ScheduleDAG.cpp revision bae5210321e0c3ea28723334841555135a61e915
1//===---- ScheduleDAG.cpp - Implement the ScheduleDAG class ---------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This implements the ScheduleDAG class, which is a base class used by 11// scheduling implementation classes. 12// 13//===----------------------------------------------------------------------===// 14 15#define DEBUG_TYPE "pre-RA-sched" 16#include "llvm/CodeGen/ScheduleDAG.h" 17#include "llvm/CodeGen/ScheduleHazardRecognizer.h" 18#include "llvm/Target/TargetMachine.h" 19#include "llvm/Target/TargetInstrInfo.h" 20#include "llvm/Target/TargetRegisterInfo.h" 21#include "llvm/Support/Debug.h" 22#include "llvm/Support/raw_ostream.h" 23#include <climits> 24using namespace llvm; 25 26ScheduleDAG::ScheduleDAG(MachineFunction &mf) 27 : TM(mf.getTarget()), 28 TII(TM.getInstrInfo()), 29 TRI(TM.getRegisterInfo()), 30 MF(mf), MRI(mf.getRegInfo()), 31 EntrySU(), ExitSU() { 32} 33 34ScheduleDAG::~ScheduleDAG() {} 35 36/// dump - dump the schedule. 37void ScheduleDAG::dumpSchedule() const { 38 for (unsigned i = 0, e = Sequence.size(); i != e; i++) { 39 if (SUnit *SU = Sequence[i]) 40 SU->dump(this); 41 else 42 dbgs() << "**** NOOP ****\n"; 43 } 44} 45 46 47/// Run - perform scheduling. 48/// 49void ScheduleDAG::Run(MachineBasicBlock *bb, 50 MachineBasicBlock::iterator insertPos) { 51 BB = bb; 52 InsertPos = insertPos; 53 54 SUnits.clear(); 55 Sequence.clear(); 56 EntrySU = SUnit(); 57 ExitSU = SUnit(); 58 59 Schedule(); 60 61 DEBUG({ 62 dbgs() << "*** Final schedule ***\n"; 63 dumpSchedule(); 64 dbgs() << '\n'; 65 }); 66} 67 68/// addPred - This adds the specified edge as a pred of the current node if 69/// not already. It also adds the current node as a successor of the 70/// specified node. 71void SUnit::addPred(const SDep &D) { 72 // If this node already has this depenence, don't add a redundant one. 73 for (SmallVector<SDep, 4>::const_iterator I = Preds.begin(), E = Preds.end(); 74 I != E; ++I) 75 if (*I == D) 76 return; 77 // Now add a corresponding succ to N. 78 SDep P = D; 79 P.setSUnit(this); 80 SUnit *N = D.getSUnit(); 81 // Update the bookkeeping. 82 if (D.getKind() == SDep::Data) { 83 assert(NumPreds < UINT_MAX && "NumPreds will overflow!"); 84 assert(N->NumSuccs < UINT_MAX && "NumSuccs will overflow!"); 85 ++NumPreds; 86 ++N->NumSuccs; 87 } 88 if (!N->isScheduled) { 89 assert(NumPredsLeft < UINT_MAX && "NumPredsLeft will overflow!"); 90 ++NumPredsLeft; 91 } 92 if (!isScheduled) { 93 assert(N->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!"); 94 ++N->NumSuccsLeft; 95 } 96 Preds.push_back(D); 97 N->Succs.push_back(P); 98 if (P.getLatency() != 0) { 99 this->setDepthDirty(); 100 N->setHeightDirty(); 101 } 102} 103 104/// removePred - This removes the specified edge as a pred of the current 105/// node if it exists. It also removes the current node as a successor of 106/// the specified node. 107void SUnit::removePred(const SDep &D) { 108 // Find the matching predecessor. 109 for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end(); 110 I != E; ++I) 111 if (*I == D) { 112 bool FoundSucc = false; 113 // Find the corresponding successor in N. 114 SDep P = D; 115 P.setSUnit(this); 116 SUnit *N = D.getSUnit(); 117 for (SmallVector<SDep, 4>::iterator II = N->Succs.begin(), 118 EE = N->Succs.end(); II != EE; ++II) 119 if (*II == P) { 120 FoundSucc = true; 121 N->Succs.erase(II); 122 break; 123 } 124 assert(FoundSucc && "Mismatching preds / succs lists!"); 125 Preds.erase(I); 126 // Update the bookkeeping. 127 if (P.getKind() == SDep::Data) { 128 assert(NumPreds > 0 && "NumPreds will underflow!"); 129 assert(N->NumSuccs > 0 && "NumSuccs will underflow!"); 130 --NumPreds; 131 --N->NumSuccs; 132 } 133 if (!N->isScheduled) { 134 assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!"); 135 --NumPredsLeft; 136 } 137 if (!isScheduled) { 138 assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!"); 139 --N->NumSuccsLeft; 140 } 141 if (P.getLatency() != 0) { 142 this->setDepthDirty(); 143 N->setHeightDirty(); 144 } 145 return; 146 } 147} 148 149void SUnit::setDepthDirty() { 150 if (!isDepthCurrent) return; 151 SmallVector<SUnit*, 8> WorkList; 152 WorkList.push_back(this); 153 do { 154 SUnit *SU = WorkList.pop_back_val(); 155 SU->isDepthCurrent = false; 156 for (SUnit::const_succ_iterator I = SU->Succs.begin(), 157 E = SU->Succs.end(); I != E; ++I) { 158 SUnit *SuccSU = I->getSUnit(); 159 if (SuccSU->isDepthCurrent) 160 WorkList.push_back(SuccSU); 161 } 162 } while (!WorkList.empty()); 163} 164 165void SUnit::setHeightDirty() { 166 if (!isHeightCurrent) return; 167 SmallVector<SUnit*, 8> WorkList; 168 WorkList.push_back(this); 169 do { 170 SUnit *SU = WorkList.pop_back_val(); 171 SU->isHeightCurrent = false; 172 for (SUnit::const_pred_iterator I = SU->Preds.begin(), 173 E = SU->Preds.end(); I != E; ++I) { 174 SUnit *PredSU = I->getSUnit(); 175 if (PredSU->isHeightCurrent) 176 WorkList.push_back(PredSU); 177 } 178 } while (!WorkList.empty()); 179} 180 181/// setDepthToAtLeast - Update this node's successors to reflect the 182/// fact that this node's depth just increased. 183/// 184void SUnit::setDepthToAtLeast(unsigned NewDepth) { 185 if (NewDepth <= getDepth()) 186 return; 187 setDepthDirty(); 188 Depth = NewDepth; 189 isDepthCurrent = true; 190} 191 192/// setHeightToAtLeast - Update this node's predecessors to reflect the 193/// fact that this node's height just increased. 194/// 195void SUnit::setHeightToAtLeast(unsigned NewHeight) { 196 if (NewHeight <= getHeight()) 197 return; 198 setHeightDirty(); 199 Height = NewHeight; 200 isHeightCurrent = true; 201} 202 203/// ComputeDepth - Calculate the maximal path from the node to the exit. 204/// 205void SUnit::ComputeDepth() { 206 SmallVector<SUnit*, 8> WorkList; 207 WorkList.push_back(this); 208 do { 209 SUnit *Cur = WorkList.back(); 210 211 bool Done = true; 212 unsigned MaxPredDepth = 0; 213 for (SUnit::const_pred_iterator I = Cur->Preds.begin(), 214 E = Cur->Preds.end(); I != E; ++I) { 215 SUnit *PredSU = I->getSUnit(); 216 if (PredSU->isDepthCurrent) 217 MaxPredDepth = std::max(MaxPredDepth, 218 PredSU->Depth + I->getLatency()); 219 else { 220 Done = false; 221 WorkList.push_back(PredSU); 222 } 223 } 224 225 if (Done) { 226 WorkList.pop_back(); 227 if (MaxPredDepth != Cur->Depth) { 228 Cur->setDepthDirty(); 229 Cur->Depth = MaxPredDepth; 230 } 231 Cur->isDepthCurrent = true; 232 } 233 } while (!WorkList.empty()); 234} 235 236/// ComputeHeight - Calculate the maximal path from the node to the entry. 237/// 238void SUnit::ComputeHeight() { 239 SmallVector<SUnit*, 8> WorkList; 240 WorkList.push_back(this); 241 do { 242 SUnit *Cur = WorkList.back(); 243 244 bool Done = true; 245 unsigned MaxSuccHeight = 0; 246 for (SUnit::const_succ_iterator I = Cur->Succs.begin(), 247 E = Cur->Succs.end(); I != E; ++I) { 248 SUnit *SuccSU = I->getSUnit(); 249 if (SuccSU->isHeightCurrent) 250 MaxSuccHeight = std::max(MaxSuccHeight, 251 SuccSU->Height + I->getLatency()); 252 else { 253 Done = false; 254 WorkList.push_back(SuccSU); 255 } 256 } 257 258 if (Done) { 259 WorkList.pop_back(); 260 if (MaxSuccHeight != Cur->Height) { 261 Cur->setHeightDirty(); 262 Cur->Height = MaxSuccHeight; 263 } 264 Cur->isHeightCurrent = true; 265 } 266 } while (!WorkList.empty()); 267} 268 269/// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or 270/// a group of nodes flagged together. 271void SUnit::dump(const ScheduleDAG *G) const { 272 dbgs() << "SU(" << NodeNum << "): "; 273 G->dumpNode(this); 274} 275 276void SUnit::dumpAll(const ScheduleDAG *G) const { 277 dump(G); 278 279 dbgs() << " # preds left : " << NumPredsLeft << "\n"; 280 dbgs() << " # succs left : " << NumSuccsLeft << "\n"; 281 dbgs() << " Latency : " << Latency << "\n"; 282 dbgs() << " Depth : " << Depth << "\n"; 283 dbgs() << " Height : " << Height << "\n"; 284 285 if (Preds.size() != 0) { 286 dbgs() << " Predecessors:\n"; 287 for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end(); 288 I != E; ++I) { 289 dbgs() << " "; 290 switch (I->getKind()) { 291 case SDep::Data: dbgs() << "val "; break; 292 case SDep::Anti: dbgs() << "anti"; break; 293 case SDep::Output: dbgs() << "out "; break; 294 case SDep::Order: dbgs() << "ch "; break; 295 } 296 dbgs() << "#"; 297 dbgs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")"; 298 if (I->isArtificial()) 299 dbgs() << " *"; 300 dbgs() << ": Latency=" << I->getLatency(); 301 dbgs() << "\n"; 302 } 303 } 304 if (Succs.size() != 0) { 305 dbgs() << " Successors:\n"; 306 for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end(); 307 I != E; ++I) { 308 dbgs() << " "; 309 switch (I->getKind()) { 310 case SDep::Data: dbgs() << "val "; break; 311 case SDep::Anti: dbgs() << "anti"; break; 312 case SDep::Output: dbgs() << "out "; break; 313 case SDep::Order: dbgs() << "ch "; break; 314 } 315 dbgs() << "#"; 316 dbgs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")"; 317 if (I->isArtificial()) 318 dbgs() << " *"; 319 dbgs() << ": Latency=" << I->getLatency(); 320 dbgs() << "\n"; 321 } 322 } 323 dbgs() << "\n"; 324} 325 326#ifndef NDEBUG 327/// VerifySchedule - Verify that all SUnits were scheduled and that 328/// their state is consistent. 329/// 330void ScheduleDAG::VerifySchedule(bool isBottomUp) { 331 bool AnyNotSched = false; 332 unsigned DeadNodes = 0; 333 unsigned Noops = 0; 334 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) { 335 if (!SUnits[i].isScheduled) { 336 if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) { 337 ++DeadNodes; 338 continue; 339 } 340 if (!AnyNotSched) 341 dbgs() << "*** Scheduling failed! ***\n"; 342 SUnits[i].dump(this); 343 dbgs() << "has not been scheduled!\n"; 344 AnyNotSched = true; 345 } 346 if (SUnits[i].isScheduled && 347 (isBottomUp ? SUnits[i].getHeight() : SUnits[i].getDepth()) > 348 unsigned(INT_MAX)) { 349 if (!AnyNotSched) 350 dbgs() << "*** Scheduling failed! ***\n"; 351 SUnits[i].dump(this); 352 dbgs() << "has an unexpected " 353 << (isBottomUp ? "Height" : "Depth") << " value!\n"; 354 AnyNotSched = true; 355 } 356 if (isBottomUp) { 357 if (SUnits[i].NumSuccsLeft != 0) { 358 if (!AnyNotSched) 359 dbgs() << "*** Scheduling failed! ***\n"; 360 SUnits[i].dump(this); 361 dbgs() << "has successors left!\n"; 362 AnyNotSched = true; 363 } 364 } else { 365 if (SUnits[i].NumPredsLeft != 0) { 366 if (!AnyNotSched) 367 dbgs() << "*** Scheduling failed! ***\n"; 368 SUnits[i].dump(this); 369 dbgs() << "has predecessors left!\n"; 370 AnyNotSched = true; 371 } 372 } 373 } 374 for (unsigned i = 0, e = Sequence.size(); i != e; ++i) 375 if (!Sequence[i]) 376 ++Noops; 377 assert(!AnyNotSched); 378 assert(Sequence.size() + DeadNodes - Noops == SUnits.size() && 379 "The number of nodes scheduled doesn't match the expected number!"); 380} 381#endif 382 383/// InitDAGTopologicalSorting - create the initial topological 384/// ordering from the DAG to be scheduled. 385/// 386/// The idea of the algorithm is taken from 387/// "Online algorithms for managing the topological order of 388/// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly 389/// This is the MNR algorithm, which was first introduced by 390/// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in 391/// "Maintaining a topological order under edge insertions". 392/// 393/// Short description of the algorithm: 394/// 395/// Topological ordering, ord, of a DAG maps each node to a topological 396/// index so that for all edges X->Y it is the case that ord(X) < ord(Y). 397/// 398/// This means that if there is a path from the node X to the node Z, 399/// then ord(X) < ord(Z). 400/// 401/// This property can be used to check for reachability of nodes: 402/// if Z is reachable from X, then an insertion of the edge Z->X would 403/// create a cycle. 404/// 405/// The algorithm first computes a topological ordering for the DAG by 406/// initializing the Index2Node and Node2Index arrays and then tries to keep 407/// the ordering up-to-date after edge insertions by reordering the DAG. 408/// 409/// On insertion of the edge X->Y, the algorithm first marks by calling DFS 410/// the nodes reachable from Y, and then shifts them using Shift to lie 411/// immediately after X in Index2Node. 412void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() { 413 unsigned DAGSize = SUnits.size(); 414 std::vector<SUnit*> WorkList; 415 WorkList.reserve(DAGSize); 416 417 Index2Node.resize(DAGSize); 418 Node2Index.resize(DAGSize); 419 420 // Initialize the data structures. 421 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 422 SUnit *SU = &SUnits[i]; 423 int NodeNum = SU->NodeNum; 424 unsigned Degree = SU->Succs.size(); 425 // Temporarily use the Node2Index array as scratch space for degree counts. 426 Node2Index[NodeNum] = Degree; 427 428 // Is it a node without dependencies? 429 if (Degree == 0) { 430 assert(SU->Succs.empty() && "SUnit should have no successors"); 431 // Collect leaf nodes. 432 WorkList.push_back(SU); 433 } 434 } 435 436 int Id = DAGSize; 437 while (!WorkList.empty()) { 438 SUnit *SU = WorkList.back(); 439 WorkList.pop_back(); 440 Allocate(SU->NodeNum, --Id); 441 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 442 I != E; ++I) { 443 SUnit *SU = I->getSUnit(); 444 if (!--Node2Index[SU->NodeNum]) 445 // If all dependencies of the node are processed already, 446 // then the node can be computed now. 447 WorkList.push_back(SU); 448 } 449 } 450 451 Visited.resize(DAGSize); 452 453#ifndef NDEBUG 454 // Check correctness of the ordering 455 for (unsigned i = 0, e = DAGSize; i != e; ++i) { 456 SUnit *SU = &SUnits[i]; 457 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 458 I != E; ++I) { 459 assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] && 460 "Wrong topological sorting"); 461 } 462 } 463#endif 464} 465 466/// AddPred - Updates the topological ordering to accomodate an edge 467/// to be added from SUnit X to SUnit Y. 468void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) { 469 int UpperBound, LowerBound; 470 LowerBound = Node2Index[Y->NodeNum]; 471 UpperBound = Node2Index[X->NodeNum]; 472 bool HasLoop = false; 473 // Is Ord(X) < Ord(Y) ? 474 if (LowerBound < UpperBound) { 475 // Update the topological order. 476 Visited.reset(); 477 DFS(Y, UpperBound, HasLoop); 478 assert(!HasLoop && "Inserted edge creates a loop!"); 479 // Recompute topological indexes. 480 Shift(Visited, LowerBound, UpperBound); 481 } 482} 483 484/// RemovePred - Updates the topological ordering to accomodate an 485/// an edge to be removed from the specified node N from the predecessors 486/// of the current node M. 487void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) { 488 // InitDAGTopologicalSorting(); 489} 490 491/// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark 492/// all nodes affected by the edge insertion. These nodes will later get new 493/// topological indexes by means of the Shift method. 494void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound, 495 bool& HasLoop) { 496 std::vector<const SUnit*> WorkList; 497 WorkList.reserve(SUnits.size()); 498 499 WorkList.push_back(SU); 500 do { 501 SU = WorkList.back(); 502 WorkList.pop_back(); 503 Visited.set(SU->NodeNum); 504 for (int I = SU->Succs.size()-1; I >= 0; --I) { 505 int s = SU->Succs[I].getSUnit()->NodeNum; 506 if (Node2Index[s] == UpperBound) { 507 HasLoop = true; 508 return; 509 } 510 // Visit successors if not already and in affected region. 511 if (!Visited.test(s) && Node2Index[s] < UpperBound) { 512 WorkList.push_back(SU->Succs[I].getSUnit()); 513 } 514 } 515 } while (!WorkList.empty()); 516} 517 518/// Shift - Renumber the nodes so that the topological ordering is 519/// preserved. 520void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound, 521 int UpperBound) { 522 std::vector<int> L; 523 int shift = 0; 524 int i; 525 526 for (i = LowerBound; i <= UpperBound; ++i) { 527 // w is node at topological index i. 528 int w = Index2Node[i]; 529 if (Visited.test(w)) { 530 // Unmark. 531 Visited.reset(w); 532 L.push_back(w); 533 shift = shift + 1; 534 } else { 535 Allocate(w, i - shift); 536 } 537 } 538 539 for (unsigned j = 0; j < L.size(); ++j) { 540 Allocate(L[j], i - shift); 541 i = i + 1; 542 } 543} 544 545 546/// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will 547/// create a cycle. 548bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *SU, SUnit *TargetSU) { 549 if (IsReachable(TargetSU, SU)) 550 return true; 551 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 552 I != E; ++I) 553 if (I->isAssignedRegDep() && 554 IsReachable(TargetSU, I->getSUnit())) 555 return true; 556 return false; 557} 558 559/// IsReachable - Checks if SU is reachable from TargetSU. 560bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU, 561 const SUnit *TargetSU) { 562 // If insertion of the edge SU->TargetSU would create a cycle 563 // then there is a path from TargetSU to SU. 564 int UpperBound, LowerBound; 565 LowerBound = Node2Index[TargetSU->NodeNum]; 566 UpperBound = Node2Index[SU->NodeNum]; 567 bool HasLoop = false; 568 // Is Ord(TargetSU) < Ord(SU) ? 569 if (LowerBound < UpperBound) { 570 Visited.reset(); 571 // There may be a path from TargetSU to SU. Check for it. 572 DFS(TargetSU, UpperBound, HasLoop); 573 } 574 return HasLoop; 575} 576 577/// Allocate - assign the topological index to the node n. 578void ScheduleDAGTopologicalSort::Allocate(int n, int index) { 579 Node2Index[n] = index; 580 Index2Node[index] = n; 581} 582 583ScheduleDAGTopologicalSort::ScheduleDAGTopologicalSort( 584 std::vector<SUnit> &sunits) 585 : SUnits(sunits) {} 586 587ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {} 588