lib/CodeGen/ScheduleDAG.cpp

   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 /// \file Implements the ScheduleDAG class, which is a base class used by
  11 /// scheduling implementation classes.
  12 //
  13 //===----------------------------------------------------------------------===//
  14
  15 #include "llvm/CodeGen/ScheduleDAG.h"
  16 #include "llvm/ADT/STLExtras.h"
  17 #include "llvm/ADT/SmallVector.h"
  18 #include "llvm/ADT/iterator_range.h"
  19 #include "llvm/CodeGen/MachineFunction.h"
  20 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
  21 #include "llvm/CodeGen/SelectionDAGNodes.h"
  22 #include "llvm/CodeGen/TargetInstrInfo.h"
  23 #include "llvm/CodeGen/TargetRegisterInfo.h"
  24 #include "llvm/CodeGen/TargetSubtargetInfo.h"
  25 #include "llvm/Config/llvm-config.h"
  26 #include "llvm/Support/CommandLine.h"
  27 #include "llvm/Support/Compiler.h"
  28 #include "llvm/Support/Debug.h"
  29 #include "llvm/Support/raw_ostream.h"
  30 #include <algorithm>
  31 #include <cassert>
  32 #include <iterator>
  33 #include <limits>
  34 #include <utility>
  35 #include <vector>
  36
  37 using namespace llvm;
  38
  39 #define DEBUG_TYPE "pre-RA-sched"
  40
  41 #ifndef NDEBUG
  42 static cl::opt<bool> StressSchedOpt(
  43   "stress-sched", cl::Hidden, cl::init(false),
  44   cl::desc("Stress test instruction scheduling"));
  45 #endif
  46
  47 void SchedulingPriorityQueue::anchor() {}
  48
  49 ScheduleDAG::ScheduleDAG(MachineFunction &mf)
  50     : TM(mf.getTarget()), TII(mf.getSubtarget().getInstrInfo()),
  51       TRI(mf.getSubtarget().getRegisterInfo()), MF(mf),
  52       MRI(mf.getRegInfo()) {
  53 #ifndef NDEBUG
  54   StressSched = StressSchedOpt;
  55 #endif
  56 }
  57
  58 ScheduleDAG::~ScheduleDAG() = default;
  59
  60 void ScheduleDAG::clearDAG() {
  61   SUnits.clear();
  62   EntrySU = SUnit();
  63   ExitSU = SUnit();
  64 }
  65
  66 const MCInstrDesc *ScheduleDAG::getNodeDesc(const SDNode *Node) const {
  67   if (!Node || !Node->isMachineOpcode()) return nullptr;
  68   return &TII->get(Node->getMachineOpcode());
  69 }
  70
  71 LLVM_DUMP_METHOD void SDep::dump(const TargetRegisterInfo *TRI) const {
  72   switch (getKind()) {
  73   case Data:   dbgs() << "Data"; break;
  74   case Anti:   dbgs() << "Anti"; break;
  75   case Output: dbgs() << "Out "; break;
  76   case Order:  dbgs() << "Ord "; break;
  77   }
  78
  79   switch (getKind()) {
  80   case Data:
  81     dbgs() << " Latency=" << getLatency();
  82     if (TRI && isAssignedRegDep())
  83       dbgs() << " Reg=" << printReg(getReg(), TRI);
  84     break;
  85   case Anti:
  86   case Output:
  87     dbgs() << " Latency=" << getLatency();
  88     break;
  89   case Order:
  90     dbgs() << " Latency=" << getLatency();
  91     switch(Contents.OrdKind) {
  92     case Barrier:      dbgs() << " Barrier"; break;
  93     case MayAliasMem:
  94     case MustAliasMem: dbgs() << " Memory"; break;
  95     case Artificial:   dbgs() << " Artificial"; break;
  96     case Weak:         dbgs() << " Weak"; break;
  97     case Cluster:      dbgs() << " Cluster"; break;
  98     }
  99     break;
 100   }
 101 }
 102
 103 bool SUnit::addPred(const SDep &D, bool Required) {
 104   // If this node already has this dependence, don't add a redundant one.
 105   for (SDep &PredDep : Preds) {
 106     // Zero-latency weak edges may be added purely for heuristic ordering. Don't
 107     // add them if another kind of edge already exists.
 108     if (!Required && PredDep.getSUnit() == D.getSUnit())
 109       return false;
 110     if (PredDep.overlaps(D)) {
 111       // Extend the latency if needed. Equivalent to
 112       // removePred(PredDep) + addPred(D).
 113       if (PredDep.getLatency() < D.getLatency()) {
 114         SUnit *PredSU = PredDep.getSUnit();
 115         // Find the corresponding successor in N.
 116         SDep ForwardD = PredDep;
 117         ForwardD.setSUnit(this);
 118         for (SDep &SuccDep : PredSU->Succs) {
 119           if (SuccDep == ForwardD) {
 120             SuccDep.setLatency(D.getLatency());
 121             break;
 122           }
 123         }
 124         PredDep.setLatency(D.getLatency());
 125       }
 126       return false;
 127     }
 128   }
 129   // Now add a corresponding succ to N.
 130   SDep P = D;
 131   P.setSUnit(this);
 132   SUnit *N = D.getSUnit();
 133   // Update the bookkeeping.
 134   if (D.getKind() == SDep::Data) {
 135     assert(NumPreds < std::numeric_limits<unsigned>::max() &&
 136            "NumPreds will overflow!");
 137     assert(N->NumSuccs < std::numeric_limits<unsigned>::max() &&
 138            "NumSuccs will overflow!");
 139     ++NumPreds;
 140     ++N->NumSuccs;
 141   }
 142   if (!N->isScheduled) {
 143     if (D.isWeak()) {
 144       ++WeakPredsLeft;
 145     }
 146     else {
 147       assert(NumPredsLeft < std::numeric_limits<unsigned>::max() &&
 148              "NumPredsLeft will overflow!");
 149       ++NumPredsLeft;
 150     }
 151   }
 152   if (!isScheduled) {
 153     if (D.isWeak()) {
 154       ++N->WeakSuccsLeft;
 155     }
 156     else {
 157       assert(N->NumSuccsLeft < std::numeric_limits<unsigned>::max() &&
 158              "NumSuccsLeft will overflow!");
 159       ++N->NumSuccsLeft;
 160     }
 161   }
 162   Preds.push_back(D);
 163   N->Succs.push_back(P);
 164   if (P.getLatency() != 0) {
 165     this->setDepthDirty();
 166     N->setHeightDirty();
 167   }
 168   return true;
 169 }
 170
 171 void SUnit::removePred(const SDep &D) {
 172   // Find the matching predecessor.
 173   SmallVectorImpl<SDep>::iterator I = llvm::find(Preds, D);
 174   if (I == Preds.end())
 175     return;
 176   // Find the corresponding successor in N.
 177   SDep P = D;
 178   P.setSUnit(this);
 179   SUnit *N = D.getSUnit();
 180   SmallVectorImpl<SDep>::iterator Succ = llvm::find(N->Succs, P);
 181   assert(Succ != N->Succs.end() && "Mismatching preds / succs lists!");
 182   N->Succs.erase(Succ);
 183   Preds.erase(I);
 184   // Update the bookkeeping.
 185   if (P.getKind() == SDep::Data) {
 186     assert(NumPreds > 0 && "NumPreds will underflow!");
 187     assert(N->NumSuccs > 0 && "NumSuccs will underflow!");
 188     --NumPreds;
 189     --N->NumSuccs;
 190   }
 191   if (!N->isScheduled) {
 192     if (D.isWeak())
 193       --WeakPredsLeft;
 194     else {
 195       assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!");
 196       --NumPredsLeft;
 197     }
 198   }
 199   if (!isScheduled) {
 200     if (D.isWeak())
 201       --N->WeakSuccsLeft;
 202     else {
 203       assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!");
 204       --N->NumSuccsLeft;
 205     }
 206   }
 207   if (P.getLatency() != 0) {
 208     this->setDepthDirty();
 209     N->setHeightDirty();
 210   }
 211 }
 212
 213 void SUnit::setDepthDirty() {
 214   if (!isDepthCurrent) return;
 215   SmallVector<SUnit*, 8> WorkList;
 216   WorkList.push_back(this);
 217   do {
 218     SUnit *SU = WorkList.pop_back_val();
 219     SU->isDepthCurrent = false;
 220     for (SDep &SuccDep : SU->Succs) {
 221       SUnit *SuccSU = SuccDep.getSUnit();
 222       if (SuccSU->isDepthCurrent)
 223         WorkList.push_back(SuccSU);
 224     }
 225   } while (!WorkList.empty());
 226 }
 227
 228 void SUnit::setHeightDirty() {
 229   if (!isHeightCurrent) return;
 230   SmallVector<SUnit*, 8> WorkList;
 231   WorkList.push_back(this);
 232   do {
 233     SUnit *SU = WorkList.pop_back_val();
 234     SU->isHeightCurrent = false;
 235     for (SDep &PredDep : SU->Preds) {
 236       SUnit *PredSU = PredDep.getSUnit();
 237       if (PredSU->isHeightCurrent)
 238         WorkList.push_back(PredSU);
 239     }
 240   } while (!WorkList.empty());
 241 }
 242
 243 void SUnit::setDepthToAtLeast(unsigned NewDepth) {
 244   if (NewDepth <= getDepth())
 245     return;
 246   setDepthDirty();
 247   Depth = NewDepth;
 248   isDepthCurrent = true;
 249 }
 250
 251 void SUnit::setHeightToAtLeast(unsigned NewHeight) {
 252   if (NewHeight <= getHeight())
 253     return;
 254   setHeightDirty();
 255   Height = NewHeight;
 256   isHeightCurrent = true;
 257 }
 258
 259 /// Calculates the maximal path from the node to the exit.
 260 void SUnit::ComputeDepth() {
 261   SmallVector<SUnit*, 8> WorkList;
 262   WorkList.push_back(this);
 263   do {
 264     SUnit *Cur = WorkList.back();
 265
 266     bool Done = true;
 267     unsigned MaxPredDepth = 0;
 268     for (const SDep &PredDep : Cur->Preds) {
 269       SUnit *PredSU = PredDep.getSUnit();
 270       if (PredSU->isDepthCurrent)
 271         MaxPredDepth = std::max(MaxPredDepth,
 272                                 PredSU->Depth + PredDep.getLatency());
 273       else {
 274         Done = false;
 275         WorkList.push_back(PredSU);
 276       }
 277     }
 278
 279     if (Done) {
 280       WorkList.pop_back();
 281       if (MaxPredDepth != Cur->Depth) {
 282         Cur->setDepthDirty();
 283         Cur->Depth = MaxPredDepth;
 284       }
 285       Cur->isDepthCurrent = true;
 286     }
 287   } while (!WorkList.empty());
 288 }
 289
 290 /// Calculates the maximal path from the node to the entry.
 291 void SUnit::ComputeHeight() {
 292   SmallVector<SUnit*, 8> WorkList;
 293   WorkList.push_back(this);
 294   do {
 295     SUnit *Cur = WorkList.back();
 296
 297     bool Done = true;
 298     unsigned MaxSuccHeight = 0;
 299     for (const SDep &SuccDep : Cur->Succs) {
 300       SUnit *SuccSU = SuccDep.getSUnit();
 301       if (SuccSU->isHeightCurrent)
 302         MaxSuccHeight = std::max(MaxSuccHeight,
 303                                  SuccSU->Height + SuccDep.getLatency());
 304       else {
 305         Done = false;
 306         WorkList.push_back(SuccSU);
 307       }
 308     }
 309
 310     if (Done) {
 311       WorkList.pop_back();
 312       if (MaxSuccHeight != Cur->Height) {
 313         Cur->setHeightDirty();
 314         Cur->Height = MaxSuccHeight;
 315       }
 316       Cur->isHeightCurrent = true;
 317     }
 318   } while (!WorkList.empty());
 319 }
 320
 321 void SUnit::biasCriticalPath() {
 322   if (NumPreds < 2)
 323     return;
 324
 325   SUnit::pred_iterator BestI = Preds.begin();
 326   unsigned MaxDepth = BestI->getSUnit()->getDepth();
 327   for (SUnit::pred_iterator I = std::next(BestI), E = Preds.end(); I != E;
 328        ++I) {
 329     if (I->getKind() == SDep::Data && I->getSUnit()->getDepth() > MaxDepth)
 330       BestI = I;
 331   }
 332   if (BestI != Preds.begin())
 333     std::swap(*Preds.begin(), *BestI);
 334 }
 335
 336 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
 337 LLVM_DUMP_METHOD void SUnit::dumpAttributes() const {
 338   dbgs() << "  # preds left       : " << NumPredsLeft << "\n";
 339   dbgs() << "  # succs left       : " << NumSuccsLeft << "\n";
 340   if (WeakPredsLeft)
 341     dbgs() << "  # weak preds left  : " << WeakPredsLeft << "\n";
 342   if (WeakSuccsLeft)
 343     dbgs() << "  # weak succs left  : " << WeakSuccsLeft << "\n";
 344   dbgs() << "  # rdefs left       : " << NumRegDefsLeft << "\n";
 345   dbgs() << "  Latency            : " << Latency << "\n";
 346   dbgs() << "  Depth              : " << getDepth() << "\n";
 347   dbgs() << "  Height             : " << getHeight() << "\n";
 348 }
 349
 350 LLVM_DUMP_METHOD void ScheduleDAG::dumpNodeName(const SUnit &SU) const {
 351   if (&SU == &EntrySU)
 352     dbgs() << "EntrySU";
 353   else if (&SU == &ExitSU)
 354     dbgs() << "ExitSU";
 355   else
 356     dbgs() << "SU(" << SU.NodeNum << ")";
 357 }
 358
 359 LLVM_DUMP_METHOD void ScheduleDAG::dumpNodeAll(const SUnit &SU) const {
 360   dumpNode(SU);
 361   SU.dumpAttributes();
 362   if (SU.Preds.size() > 0) {
 363     dbgs() << "  Predecessors:\n";
 364     for (const SDep &Dep : SU.Preds) {
 365       dbgs() << "    ";
 366       dumpNodeName(*Dep.getSUnit());
 367       dbgs() << ": ";
 368       Dep.dump(TRI);
 369       dbgs() << '\n';
 370     }
 371   }
 372   if (SU.Succs.size() > 0) {
 373     dbgs() << "  Successors:\n";
 374     for (const SDep &Dep : SU.Succs) {
 375       dbgs() << "    ";
 376       dumpNodeName(*Dep.getSUnit());
 377       dbgs() << ": ";
 378       Dep.dump(TRI);
 379       dbgs() << '\n';
 380     }
 381   }
 382 }
 383 #endif
 384
 385 #ifndef NDEBUG
 386 unsigned ScheduleDAG::VerifyScheduledDAG(bool isBottomUp) {
 387   bool AnyNotSched = false;
 388   unsigned DeadNodes = 0;
 389   for (const SUnit &SUnit : SUnits) {
 390     if (!SUnit.isScheduled) {
 391       if (SUnit.NumPreds == 0 && SUnit.NumSuccs == 0) {
 392         ++DeadNodes;
 393         continue;
 394       }
 395       if (!AnyNotSched)
 396         dbgs() << "*** Scheduling failed! ***\n";
 397       dumpNode(SUnit);
 398       dbgs() << "has not been scheduled!\n";
 399       AnyNotSched = true;
 400     }
 401     if (SUnit.isScheduled &&
 402         (isBottomUp ? SUnit.getHeight() : SUnit.getDepth()) >
 403           unsigned(std::numeric_limits<int>::max())) {
 404       if (!AnyNotSched)
 405         dbgs() << "*** Scheduling failed! ***\n";
 406       dumpNode(SUnit);
 407       dbgs() << "has an unexpected "
 408            << (isBottomUp ? "Height" : "Depth") << " value!\n";
 409       AnyNotSched = true;
 410     }
 411     if (isBottomUp) {
 412       if (SUnit.NumSuccsLeft != 0) {
 413         if (!AnyNotSched)
 414           dbgs() << "*** Scheduling failed! ***\n";
 415         dumpNode(SUnit);
 416         dbgs() << "has successors left!\n";
 417         AnyNotSched = true;
 418       }
 419     } else {
 420       if (SUnit.NumPredsLeft != 0) {
 421         if (!AnyNotSched)
 422           dbgs() << "*** Scheduling failed! ***\n";
 423         dumpNode(SUnit);
 424         dbgs() << "has predecessors left!\n";
 425         AnyNotSched = true;
 426       }
 427     }
 428   }
 429   assert(!AnyNotSched);
 430   return SUnits.size() - DeadNodes;
 431 }
 432 #endif
 433
 434 void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() {
 435   // The idea of the algorithm is taken from
 436   // "Online algorithms for managing the topological order of
 437   // a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly
 438   // This is the MNR algorithm, which was first introduced by
 439   // A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in
 440   // "Maintaining a topological order under edge insertions".
 441   //
 442   // Short description of the algorithm:
 443   //
 444   // Topological ordering, ord, of a DAG maps each node to a topological
 445   // index so that for all edges X->Y it is the case that ord(X) < ord(Y).
 446   //
 447   // This means that if there is a path from the node X to the node Z,
 448   // then ord(X) < ord(Z).
 449   //
 450   // This property can be used to check for reachability of nodes:
 451   // if Z is reachable from X, then an insertion of the edge Z->X would
 452   // create a cycle.
 453   //
 454   // The algorithm first computes a topological ordering for the DAG by
 455   // initializing the Index2Node and Node2Index arrays and then tries to keep
 456   // the ordering up-to-date after edge insertions by reordering the DAG.
 457   //
 458   // On insertion of the edge X->Y, the algorithm first marks by calling DFS
 459   // the nodes reachable from Y, and then shifts them using Shift to lie
 460   // immediately after X in Index2Node.
 461   unsigned DAGSize = SUnits.size();
 462   std::vector<SUnit*> WorkList;
 463   WorkList.reserve(DAGSize);
 464
 465   Index2Node.resize(DAGSize);
 466   Node2Index.resize(DAGSize);
 467
 468   // Initialize the data structures.
 469   if (ExitSU)
 470     WorkList.push_back(ExitSU);
 471   for (SUnit &SU : SUnits) {
 472     int NodeNum = SU.NodeNum;
 473     unsigned Degree = SU.Succs.size();
 474     // Temporarily use the Node2Index array as scratch space for degree counts.
 475     Node2Index[NodeNum] = Degree;
 476
 477     // Is it a node without dependencies?
 478     if (Degree == 0) {
 479       assert(SU.Succs.empty() && "SUnit should have no successors");
 480       // Collect leaf nodes.
 481       WorkList.push_back(&SU);
 482     }
 483   }
 484
 485   int Id = DAGSize;
 486   while (!WorkList.empty()) {
 487     SUnit *SU = WorkList.back();
 488     WorkList.pop_back();
 489     if (SU->NodeNum < DAGSize)
 490       Allocate(SU->NodeNum, --Id);
 491     for (const SDep &PredDep : SU->Preds) {
 492       SUnit *SU = PredDep.getSUnit();
 493       if (SU->NodeNum < DAGSize && !--Node2Index[SU->NodeNum])
 494         // If all dependencies of the node are processed already,
 495         // then the node can be computed now.
 496         WorkList.push_back(SU);
 497     }
 498   }
 499
 500   Visited.resize(DAGSize);
 501
 502 #ifndef NDEBUG
 503   // Check correctness of the ordering
 504   for (SUnit &SU : SUnits)  {
 505     for (const SDep &PD : SU.Preds) {
 506       assert(Node2Index[SU.NodeNum] > Node2Index[PD.getSUnit()->NodeNum] &&
 507       "Wrong topological sorting");
 508     }
 509   }
 510 #endif
 511 }
 512
 513 void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) {
 514   int UpperBound, LowerBound;
 515   LowerBound = Node2Index[Y->NodeNum];
 516   UpperBound = Node2Index[X->NodeNum];
 517   bool HasLoop = false;
 518   // Is Ord(X) < Ord(Y) ?
 519   if (LowerBound < UpperBound) {
 520     // Update the topological order.
 521     Visited.reset();
 522     DFS(Y, UpperBound, HasLoop);
 523     assert(!HasLoop && "Inserted edge creates a loop!");
 524     // Recompute topological indexes.
 525     Shift(Visited, LowerBound, UpperBound);
 526   }
 527 }
 528
 529 void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) {
 530   // InitDAGTopologicalSorting();
 531 }
 532
 533 void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound,
 534                                      bool &HasLoop) {
 535   std::vector<const SUnit*> WorkList;
 536   WorkList.reserve(SUnits.size());
 537
 538   WorkList.push_back(SU);
 539   do {
 540     SU = WorkList.back();
 541     WorkList.pop_back();
 542     Visited.set(SU->NodeNum);
 543     for (const SDep &SuccDep
 544          : make_range(SU->Succs.rbegin(), SU->Succs.rend())) {
 545       unsigned s = SuccDep.getSUnit()->NodeNum;
 546       // Edges to non-SUnits are allowed but ignored (e.g. ExitSU).
 547       if (s >= Node2Index.size())
 548         continue;
 549       if (Node2Index[s] == UpperBound) {
 550         HasLoop = true;
 551         return;
 552       }
 553       // Visit successors if not already and in affected region.
 554       if (!Visited.test(s) && Node2Index[s] < UpperBound) {
 555         WorkList.push_back(SuccDep.getSUnit());
 556       }
 557     }
 558   } while (!WorkList.empty());
 559 }
 560
 561 std::vector<int> ScheduleDAGTopologicalSort::GetSubGraph(const SUnit &StartSU,
 562                                                          const SUnit &TargetSU,
 563                                                          bool &Success) {
 564   std::vector<const SUnit*> WorkList;
 565   int LowerBound = Node2Index[StartSU.NodeNum];
 566   int UpperBound = Node2Index[TargetSU.NodeNum];
 567   bool Found = false;
 568   BitVector VisitedBack;
 569   std::vector<int> Nodes;
 570
 571   if (LowerBound > UpperBound) {
 572     Success = false;
 573     return Nodes;
 574   }
 575
 576   WorkList.reserve(SUnits.size());
 577   Visited.reset();
 578
 579   // Starting from StartSU, visit all successors up
 580   // to UpperBound.
 581   WorkList.push_back(&StartSU);
 582   do {
 583     const SUnit *SU = WorkList.back();
 584     WorkList.pop_back();
 585     for (int I = SU->Succs.size()-1; I >= 0; --I) {
 586       const SUnit *Succ = SU->Succs[I].getSUnit();
 587       unsigned s = Succ->NodeNum;
 588       // Edges to non-SUnits are allowed but ignored (e.g. ExitSU).
 589       if (Succ->isBoundaryNode())
 590         continue;
 591       if (Node2Index[s] == UpperBound) {
 592         Found = true;
 593         continue;
 594       }
 595       // Visit successors if not already and in affected region.
 596       if (!Visited.test(s) && Node2Index[s] < UpperBound) {
 597         Visited.set(s);
 598         WorkList.push_back(Succ);
 599       }
 600     }
 601   } while (!WorkList.empty());
 602
 603   if (!Found) {
 604     Success = false;
 605     return Nodes;
 606   }
 607
 608   WorkList.clear();
 609   VisitedBack.resize(SUnits.size());
 610   Found = false;
 611
 612   // Starting from TargetSU, visit all predecessors up
 613   // to LowerBound. SUs that are visited by the two
 614   // passes are added to Nodes.
 615   WorkList.push_back(&TargetSU);
 616   do {
 617     const SUnit *SU = WorkList.back();
 618     WorkList.pop_back();
 619     for (int I = SU->Preds.size()-1; I >= 0; --I) {
 620       const SUnit *Pred = SU->Preds[I].getSUnit();
 621       unsigned s = Pred->NodeNum;
 622       // Edges to non-SUnits are allowed but ignored (e.g. EntrySU).
 623       if (Pred->isBoundaryNode())
 624         continue;
 625       if (Node2Index[s] == LowerBound) {
 626         Found = true;
 627         continue;
 628       }
 629       if (!VisitedBack.test(s) && Visited.test(s)) {
 630         VisitedBack.set(s);
 631         WorkList.push_back(Pred);
 632         Nodes.push_back(s);
 633       }
 634     }
 635   } while (!WorkList.empty());
 636
 637   assert(Found && "Error in SUnit Graph!");
 638   Success = true;
 639   return Nodes;
 640 }
 641
 642 void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound,
 643                                        int UpperBound) {
 644   std::vector<int> L;
 645   int shift = 0;
 646   int i;
 647
 648   for (i = LowerBound; i <= UpperBound; ++i) {
 649     // w is node at topological index i.
 650     int w = Index2Node[i];
 651     if (Visited.test(w)) {
 652       // Unmark.
 653       Visited.reset(w);
 654       L.push_back(w);
 655       shift = shift + 1;
 656     } else {
 657       Allocate(w, i - shift);
 658     }
 659   }
 660
 661   for (unsigned LI : L) {
 662     Allocate(LI, i - shift);
 663     i = i + 1;
 664   }
 665 }
 666
 667 bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *TargetSU, SUnit *SU) {
 668   // Is SU reachable from TargetSU via successor edges?
 669   if (IsReachable(SU, TargetSU))
 670     return true;
 671   for (const SDep &PredDep : TargetSU->Preds)
 672     if (PredDep.isAssignedRegDep() &&
 673         IsReachable(SU, PredDep.getSUnit()))
 674       return true;
 675   return false;
 676 }
 677
 678 bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU,
 679                                              const SUnit *TargetSU) {
 680   // If insertion of the edge SU->TargetSU would create a cycle
 681   // then there is a path from TargetSU to SU.
 682   int UpperBound, LowerBound;
 683   LowerBound = Node2Index[TargetSU->NodeNum];
 684   UpperBound = Node2Index[SU->NodeNum];
 685   bool HasLoop = false;
 686   // Is Ord(TargetSU) < Ord(SU) ?
 687   if (LowerBound < UpperBound) {
 688     Visited.reset();
 689     // There may be a path from TargetSU to SU. Check for it.
 690     DFS(TargetSU, UpperBound, HasLoop);
 691   }
 692   return HasLoop;
 693 }
 694
 695 void ScheduleDAGTopologicalSort::Allocate(int n, int index) {
 696   Node2Index[n] = index;
 697   Index2Node[index] = n;
 698 }
 699
 700 ScheduleDAGTopologicalSort::
 701 ScheduleDAGTopologicalSort(std::vector<SUnit> &sunits, SUnit *exitsu)
 702   : SUnits(sunits), ExitSU(exitsu) {}
 703
 704 ScheduleHazardRecognizer::~ScheduleHazardRecognizer() = default;