Fix comment for consistency sake.
[llvm/avr.git] / lib / CodeGen / ScheduleDAG.cpp
blobff5c236e3797cecbc04e180efc8bccb64f9fd610
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.
13 //===----------------------------------------------------------------------===//
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>
24 using namespace llvm;
26 ScheduleDAG::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 ConstPool(MF.getConstantPool()),
33 EntrySU(), ExitSU() {
36 ScheduleDAG::~ScheduleDAG() {}
38 /// dump - dump the schedule.
39 void ScheduleDAG::dumpSchedule() const {
40 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
41 if (SUnit *SU = Sequence[i])
42 SU->dump(this);
43 else
44 errs() << "**** NOOP ****\n";
49 /// Run - perform scheduling.
50 ///
51 void ScheduleDAG::Run(MachineBasicBlock *bb,
52 MachineBasicBlock::iterator insertPos) {
53 BB = bb;
54 InsertPos = insertPos;
56 SUnits.clear();
57 Sequence.clear();
58 EntrySU = SUnit();
59 ExitSU = SUnit();
61 Schedule();
63 DEBUG({
64 errs() << "*** Final schedule ***\n";
65 dumpSchedule();
66 errs() << '\n';
67 });
70 /// addPred - This adds the specified edge as a pred of the current node if
71 /// not already. It also adds the current node as a successor of the
72 /// specified node.
73 void SUnit::addPred(const SDep &D) {
74 // If this node already has this depenence, don't add a redundant one.
75 for (SmallVector<SDep, 4>::const_iterator I = Preds.begin(), E = Preds.end();
76 I != E; ++I)
77 if (*I == D)
78 return;
79 // Now add a corresponding succ to N.
80 SDep P = D;
81 P.setSUnit(this);
82 SUnit *N = D.getSUnit();
83 // Update the bookkeeping.
84 if (D.getKind() == SDep::Data) {
85 ++NumPreds;
86 ++N->NumSuccs;
88 if (!N->isScheduled)
89 ++NumPredsLeft;
90 if (!isScheduled)
91 ++N->NumSuccsLeft;
92 Preds.push_back(D);
93 N->Succs.push_back(P);
94 if (P.getLatency() != 0) {
95 this->setDepthDirty();
96 N->setHeightDirty();
100 /// removePred - This removes the specified edge as a pred of the current
101 /// node if it exists. It also removes the current node as a successor of
102 /// the specified node.
103 void SUnit::removePred(const SDep &D) {
104 // Find the matching predecessor.
105 for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end();
106 I != E; ++I)
107 if (*I == D) {
108 bool FoundSucc = false;
109 // Find the corresponding successor in N.
110 SDep P = D;
111 P.setSUnit(this);
112 SUnit *N = D.getSUnit();
113 for (SmallVector<SDep, 4>::iterator II = N->Succs.begin(),
114 EE = N->Succs.end(); II != EE; ++II)
115 if (*II == P) {
116 FoundSucc = true;
117 N->Succs.erase(II);
118 break;
120 assert(FoundSucc && "Mismatching preds / succs lists!");
121 Preds.erase(I);
122 // Update the bookkeeping.
123 if (P.getKind() == SDep::Data) {
124 --NumPreds;
125 --N->NumSuccs;
127 if (!N->isScheduled)
128 --NumPredsLeft;
129 if (!isScheduled)
130 --N->NumSuccsLeft;
131 if (P.getLatency() != 0) {
132 this->setDepthDirty();
133 N->setHeightDirty();
135 return;
139 void SUnit::setDepthDirty() {
140 if (!isDepthCurrent) return;
141 SmallVector<SUnit*, 8> WorkList;
142 WorkList.push_back(this);
143 do {
144 SUnit *SU = WorkList.pop_back_val();
145 SU->isDepthCurrent = false;
146 for (SUnit::const_succ_iterator I = SU->Succs.begin(),
147 E = SU->Succs.end(); I != E; ++I) {
148 SUnit *SuccSU = I->getSUnit();
149 if (SuccSU->isDepthCurrent)
150 WorkList.push_back(SuccSU);
152 } while (!WorkList.empty());
155 void SUnit::setHeightDirty() {
156 if (!isHeightCurrent) return;
157 SmallVector<SUnit*, 8> WorkList;
158 WorkList.push_back(this);
159 do {
160 SUnit *SU = WorkList.pop_back_val();
161 SU->isHeightCurrent = false;
162 for (SUnit::const_pred_iterator I = SU->Preds.begin(),
163 E = SU->Preds.end(); I != E; ++I) {
164 SUnit *PredSU = I->getSUnit();
165 if (PredSU->isHeightCurrent)
166 WorkList.push_back(PredSU);
168 } while (!WorkList.empty());
171 /// setDepthToAtLeast - Update this node's successors to reflect the
172 /// fact that this node's depth just increased.
174 void SUnit::setDepthToAtLeast(unsigned NewDepth) {
175 if (NewDepth <= getDepth())
176 return;
177 setDepthDirty();
178 Depth = NewDepth;
179 isDepthCurrent = true;
182 /// setHeightToAtLeast - Update this node's predecessors to reflect the
183 /// fact that this node's height just increased.
185 void SUnit::setHeightToAtLeast(unsigned NewHeight) {
186 if (NewHeight <= getHeight())
187 return;
188 setHeightDirty();
189 Height = NewHeight;
190 isHeightCurrent = true;
193 /// ComputeDepth - Calculate the maximal path from the node to the exit.
195 void SUnit::ComputeDepth() {
196 SmallVector<SUnit*, 8> WorkList;
197 WorkList.push_back(this);
198 do {
199 SUnit *Cur = WorkList.back();
201 bool Done = true;
202 unsigned MaxPredDepth = 0;
203 for (SUnit::const_pred_iterator I = Cur->Preds.begin(),
204 E = Cur->Preds.end(); I != E; ++I) {
205 SUnit *PredSU = I->getSUnit();
206 if (PredSU->isDepthCurrent)
207 MaxPredDepth = std::max(MaxPredDepth,
208 PredSU->Depth + I->getLatency());
209 else {
210 Done = false;
211 WorkList.push_back(PredSU);
215 if (Done) {
216 WorkList.pop_back();
217 if (MaxPredDepth != Cur->Depth) {
218 Cur->setDepthDirty();
219 Cur->Depth = MaxPredDepth;
221 Cur->isDepthCurrent = true;
223 } while (!WorkList.empty());
226 /// ComputeHeight - Calculate the maximal path from the node to the entry.
228 void SUnit::ComputeHeight() {
229 SmallVector<SUnit*, 8> WorkList;
230 WorkList.push_back(this);
231 do {
232 SUnit *Cur = WorkList.back();
234 bool Done = true;
235 unsigned MaxSuccHeight = 0;
236 for (SUnit::const_succ_iterator I = Cur->Succs.begin(),
237 E = Cur->Succs.end(); I != E; ++I) {
238 SUnit *SuccSU = I->getSUnit();
239 if (SuccSU->isHeightCurrent)
240 MaxSuccHeight = std::max(MaxSuccHeight,
241 SuccSU->Height + I->getLatency());
242 else {
243 Done = false;
244 WorkList.push_back(SuccSU);
248 if (Done) {
249 WorkList.pop_back();
250 if (MaxSuccHeight != Cur->Height) {
251 Cur->setHeightDirty();
252 Cur->Height = MaxSuccHeight;
254 Cur->isHeightCurrent = true;
256 } while (!WorkList.empty());
259 /// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
260 /// a group of nodes flagged together.
261 void SUnit::dump(const ScheduleDAG *G) const {
262 errs() << "SU(" << NodeNum << "): ";
263 G->dumpNode(this);
266 void SUnit::dumpAll(const ScheduleDAG *G) const {
267 dump(G);
269 errs() << " # preds left : " << NumPredsLeft << "\n";
270 errs() << " # succs left : " << NumSuccsLeft << "\n";
271 errs() << " Latency : " << Latency << "\n";
272 errs() << " Depth : " << Depth << "\n";
273 errs() << " Height : " << Height << "\n";
275 if (Preds.size() != 0) {
276 errs() << " Predecessors:\n";
277 for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
278 I != E; ++I) {
279 errs() << " ";
280 switch (I->getKind()) {
281 case SDep::Data: errs() << "val "; break;
282 case SDep::Anti: errs() << "anti"; break;
283 case SDep::Output: errs() << "out "; break;
284 case SDep::Order: errs() << "ch "; break;
286 errs() << "#";
287 errs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
288 if (I->isArtificial())
289 errs() << " *";
290 errs() << ": Latency=" << I->getLatency();
291 errs() << "\n";
294 if (Succs.size() != 0) {
295 errs() << " Successors:\n";
296 for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
297 I != E; ++I) {
298 errs() << " ";
299 switch (I->getKind()) {
300 case SDep::Data: errs() << "val "; break;
301 case SDep::Anti: errs() << "anti"; break;
302 case SDep::Output: errs() << "out "; break;
303 case SDep::Order: errs() << "ch "; break;
305 errs() << "#";
306 errs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
307 if (I->isArtificial())
308 errs() << " *";
309 errs() << ": Latency=" << I->getLatency();
310 errs() << "\n";
313 errs() << "\n";
316 #ifndef NDEBUG
317 /// VerifySchedule - Verify that all SUnits were scheduled and that
318 /// their state is consistent.
320 void ScheduleDAG::VerifySchedule(bool isBottomUp) {
321 bool AnyNotSched = false;
322 unsigned DeadNodes = 0;
323 unsigned Noops = 0;
324 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
325 if (!SUnits[i].isScheduled) {
326 if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) {
327 ++DeadNodes;
328 continue;
330 if (!AnyNotSched)
331 errs() << "*** Scheduling failed! ***\n";
332 SUnits[i].dump(this);
333 errs() << "has not been scheduled!\n";
334 AnyNotSched = true;
336 if (SUnits[i].isScheduled &&
337 (isBottomUp ? SUnits[i].getHeight() : SUnits[i].getHeight()) >
338 unsigned(INT_MAX)) {
339 if (!AnyNotSched)
340 errs() << "*** Scheduling failed! ***\n";
341 SUnits[i].dump(this);
342 errs() << "has an unexpected "
343 << (isBottomUp ? "Height" : "Depth") << " value!\n";
344 AnyNotSched = true;
346 if (isBottomUp) {
347 if (SUnits[i].NumSuccsLeft != 0) {
348 if (!AnyNotSched)
349 errs() << "*** Scheduling failed! ***\n";
350 SUnits[i].dump(this);
351 errs() << "has successors left!\n";
352 AnyNotSched = true;
354 } else {
355 if (SUnits[i].NumPredsLeft != 0) {
356 if (!AnyNotSched)
357 errs() << "*** Scheduling failed! ***\n";
358 SUnits[i].dump(this);
359 errs() << "has predecessors left!\n";
360 AnyNotSched = true;
364 for (unsigned i = 0, e = Sequence.size(); i != e; ++i)
365 if (!Sequence[i])
366 ++Noops;
367 assert(!AnyNotSched);
368 assert(Sequence.size() + DeadNodes - Noops == SUnits.size() &&
369 "The number of nodes scheduled doesn't match the expected number!");
371 #endif
373 /// InitDAGTopologicalSorting - create the initial topological
374 /// ordering from the DAG to be scheduled.
376 /// The idea of the algorithm is taken from
377 /// "Online algorithms for managing the topological order of
378 /// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly
379 /// This is the MNR algorithm, which was first introduced by
380 /// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in
381 /// "Maintaining a topological order under edge insertions".
383 /// Short description of the algorithm:
385 /// Topological ordering, ord, of a DAG maps each node to a topological
386 /// index so that for all edges X->Y it is the case that ord(X) < ord(Y).
388 /// This means that if there is a path from the node X to the node Z,
389 /// then ord(X) < ord(Z).
391 /// This property can be used to check for reachability of nodes:
392 /// if Z is reachable from X, then an insertion of the edge Z->X would
393 /// create a cycle.
395 /// The algorithm first computes a topological ordering for the DAG by
396 /// initializing the Index2Node and Node2Index arrays and then tries to keep
397 /// the ordering up-to-date after edge insertions by reordering the DAG.
399 /// On insertion of the edge X->Y, the algorithm first marks by calling DFS
400 /// the nodes reachable from Y, and then shifts them using Shift to lie
401 /// immediately after X in Index2Node.
402 void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() {
403 unsigned DAGSize = SUnits.size();
404 std::vector<SUnit*> WorkList;
405 WorkList.reserve(DAGSize);
407 Index2Node.resize(DAGSize);
408 Node2Index.resize(DAGSize);
410 // Initialize the data structures.
411 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
412 SUnit *SU = &SUnits[i];
413 int NodeNum = SU->NodeNum;
414 unsigned Degree = SU->Succs.size();
415 // Temporarily use the Node2Index array as scratch space for degree counts.
416 Node2Index[NodeNum] = Degree;
418 // Is it a node without dependencies?
419 if (Degree == 0) {
420 assert(SU->Succs.empty() && "SUnit should have no successors");
421 // Collect leaf nodes.
422 WorkList.push_back(SU);
426 int Id = DAGSize;
427 while (!WorkList.empty()) {
428 SUnit *SU = WorkList.back();
429 WorkList.pop_back();
430 Allocate(SU->NodeNum, --Id);
431 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
432 I != E; ++I) {
433 SUnit *SU = I->getSUnit();
434 if (!--Node2Index[SU->NodeNum])
435 // If all dependencies of the node are processed already,
436 // then the node can be computed now.
437 WorkList.push_back(SU);
441 Visited.resize(DAGSize);
443 #ifndef NDEBUG
444 // Check correctness of the ordering
445 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
446 SUnit *SU = &SUnits[i];
447 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
448 I != E; ++I) {
449 assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] &&
450 "Wrong topological sorting");
453 #endif
456 /// AddPred - Updates the topological ordering to accomodate an edge
457 /// to be added from SUnit X to SUnit Y.
458 void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) {
459 int UpperBound, LowerBound;
460 LowerBound = Node2Index[Y->NodeNum];
461 UpperBound = Node2Index[X->NodeNum];
462 bool HasLoop = false;
463 // Is Ord(X) < Ord(Y) ?
464 if (LowerBound < UpperBound) {
465 // Update the topological order.
466 Visited.reset();
467 DFS(Y, UpperBound, HasLoop);
468 assert(!HasLoop && "Inserted edge creates a loop!");
469 // Recompute topological indexes.
470 Shift(Visited, LowerBound, UpperBound);
474 /// RemovePred - Updates the topological ordering to accomodate an
475 /// an edge to be removed from the specified node N from the predecessors
476 /// of the current node M.
477 void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) {
478 // InitDAGTopologicalSorting();
481 /// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark
482 /// all nodes affected by the edge insertion. These nodes will later get new
483 /// topological indexes by means of the Shift method.
484 void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound,
485 bool& HasLoop) {
486 std::vector<const SUnit*> WorkList;
487 WorkList.reserve(SUnits.size());
489 WorkList.push_back(SU);
490 do {
491 SU = WorkList.back();
492 WorkList.pop_back();
493 Visited.set(SU->NodeNum);
494 for (int I = SU->Succs.size()-1; I >= 0; --I) {
495 int s = SU->Succs[I].getSUnit()->NodeNum;
496 if (Node2Index[s] == UpperBound) {
497 HasLoop = true;
498 return;
500 // Visit successors if not already and in affected region.
501 if (!Visited.test(s) && Node2Index[s] < UpperBound) {
502 WorkList.push_back(SU->Succs[I].getSUnit());
505 } while (!WorkList.empty());
508 /// Shift - Renumber the nodes so that the topological ordering is
509 /// preserved.
510 void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound,
511 int UpperBound) {
512 std::vector<int> L;
513 int shift = 0;
514 int i;
516 for (i = LowerBound; i <= UpperBound; ++i) {
517 // w is node at topological index i.
518 int w = Index2Node[i];
519 if (Visited.test(w)) {
520 // Unmark.
521 Visited.reset(w);
522 L.push_back(w);
523 shift = shift + 1;
524 } else {
525 Allocate(w, i - shift);
529 for (unsigned j = 0; j < L.size(); ++j) {
530 Allocate(L[j], i - shift);
531 i = i + 1;
536 /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
537 /// create a cycle.
538 bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
539 if (IsReachable(TargetSU, SU))
540 return true;
541 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
542 I != E; ++I)
543 if (I->isAssignedRegDep() &&
544 IsReachable(TargetSU, I->getSUnit()))
545 return true;
546 return false;
549 /// IsReachable - Checks if SU is reachable from TargetSU.
550 bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU,
551 const SUnit *TargetSU) {
552 // If insertion of the edge SU->TargetSU would create a cycle
553 // then there is a path from TargetSU to SU.
554 int UpperBound, LowerBound;
555 LowerBound = Node2Index[TargetSU->NodeNum];
556 UpperBound = Node2Index[SU->NodeNum];
557 bool HasLoop = false;
558 // Is Ord(TargetSU) < Ord(SU) ?
559 if (LowerBound < UpperBound) {
560 Visited.reset();
561 // There may be a path from TargetSU to SU. Check for it.
562 DFS(TargetSU, UpperBound, HasLoop);
564 return HasLoop;
567 /// Allocate - assign the topological index to the node n.
568 void ScheduleDAGTopologicalSort::Allocate(int n, int index) {
569 Node2Index[n] = index;
570 Index2Node[index] = n;
573 ScheduleDAGTopologicalSort::ScheduleDAGTopologicalSort(
574 std::vector<SUnit> &sunits)
575 : SUnits(sunits) {}
577 ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {}