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/Support/CommandLine.h"
20 #include "llvm/Support/Debug.h"
21 #include "llvm/Support/raw_ostream.h"
22 #include "llvm/Target/TargetInstrInfo.h"
23 #include "llvm/Target/TargetMachine.h"
24 #include "llvm/Target/TargetRegisterInfo.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 /// Clear the DAG state (e.g. between scheduling regions).
50 void ScheduleDAG::clearDAG() {
56 /// getInstrDesc helper to handle SDNodes.
57 const MCInstrDesc *ScheduleDAG::getNodeDesc(const SDNode *Node) const {
58 if (!Node || !Node->isMachineOpcode()) return NULL;
59 return &TII->get(Node->getMachineOpcode());
62 /// addPred - This adds the specified edge as a pred of the current node if
63 /// not already. It also adds the current node as a successor of the
65 bool SUnit::addPred(const SDep &D, bool Required) {
66 // If this node already has this depenence, don't add a redundant one.
67 for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end();
69 // Zero-latency weak edges may be added purely for heuristic ordering. Don't
70 // add them if another kind of edge already exists.
71 if (!Required && I->getSUnit() == D.getSUnit())
74 // Extend the latency if needed. Equivalent to removePred(I) + addPred(D).
75 if (I->getLatency() < D.getLatency()) {
76 SUnit *PredSU = I->getSUnit();
77 // Find the corresponding successor in N.
79 ForwardD.setSUnit(this);
80 for (SmallVector<SDep, 4>::iterator II = PredSU->Succs.begin(),
81 EE = PredSU->Succs.end(); II != EE; ++II) {
82 if (*II == ForwardD) {
83 II->setLatency(D.getLatency());
87 I->setLatency(D.getLatency());
92 // Now add a corresponding succ to N.
95 SUnit *N = D.getSUnit();
96 // Update the bookkeeping.
97 if (D.getKind() == SDep::Data) {
98 assert(NumPreds < UINT_MAX && "NumPreds will overflow!");
99 assert(N->NumSuccs < UINT_MAX && "NumSuccs will overflow!");
103 if (!N->isScheduled) {
108 assert(NumPredsLeft < UINT_MAX && "NumPredsLeft will overflow!");
117 assert(N->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!");
122 N->Succs.push_back(P);
123 if (P.getLatency() != 0) {
124 this->setDepthDirty();
130 /// removePred - This removes the specified edge as a pred of the current
131 /// node if it exists. It also removes the current node as a successor of
132 /// the specified node.
133 void SUnit::removePred(const SDep &D) {
134 // Find the matching predecessor.
135 for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end();
138 // Find the corresponding successor in N.
141 SUnit *N = D.getSUnit();
142 SmallVectorImpl<SDep>::iterator Succ = std::find(N->Succs.begin(),
144 assert(Succ != N->Succs.end() && "Mismatching preds / succs lists!");
145 N->Succs.erase(Succ);
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) {
158 assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!");
166 assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!");
170 if (P.getLatency() != 0) {
171 this->setDepthDirty();
178 void SUnit::setDepthDirty() {
179 if (!isDepthCurrent) return;
180 SmallVector<SUnit*, 8> WorkList;
181 WorkList.push_back(this);
183 SUnit *SU = WorkList.pop_back_val();
184 SU->isDepthCurrent = false;
185 for (SUnit::const_succ_iterator I = SU->Succs.begin(),
186 E = SU->Succs.end(); I != E; ++I) {
187 SUnit *SuccSU = I->getSUnit();
188 if (SuccSU->isDepthCurrent)
189 WorkList.push_back(SuccSU);
191 } while (!WorkList.empty());
194 void SUnit::setHeightDirty() {
195 if (!isHeightCurrent) return;
196 SmallVector<SUnit*, 8> WorkList;
197 WorkList.push_back(this);
199 SUnit *SU = WorkList.pop_back_val();
200 SU->isHeightCurrent = false;
201 for (SUnit::const_pred_iterator I = SU->Preds.begin(),
202 E = SU->Preds.end(); I != E; ++I) {
203 SUnit *PredSU = I->getSUnit();
204 if (PredSU->isHeightCurrent)
205 WorkList.push_back(PredSU);
207 } while (!WorkList.empty());
210 /// setDepthToAtLeast - Update this node's successors to reflect the
211 /// fact that this node's depth just increased.
213 void SUnit::setDepthToAtLeast(unsigned NewDepth) {
214 if (NewDepth <= getDepth())
218 isDepthCurrent = true;
221 /// setHeightToAtLeast - Update this node's predecessors to reflect the
222 /// fact that this node's height just increased.
224 void SUnit::setHeightToAtLeast(unsigned NewHeight) {
225 if (NewHeight <= getHeight())
229 isHeightCurrent = true;
232 /// ComputeDepth - Calculate the maximal path from the node to the exit.
234 void SUnit::ComputeDepth() {
235 SmallVector<SUnit*, 8> WorkList;
236 WorkList.push_back(this);
238 SUnit *Cur = WorkList.back();
241 unsigned MaxPredDepth = 0;
242 for (SUnit::const_pred_iterator I = Cur->Preds.begin(),
243 E = Cur->Preds.end(); I != E; ++I) {
244 SUnit *PredSU = I->getSUnit();
245 if (PredSU->isDepthCurrent)
246 MaxPredDepth = std::max(MaxPredDepth,
247 PredSU->Depth + I->getLatency());
250 WorkList.push_back(PredSU);
256 if (MaxPredDepth != Cur->Depth) {
257 Cur->setDepthDirty();
258 Cur->Depth = MaxPredDepth;
260 Cur->isDepthCurrent = true;
262 } while (!WorkList.empty());
265 /// ComputeHeight - Calculate the maximal path from the node to the entry.
267 void SUnit::ComputeHeight() {
268 SmallVector<SUnit*, 8> WorkList;
269 WorkList.push_back(this);
271 SUnit *Cur = WorkList.back();
274 unsigned MaxSuccHeight = 0;
275 for (SUnit::const_succ_iterator I = Cur->Succs.begin(),
276 E = Cur->Succs.end(); I != E; ++I) {
277 SUnit *SuccSU = I->getSUnit();
278 if (SuccSU->isHeightCurrent)
279 MaxSuccHeight = std::max(MaxSuccHeight,
280 SuccSU->Height + I->getLatency());
283 WorkList.push_back(SuccSU);
289 if (MaxSuccHeight != Cur->Height) {
290 Cur->setHeightDirty();
291 Cur->Height = MaxSuccHeight;
293 Cur->isHeightCurrent = true;
295 } while (!WorkList.empty());
298 void SUnit::biasCriticalPath() {
302 SUnit::pred_iterator BestI = Preds.begin();
303 unsigned MaxDepth = BestI->getSUnit()->getDepth();
304 for (SUnit::pred_iterator
305 I = llvm::next(BestI), E = Preds.end(); I != E; ++I) {
306 if (I->getKind() == SDep::Data && I->getSUnit()->getDepth() > MaxDepth)
309 if (BestI != Preds.begin())
310 std::swap(*Preds.begin(), *BestI);
313 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
314 /// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
315 /// a group of nodes flagged together.
316 void SUnit::dump(const ScheduleDAG *G) const {
317 dbgs() << "SU(" << NodeNum << "): ";
321 void SUnit::dumpAll(const ScheduleDAG *G) const {
324 dbgs() << " # preds left : " << NumPredsLeft << "\n";
325 dbgs() << " # succs left : " << NumSuccsLeft << "\n";
327 dbgs() << " # weak preds left : " << WeakPredsLeft << "\n";
329 dbgs() << " # weak succs left : " << WeakSuccsLeft << "\n";
330 dbgs() << " # rdefs left : " << NumRegDefsLeft << "\n";
331 dbgs() << " Latency : " << Latency << "\n";
332 dbgs() << " Depth : " << Depth << "\n";
333 dbgs() << " Height : " << Height << "\n";
335 if (Preds.size() != 0) {
336 dbgs() << " Predecessors:\n";
337 for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
340 switch (I->getKind()) {
341 case SDep::Data: dbgs() << "val "; break;
342 case SDep::Anti: dbgs() << "anti"; break;
343 case SDep::Output: dbgs() << "out "; break;
344 case SDep::Order: dbgs() << "ch "; break;
346 dbgs() << "SU(" << I->getSUnit()->NodeNum << ")";
347 if (I->isArtificial())
349 dbgs() << ": Latency=" << I->getLatency();
350 if (I->isAssignedRegDep())
351 dbgs() << " Reg=" << PrintReg(I->getReg(), G->TRI);
355 if (Succs.size() != 0) {
356 dbgs() << " Successors:\n";
357 for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
360 switch (I->getKind()) {
361 case SDep::Data: dbgs() << "val "; break;
362 case SDep::Anti: dbgs() << "anti"; break;
363 case SDep::Output: dbgs() << "out "; break;
364 case SDep::Order: dbgs() << "ch "; break;
366 dbgs() << "SU(" << I->getSUnit()->NodeNum << ")";
367 if (I->isArtificial())
369 dbgs() << ": Latency=" << I->getLatency();
378 /// VerifyScheduledDAG - Verify that all SUnits were scheduled and that
379 /// their state is consistent. Return the number of scheduled nodes.
381 unsigned ScheduleDAG::VerifyScheduledDAG(bool isBottomUp) {
382 bool AnyNotSched = false;
383 unsigned DeadNodes = 0;
384 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
385 if (!SUnits[i].isScheduled) {
386 if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) {
391 dbgs() << "*** Scheduling failed! ***\n";
392 SUnits[i].dump(this);
393 dbgs() << "has not been scheduled!\n";
396 if (SUnits[i].isScheduled &&
397 (isBottomUp ? SUnits[i].getHeight() : SUnits[i].getDepth()) >
400 dbgs() << "*** Scheduling failed! ***\n";
401 SUnits[i].dump(this);
402 dbgs() << "has an unexpected "
403 << (isBottomUp ? "Height" : "Depth") << " value!\n";
407 if (SUnits[i].NumSuccsLeft != 0) {
409 dbgs() << "*** Scheduling failed! ***\n";
410 SUnits[i].dump(this);
411 dbgs() << "has successors left!\n";
415 if (SUnits[i].NumPredsLeft != 0) {
417 dbgs() << "*** Scheduling failed! ***\n";
418 SUnits[i].dump(this);
419 dbgs() << "has predecessors left!\n";
424 assert(!AnyNotSched);
425 return SUnits.size() - DeadNodes;
429 /// InitDAGTopologicalSorting - create the initial topological
430 /// ordering from the DAG to be scheduled.
432 /// The idea of the algorithm is taken from
433 /// "Online algorithms for managing the topological order of
434 /// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly
435 /// This is the MNR algorithm, which was first introduced by
436 /// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in
437 /// "Maintaining a topological order under edge insertions".
439 /// Short description of the algorithm:
441 /// Topological ordering, ord, of a DAG maps each node to a topological
442 /// index so that for all edges X->Y it is the case that ord(X) < ord(Y).
444 /// This means that if there is a path from the node X to the node Z,
445 /// then ord(X) < ord(Z).
447 /// This property can be used to check for reachability of nodes:
448 /// if Z is reachable from X, then an insertion of the edge Z->X would
451 /// The algorithm first computes a topological ordering for the DAG by
452 /// initializing the Index2Node and Node2Index arrays and then tries to keep
453 /// the ordering up-to-date after edge insertions by reordering the DAG.
455 /// On insertion of the edge X->Y, the algorithm first marks by calling DFS
456 /// the nodes reachable from Y, and then shifts them using Shift to lie
457 /// immediately after X in Index2Node.
458 void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() {
459 unsigned DAGSize = SUnits.size();
460 std::vector<SUnit*> WorkList;
461 WorkList.reserve(DAGSize);
463 Index2Node.resize(DAGSize);
464 Node2Index.resize(DAGSize);
466 // Initialize the data structures.
468 WorkList.push_back(ExitSU);
469 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
470 SUnit *SU = &SUnits[i];
471 int NodeNum = SU->NodeNum;
472 unsigned Degree = SU->Succs.size();
473 // Temporarily use the Node2Index array as scratch space for degree counts.
474 Node2Index[NodeNum] = Degree;
476 // Is it a node without dependencies?
478 assert(SU->Succs.empty() && "SUnit should have no successors");
479 // Collect leaf nodes.
480 WorkList.push_back(SU);
485 while (!WorkList.empty()) {
486 SUnit *SU = WorkList.back();
488 if (SU->NodeNum < DAGSize)
489 Allocate(SU->NodeNum, --Id);
490 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
492 SUnit *SU = I->getSUnit();
493 if (SU->NodeNum < DAGSize && !--Node2Index[SU->NodeNum])
494 // If all dependencies of the node are processed already,
495 // then the node can be computed now.
496 WorkList.push_back(SU);
500 Visited.resize(DAGSize);
503 // Check correctness of the ordering
504 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
505 SUnit *SU = &SUnits[i];
506 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
508 assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] &&
509 "Wrong topological sorting");
515 /// AddPred - Updates the topological ordering to accommodate an edge
516 /// to be added from SUnit X to SUnit Y.
517 void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) {
518 int UpperBound, LowerBound;
519 LowerBound = Node2Index[Y->NodeNum];
520 UpperBound = Node2Index[X->NodeNum];
521 bool HasLoop = false;
522 // Is Ord(X) < Ord(Y) ?
523 if (LowerBound < UpperBound) {
524 // Update the topological order.
526 DFS(Y, UpperBound, HasLoop);
527 assert(!HasLoop && "Inserted edge creates a loop!");
528 // Recompute topological indexes.
529 Shift(Visited, LowerBound, UpperBound);
533 /// RemovePred - Updates the topological ordering to accommodate an
534 /// an edge to be removed from the specified node N from the predecessors
535 /// of the current node M.
536 void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) {
537 // InitDAGTopologicalSorting();
540 /// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark
541 /// all nodes affected by the edge insertion. These nodes will later get new
542 /// topological indexes by means of the Shift method.
543 void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound,
545 std::vector<const SUnit*> WorkList;
546 WorkList.reserve(SUnits.size());
548 WorkList.push_back(SU);
550 SU = WorkList.back();
552 Visited.set(SU->NodeNum);
553 for (int I = SU->Succs.size()-1; I >= 0; --I) {
554 unsigned s = SU->Succs[I].getSUnit()->NodeNum;
555 // Edges to non-SUnits are allowed but ignored (e.g. ExitSU).
556 if (s >= Node2Index.size())
558 if (Node2Index[s] == UpperBound) {
562 // Visit successors if not already and in affected region.
563 if (!Visited.test(s) && Node2Index[s] < UpperBound) {
564 WorkList.push_back(SU->Succs[I].getSUnit());
567 } while (!WorkList.empty());
570 /// Shift - Renumber the nodes so that the topological ordering is
572 void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound,
578 for (i = LowerBound; i <= UpperBound; ++i) {
579 // w is node at topological index i.
580 int w = Index2Node[i];
581 if (Visited.test(w)) {
587 Allocate(w, i - shift);
591 for (unsigned j = 0; j < L.size(); ++j) {
592 Allocate(L[j], i - shift);
598 /// WillCreateCycle - Returns true if adding an edge to TargetSU from SU will
599 /// create a cycle. If so, it is not safe to call AddPred(TargetSU, SU).
600 bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *TargetSU, SUnit *SU) {
601 // Is SU reachable from TargetSU via successor edges?
602 if (IsReachable(SU, TargetSU))
604 for (SUnit::pred_iterator
605 I = TargetSU->Preds.begin(), E = TargetSU->Preds.end(); I != E; ++I)
606 if (I->isAssignedRegDep() &&
607 IsReachable(SU, I->getSUnit()))
612 /// IsReachable - Checks if SU is reachable from TargetSU.
613 bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU,
614 const SUnit *TargetSU) {
615 // If insertion of the edge SU->TargetSU would create a cycle
616 // then there is a path from TargetSU to SU.
617 int UpperBound, LowerBound;
618 LowerBound = Node2Index[TargetSU->NodeNum];
619 UpperBound = Node2Index[SU->NodeNum];
620 bool HasLoop = false;
621 // Is Ord(TargetSU) < Ord(SU) ?
622 if (LowerBound < UpperBound) {
624 // There may be a path from TargetSU to SU. Check for it.
625 DFS(TargetSU, UpperBound, HasLoop);
630 /// Allocate - assign the topological index to the node n.
631 void ScheduleDAGTopologicalSort::Allocate(int n, int index) {
632 Node2Index[n] = index;
633 Index2Node[index] = n;
636 ScheduleDAGTopologicalSort::
637 ScheduleDAGTopologicalSort(std::vector<SUnit> &sunits, SUnit *exitsu)
638 : SUnits(sunits), ExitSU(exitsu) {}
640 ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {}