1 //===---- ScheduleDAG.cpp - Implement the ScheduleDAG class ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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"
26 ScheduleDAG::ScheduleDAG(MachineFunction &mf)
28 TII(TM.getInstrInfo()),
29 TRI(TM.getRegisterInfo()),
30 TLI(TM.getTargetLowering()),
31 MF(mf), MRI(mf.getRegInfo()),
32 ConstPool(MF.getConstantPool()),
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])
44 errs() << "**** NOOP ****\n";
49 /// Run - perform scheduling.
51 void ScheduleDAG::Run(MachineBasicBlock *bb,
52 MachineBasicBlock::iterator insertPos) {
54 InsertPos = insertPos;
64 errs() << "*** Final schedule ***\n";
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
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();
79 // Now add a corresponding succ to N.
82 SUnit *N = D.getSUnit();
83 // Update the bookkeeping.
84 if (D.getKind() == SDep::Data) {
85 assert(NumPreds < UINT_MAX && "NumPreds will overflow!");
86 assert(N->NumSuccs < UINT_MAX && "NumSuccs will overflow!");
90 if (!N->isScheduled) {
91 assert(NumPredsLeft < UINT_MAX && "NumPredsLeft will overflow!");
95 assert(N->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!");
99 N->Succs.push_back(P);
100 if (P.getLatency() != 0) {
101 this->setDepthDirty();
106 /// removePred - This removes the specified edge as a pred of the current
107 /// node if it exists. It also removes the current node as a successor of
108 /// the specified node.
109 void SUnit::removePred(const SDep &D) {
110 // Find the matching predecessor.
111 for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end();
114 bool FoundSucc = false;
115 // Find the corresponding successor in N.
118 SUnit *N = D.getSUnit();
119 for (SmallVector<SDep, 4>::iterator II = N->Succs.begin(),
120 EE = N->Succs.end(); II != EE; ++II)
126 assert(FoundSucc && "Mismatching preds / succs lists!");
128 // Update the bookkeeping.
129 if (P.getKind() == SDep::Data) {
130 assert(NumPreds > 0 && "NumPreds will underflow!");
131 assert(N->NumSuccs > 0 && "NumSuccs will underflow!");
135 if (!N->isScheduled) {
136 assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!");
140 assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!");
143 if (P.getLatency() != 0) {
144 this->setDepthDirty();
151 void SUnit::setDepthDirty() {
152 if (!isDepthCurrent) return;
153 SmallVector<SUnit*, 8> WorkList;
154 WorkList.push_back(this);
156 SUnit *SU = WorkList.pop_back_val();
157 SU->isDepthCurrent = false;
158 for (SUnit::const_succ_iterator I = SU->Succs.begin(),
159 E = SU->Succs.end(); I != E; ++I) {
160 SUnit *SuccSU = I->getSUnit();
161 if (SuccSU->isDepthCurrent)
162 WorkList.push_back(SuccSU);
164 } while (!WorkList.empty());
167 void SUnit::setHeightDirty() {
168 if (!isHeightCurrent) return;
169 SmallVector<SUnit*, 8> WorkList;
170 WorkList.push_back(this);
172 SUnit *SU = WorkList.pop_back_val();
173 SU->isHeightCurrent = false;
174 for (SUnit::const_pred_iterator I = SU->Preds.begin(),
175 E = SU->Preds.end(); I != E; ++I) {
176 SUnit *PredSU = I->getSUnit();
177 if (PredSU->isHeightCurrent)
178 WorkList.push_back(PredSU);
180 } while (!WorkList.empty());
183 /// setDepthToAtLeast - Update this node's successors to reflect the
184 /// fact that this node's depth just increased.
186 void SUnit::setDepthToAtLeast(unsigned NewDepth) {
187 if (NewDepth <= getDepth())
191 isDepthCurrent = true;
194 /// setHeightToAtLeast - Update this node's predecessors to reflect the
195 /// fact that this node's height just increased.
197 void SUnit::setHeightToAtLeast(unsigned NewHeight) {
198 if (NewHeight <= getHeight())
202 isHeightCurrent = true;
205 /// ComputeDepth - Calculate the maximal path from the node to the exit.
207 void SUnit::ComputeDepth() {
208 SmallVector<SUnit*, 8> WorkList;
209 WorkList.push_back(this);
211 SUnit *Cur = WorkList.back();
214 unsigned MaxPredDepth = 0;
215 for (SUnit::const_pred_iterator I = Cur->Preds.begin(),
216 E = Cur->Preds.end(); I != E; ++I) {
217 SUnit *PredSU = I->getSUnit();
218 if (PredSU->isDepthCurrent)
219 MaxPredDepth = std::max(MaxPredDepth,
220 PredSU->Depth + I->getLatency());
223 WorkList.push_back(PredSU);
229 if (MaxPredDepth != Cur->Depth) {
230 Cur->setDepthDirty();
231 Cur->Depth = MaxPredDepth;
233 Cur->isDepthCurrent = true;
235 } while (!WorkList.empty());
238 /// ComputeHeight - Calculate the maximal path from the node to the entry.
240 void SUnit::ComputeHeight() {
241 SmallVector<SUnit*, 8> WorkList;
242 WorkList.push_back(this);
244 SUnit *Cur = WorkList.back();
247 unsigned MaxSuccHeight = 0;
248 for (SUnit::const_succ_iterator I = Cur->Succs.begin(),
249 E = Cur->Succs.end(); I != E; ++I) {
250 SUnit *SuccSU = I->getSUnit();
251 if (SuccSU->isHeightCurrent)
252 MaxSuccHeight = std::max(MaxSuccHeight,
253 SuccSU->Height + I->getLatency());
256 WorkList.push_back(SuccSU);
262 if (MaxSuccHeight != Cur->Height) {
263 Cur->setHeightDirty();
264 Cur->Height = MaxSuccHeight;
266 Cur->isHeightCurrent = true;
268 } while (!WorkList.empty());
271 /// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
272 /// a group of nodes flagged together.
273 void SUnit::dump(const ScheduleDAG *G) const {
274 errs() << "SU(" << NodeNum << "): ";
278 void SUnit::dumpAll(const ScheduleDAG *G) const {
281 errs() << " # preds left : " << NumPredsLeft << "\n";
282 errs() << " # succs left : " << NumSuccsLeft << "\n";
283 errs() << " Latency : " << Latency << "\n";
284 errs() << " Depth : " << Depth << "\n";
285 errs() << " Height : " << Height << "\n";
287 if (Preds.size() != 0) {
288 errs() << " Predecessors:\n";
289 for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
292 switch (I->getKind()) {
293 case SDep::Data: errs() << "val "; break;
294 case SDep::Anti: errs() << "anti"; break;
295 case SDep::Output: errs() << "out "; break;
296 case SDep::Order: errs() << "ch "; break;
299 errs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
300 if (I->isArtificial())
302 errs() << ": Latency=" << I->getLatency();
306 if (Succs.size() != 0) {
307 errs() << " Successors:\n";
308 for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
311 switch (I->getKind()) {
312 case SDep::Data: errs() << "val "; break;
313 case SDep::Anti: errs() << "anti"; break;
314 case SDep::Output: errs() << "out "; break;
315 case SDep::Order: errs() << "ch "; break;
318 errs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
319 if (I->isArtificial())
321 errs() << ": Latency=" << I->getLatency();
329 /// VerifySchedule - Verify that all SUnits were scheduled and that
330 /// their state is consistent.
332 void ScheduleDAG::VerifySchedule(bool isBottomUp) {
333 bool AnyNotSched = false;
334 unsigned DeadNodes = 0;
336 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
337 if (!SUnits[i].isScheduled) {
338 if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) {
343 errs() << "*** Scheduling failed! ***\n";
344 SUnits[i].dump(this);
345 errs() << "has not been scheduled!\n";
348 if (SUnits[i].isScheduled &&
349 (isBottomUp ? SUnits[i].getHeight() : SUnits[i].getHeight()) >
352 errs() << "*** Scheduling failed! ***\n";
353 SUnits[i].dump(this);
354 errs() << "has an unexpected "
355 << (isBottomUp ? "Height" : "Depth") << " value!\n";
359 if (SUnits[i].NumSuccsLeft != 0) {
361 errs() << "*** Scheduling failed! ***\n";
362 SUnits[i].dump(this);
363 errs() << "has successors left!\n";
367 if (SUnits[i].NumPredsLeft != 0) {
369 errs() << "*** Scheduling failed! ***\n";
370 SUnits[i].dump(this);
371 errs() << "has predecessors left!\n";
376 for (unsigned i = 0, e = Sequence.size(); i != e; ++i)
379 assert(!AnyNotSched);
380 assert(Sequence.size() + DeadNodes - Noops == SUnits.size() &&
381 "The number of nodes scheduled doesn't match the expected number!");
385 /// InitDAGTopologicalSorting - create the initial topological
386 /// ordering from the DAG to be scheduled.
388 /// The idea of the algorithm is taken from
389 /// "Online algorithms for managing the topological order of
390 /// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly
391 /// This is the MNR algorithm, which was first introduced by
392 /// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in
393 /// "Maintaining a topological order under edge insertions".
395 /// Short description of the algorithm:
397 /// Topological ordering, ord, of a DAG maps each node to a topological
398 /// index so that for all edges X->Y it is the case that ord(X) < ord(Y).
400 /// This means that if there is a path from the node X to the node Z,
401 /// then ord(X) < ord(Z).
403 /// This property can be used to check for reachability of nodes:
404 /// if Z is reachable from X, then an insertion of the edge Z->X would
407 /// The algorithm first computes a topological ordering for the DAG by
408 /// initializing the Index2Node and Node2Index arrays and then tries to keep
409 /// the ordering up-to-date after edge insertions by reordering the DAG.
411 /// On insertion of the edge X->Y, the algorithm first marks by calling DFS
412 /// the nodes reachable from Y, and then shifts them using Shift to lie
413 /// immediately after X in Index2Node.
414 void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() {
415 unsigned DAGSize = SUnits.size();
416 std::vector<SUnit*> WorkList;
417 WorkList.reserve(DAGSize);
419 Index2Node.resize(DAGSize);
420 Node2Index.resize(DAGSize);
422 // Initialize the data structures.
423 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
424 SUnit *SU = &SUnits[i];
425 int NodeNum = SU->NodeNum;
426 unsigned Degree = SU->Succs.size();
427 // Temporarily use the Node2Index array as scratch space for degree counts.
428 Node2Index[NodeNum] = Degree;
430 // Is it a node without dependencies?
432 assert(SU->Succs.empty() && "SUnit should have no successors");
433 // Collect leaf nodes.
434 WorkList.push_back(SU);
439 while (!WorkList.empty()) {
440 SUnit *SU = WorkList.back();
442 Allocate(SU->NodeNum, --Id);
443 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
445 SUnit *SU = I->getSUnit();
446 if (!--Node2Index[SU->NodeNum])
447 // If all dependencies of the node are processed already,
448 // then the node can be computed now.
449 WorkList.push_back(SU);
453 Visited.resize(DAGSize);
456 // Check correctness of the ordering
457 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
458 SUnit *SU = &SUnits[i];
459 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
461 assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] &&
462 "Wrong topological sorting");
468 /// AddPred - Updates the topological ordering to accomodate an edge
469 /// to be added from SUnit X to SUnit Y.
470 void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) {
471 int UpperBound, LowerBound;
472 LowerBound = Node2Index[Y->NodeNum];
473 UpperBound = Node2Index[X->NodeNum];
474 bool HasLoop = false;
475 // Is Ord(X) < Ord(Y) ?
476 if (LowerBound < UpperBound) {
477 // Update the topological order.
479 DFS(Y, UpperBound, HasLoop);
480 assert(!HasLoop && "Inserted edge creates a loop!");
481 // Recompute topological indexes.
482 Shift(Visited, LowerBound, UpperBound);
486 /// RemovePred - Updates the topological ordering to accomodate an
487 /// an edge to be removed from the specified node N from the predecessors
488 /// of the current node M.
489 void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) {
490 // InitDAGTopologicalSorting();
493 /// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark
494 /// all nodes affected by the edge insertion. These nodes will later get new
495 /// topological indexes by means of the Shift method.
496 void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound,
498 std::vector<const SUnit*> WorkList;
499 WorkList.reserve(SUnits.size());
501 WorkList.push_back(SU);
503 SU = WorkList.back();
505 Visited.set(SU->NodeNum);
506 for (int I = SU->Succs.size()-1; I >= 0; --I) {
507 int s = SU->Succs[I].getSUnit()->NodeNum;
508 if (Node2Index[s] == UpperBound) {
512 // Visit successors if not already and in affected region.
513 if (!Visited.test(s) && Node2Index[s] < UpperBound) {
514 WorkList.push_back(SU->Succs[I].getSUnit());
517 } while (!WorkList.empty());
520 /// Shift - Renumber the nodes so that the topological ordering is
522 void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound,
528 for (i = LowerBound; i <= UpperBound; ++i) {
529 // w is node at topological index i.
530 int w = Index2Node[i];
531 if (Visited.test(w)) {
537 Allocate(w, i - shift);
541 for (unsigned j = 0; j < L.size(); ++j) {
542 Allocate(L[j], i - shift);
548 /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
550 bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
551 if (IsReachable(TargetSU, SU))
553 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
555 if (I->isAssignedRegDep() &&
556 IsReachable(TargetSU, I->getSUnit()))
561 /// IsReachable - Checks if SU is reachable from TargetSU.
562 bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU,
563 const SUnit *TargetSU) {
564 // If insertion of the edge SU->TargetSU would create a cycle
565 // then there is a path from TargetSU to SU.
566 int UpperBound, LowerBound;
567 LowerBound = Node2Index[TargetSU->NodeNum];
568 UpperBound = Node2Index[SU->NodeNum];
569 bool HasLoop = false;
570 // Is Ord(TargetSU) < Ord(SU) ?
571 if (LowerBound < UpperBound) {
573 // There may be a path from TargetSU to SU. Check for it.
574 DFS(TargetSU, UpperBound, HasLoop);
579 /// Allocate - assign the topological index to the node n.
580 void ScheduleDAGTopologicalSort::Allocate(int n, int index) {
581 Node2Index[n] = index;
582 Index2Node[index] = n;
585 ScheduleDAGTopologicalSort::ScheduleDAGTopologicalSort(
586 std::vector<SUnit> &sunits)
589 ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {}