ScheduleDAG.cpp revision 5c982949bf87ffda129fb7fe372e414e255f4630 
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    TLI(TM.getTargetLowering()),
31    MF(mf), MRI(mf.getRegInfo()),
32    EntrySU(), ExitSU() {
33}
34
35ScheduleDAG::~ScheduleDAG() {}
36
37/// dump - dump the schedule.
38void ScheduleDAG::dumpSchedule() const {
39  for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
40    if (SUnit *SU = Sequence[i])
41      SU->dump(this);
42    else
43      dbgs() << "**** NOOP ****\n";
44  }
45}
46
47
48/// Run - perform scheduling.
49///
50void ScheduleDAG::Run(MachineBasicBlock *bb,
51                      MachineBasicBlock::iterator insertPos) {
52  BB = bb;
53  InsertPos = insertPos;
54
55  SUnits.clear();
56  Sequence.clear();
57  EntrySU = SUnit();
58  ExitSU = SUnit();
59
60  Schedule();
61
62  DEBUG({
63      dbgs() << "*** Final schedule ***\n";
64      dumpSchedule();
65      dbgs() << '\n';
66    });
67}
68
69/// addPred - This adds the specified edge as a pred of the current node if
70/// not already.  It also adds the current node as a successor of the
71/// specified node.
72void SUnit::addPred(const SDep &D) {
73  // If this node already has this depenence, don't add a redundant one.
74  for (SmallVector<SDep, 4>::const_iterator I = Preds.begin(), E = Preds.end();
75       I != E; ++I)
76    if (*I == D)
77      return;
78  // Now add a corresponding succ to N.
79  SDep P = D;
80  P.setSUnit(this);
81  SUnit *N = D.getSUnit();
82  // Update the bookkeeping.
83  if (D.getKind() == SDep::Data) {
84    assert(NumPreds < UINT_MAX && "NumPreds will overflow!");
85    assert(N->NumSuccs < UINT_MAX && "NumSuccs will overflow!");
86    ++NumPreds;
87    ++N->NumSuccs;
88  }
89  if (!N->isScheduled) {
90    assert(NumPredsLeft < UINT_MAX && "NumPredsLeft will overflow!");
91    ++NumPredsLeft;
92  }
93  if (!isScheduled) {
94    assert(N->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!");
95    ++N->NumSuccsLeft;
96  }
97  Preds.push_back(D);
98  N->Succs.push_back(P);
99  if (P.getLatency() != 0) {
100    this->setDepthDirty();
101    N->setHeightDirty();
102  }
103}
104
105/// removePred - This removes the specified edge as a pred of the current
106/// node if it exists.  It also removes the current node as a successor of
107/// the specified node.
108void SUnit::removePred(const SDep &D) {
109  // Find the matching predecessor.
110  for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end();
111       I != E; ++I)
112    if (*I == D) {
113      bool FoundSucc = false;
114      // Find the corresponding successor in N.
115      SDep P = D;
116      P.setSUnit(this);
117      SUnit *N = D.getSUnit();
118      for (SmallVector<SDep, 4>::iterator II = N->Succs.begin(),
119             EE = N->Succs.end(); II != EE; ++II)
120        if (*II == P) {
121          FoundSucc = true;
122          N->Succs.erase(II);
123          break;
124        }
125      assert(FoundSucc && "Mismatching preds / succs lists!");
126      Preds.erase(I);
127      // Update the bookkeeping.
128      if (P.getKind() == SDep::Data) {
129        assert(NumPreds > 0 && "NumPreds will underflow!");
130        assert(N->NumSuccs > 0 && "NumSuccs will underflow!");
131        --NumPreds;
132        --N->NumSuccs;
133      }
134      if (!N->isScheduled) {
135        assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!");
136        --NumPredsLeft;
137      }
138      if (!isScheduled) {
139        assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!");
140        --N->NumSuccsLeft;
141      }
142      if (P.getLatency() != 0) {
143        this->setDepthDirty();
144        N->setHeightDirty();
145      }
146      return;
147    }
148}
149
150void SUnit::setDepthDirty() {
151  if (!isDepthCurrent) return;
152  SmallVector<SUnit*, 8> WorkList;
153  WorkList.push_back(this);
154  do {
155    SUnit *SU = WorkList.pop_back_val();
156    SU->isDepthCurrent = false;
157    for (SUnit::const_succ_iterator I = SU->Succs.begin(),
158         E = SU->Succs.end(); I != E; ++I) {
159      SUnit *SuccSU = I->getSUnit();
160      if (SuccSU->isDepthCurrent)
161        WorkList.push_back(SuccSU);
162    }
163  } while (!WorkList.empty());
164}
165
166void SUnit::setHeightDirty() {
167  if (!isHeightCurrent) return;
168  SmallVector<SUnit*, 8> WorkList;
169  WorkList.push_back(this);
170  do {
171    SUnit *SU = WorkList.pop_back_val();
172    SU->isHeightCurrent = false;
173    for (SUnit::const_pred_iterator I = SU->Preds.begin(),
174         E = SU->Preds.end(); I != E; ++I) {
175      SUnit *PredSU = I->getSUnit();
176      if (PredSU->isHeightCurrent)
177        WorkList.push_back(PredSU);
178    }
179  } while (!WorkList.empty());
180}
181
182/// setDepthToAtLeast - Update this node's successors to reflect the
183/// fact that this node's depth just increased.
184///
185void SUnit::setDepthToAtLeast(unsigned NewDepth) {
186  if (NewDepth <= getDepth())
187    return;
188  setDepthDirty();
189  Depth = NewDepth;
190  isDepthCurrent = true;
191}
192
193/// setHeightToAtLeast - Update this node's predecessors to reflect the
194/// fact that this node's height just increased.
195///
196void SUnit::setHeightToAtLeast(unsigned NewHeight) {
197  if (NewHeight <= getHeight())
198    return;
199  setHeightDirty();
200  Height = NewHeight;
201  isHeightCurrent = true;
202}
203
204/// ComputeDepth - Calculate the maximal path from the node to the exit.
205///
206void SUnit::ComputeDepth() {
207  SmallVector<SUnit*, 8> WorkList;
208  WorkList.push_back(this);
209  do {
210    SUnit *Cur = WorkList.back();
211
212    bool Done = true;
213    unsigned MaxPredDepth = 0;
214    for (SUnit::const_pred_iterator I = Cur->Preds.begin(),
215         E = Cur->Preds.end(); I != E; ++I) {
216      SUnit *PredSU = I->getSUnit();
217      if (PredSU->isDepthCurrent)
218        MaxPredDepth = std::max(MaxPredDepth,
219                                PredSU->Depth + I->getLatency());
220      else {
221        Done = false;
222        WorkList.push_back(PredSU);
223      }
224    }
225
226    if (Done) {
227      WorkList.pop_back();
228      if (MaxPredDepth != Cur->Depth) {
229        Cur->setDepthDirty();
230        Cur->Depth = MaxPredDepth;
231      }
232      Cur->isDepthCurrent = true;
233    }
234  } while (!WorkList.empty());
235}
236
237/// ComputeHeight - Calculate the maximal path from the node to the entry.
238///
239void SUnit::ComputeHeight() {
240  SmallVector<SUnit*, 8> WorkList;
241  WorkList.push_back(this);
242  do {
243    SUnit *Cur = WorkList.back();
244
245    bool Done = true;
246    unsigned MaxSuccHeight = 0;
247    for (SUnit::const_succ_iterator I = Cur->Succs.begin(),
248         E = Cur->Succs.end(); I != E; ++I) {
249      SUnit *SuccSU = I->getSUnit();
250      if (SuccSU->isHeightCurrent)
251        MaxSuccHeight = std::max(MaxSuccHeight,
252                                 SuccSU->Height + I->getLatency());
253      else {
254        Done = false;
255        WorkList.push_back(SuccSU);
256      }
257    }
258
259    if (Done) {
260      WorkList.pop_back();
261      if (MaxSuccHeight != Cur->Height) {
262        Cur->setHeightDirty();
263        Cur->Height = MaxSuccHeight;
264      }
265      Cur->isHeightCurrent = true;
266    }
267  } while (!WorkList.empty());
268}
269
270/// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
271/// a group of nodes flagged together.
272void SUnit::dump(const ScheduleDAG *G) const {
273  dbgs() << "SU(" << NodeNum << "): ";
274  G->dumpNode(this);
275}
276
277void SUnit::dumpAll(const ScheduleDAG *G) const {
278  dump(G);
279
280  dbgs() << "  # preds left       : " << NumPredsLeft << "\n";
281  dbgs() << "  # succs left       : " << NumSuccsLeft << "\n";
282  dbgs() << "  Latency            : " << Latency << "\n";
283  dbgs() << "  Depth              : " << Depth << "\n";
284  dbgs() << "  Height             : " << Height << "\n";
285
286  if (Preds.size() != 0) {
287    dbgs() << "  Predecessors:\n";
288    for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
289         I != E; ++I) {
290      dbgs() << "   ";
291      switch (I->getKind()) {
292      case SDep::Data:        dbgs() << "val "; break;
293      case SDep::Anti:        dbgs() << "anti"; break;
294      case SDep::Output:      dbgs() << "out "; break;
295      case SDep::Order:       dbgs() << "ch  "; break;
296      }
297      dbgs() << "#";
298      dbgs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
299      if (I->isArtificial())
300        dbgs() << " *";
301      dbgs() << ": Latency=" << I->getLatency();
302      dbgs() << "\n";
303    }
304  }
305  if (Succs.size() != 0) {
306    dbgs() << "  Successors:\n";
307    for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
308         I != E; ++I) {
309      dbgs() << "   ";
310      switch (I->getKind()) {
311      case SDep::Data:        dbgs() << "val "; break;
312      case SDep::Anti:        dbgs() << "anti"; break;
313      case SDep::Output:      dbgs() << "out "; break;
314      case SDep::Order:       dbgs() << "ch  "; break;
315      }
316      dbgs() << "#";
317      dbgs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
318      if (I->isArtificial())
319        dbgs() << " *";
320      dbgs() << ": Latency=" << I->getLatency();
321      dbgs() << "\n";
322    }
323  }
324  dbgs() << "\n";
325}
326
327#ifndef NDEBUG
328/// VerifySchedule - Verify that all SUnits were scheduled and that
329/// their state is consistent.
330///
331void ScheduleDAG::VerifySchedule(bool isBottomUp) {
332  bool AnyNotSched = false;
333  unsigned DeadNodes = 0;
334  unsigned Noops = 0;
335  for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
336    if (!SUnits[i].isScheduled) {
337      if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) {
338        ++DeadNodes;
339        continue;
340      }
341      if (!AnyNotSched)
342        dbgs() << "*** Scheduling failed! ***\n";
343      SUnits[i].dump(this);
344      dbgs() << "has not been scheduled!\n";
345      AnyNotSched = true;
346    }
347    if (SUnits[i].isScheduled &&
348        (isBottomUp ? SUnits[i].getHeight() : SUnits[i].getDepth()) >
349          unsigned(INT_MAX)) {
350      if (!AnyNotSched)
351        dbgs() << "*** Scheduling failed! ***\n";
352      SUnits[i].dump(this);
353      dbgs() << "has an unexpected "
354           << (isBottomUp ? "Height" : "Depth") << " value!\n";
355      AnyNotSched = true;
356    }
357    if (isBottomUp) {
358      if (SUnits[i].NumSuccsLeft != 0) {
359        if (!AnyNotSched)
360          dbgs() << "*** Scheduling failed! ***\n";
361        SUnits[i].dump(this);
362        dbgs() << "has successors left!\n";
363        AnyNotSched = true;
364      }
365    } else {
366      if (SUnits[i].NumPredsLeft != 0) {
367        if (!AnyNotSched)
368          dbgs() << "*** Scheduling failed! ***\n";
369        SUnits[i].dump(this);
370        dbgs() << "has predecessors left!\n";
371        AnyNotSched = true;
372      }
373    }
374  }
375  for (unsigned i = 0, e = Sequence.size(); i != e; ++i)
376    if (!Sequence[i])
377      ++Noops;
378  assert(!AnyNotSched);
379  assert(Sequence.size() + DeadNodes - Noops == SUnits.size() &&
380         "The number of nodes scheduled doesn't match the expected number!");
381}
382#endif
383
384/// InitDAGTopologicalSorting - create the initial topological
385/// ordering from the DAG to be scheduled.
386///
387/// The idea of the algorithm is taken from
388/// "Online algorithms for managing the topological order of
389/// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly
390/// This is the MNR algorithm, which was first introduced by
391/// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in
392/// "Maintaining a topological order under edge insertions".
393///
394/// Short description of the algorithm:
395///
396/// Topological ordering, ord, of a DAG maps each node to a topological
397/// index so that for all edges X->Y it is the case that ord(X) < ord(Y).
398///
399/// This means that if there is a path from the node X to the node Z,
400/// then ord(X) < ord(Z).
401///
402/// This property can be used to check for reachability of nodes:
403/// if Z is reachable from X, then an insertion of the edge Z->X would
404/// create a cycle.
405///
406/// The algorithm first computes a topological ordering for the DAG by
407/// initializing the Index2Node and Node2Index arrays and then tries to keep
408/// the ordering up-to-date after edge insertions by reordering the DAG.
409///
410/// On insertion of the edge X->Y, the algorithm first marks by calling DFS
411/// the nodes reachable from Y, and then shifts them using Shift to lie
412/// immediately after X in Index2Node.
413void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() {
414  unsigned DAGSize = SUnits.size();
415  std::vector<SUnit*> WorkList;
416  WorkList.reserve(DAGSize);
417
418  Index2Node.resize(DAGSize);
419  Node2Index.resize(DAGSize);
420
421  // Initialize the data structures.
422  for (unsigned i = 0, e = DAGSize; i != e; ++i) {
423    SUnit *SU = &SUnits[i];
424    int NodeNum = SU->NodeNum;
425    unsigned Degree = SU->Succs.size();
426    // Temporarily use the Node2Index array as scratch space for degree counts.
427    Node2Index[NodeNum] = Degree;
428
429    // Is it a node without dependencies?
430    if (Degree == 0) {
431      assert(SU->Succs.empty() && "SUnit should have no successors");
432      // Collect leaf nodes.
433      WorkList.push_back(SU);
434    }
435  }
436
437  int Id = DAGSize;
438  while (!WorkList.empty()) {
439    SUnit *SU = WorkList.back();
440    WorkList.pop_back();
441    Allocate(SU->NodeNum, --Id);
442    for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
443         I != E; ++I) {
444      SUnit *SU = I->getSUnit();
445      if (!--Node2Index[SU->NodeNum])
446        // If all dependencies of the node are processed already,
447        // then the node can be computed now.
448        WorkList.push_back(SU);
449    }
450  }
451
452  Visited.resize(DAGSize);
453
454#ifndef NDEBUG
455  // Check correctness of the ordering
456  for (unsigned i = 0, e = DAGSize; i != e; ++i) {
457    SUnit *SU = &SUnits[i];
458    for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
459         I != E; ++I) {
460      assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] &&
461      "Wrong topological sorting");
462    }
463  }
464#endif
465}
466
467/// AddPred - Updates the topological ordering to accomodate an edge
468/// to be added from SUnit X to SUnit Y.
469void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) {
470  int UpperBound, LowerBound;
471  LowerBound = Node2Index[Y->NodeNum];
472  UpperBound = Node2Index[X->NodeNum];
473  bool HasLoop = false;
474  // Is Ord(X) < Ord(Y) ?
475  if (LowerBound < UpperBound) {
476    // Update the topological order.
477    Visited.reset();
478    DFS(Y, UpperBound, HasLoop);
479    assert(!HasLoop && "Inserted edge creates a loop!");
480    // Recompute topological indexes.
481    Shift(Visited, LowerBound, UpperBound);
482  }
483}
484
485/// RemovePred - Updates the topological ordering to accomodate an
486/// an edge to be removed from the specified node N from the predecessors
487/// of the current node M.
488void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) {
489  // InitDAGTopologicalSorting();
490}
491
492/// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark
493/// all nodes affected by the edge insertion. These nodes will later get new
494/// topological indexes by means of the Shift method.
495void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound,
496                                     bool& HasLoop) {
497  std::vector<const SUnit*> WorkList;
498  WorkList.reserve(SUnits.size());
499
500  WorkList.push_back(SU);
501  do {
502    SU = WorkList.back();
503    WorkList.pop_back();
504    Visited.set(SU->NodeNum);
505    for (int I = SU->Succs.size()-1; I >= 0; --I) {
506      int s = SU->Succs[I].getSUnit()->NodeNum;
507      if (Node2Index[s] == UpperBound) {
508        HasLoop = true;
509        return;
510      }
511      // Visit successors if not already and in affected region.
512      if (!Visited.test(s) && Node2Index[s] < UpperBound) {
513        WorkList.push_back(SU->Succs[I].getSUnit());
514      }
515    }
516  } while (!WorkList.empty());
517}
518
519/// Shift - Renumber the nodes so that the topological ordering is
520/// preserved.
521void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound,
522                                       int UpperBound) {
523  std::vector<int> L;
524  int shift = 0;
525  int i;
526
527  for (i = LowerBound; i <= UpperBound; ++i) {
528    // w is node at topological index i.
529    int w = Index2Node[i];
530    if (Visited.test(w)) {
531      // Unmark.
532      Visited.reset(w);
533      L.push_back(w);
534      shift = shift + 1;
535    } else {
536      Allocate(w, i - shift);
537    }
538  }
539
540  for (unsigned j = 0; j < L.size(); ++j) {
541    Allocate(L[j], i - shift);
542    i = i + 1;
543  }
544}
545
546
547/// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
548/// create a cycle.
549bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
550  if (IsReachable(TargetSU, SU))
551    return true;
552  for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
553       I != E; ++I)
554    if (I->isAssignedRegDep() &&
555        IsReachable(TargetSU, I->getSUnit()))
556      return true;
557  return false;
558}
559
560/// IsReachable - Checks if SU is reachable from TargetSU.
561bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU,
562                                             const SUnit *TargetSU) {
563  // If insertion of the edge SU->TargetSU would create a cycle
564  // then there is a path from TargetSU to SU.
565  int UpperBound, LowerBound;
566  LowerBound = Node2Index[TargetSU->NodeNum];
567  UpperBound = Node2Index[SU->NodeNum];
568  bool HasLoop = false;
569  // Is Ord(TargetSU) < Ord(SU) ?
570  if (LowerBound < UpperBound) {
571    Visited.reset();
572    // There may be a path from TargetSU to SU. Check for it.
573    DFS(TargetSU, UpperBound, HasLoop);
574  }
575  return HasLoop;
576}
577
578/// Allocate - assign the topological index to the node n.
579void ScheduleDAGTopologicalSort::Allocate(int n, int index) {
580  Node2Index[n] = index;
581  Index2Node[index] = n;
582}
583
584ScheduleDAGTopologicalSort::ScheduleDAGTopologicalSort(
585                                                     std::vector<SUnit> &sunits)
586 : SUnits(sunits) {}
587
588ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {}
589