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/CodeGen/SelectionDAGNodes.h"
19 #include "llvm/Target/TargetMachine.h"
20 #include "llvm/Target/TargetInstrInfo.h"
21 #include "llvm/Target/TargetRegisterInfo.h"
22 #include "llvm/Support/CommandLine.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/raw_ostream.h"
29 static cl::opt<bool> StressSchedOpt(
30 "stress-sched", cl::Hidden, cl::init(false),
31 cl::desc("Stress test instruction scheduling"));
34 void SchedulingPriorityQueue::anchor() { }
36 ScheduleDAG::ScheduleDAG(MachineFunction &mf)
38 TII(TM.getInstrInfo()),
39 TRI(TM.getRegisterInfo()),
40 MF(mf), MRI(mf.getRegInfo()),
43 StressSched = StressSchedOpt;
47 ScheduleDAG::~ScheduleDAG() {}
49 /// getInstrDesc helper to handle SDNodes.
50 const MCInstrDesc *ScheduleDAG::getNodeDesc(const SDNode *Node) const {
51 if (!Node || !Node->isMachineOpcode()) return NULL;
52 return &TII->get(Node->getMachineOpcode());
55 /// dump - dump the schedule.
56 void ScheduleDAG::dumpSchedule() const {
57 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
58 if (SUnit *SU = Sequence[i])
61 dbgs() << "**** NOOP ****\n";
66 /// Run - perform scheduling.
68 void ScheduleDAG::Run(MachineBasicBlock *bb,
69 MachineBasicBlock::iterator insertPos) {
71 InsertPos = insertPos;
81 dbgs() << "*** Final schedule ***\n";
87 /// addPred - This adds the specified edge as a pred of the current node if
88 /// not already. It also adds the current node as a successor of the
90 bool SUnit::addPred(const SDep &D) {
91 // If this node already has this depenence, don't add a redundant one.
92 for (SmallVector<SDep, 4>::const_iterator I = Preds.begin(), E = Preds.end();
96 // Now add a corresponding succ to N.
99 SUnit *N = D.getSUnit();
100 // Update the bookkeeping.
101 if (D.getKind() == SDep::Data) {
102 assert(NumPreds < UINT_MAX && "NumPreds will overflow!");
103 assert(N->NumSuccs < UINT_MAX && "NumSuccs will overflow!");
107 if (!N->isScheduled) {
108 assert(NumPredsLeft < UINT_MAX && "NumPredsLeft will overflow!");
112 assert(N->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!");
116 N->Succs.push_back(P);
117 if (P.getLatency() != 0) {
118 this->setDepthDirty();
124 /// removePred - This removes the specified edge as a pred of the current
125 /// node if it exists. It also removes the current node as a successor of
126 /// the specified node.
127 void SUnit::removePred(const SDep &D) {
128 // Find the matching predecessor.
129 for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end();
132 bool FoundSucc = false;
133 // Find the corresponding successor in N.
136 SUnit *N = D.getSUnit();
137 for (SmallVector<SDep, 4>::iterator II = N->Succs.begin(),
138 EE = N->Succs.end(); II != EE; ++II)
144 assert(FoundSucc && "Mismatching preds / succs lists!");
147 // Update the bookkeeping.
148 if (P.getKind() == SDep::Data) {
149 assert(NumPreds > 0 && "NumPreds will underflow!");
150 assert(N->NumSuccs > 0 && "NumSuccs will underflow!");
154 if (!N->isScheduled) {
155 assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!");
159 assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!");
162 if (P.getLatency() != 0) {
163 this->setDepthDirty();
170 void SUnit::setDepthDirty() {
171 if (!isDepthCurrent) return;
172 SmallVector<SUnit*, 8> WorkList;
173 WorkList.push_back(this);
175 SUnit *SU = WorkList.pop_back_val();
176 SU->isDepthCurrent = false;
177 for (SUnit::const_succ_iterator I = SU->Succs.begin(),
178 E = SU->Succs.end(); I != E; ++I) {
179 SUnit *SuccSU = I->getSUnit();
180 if (SuccSU->isDepthCurrent)
181 WorkList.push_back(SuccSU);
183 } while (!WorkList.empty());
186 void SUnit::setHeightDirty() {
187 if (!isHeightCurrent) return;
188 SmallVector<SUnit*, 8> WorkList;
189 WorkList.push_back(this);
191 SUnit *SU = WorkList.pop_back_val();
192 SU->isHeightCurrent = false;
193 for (SUnit::const_pred_iterator I = SU->Preds.begin(),
194 E = SU->Preds.end(); I != E; ++I) {
195 SUnit *PredSU = I->getSUnit();
196 if (PredSU->isHeightCurrent)
197 WorkList.push_back(PredSU);
199 } while (!WorkList.empty());
202 /// setDepthToAtLeast - Update this node's successors to reflect the
203 /// fact that this node's depth just increased.
205 void SUnit::setDepthToAtLeast(unsigned NewDepth) {
206 if (NewDepth <= getDepth())
210 isDepthCurrent = true;
213 /// setHeightToAtLeast - Update this node's predecessors to reflect the
214 /// fact that this node's height just increased.
216 void SUnit::setHeightToAtLeast(unsigned NewHeight) {
217 if (NewHeight <= getHeight())
221 isHeightCurrent = true;
224 /// ComputeDepth - Calculate the maximal path from the node to the exit.
226 void SUnit::ComputeDepth() {
227 SmallVector<SUnit*, 8> WorkList;
228 WorkList.push_back(this);
230 SUnit *Cur = WorkList.back();
233 unsigned MaxPredDepth = 0;
234 for (SUnit::const_pred_iterator I = Cur->Preds.begin(),
235 E = Cur->Preds.end(); I != E; ++I) {
236 SUnit *PredSU = I->getSUnit();
237 if (PredSU->isDepthCurrent)
238 MaxPredDepth = std::max(MaxPredDepth,
239 PredSU->Depth + I->getLatency());
242 WorkList.push_back(PredSU);
248 if (MaxPredDepth != Cur->Depth) {
249 Cur->setDepthDirty();
250 Cur->Depth = MaxPredDepth;
252 Cur->isDepthCurrent = true;
254 } while (!WorkList.empty());
257 /// ComputeHeight - Calculate the maximal path from the node to the entry.
259 void SUnit::ComputeHeight() {
260 SmallVector<SUnit*, 8> WorkList;
261 WorkList.push_back(this);
263 SUnit *Cur = WorkList.back();
266 unsigned MaxSuccHeight = 0;
267 for (SUnit::const_succ_iterator I = Cur->Succs.begin(),
268 E = Cur->Succs.end(); I != E; ++I) {
269 SUnit *SuccSU = I->getSUnit();
270 if (SuccSU->isHeightCurrent)
271 MaxSuccHeight = std::max(MaxSuccHeight,
272 SuccSU->Height + I->getLatency());
275 WorkList.push_back(SuccSU);
281 if (MaxSuccHeight != Cur->Height) {
282 Cur->setHeightDirty();
283 Cur->Height = MaxSuccHeight;
285 Cur->isHeightCurrent = true;
287 } while (!WorkList.empty());
290 /// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
291 /// a group of nodes flagged together.
292 void SUnit::dump(const ScheduleDAG *G) const {
293 dbgs() << "SU(" << NodeNum << "): ";
297 void SUnit::dumpAll(const ScheduleDAG *G) const {
300 dbgs() << " # preds left : " << NumPredsLeft << "\n";
301 dbgs() << " # succs left : " << NumSuccsLeft << "\n";
302 dbgs() << " # rdefs left : " << NumRegDefsLeft << "\n";
303 dbgs() << " Latency : " << Latency << "\n";
304 dbgs() << " Depth : " << Depth << "\n";
305 dbgs() << " Height : " << Height << "\n";
307 if (Preds.size() != 0) {
308 dbgs() << " Predecessors:\n";
309 for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
312 switch (I->getKind()) {
313 case SDep::Data: dbgs() << "val "; break;
314 case SDep::Anti: dbgs() << "anti"; break;
315 case SDep::Output: dbgs() << "out "; break;
316 case SDep::Order: dbgs() << "ch "; break;
318 dbgs() << "SU(" << I->getSUnit()->NodeNum << ")";
319 if (I->isArtificial())
321 dbgs() << ": Latency=" << I->getLatency();
322 if (I->isAssignedRegDep())
323 dbgs() << " Reg=" << PrintReg(I->getReg(), G->TRI);
327 if (Succs.size() != 0) {
328 dbgs() << " Successors:\n";
329 for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
332 switch (I->getKind()) {
333 case SDep::Data: dbgs() << "val "; break;
334 case SDep::Anti: dbgs() << "anti"; break;
335 case SDep::Output: dbgs() << "out "; break;
336 case SDep::Order: dbgs() << "ch "; break;
338 dbgs() << "SU(" << I->getSUnit()->NodeNum << ")";
339 if (I->isArtificial())
341 dbgs() << ": Latency=" << I->getLatency();
349 /// VerifyScheduledDAG - Verify that all SUnits were scheduled and that
350 /// their state is consistent. Return the number of scheduled nodes.
352 unsigned ScheduleDAG::VerifyScheduledDAG(bool isBottomUp) {
353 bool AnyNotSched = false;
354 unsigned DeadNodes = 0;
355 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
356 if (!SUnits[i].isScheduled) {
357 if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) {
362 dbgs() << "*** Scheduling failed! ***\n";
363 SUnits[i].dump(this);
364 dbgs() << "has not been scheduled!\n";
367 if (SUnits[i].isScheduled &&
368 (isBottomUp ? SUnits[i].getHeight() : SUnits[i].getDepth()) >
371 dbgs() << "*** Scheduling failed! ***\n";
372 SUnits[i].dump(this);
373 dbgs() << "has an unexpected "
374 << (isBottomUp ? "Height" : "Depth") << " value!\n";
378 if (SUnits[i].NumSuccsLeft != 0) {
380 dbgs() << "*** Scheduling failed! ***\n";
381 SUnits[i].dump(this);
382 dbgs() << "has successors left!\n";
386 if (SUnits[i].NumPredsLeft != 0) {
388 dbgs() << "*** Scheduling failed! ***\n";
389 SUnits[i].dump(this);
390 dbgs() << "has predecessors left!\n";
395 assert(!AnyNotSched);
396 return SUnits.size() - DeadNodes;
400 /// InitDAGTopologicalSorting - create the initial topological
401 /// ordering from the DAG to be scheduled.
403 /// The idea of the algorithm is taken from
404 /// "Online algorithms for managing the topological order of
405 /// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly
406 /// This is the MNR algorithm, which was first introduced by
407 /// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in
408 /// "Maintaining a topological order under edge insertions".
410 /// Short description of the algorithm:
412 /// Topological ordering, ord, of a DAG maps each node to a topological
413 /// index so that for all edges X->Y it is the case that ord(X) < ord(Y).
415 /// This means that if there is a path from the node X to the node Z,
416 /// then ord(X) < ord(Z).
418 /// This property can be used to check for reachability of nodes:
419 /// if Z is reachable from X, then an insertion of the edge Z->X would
422 /// The algorithm first computes a topological ordering for the DAG by
423 /// initializing the Index2Node and Node2Index arrays and then tries to keep
424 /// the ordering up-to-date after edge insertions by reordering the DAG.
426 /// On insertion of the edge X->Y, the algorithm first marks by calling DFS
427 /// the nodes reachable from Y, and then shifts them using Shift to lie
428 /// immediately after X in Index2Node.
429 void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() {
430 unsigned DAGSize = SUnits.size();
431 std::vector<SUnit*> WorkList;
432 WorkList.reserve(DAGSize);
434 Index2Node.resize(DAGSize);
435 Node2Index.resize(DAGSize);
437 // Initialize the data structures.
438 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
439 SUnit *SU = &SUnits[i];
440 int NodeNum = SU->NodeNum;
441 unsigned Degree = SU->Succs.size();
442 // Temporarily use the Node2Index array as scratch space for degree counts.
443 Node2Index[NodeNum] = Degree;
445 // Is it a node without dependencies?
447 assert(SU->Succs.empty() && "SUnit should have no successors");
448 // Collect leaf nodes.
449 WorkList.push_back(SU);
454 while (!WorkList.empty()) {
455 SUnit *SU = WorkList.back();
457 Allocate(SU->NodeNum, --Id);
458 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
460 SUnit *SU = I->getSUnit();
461 if (!--Node2Index[SU->NodeNum])
462 // If all dependencies of the node are processed already,
463 // then the node can be computed now.
464 WorkList.push_back(SU);
468 Visited.resize(DAGSize);
471 // Check correctness of the ordering
472 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
473 SUnit *SU = &SUnits[i];
474 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
476 assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] &&
477 "Wrong topological sorting");
483 /// AddPred - Updates the topological ordering to accommodate an edge
484 /// to be added from SUnit X to SUnit Y.
485 void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) {
486 int UpperBound, LowerBound;
487 LowerBound = Node2Index[Y->NodeNum];
488 UpperBound = Node2Index[X->NodeNum];
489 bool HasLoop = false;
490 // Is Ord(X) < Ord(Y) ?
491 if (LowerBound < UpperBound) {
492 // Update the topological order.
494 DFS(Y, UpperBound, HasLoop);
495 assert(!HasLoop && "Inserted edge creates a loop!");
496 // Recompute topological indexes.
497 Shift(Visited, LowerBound, UpperBound);
501 /// RemovePred - Updates the topological ordering to accommodate an
502 /// an edge to be removed from the specified node N from the predecessors
503 /// of the current node M.
504 void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) {
505 // InitDAGTopologicalSorting();
508 /// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark
509 /// all nodes affected by the edge insertion. These nodes will later get new
510 /// topological indexes by means of the Shift method.
511 void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound,
513 std::vector<const SUnit*> WorkList;
514 WorkList.reserve(SUnits.size());
516 WorkList.push_back(SU);
518 SU = WorkList.back();
520 Visited.set(SU->NodeNum);
521 for (int I = SU->Succs.size()-1; I >= 0; --I) {
522 int s = SU->Succs[I].getSUnit()->NodeNum;
523 if (Node2Index[s] == UpperBound) {
527 // Visit successors if not already and in affected region.
528 if (!Visited.test(s) && Node2Index[s] < UpperBound) {
529 WorkList.push_back(SU->Succs[I].getSUnit());
532 } while (!WorkList.empty());
535 /// Shift - Renumber the nodes so that the topological ordering is
537 void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound,
543 for (i = LowerBound; i <= UpperBound; ++i) {
544 // w is node at topological index i.
545 int w = Index2Node[i];
546 if (Visited.test(w)) {
552 Allocate(w, i - shift);
556 for (unsigned j = 0; j < L.size(); ++j) {
557 Allocate(L[j], i - shift);
563 /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
565 bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
566 if (IsReachable(TargetSU, SU))
568 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
570 if (I->isAssignedRegDep() &&
571 IsReachable(TargetSU, I->getSUnit()))
576 /// IsReachable - Checks if SU is reachable from TargetSU.
577 bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU,
578 const SUnit *TargetSU) {
579 // If insertion of the edge SU->TargetSU would create a cycle
580 // then there is a path from TargetSU to SU.
581 int UpperBound, LowerBound;
582 LowerBound = Node2Index[TargetSU->NodeNum];
583 UpperBound = Node2Index[SU->NodeNum];
584 bool HasLoop = false;
585 // Is Ord(TargetSU) < Ord(SU) ?
586 if (LowerBound < UpperBound) {
588 // There may be a path from TargetSU to SU. Check for it.
589 DFS(TargetSU, UpperBound, HasLoop);
594 /// Allocate - assign the topological index to the node n.
595 void ScheduleDAGTopologicalSort::Allocate(int n, int index) {
596 Node2Index[n] = index;
597 Index2Node[index] = n;
600 ScheduleDAGTopologicalSort::
601 ScheduleDAGTopologicalSort(std::vector<SUnit> &sunits) : SUnits(sunits) {}
603 ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {}