ScheduleDAG.cpp revision e07f4c0f48e17594fd39b1151fc18c164aa5b4d1
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/Target/TargetMachine.h"
18#include "llvm/Target/TargetInstrInfo.h"
19#include "llvm/Target/TargetRegisterInfo.h"
20#include "llvm/Support/Debug.h"
21#include <climits>
22using namespace llvm;
23
24ScheduleDAG::ScheduleDAG(SelectionDAG *dag, MachineBasicBlock *bb,
25                         const TargetMachine &tm)
26  : DAG(dag), BB(bb), TM(tm), MRI(BB->getParent()->getRegInfo()) {
27  TII = TM.getInstrInfo();
28  MF  = BB->getParent();
29  TRI = TM.getRegisterInfo();
30  TLI = TM.getTargetLowering();
31  ConstPool = MF->getConstantPool();
32}
33
34ScheduleDAG::~ScheduleDAG() {}
35
36/// CalculateDepths - compute depths using algorithms for the longest
37/// paths in the DAG
38void ScheduleDAG::CalculateDepths() {
39  unsigned DAGSize = SUnits.size();
40  std::vector<SUnit*> WorkList;
41  WorkList.reserve(DAGSize);
42
43  // Initialize the data structures
44  for (unsigned i = 0, e = DAGSize; i != e; ++i) {
45    SUnit *SU = &SUnits[i];
46    unsigned Degree = SU->Preds.size();
47    // Temporarily use the Depth field as scratch space for the degree count.
48    SU->Depth = Degree;
49
50    // Is it a node without dependencies?
51    if (Degree == 0) {
52      assert(SU->Preds.empty() && "SUnit should have no predecessors");
53      // Collect leaf nodes
54      WorkList.push_back(SU);
55    }
56  }
57
58  // Process nodes in the topological order
59  while (!WorkList.empty()) {
60    SUnit *SU = WorkList.back();
61    WorkList.pop_back();
62    unsigned SUDepth = 0;
63
64    // Use dynamic programming:
65    // When current node is being processed, all of its dependencies
66    // are already processed.
67    // So, just iterate over all predecessors and take the longest path
68    for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
69         I != E; ++I) {
70      unsigned PredDepth = I->Dep->Depth;
71      if (PredDepth+1 > SUDepth) {
72        SUDepth = PredDepth + 1;
73      }
74    }
75
76    SU->Depth = SUDepth;
77
78    // Update degrees of all nodes depending on current SUnit
79    for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
80         I != E; ++I) {
81      SUnit *SU = I->Dep;
82      if (!--SU->Depth)
83        // If all dependencies of the node are processed already,
84        // then the longest path for the node can be computed now
85        WorkList.push_back(SU);
86    }
87  }
88}
89
90/// CalculateHeights - compute heights using algorithms for the longest
91/// paths in the DAG
92void ScheduleDAG::CalculateHeights() {
93  unsigned DAGSize = SUnits.size();
94  std::vector<SUnit*> WorkList;
95  WorkList.reserve(DAGSize);
96
97  // Initialize the data structures
98  for (unsigned i = 0, e = DAGSize; i != e; ++i) {
99    SUnit *SU = &SUnits[i];
100    unsigned Degree = SU->Succs.size();
101    // Temporarily use the Height field as scratch space for the degree count.
102    SU->Height = Degree;
103
104    // Is it a node without dependencies?
105    if (Degree == 0) {
106      assert(SU->Succs.empty() && "Something wrong");
107      assert(WorkList.empty() && "Should be empty");
108      // Collect leaf nodes
109      WorkList.push_back(SU);
110    }
111  }
112
113  // Process nodes in the topological order
114  while (!WorkList.empty()) {
115    SUnit *SU = WorkList.back();
116    WorkList.pop_back();
117    unsigned SUHeight = 0;
118
119    // Use dynamic programming:
120    // When current node is being processed, all of its dependencies
121    // are already processed.
122    // So, just iterate over all successors and take the longest path
123    for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
124         I != E; ++I) {
125      unsigned SuccHeight = I->Dep->Height;
126      if (SuccHeight+1 > SUHeight) {
127        SUHeight = SuccHeight + 1;
128      }
129    }
130
131    SU->Height = SUHeight;
132
133    // Update degrees of all nodes depending on current SUnit
134    for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
135         I != E; ++I) {
136      SUnit *SU = I->Dep;
137      if (!--SU->Height)
138        // If all dependencies of the node are processed already,
139        // then the longest path for the node can be computed now
140        WorkList.push_back(SU);
141    }
142  }
143}
144
145/// dump - dump the schedule.
146void ScheduleDAG::dumpSchedule() const {
147  for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
148    if (SUnit *SU = Sequence[i])
149      SU->dump(this);
150    else
151      cerr << "**** NOOP ****\n";
152  }
153}
154
155
156/// Run - perform scheduling.
157///
158void ScheduleDAG::Run() {
159  Schedule();
160
161  DOUT << "*** Final schedule ***\n";
162  DEBUG(dumpSchedule());
163  DOUT << "\n";
164}
165
166/// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
167/// a group of nodes flagged together.
168void SUnit::dump(const ScheduleDAG *G) const {
169  cerr << "SU(" << NodeNum << "): ";
170  G->dumpNode(this);
171}
172
173void SUnit::dumpAll(const ScheduleDAG *G) const {
174  dump(G);
175
176  cerr << "  # preds left       : " << NumPredsLeft << "\n";
177  cerr << "  # succs left       : " << NumSuccsLeft << "\n";
178  cerr << "  Latency            : " << Latency << "\n";
179  cerr << "  Depth              : " << Depth << "\n";
180  cerr << "  Height             : " << Height << "\n";
181
182  if (Preds.size() != 0) {
183    cerr << "  Predecessors:\n";
184    for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
185         I != E; ++I) {
186      if (I->isCtrl)
187        cerr << "   ch  #";
188      else
189        cerr << "   val #";
190      cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")";
191      if (I->isArtificial)
192        cerr << " *";
193      cerr << "\n";
194    }
195  }
196  if (Succs.size() != 0) {
197    cerr << "  Successors:\n";
198    for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
199         I != E; ++I) {
200      if (I->isCtrl)
201        cerr << "   ch  #";
202      else
203        cerr << "   val #";
204      cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")";
205      if (I->isArtificial)
206        cerr << " *";
207      cerr << "\n";
208    }
209  }
210  cerr << "\n";
211}
212
213#ifndef NDEBUG
214/// VerifySchedule - Verify that all SUnits were scheduled and that
215/// their state is consistent.
216///
217void ScheduleDAG::VerifySchedule(bool isBottomUp) {
218  bool AnyNotSched = false;
219  unsigned DeadNodes = 0;
220  unsigned Noops = 0;
221  for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
222    if (!SUnits[i].isScheduled) {
223      if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) {
224        ++DeadNodes;
225        continue;
226      }
227      if (!AnyNotSched)
228        cerr << "*** Scheduling failed! ***\n";
229      SUnits[i].dump(this);
230      cerr << "has not been scheduled!\n";
231      AnyNotSched = true;
232    }
233    if (SUnits[i].isScheduled && SUnits[i].Cycle > (unsigned)INT_MAX) {
234      if (!AnyNotSched)
235        cerr << "*** Scheduling failed! ***\n";
236      SUnits[i].dump(this);
237      cerr << "has an unexpected Cycle value!\n";
238      AnyNotSched = true;
239    }
240    if (isBottomUp) {
241      if (SUnits[i].NumSuccsLeft != 0) {
242        if (!AnyNotSched)
243          cerr << "*** Scheduling failed! ***\n";
244        SUnits[i].dump(this);
245        cerr << "has successors left!\n";
246        AnyNotSched = true;
247      }
248    } else {
249      if (SUnits[i].NumPredsLeft != 0) {
250        if (!AnyNotSched)
251          cerr << "*** Scheduling failed! ***\n";
252        SUnits[i].dump(this);
253        cerr << "has predecessors left!\n";
254        AnyNotSched = true;
255      }
256    }
257  }
258  for (unsigned i = 0, e = Sequence.size(); i != e; ++i)
259    if (!Sequence[i])
260      ++Noops;
261  assert(!AnyNotSched);
262  assert(Sequence.size() + DeadNodes - Noops == SUnits.size() &&
263         "The number of nodes scheduled doesn't match the expected number!");
264}
265#endif
266
267/// InitDAGTopologicalSorting - create the initial topological
268/// ordering from the DAG to be scheduled.
269///
270/// The idea of the algorithm is taken from
271/// "Online algorithms for managing the topological order of
272/// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly
273/// This is the MNR algorithm, which was first introduced by
274/// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in
275/// "Maintaining a topological order under edge insertions".
276///
277/// Short description of the algorithm:
278///
279/// Topological ordering, ord, of a DAG maps each node to a topological
280/// index so that for all edges X->Y it is the case that ord(X) < ord(Y).
281///
282/// This means that if there is a path from the node X to the node Z,
283/// then ord(X) < ord(Z).
284///
285/// This property can be used to check for reachability of nodes:
286/// if Z is reachable from X, then an insertion of the edge Z->X would
287/// create a cycle.
288///
289/// The algorithm first computes a topological ordering for the DAG by
290/// initializing the Index2Node and Node2Index arrays and then tries to keep
291/// the ordering up-to-date after edge insertions by reordering the DAG.
292///
293/// On insertion of the edge X->Y, the algorithm first marks by calling DFS
294/// the nodes reachable from Y, and then shifts them using Shift to lie
295/// immediately after X in Index2Node.
296void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() {
297  unsigned DAGSize = SUnits.size();
298  std::vector<SUnit*> WorkList;
299  WorkList.reserve(DAGSize);
300
301  Index2Node.resize(DAGSize);
302  Node2Index.resize(DAGSize);
303
304  // Initialize the data structures.
305  for (unsigned i = 0, e = DAGSize; i != e; ++i) {
306    SUnit *SU = &SUnits[i];
307    int NodeNum = SU->NodeNum;
308    unsigned Degree = SU->Succs.size();
309    // Temporarily use the Node2Index array as scratch space for degree counts.
310    Node2Index[NodeNum] = Degree;
311
312    // Is it a node without dependencies?
313    if (Degree == 0) {
314      assert(SU->Succs.empty() && "SUnit should have no successors");
315      // Collect leaf nodes.
316      WorkList.push_back(SU);
317    }
318  }
319
320  int Id = DAGSize;
321  while (!WorkList.empty()) {
322    SUnit *SU = WorkList.back();
323    WorkList.pop_back();
324    Allocate(SU->NodeNum, --Id);
325    for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
326         I != E; ++I) {
327      SUnit *SU = I->Dep;
328      if (!--Node2Index[SU->NodeNum])
329        // If all dependencies of the node are processed already,
330        // then the node can be computed now.
331        WorkList.push_back(SU);
332    }
333  }
334
335  Visited.resize(DAGSize);
336
337#ifndef NDEBUG
338  // Check correctness of the ordering
339  for (unsigned i = 0, e = DAGSize; i != e; ++i) {
340    SUnit *SU = &SUnits[i];
341    for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
342         I != E; ++I) {
343      assert(Node2Index[SU->NodeNum] > Node2Index[I->Dep->NodeNum] &&
344      "Wrong topological sorting");
345    }
346  }
347#endif
348}
349
350/// AddPred - Updates the topological ordering to accomodate an edge
351/// to be added from SUnit X to SUnit Y.
352void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) {
353  int UpperBound, LowerBound;
354  LowerBound = Node2Index[Y->NodeNum];
355  UpperBound = Node2Index[X->NodeNum];
356  bool HasLoop = false;
357  // Is Ord(X) < Ord(Y) ?
358  if (LowerBound < UpperBound) {
359    // Update the topological order.
360    Visited.reset();
361    DFS(Y, UpperBound, HasLoop);
362    assert(!HasLoop && "Inserted edge creates a loop!");
363    // Recompute topological indexes.
364    Shift(Visited, LowerBound, UpperBound);
365  }
366}
367
368/// RemovePred - Updates the topological ordering to accomodate an
369/// an edge to be removed from the specified node N from the predecessors
370/// of the current node M.
371void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) {
372  // InitDAGTopologicalSorting();
373}
374
375/// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark
376/// all nodes affected by the edge insertion. These nodes will later get new
377/// topological indexes by means of the Shift method.
378void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound,
379                                     bool& HasLoop) {
380  std::vector<const SUnit*> WorkList;
381  WorkList.reserve(SUnits.size());
382
383  WorkList.push_back(SU);
384  while (!WorkList.empty()) {
385    SU = WorkList.back();
386    WorkList.pop_back();
387    Visited.set(SU->NodeNum);
388    for (int I = SU->Succs.size()-1; I >= 0; --I) {
389      int s = SU->Succs[I].Dep->NodeNum;
390      if (Node2Index[s] == UpperBound) {
391        HasLoop = true;
392        return;
393      }
394      // Visit successors if not already and in affected region.
395      if (!Visited.test(s) && Node2Index[s] < UpperBound) {
396        WorkList.push_back(SU->Succs[I].Dep);
397      }
398    }
399  }
400}
401
402/// Shift - Renumber the nodes so that the topological ordering is
403/// preserved.
404void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound,
405                                       int UpperBound) {
406  std::vector<int> L;
407  int shift = 0;
408  int i;
409
410  for (i = LowerBound; i <= UpperBound; ++i) {
411    // w is node at topological index i.
412    int w = Index2Node[i];
413    if (Visited.test(w)) {
414      // Unmark.
415      Visited.reset(w);
416      L.push_back(w);
417      shift = shift + 1;
418    } else {
419      Allocate(w, i - shift);
420    }
421  }
422
423  for (unsigned j = 0; j < L.size(); ++j) {
424    Allocate(L[j], i - shift);
425    i = i + 1;
426  }
427}
428
429
430/// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
431/// create a cycle.
432bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
433  if (IsReachable(TargetSU, SU))
434    return true;
435  for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
436       I != E; ++I)
437    if (I->Cost < 0 && IsReachable(TargetSU, I->Dep))
438      return true;
439  return false;
440}
441
442/// IsReachable - Checks if SU is reachable from TargetSU.
443bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU,
444                                             const SUnit *TargetSU) {
445  // If insertion of the edge SU->TargetSU would create a cycle
446  // then there is a path from TargetSU to SU.
447  int UpperBound, LowerBound;
448  LowerBound = Node2Index[TargetSU->NodeNum];
449  UpperBound = Node2Index[SU->NodeNum];
450  bool HasLoop = false;
451  // Is Ord(TargetSU) < Ord(SU) ?
452  if (LowerBound < UpperBound) {
453    Visited.reset();
454    // There may be a path from TargetSU to SU. Check for it.
455    DFS(TargetSU, UpperBound, HasLoop);
456  }
457  return HasLoop;
458}
459
460/// Allocate - assign the topological index to the node n.
461void ScheduleDAGTopologicalSort::Allocate(int n, int index) {
462  Node2Index[n] = index;
463  Index2Node[index] = n;
464}
465
466ScheduleDAGTopologicalSort::ScheduleDAGTopologicalSort(
467                                                     std::vector<SUnit> &sunits)
468 : SUnits(sunits) {}
469