); 967 if (pred != header) { 968 if (!loops_[loop_num].members->Contains(pred->id().ToInt())) { 969 loops_[loop_num].members->Add(pred->id().ToInt()); 970 queue[queue_length++].block = pred; 971 } 972 } 973 } 974 } 975 } 976 } 977 978 #if DEBUG 979 void PrintRPO() { 980 OFStream os(stdout); 981 os << "RPO with " << loops_.size() << " loops"; 982 if (loops_.size() > 0) { 983 os << " ("; 984 for (size_t i = 0; i < loops_.size(); i++) { 985 if (i > 0) os << " "; 986 os << "id:" << loops_[i].header->id(); 987 } 988 os << ")"; 989 } 990 os << ":\n"; 991 992 for (BasicBlock* block = order_; block != nullptr; 993 block = block->rpo_next()) { 994 os << std::setw(5) << "B" << block->rpo_number() << ":"; 995 for (size_t i = 0; i < loops_.size(); i++) { 996 bool range = loops_[i].header->LoopContains(block); 997 bool membership = loops_[i].header != block && range; 998 os << (membership ? " |" : " "); 999 os << (range ? "x" : " "); 1000 } 1001 os << " id:" << block->id() << ": "; 1002 if (block->loop_end() != nullptr) { 1003 os << " range: [B" << block->rpo_number() << ", B" 1004 << block->loop_end()->rpo_number() << ")"; 1005 } 1006 if (block->loop_header() != nullptr) { 1007 os << " header: id:" << block->loop_header()->id(); 1008 } 1009 if (block->loop_depth() > 0) { 1010 os << " depth: " << block->loop_depth(); 1011 } 1012 os << "\n"; 1013 } 1014 } 1015 1016 void VerifySpecialRPO() { 1017 BasicBlockVector* order = schedule_->rpo_order(); 1018 DCHECK(order->size() > 0); 1019 DCHECK((*order)[0]->id().ToInt() == 0); // entry should be first. 1020 1021 for (size_t i = 0; i < loops_.size(); i++) { 1022 LoopInfo* loop = &loops_[i]; 1023 BasicBlock* header = loop->header; 1024 BasicBlock* end = header->loop_end(); 1025 1026 DCHECK_NOT_NULL(header); 1027 DCHECK(header->rpo_number() >= 0); 1028 DCHECK(header->rpo_number() < static_cast<int>(order->size())); 1029 DCHECK_NOT_NULL(end); 1030 DCHECK(end->rpo_number() <= static_cast<int>(order->size())); 1031 DCHECK(end->rpo_number() > header->rpo_number()); 1032 DCHECK(header->loop_header() != header); 1033 1034 // Verify the start ... end list relationship. 1035 int links = 0; 1036 BasicBlock* block = loop->start; 1037 DCHECK_EQ(header, block); 1038 bool end_found; 1039 while (true) { 1040 if (block == nullptr || block == loop->end) { 1041 end_found = (loop->end == block); 1042 break; 1043 } 1044 // The list should be in same order as the final result. 1045 DCHECK(block->rpo_number() == links + header->rpo_number()); 1046 links++; 1047 block = block->rpo_next(); 1048 DCHECK_LT(links, static_cast<int>(2 * order->size())); // cycle? 1049 } 1050 DCHECK(links > 0); 1051 DCHECK(links == end->rpo_number() - header->rpo_number()); 1052 DCHECK(end_found); 1053 1054 // Check loop depth of the header. 1055 int loop_depth = 0; 1056 for (LoopInfo* outer = loop; outer != nullptr; outer = outer->prev) { 1057 loop_depth++; 1058 } 1059 DCHECK_EQ(loop_depth, header->loop_depth()); 1060 1061 // Check the contiguousness of loops. 1062 int count = 0; 1063 for (int j = 0; j < static_cast<int>(order->size()); j++) { 1064 BasicBlock* block = order->at(j); 1065 DCHECK(block->rpo_number() == j); 1066 if (j < header->rpo_number() || j >= end->rpo_number()) { 1067 DCHECK(!header->LoopContains(block)); 1068 } else { 1069 DCHECK(header->LoopContains(block)); 1070 DCHECK_GE(block->loop_depth(), loop_depth); 1071 count++; 1072 } 1073 } 1074 DCHECK(links == count); 1075 } 1076 } 1077 #endif // DEBUG 1078 1079 Zone* zone_; 1080 Schedule* schedule_; 1081 BasicBlock* order_; 1082 BasicBlock* beyond_end_; 1083 ZoneVector<LoopInfo> loops_; 1084 ZoneVector<Backedge> backedges_; 1085 ZoneVector<SpecialRPOStackFrame> stack_; 1086 size_t previous_block_count_; 1087 ZoneVector<BasicBlock*> const empty_; 1088 }; 1089 1090 1091 BasicBlockVector* Scheduler::ComputeSpecialRPO(Zone* zone, Schedule* schedule) { 1092 SpecialRPONumberer numberer(zone, schedule); 1093 numberer.ComputeSpecialRPO(); 1094 numberer.SerializeRPOIntoSchedule(); 1095 numberer.PrintAndVerifySpecialRPO(); 1096 return schedule->rpo_order(); 1097 } 1098 1099 1100 void Scheduler::ComputeSpecialRPONumbering() { 1101 TRACE("--- COMPUTING SPECIAL RPO ----------------------------------\n"); 1102 1103 // Compute the special reverse-post-order for basic blocks. 1104 special_rpo_ = new (zone_) SpecialRPONumberer(zone_, schedule_); 1105 special_rpo_->ComputeSpecialRPO(); 1106 } 1107 1108 1109 void Scheduler::PropagateImmediateDominators(BasicBlock* block) { 1110 for (/*nop*/; block != nullptr; block = block->rpo_next()) { 1111 auto pred = block->predecessors().begin(); 1112 auto end = block->predecessors().end(); 1113 DCHECK(pred != end); // All blocks except start have predecessors. 1114 BasicBlock* dominator = *pred; 1115 bool deferred = dominator->deferred(); 1116 // For multiple predecessors, walk up the dominator tree until a common 1117 // dominator is found. Visitation order guarantees that all predecessors 1118 // except for backwards edges have been visited. 1119 for (++pred; pred != end; ++pred) { 1120 // Don't examine backwards edges. 1121 if ((*pred)->dominator_depth() < 0) continue; 1122 dominator = BasicBlock::GetCommonDominator(dominator, *pred); 1123 deferred = deferred & (*pred)->deferred(); 1124 } 1125 block->set_dominator(dominator); 1126 block->set_dominator_depth(dominator->dominator_depth() + 1); 1127 block->set_deferred(deferred | block->deferred()); 1128 TRACE("Block id:%d's idom is id:%d, depth = %d\n", block->id().ToInt(), 1129 dominator->id().ToInt(), block->dominator_depth()); 1130 } 1131 } 1132 1133 1134 void Scheduler::GenerateImmediateDominatorTree() { 1135 TRACE("--- IMMEDIATE BLOCK DOMINATORS -----------------------------\n"); 1136 1137 // Seed start block to be the first dominator. 1138 schedule_->start()->set_dominator_depth(0); 1139 1140 // Build the block dominator tree resulting from the above seed. 1141 PropagateImmediateDominators(schedule_->start()->rpo_next()); 1142 } 1143 1144 1145 // ----------------------------------------------------------------------------- 1146 // Phase 3: Prepare use counts for nodes. 1147 1148 1149 class PrepareUsesVisitor { 1150 public: 1151 explicit PrepareUsesVisitor(Scheduler* scheduler) 1152 : scheduler_(scheduler), schedule_(scheduler->schedule_) {} 1153 1154 void Pre(Node* node) { 1155 if (scheduler_->GetPlacement(node) == Scheduler::kFixed) { 1156 // Fixed nodes are always roots for schedule late. 1157 scheduler_->schedule_root_nodes_.push_back(node); 1158 if (!schedule_->IsScheduled(node)) { 1159 // Make sure root nodes are scheduled in their respective blocks. 1160 TRACE("Scheduling fixed position node #%d:%s\n", node->id(), 1161 node->op()->mnemonic()); 1162 IrOpcode::Value opcode = node->opcode(); 1163 BasicBlock* block = 1164 opcode == IrOpcode::kParameter 1165 ? schedule_->start() 1166 : schedule_->block(NodeProperties::GetControlInput(node)); 1167 DCHECK_NOT_NULL(block); 1168 schedule_->AddNode(block, node); 1169 } 1170 } 1171 } 1172 1173 void PostEdge(Node* from, int index, Node* to) { 1174 // If the edge is from an unscheduled node, then tally it in the use count 1175 // for all of its inputs. The same criterion will be used in ScheduleLate 1176 // for decrementing use counts. 1177 if (!schedule_->IsScheduled(from)) { 1178 DCHECK_NE(Scheduler::kFixed, scheduler_->GetPlacement(from)); 1179 scheduler_->IncrementUnscheduledUseCount(to, index, from); 1180 } 1181 } 1182 1183 private: 1184 Scheduler* scheduler_; 1185 Schedule* schedule_; 1186 }; 1187 1188 1189 void Scheduler::PrepareUses() { 1190 TRACE("--- PREPARE USES -------------------------------------------\n"); 1191 1192 // Count the uses of every node, which is used to ensure that all of a 1193 // node's uses are scheduled before the node itself. 1194 PrepareUsesVisitor prepare_uses(this); 1195 1196 // TODO(turbofan): simplify the careful pre/post ordering here. 1197 BoolVector visited(graph_->NodeCount(), false, zone_); 1198 ZoneStack<Node::InputEdges::iterator> stack(zone_); 1199 Node* node = graph_->end(); 1200 prepare_uses.Pre(node); 1201 visited[node->id()] = true; 1202 stack.push(node->input_edges().begin()); 1203 while (!stack.empty()) { 1204 Edge edge = *stack.top(); 1205 Node* node = edge.to(); 1206 if (visited[node->id()]) { 1207 prepare_uses.PostEdge(edge.from(), edge.index(), edge.to()); 1208 if (++stack.top() == edge.from()->input_edges().end()) stack.pop(); 1209 } else { 1210 prepare_uses.Pre(node); 1211 visited[node->id()] = true; 1212 if (node->InputCount() > 0) stack.push(node->input_edges().begin()); 1213 } 1214 } 1215 } 1216 1217 1218 // ----------------------------------------------------------------------------- 1219 // Phase 4: Schedule nodes early. 1220 1221 1222 class ScheduleEarlyNodeVisitor { 1223 public: 1224 ScheduleEarlyNodeVisitor(Zone* zone, Scheduler* scheduler) 1225 : scheduler_(scheduler), schedule_(scheduler->schedule_), queue_(zone) {} 1226 1227 // Run the schedule early algorithm on a set of fixed root nodes. 1228 void Run(NodeVector* roots) { 1229 for (Node* const root : *roots) { 1230 queue_.push(root); 1231 while (!queue_.empty()) { 1232 VisitNode(queue_.front()); 1233 queue_.pop(); 1234 } 1235 } 1236 } 1237 1238 private: 1239 // Visits one node from the queue and propagates its current schedule early 1240 // position to all uses. This in turn might push more nodes onto the queue. 1241 void VisitNode(Node* node) { 1242 Scheduler::SchedulerData* data = scheduler_->GetData(node); 1243 1244 // Fixed nodes already know their schedule early position. 1245 if (scheduler_->GetPlacement(node) == Scheduler::kFixed) { 1246 data->minimum_block_ = schedule_->block(node); 1247 TRACE("Fixing #%d:%s minimum_block = id:%d, dominator_depth = %d\n", 1248 node->id(), node->op()->mnemonic(), 1249 data->minimum_block_->id().ToInt(), 1250 data->minimum_block_->dominator_depth()); 1251 } 1252 1253 // No need to propagate unconstrained schedule early positions. 1254 if (data->minimum_block_ == schedule_->start()) return; 1255 1256 // Propagate schedule early position. 1257 DCHECK_NOT_NULL(data->minimum_block_); 1258 for (auto use : node->uses()) { 1259 PropagateMinimumPositionToNode(data->minimum_block_, use); 1260 } 1261 } 1262 1263 // Propagates {block} as another minimum position into the given {node}. This 1264 // has the net effect of computing the minimum dominator block of {node} that 1265 // still post-dominates all inputs to {node} when the queue is processed. 1266 void PropagateMinimumPositionToNode(BasicBlock* block, Node* node) { 1267 Scheduler::SchedulerData* data = scheduler_->GetData(node); 1268 1269 // No need to propagate to fixed node, it's guaranteed to be a root. 1270 if (scheduler_->GetPlacement(node) == Scheduler::kFixed) return; 1271 1272 // Coupled nodes influence schedule early position of their control. 1273 if (scheduler_->GetPlacement(node) == Scheduler::kCoupled) { 1274 Node* control = NodeProperties::GetControlInput(node); 1275 PropagateMinimumPositionToNode(block, control); 1276 } 1277 1278 // Propagate new position if it is deeper down the dominator tree than the 1279 // current. Note that all inputs need to have minimum block position inside 1280 // the dominator chain of {node}'s minimum block position. 1281 DCHECK(InsideSameDominatorChain(block, data->minimum_block_)); 1282 if (block->dominator_depth() > data->minimum_block_->dominator_depth()) { 1283 data->minimum_block_ = block; 1284 queue_.push(node); 1285 TRACE("Propagating #%d:%s minimum_block = id:%d, dominator_depth = %d\n", 1286 node->id(), node->op()->mnemonic(), 1287 data->minimum_block_->id().ToInt(), 1288 data->minimum_block_->dominator_depth()); 1289 } 1290 } 1291 1292 #if DEBUG 1293 bool InsideSameDominatorChain(BasicBlock* b1, BasicBlock* b2) { 1294 BasicBlock* dominator = BasicBlock::GetCommonDominator(b1, b2); 1295 return dominator == b1 || dominator == b2; 1296 } 1297 #endif 1298 1299 Scheduler* scheduler_; 1300 Schedule* schedule_; 1301 ZoneQueue<Node*> queue_; 1302 }; 1303 1304 1305 void Scheduler::ScheduleEarly() { 1306 TRACE("--- SCHEDULE EARLY -----------------------------------------\n"); 1307 if (FLAG_trace_turbo_scheduler) { 1308 TRACE("roots: "); 1309 for (Node* node : schedule_root_nodes_) { 1310 TRACE("#%d:%s ", node->id(), node->op()->mnemonic()); 1311 } 1312 TRACE("\n"); 1313 } 1314 1315 // Compute the minimum block for each node thereby determining the earliest 1316 // position each node could be placed within a valid schedule. 1317 ScheduleEarlyNodeVisitor schedule_early_visitor(zone_, this); 1318 schedule_early_visitor.Run(&schedule_root_nodes_); 1319 } 1320 1321 1322 // ----------------------------------------------------------------------------- 1323 // Phase 5: Schedule nodes late. 1324 1325 1326 class ScheduleLateNodeVisitor { 1327 public: 1328 ScheduleLateNodeVisitor(Zone* zone, Scheduler* scheduler) 1329 : scheduler_(scheduler), 1330 schedule_(scheduler_->schedule_), 1331 marked_(scheduler->zone_), 1332 marking_queue_(scheduler->zone_) {} 1333 1334 // Run the schedule late algorithm on a set of fixed root nodes. 1335 void Run(NodeVector* roots) { 1336 for (Node* const root : *roots) { 1337 ProcessQueue(root); 1338 } 1339 } 1340 1341 private: 1342 void ProcessQueue(Node* root) { 1343 ZoneQueue<Node*>* queue = &(scheduler_->schedule_queue_); 1344 for (Node* node : root->inputs()) { 1345 // Don't schedule coupled nodes on their own. 1346 if (scheduler_->GetPlacement(node) == Scheduler::kCoupled) { 1347 node = NodeProperties::GetControlInput(node); 1348 } 1349 1350 // Test schedulability condition by looking at unscheduled use count. 1351 if (scheduler_->GetData(node)->unscheduled_count_ != 0) continue; 1352 1353 queue->push(node); 1354 do { 1355 Node* const node = queue->front(); 1356 queue->pop(); 1357 VisitNode(node); 1358 } while (!queue->empty()); 1359 } 1360 } 1361 1362 // Visits one node from the queue of schedulable nodes and determines its 1363 // schedule late position. Also hoists nodes out of loops to find a more 1364 // optimal scheduling position. 1365 void VisitNode(Node* node) { 1366 DCHECK_EQ(0, scheduler_->GetData(node)->unscheduled_count_); 1367 1368 // Don't schedule nodes that are already scheduled. 1369 if (schedule_->IsScheduled(node)) return; 1370 DCHECK_EQ(Scheduler::kSchedulable, scheduler_->GetPlacement(node)); 1371 1372 // Determine the dominating block for all of the uses of this node. It is 1373 // the latest block that this node can be scheduled in. 1374 TRACE("Scheduling #%d:%s\n", node->id(), node->op()->mnemonic()); 1375 BasicBlock* block = GetCommonDominatorOfUses(node); 1376 DCHECK_NOT_NULL(block); 1377 1378 // The schedule early block dominates the schedule late block. 1379 BasicBlock* min_block = scheduler_->GetData(node)->minimum_block_; 1380 DCHECK_EQ(min_block, BasicBlock::GetCommonDominator(block, min_block)); 1381 TRACE( 1382 "Schedule late of #%d:%s is id:%d at loop depth %d, minimum = id:%d\n", 1383 node->id(), node->op()->mnemonic(), block->id().ToInt(), 1384 block->loop_depth(), min_block->id().ToInt()); 1385 1386 // Hoist nodes out of loops if possible. Nodes can be hoisted iteratively 1387 // into enclosing loop pre-headers until they would preceed their schedule 1388 // early position. 1389 BasicBlock* hoist_block = GetHoistBlock(block); 1390 if (hoist_block && 1391 hoist_block->dominator_depth() >= min_block->dominator_depth()) { 1392 do { 1393 TRACE(" hoisting #%d:%s to block id:%d\n", node->id(), 1394 node->op()->mnemonic(), hoist_block->id().ToInt()); 1395 DCHECK_LT(hoist_block->loop_depth(), block->loop_depth()); 1396 block = hoist_block; 1397 hoist_block = GetHoistBlock(hoist_block); 1398 } while (hoist_block && 1399 hoist_block->dominator_depth() >= min_block->dominator_depth()); 1400 } else if (scheduler_->flags_ & Scheduler::kSplitNodes) { 1401 // Split the {node} if beneficial and return the new {block} for it. 1402 block = SplitNode(block, node); 1403 } 1404 1405 // Schedule the node or a floating control structure. 1406 if (IrOpcode::IsMergeOpcode(node->opcode())) { 1407 ScheduleFloatingControl(block, node); 1408 } else if (node->opcode() == IrOpcode::kFinishRegion) { 1409 ScheduleRegion(block, node); 1410 } else { 1411 ScheduleNode(block, node); 1412 } 1413 } 1414 1415 // Mark {block} and push its non-marked predecessor on the marking queue. 1416 void MarkBlock(BasicBlock* block) { 1417 DCHECK_LT(block->id().ToSize(), marked_.size()); 1418 marked_[block->id().ToSize()] = true; 1419 for (BasicBlock* pred_block : block->predecessors()) { 1420 DCHECK_LT(pred_block->id().ToSize(), marked_.size()); 1421 if (marked_[pred_block->id().ToSize()]) continue; 1422 marking_queue_.push_back(pred_block); 1423 } 1424 } 1425 1426 BasicBlock* SplitNode(BasicBlock* block, Node* node) { 1427 // For now, we limit splitting to pure nodes. 1428 if (!node->op()->HasProperty(Operator::kPure)) return block; 1429 // TODO(titzer): fix the special case of splitting of projections. 1430 if (node->opcode() == IrOpcode::kProjection) return block; 1431 1432 // The {block} is common dominator of all uses of {node}, so we cannot 1433 // split anything unless the {block} has at least two successors. 1434 DCHECK_EQ(block, GetCommonDominatorOfUses(node)); 1435 if (block->SuccessorCount() < 2) return block; 1436 1437 // Clear marking bits. 1438 DCHECK(marking_queue_.empty()); 1439 std::fill(marked_.begin(), marked_.end(), false); 1440 marked_.resize(schedule_->BasicBlockCount() + 1, false); 1441 1442 // Check if the {node} has uses in {block}. 1443 for (Edge edge : node->use_edges()) { 1444 BasicBlock* use_block = GetBlockForUse(edge); 1445 if (use_block == nullptr || marked_[use_block->id().ToSize()]) continue; 1446 if (use_block == block) { 1447 TRACE(" not splitting #%d:%s, it is used in id:%d\n", node->id(), 1448 node->op()->mnemonic(), block->id().ToInt()); 1449 marking_queue_.clear(); 1450 return block; 1451 } 1452 MarkBlock(use_block); 1453 } 1454 1455 // Compute transitive marking closure; a block is marked if all its 1456 // successors are marked. 1457 do { 1458 BasicBlock* top_block = marking_queue_.front(); 1459 marking_queue_.pop_front(); 1460 if (marked_[top_block->id().ToSize()]) continue; 1461 bool marked = true; 1462 for (BasicBlock* successor : top_block->successors()) { 1463 if (!marked_[successor->id().ToSize()]) { 1464 marked = false; 1465 break; 1466 } 1467 } 1468 if (marked) MarkBlock(top_block); 1469 } while (!marking_queue_.empty()); 1470 1471 // If the (common dominator) {block} is marked, we know that all paths from 1472 // {block} to the end contain at least one use of {node}, and hence there's 1473 // no point in splitting the {node} in this case. 1474 if (marked_[block->id().ToSize()]) { 1475 TRACE(" not splitting #%d:%s, its common dominator id:%d is perfect\n", 1476 node->id(), node->op()->mnemonic(), block->id().ToInt()); 1477 return block; 1478 } 1479 1480 // Split {node} for uses according to the previously computed marking 1481 // closure. Every marking partition has a unique dominator, which get's a 1482 // copy of the {node} with the exception of the first partition, which get's 1483 // the {node} itself. 1484 ZoneMap<BasicBlock*, Node*> dominators(scheduler_->zone_); 1485 for (Edge edge : node->use_edges()) { 1486 BasicBlock* use_block = GetBlockForUse(edge); 1487 if (use_block == nullptr) continue; 1488 while (marked_[use_block->dominator()->id().ToSize()]) { 1489 use_block = use_block->dominator(); 1490 } 1491 auto& use_node = dominators[use_block]; 1492 if (use_node == nullptr) { 1493 if (dominators.size() == 1u) { 1494 // Place the {node} at {use_block}. 1495 block = use_block; 1496 use_node = node; 1497 TRACE(" pushing #%d:%s down to id:%d\n", node->id(), 1498 node->op()->mnemonic(), block->id().ToInt()); 1499 } else { 1500 // Place a copy of {node} at {use_block}. 1501 use_node = CloneNode(node); 1502 TRACE(" cloning #%d:%s for id:%d\n", use_node->id(), 1503 use_node->op()->mnemonic(), use_block->id().ToInt()); 1504 scheduler_->schedule_queue_.push(use_node); 1505 } 1506 } 1507 edge.UpdateTo(use_node); 1508 } 1509 return block; 1510 } 1511 1512 BasicBlock* GetHoistBlock(BasicBlock* block) { 1513 if (block->IsLoopHeader()) return block->dominator(); 1514 // We have to check to make sure that the {block} dominates all 1515 // of the outgoing blocks. If it doesn't, then there is a path 1516 // out of the loop which does not execute this {block}, so we 1517 // can't hoist operations from this {block} out of the loop, as 1518 // that would introduce additional computations. 1519 if (BasicBlock* header_block = block->loop_header()) { 1520 for (BasicBlock* outgoing_block : 1521 scheduler_->special_rpo_->GetOutgoingBlocks(header_block)) { 1522 if (BasicBlock::GetCommonDominator(block, outgoing_block) != block) { 1523 return nullptr; 1524 } 1525 } 1526 return header_block->dominator(); 1527 } 1528 return nullptr; 1529 } 1530 1531 BasicBlock* GetCommonDominatorOfUses(Node* node) { 1532 BasicBlock* block = nullptr; 1533 for (Edge edge : node->use_edges()) { 1534 BasicBlock* use_block = GetBlockForUse(edge); 1535 block = block == nullptr 1536 ? use_block 1537 : use_block == nullptr 1538 ? block 1539 : BasicBlock::GetCommonDominator(block, use_block); 1540 } 1541 return block; 1542 } 1543 1544 BasicBlock* FindPredecessorBlock(Node* node) { 1545 return scheduler_->control_flow_builder_->FindPredecessorBlock(node); 1546 } 1547 1548 BasicBlock* GetBlockForUse(Edge edge) { 1549 // TODO(titzer): ignore uses from dead nodes (not visited in PrepareUses()). 1550 // Dead uses only occur if the graph is not trimmed before scheduling. 1551 Node* use = edge.from(); 1552 if (IrOpcode::IsPhiOpcode(use->opcode())) { 1553 // If the use is from a coupled (i.e. floating) phi, compute the common 1554 // dominator of its uses. This will not recurse more than one level. 1555 if (scheduler_->GetPlacement(use) == Scheduler::kCoupled) { 1556 TRACE(" inspecting uses of coupled #%d:%s\n", use->id(), 1557 use->op()->mnemonic()); 1558 // TODO(titzer): reenable once above TODO is addressed. 1559 // DCHECK_EQ(edge.to(), NodeProperties::GetControlInput(use)); 1560 return GetCommonDominatorOfUses(use); 1561 } 1562 // If the use is from a fixed (i.e. non-floating) phi, we use the 1563 // predecessor block of the corresponding control input to the merge. 1564 if (scheduler_->GetPlacement(use) == Scheduler::kFixed) { 1565 TRACE(" input@%d into a fixed phi #%d:%s\n", edge.index(), use->id(), 1566 use->op()->mnemonic()); 1567 Node* merge = NodeProperties::GetControlInput(use, 0); 1568 DCHECK(IrOpcode::IsMergeOpcode(merge->opcode())); 1569 Node* input = NodeProperties::GetControlInput(merge, edge.index()); 1570 return FindPredecessorBlock(input); 1571 } 1572 } else if (IrOpcode::IsMergeOpcode(use->opcode())) { 1573 // If the use is from a fixed (i.e. non-floating) merge, we use the 1574 // predecessor block of the current input to the merge. 1575 if (scheduler_->GetPlacement(use) == Scheduler::kFixed) { 1576 TRACE(" input@%d into a fixed merge #%d:%s\n", edge.index(), use->id(), 1577 use->op()->mnemonic()); 1578 return FindPredecessorBlock(edge.to()); 1579 } 1580 } 1581 BasicBlock* result = schedule_->block(use); 1582 if (result == nullptr) return nullptr; 1583 TRACE(" must dominate use #%d:%s in id:%d\n", use->id(), 1584 use->op()->mnemonic(), result->id().ToInt()); 1585 return result; 1586 } 1587 1588 void ScheduleFloatingControl(BasicBlock* block, Node* node) { 1589 scheduler_->FuseFloatingControl(block, node); 1590 } 1591 1592 void ScheduleRegion(BasicBlock* block, Node* region_end) { 1593 // We only allow regions of instructions connected into a linear 1594 // effect chain. The only value allowed to be produced by a node 1595 // in the chain must be the value consumed by the FinishRegion node. 1596 1597 // We schedule back to front; we first schedule FinishRegion. 1598 CHECK_EQ(IrOpcode::kFinishRegion, region_end->opcode()); 1599 ScheduleNode(block, region_end); 1600 1601 // Schedule the chain. 1602 Node* node = NodeProperties::GetEffectInput(region_end); 1603 while (node->opcode() != IrOpcode::kBeginRegion) { 1604 DCHECK_EQ(0, scheduler_->GetData(node)->unscheduled_count_); 1605 DCHECK_EQ(1, node->op()->EffectInputCount()); 1606 DCHECK_EQ(1, node->op()->EffectOutputCount()); 1607 DCHECK_EQ(0, node->op()->ControlOutputCount()); 1608 // The value output (if there is any) must be consumed 1609 // by the EndRegion node. 1610 DCHECK(node->op()->ValueOutputCount() == 0 || 1611 node == region_end->InputAt(0)); 1612 ScheduleNode(block, node); 1613 node = NodeProperties::GetEffectInput(node); 1614 } 1615 // Schedule the BeginRegion node. 1616 DCHECK_EQ(0, scheduler_->GetData(node)->unscheduled_count_); 1617 ScheduleNode(block, node); 1618 } 1619 1620 void ScheduleNode(BasicBlock* block, Node* node) { 1621 schedule_->PlanNode(block, node); 1622 scheduler_->scheduled_nodes_[block->id().ToSize()].push_back(node); 1623 scheduler_->UpdatePlacement(node, Scheduler::kScheduled); 1624 } 1625 1626 Node* CloneNode(Node* node) { 1627 int const input_count = node->InputCount(); 1628 for (int index = 0; index < input_count; ++index) { 1629 Node* const input = node->InputAt(index); 1630 scheduler_->IncrementUnscheduledUseCount(input, index, node); 1631 } 1632 Node* const copy = scheduler_->graph_->CloneNode(node); 1633 TRACE(("clone #%d:%s -> #%d\n"), node->id(), node->op()->mnemonic(), 1634 copy->id()); 1635 scheduler_->node_data_.resize(copy->id() + 1, 1636 scheduler_->DefaultSchedulerData()); 1637 scheduler_->node_data_[copy->id()] = scheduler_->node_data_[node->id()]; 1638 return copy; 1639 } 1640 1641 Scheduler* scheduler_; 1642 Schedule* schedule_; 1643 BoolVector marked_; 1644 ZoneDeque<BasicBlock*> marking_queue_; 1645 }; 1646 1647 1648 void Scheduler::ScheduleLate() { 1649 TRACE("--- SCHEDULE LATE ------------------------------------------\n"); 1650 if (FLAG_trace_turbo_scheduler) { 1651 TRACE("roots: "); 1652 for (Node* node : schedule_root_nodes_) { 1653 TRACE("#%d:%s ", node->id(), node->op()->mnemonic()); 1654 } 1655 TRACE("\n"); 1656 } 1657 1658 // Schedule: Places nodes in dominator block of all their uses. 1659 ScheduleLateNodeVisitor schedule_late_visitor(zone_, this); 1660 schedule_late_visitor.Run(&schedule_root_nodes_); 1661 } 1662 1663 1664 // ----------------------------------------------------------------------------- 1665 // Phase 6: Seal the final schedule. 1666 1667 1668 void Scheduler::SealFinalSchedule() { 1669 TRACE("--- SEAL FINAL SCHEDULE ------------------------------------\n"); 1670 1671 // Serialize the assembly order and reverse-post-order numbering. 1672 special_rpo_->SerializeRPOIntoSchedule(); 1673 special_rpo_->PrintAndVerifySpecialRPO(); 1674 1675 // Add collected nodes for basic blocks to their blocks in the right order. 1676 int block_num = 0; 1677 for (NodeVector& nodes : scheduled_nodes_) { 1678 BasicBlock::Id id = BasicBlock::Id::FromInt(block_num++); 1679 BasicBlock* block = schedule_->GetBlockById(id); 1680 for (Node* node : base::Reversed(nodes)) { 1681 schedule_->AddNode(block, node); 1682 } 1683 } 1684 } 1685 1686 1687 // ----------------------------------------------------------------------------- 1688 1689 1690 void Scheduler::FuseFloatingControl(BasicBlock* block, Node* node) { 1691 TRACE("--- FUSE FLOATING CONTROL ----------------------------------\n"); 1692 if (FLAG_trace_turbo_scheduler) { 1693 OFStream os(stdout); 1694 os << "Schedule before control flow fusion:\n" << *schedule_; 1695 } 1696 1697 // Iterate on phase 1: Build control-flow graph. 1698 control_flow_builder_->Run(block, node); 1699 1700 // Iterate on phase 2: Compute special RPO and dominator tree. 1701 special_rpo_->UpdateSpecialRPO(block, schedule_->block(node)); 1702 // TODO(mstarzinger): Currently "iterate on" means "re-run". Fix that. 1703 for (BasicBlock* b = block->rpo_next(); b != nullptr; b = b->rpo_next()) { 1704 b->set_dominator_depth(-1); 1705 b->set_dominator(nullptr); 1706 } 1707 PropagateImmediateDominators(block->rpo_next()); 1708 1709 // Iterate on phase 4: Schedule nodes early. 1710 // TODO(mstarzinger): The following loop gathering the propagation roots is a 1711 // temporary solution and should be merged into the rest of the scheduler as 1712 // soon as the approach settled for all floating loops. 1713 NodeVector propagation_roots(control_flow_builder_->control_); 1714 for (Node* node : control_flow_builder_->control_) { 1715 for (Node* use : node->uses()) { 1716 if (NodeProperties::IsPhi(use)) propagation_roots.push_back(use); 1717 } 1718 } 1719 if (FLAG_trace_turbo_scheduler) { 1720 TRACE("propagation roots: "); 1721 for (Node* node : propagation_roots) { 1722 TRACE("#%d:%s ", node->id(), node->op()->mnemonic()); 1723 } 1724 TRACE("\n"); 1725 } 1726 ScheduleEarlyNodeVisitor schedule_early_visitor(zone_, this); 1727 schedule_early_visitor.Run(&propagation_roots); 1728 1729 // Move previously planned nodes. 1730 // TODO(mstarzinger): Improve that by supporting bulk moves. 1731 scheduled_nodes_.resize(schedule_->BasicBlockCount(), NodeVector(zone_)); 1732 MovePlannedNodes(block, schedule_->block(node)); 1733 1734 if (FLAG_trace_turbo_scheduler) { 1735 OFStream os(stdout); 1736 os << "Schedule after control flow fusion:\n" << *schedule_; 1737 } 1738 } 1739 1740 1741 void Scheduler::MovePlannedNodes(BasicBlock* from, BasicBlock* to) { 1742 TRACE("Move planned nodes from id:%d to id:%d\n", from->id().ToInt(), 1743 to->id().ToInt()); 1744 NodeVector* nodes = &(scheduled_nodes_[from->id().ToSize()]); 1745 for (Node* const node : *nodes) { 1746 schedule_->SetBlockForNode(to, node); 1747 scheduled_nodes_[to->id().ToSize()].push_back(node); 1748 } 1749 nodes->clear(); 1750 } 1751 1752 } // namespace compiler 1753 } // namespace internal 1754 } // namespace v8 1755