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 /// \file Implements the ScheduleDAG class, which is a base class used by
11 /// scheduling implementation classes.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/CodeGen/ScheduleDAG.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/iterator_range.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
21 #include "llvm/CodeGen/SelectionDAGNodes.h"
22 #include "llvm/CodeGen/TargetInstrInfo.h"
23 #include "llvm/CodeGen/TargetRegisterInfo.h"
24 #include "llvm/CodeGen/TargetSubtargetInfo.h"
25 #include "llvm/Config/llvm-config.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/Compiler.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/raw_ostream.h"
39 #define DEBUG_TYPE "pre-RA-sched"
42 static cl::opt<bool> StressSchedOpt(
43 "stress-sched", cl::Hidden, cl::init(false),
44 cl::desc("Stress test instruction scheduling"));
47 void SchedulingPriorityQueue::anchor() {}
49 ScheduleDAG::ScheduleDAG(MachineFunction &mf)
50 : TM(mf.getTarget()), TII(mf.getSubtarget().getInstrInfo()),
51 TRI(mf.getSubtarget().getRegisterInfo()), MF(mf),
52 MRI(mf.getRegInfo()) {
54 StressSched = StressSchedOpt;
58 ScheduleDAG::~ScheduleDAG() = default;
60 void ScheduleDAG::clearDAG() {
66 const MCInstrDesc *ScheduleDAG::getNodeDesc(const SDNode *Node) const {
67 if (!Node || !Node->isMachineOpcode()) return nullptr;
68 return &TII->get(Node->getMachineOpcode());
72 raw_ostream &SDep::print(raw_ostream &OS, const TargetRegisterInfo *TRI) const {
74 case Data: OS << "Data"; break;
75 case Anti: OS << "Anti"; break;
76 case Output: OS << "Out "; break;
77 case Order: OS << "Ord "; break;
82 OS << " Latency=" << getLatency();
83 if (TRI && isAssignedRegDep())
84 OS << " Reg=" << printReg(getReg(), TRI);
88 OS << " Latency=" << getLatency();
91 OS << " Latency=" << getLatency();
92 switch(Contents.OrdKind) {
93 case Barrier: OS << " Barrier"; break;
95 case MustAliasMem: OS << " Memory"; break;
96 case Artificial: OS << " Artificial"; break;
97 case Weak: OS << " Weak"; break;
98 case Cluster: OS << " Cluster"; break;
106 bool SUnit::addPred(const SDep &D, bool Required) {
107 // If this node already has this dependence, don't add a redundant one.
108 for (SDep &PredDep : Preds) {
109 // Zero-latency weak edges may be added purely for heuristic ordering. Don't
110 // add them if another kind of edge already exists.
111 if (!Required && PredDep.getSUnit() == D.getSUnit())
113 if (PredDep.overlaps(D)) {
114 // Extend the latency if needed. Equivalent to
115 // removePred(PredDep) + addPred(D).
116 if (PredDep.getLatency() < D.getLatency()) {
117 SUnit *PredSU = PredDep.getSUnit();
118 // Find the corresponding successor in N.
119 SDep ForwardD = PredDep;
120 ForwardD.setSUnit(this);
121 for (SDep &SuccDep : PredSU->Succs) {
122 if (SuccDep == ForwardD) {
123 SuccDep.setLatency(D.getLatency());
127 PredDep.setLatency(D.getLatency());
132 // Now add a corresponding succ to N.
135 SUnit *N = D.getSUnit();
136 // Update the bookkeeping.
137 if (D.getKind() == SDep::Data) {
138 assert(NumPreds < std::numeric_limits<unsigned>::max() &&
139 "NumPreds will overflow!");
140 assert(N->NumSuccs < std::numeric_limits<unsigned>::max() &&
141 "NumSuccs will overflow!");
145 if (!N->isScheduled) {
150 assert(NumPredsLeft < std::numeric_limits<unsigned>::max() &&
151 "NumPredsLeft will overflow!");
160 assert(N->NumSuccsLeft < std::numeric_limits<unsigned>::max() &&
161 "NumSuccsLeft will overflow!");
166 N->Succs.push_back(P);
167 if (P.getLatency() != 0) {
168 this->setDepthDirty();
174 void SUnit::removePred(const SDep &D) {
175 // Find the matching predecessor.
176 SmallVectorImpl<SDep>::iterator I = llvm::find(Preds, D);
177 if (I == Preds.end())
179 // Find the corresponding successor in N.
182 SUnit *N = D.getSUnit();
183 SmallVectorImpl<SDep>::iterator Succ = llvm::find(N->Succs, P);
184 assert(Succ != N->Succs.end() && "Mismatching preds / succs lists!");
185 N->Succs.erase(Succ);
187 // Update the bookkeeping.
188 if (P.getKind() == SDep::Data) {
189 assert(NumPreds > 0 && "NumPreds will underflow!");
190 assert(N->NumSuccs > 0 && "NumSuccs will underflow!");
194 if (!N->isScheduled) {
198 assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!");
206 assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!");
210 if (P.getLatency() != 0) {
211 this->setDepthDirty();
216 void SUnit::setDepthDirty() {
217 if (!isDepthCurrent) return;
218 SmallVector<SUnit*, 8> WorkList;
219 WorkList.push_back(this);
221 SUnit *SU = WorkList.pop_back_val();
222 SU->isDepthCurrent = false;
223 for (SDep &SuccDep : SU->Succs) {
224 SUnit *SuccSU = SuccDep.getSUnit();
225 if (SuccSU->isDepthCurrent)
226 WorkList.push_back(SuccSU);
228 } while (!WorkList.empty());
231 void SUnit::setHeightDirty() {
232 if (!isHeightCurrent) return;
233 SmallVector<SUnit*, 8> WorkList;
234 WorkList.push_back(this);
236 SUnit *SU = WorkList.pop_back_val();
237 SU->isHeightCurrent = false;
238 for (SDep &PredDep : SU->Preds) {
239 SUnit *PredSU = PredDep.getSUnit();
240 if (PredSU->isHeightCurrent)
241 WorkList.push_back(PredSU);
243 } while (!WorkList.empty());
246 void SUnit::setDepthToAtLeast(unsigned NewDepth) {
247 if (NewDepth <= getDepth())
251 isDepthCurrent = true;
254 void SUnit::setHeightToAtLeast(unsigned NewHeight) {
255 if (NewHeight <= getHeight())
259 isHeightCurrent = true;
262 /// Calculates the maximal path from the node to the exit.
263 void SUnit::ComputeDepth() {
264 SmallVector<SUnit*, 8> WorkList;
265 WorkList.push_back(this);
267 SUnit *Cur = WorkList.back();
270 unsigned MaxPredDepth = 0;
271 for (const SDep &PredDep : Cur->Preds) {
272 SUnit *PredSU = PredDep.getSUnit();
273 if (PredSU->isDepthCurrent)
274 MaxPredDepth = std::max(MaxPredDepth,
275 PredSU->Depth + PredDep.getLatency());
278 WorkList.push_back(PredSU);
284 if (MaxPredDepth != Cur->Depth) {
285 Cur->setDepthDirty();
286 Cur->Depth = MaxPredDepth;
288 Cur->isDepthCurrent = true;
290 } while (!WorkList.empty());
293 /// Calculates the maximal path from the node to the entry.
294 void SUnit::ComputeHeight() {
295 SmallVector<SUnit*, 8> WorkList;
296 WorkList.push_back(this);
298 SUnit *Cur = WorkList.back();
301 unsigned MaxSuccHeight = 0;
302 for (const SDep &SuccDep : Cur->Succs) {
303 SUnit *SuccSU = SuccDep.getSUnit();
304 if (SuccSU->isHeightCurrent)
305 MaxSuccHeight = std::max(MaxSuccHeight,
306 SuccSU->Height + SuccDep.getLatency());
309 WorkList.push_back(SuccSU);
315 if (MaxSuccHeight != Cur->Height) {
316 Cur->setHeightDirty();
317 Cur->Height = MaxSuccHeight;
319 Cur->isHeightCurrent = true;
321 } while (!WorkList.empty());
324 void SUnit::biasCriticalPath() {
328 SUnit::pred_iterator BestI = Preds.begin();
329 unsigned MaxDepth = BestI->getSUnit()->getDepth();
330 for (SUnit::pred_iterator I = std::next(BestI), E = Preds.end(); I != E;
332 if (I->getKind() == SDep::Data && I->getSUnit()->getDepth() > MaxDepth)
335 if (BestI != Preds.begin())
336 std::swap(*Preds.begin(), *BestI);
339 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
341 raw_ostream &SUnit::print(raw_ostream &OS,
342 const SUnit *Entry, const SUnit *Exit) const {
345 else if (this == Exit)
348 OS << "SU(" << NodeNum << ")";
353 raw_ostream &SUnit::print(raw_ostream &OS, const ScheduleDAG *G) const {
354 return print(OS, &G->EntrySU, &G->ExitSU);
358 void SUnit::dump(const ScheduleDAG *G) const {
364 LLVM_DUMP_METHOD void SUnit::dumpAll(const ScheduleDAG *G) const {
367 dbgs() << " # preds left : " << NumPredsLeft << "\n";
368 dbgs() << " # succs left : " << NumSuccsLeft << "\n";
370 dbgs() << " # weak preds left : " << WeakPredsLeft << "\n";
372 dbgs() << " # weak succs left : " << WeakSuccsLeft << "\n";
373 dbgs() << " # rdefs left : " << NumRegDefsLeft << "\n";
374 dbgs() << " Latency : " << Latency << "\n";
375 dbgs() << " Depth : " << getDepth() << "\n";
376 dbgs() << " Height : " << getHeight() << "\n";
378 if (Preds.size() != 0) {
379 dbgs() << " Predecessors:\n";
380 for (const SDep &Dep : Preds) {
382 Dep.getSUnit()->print(dbgs(), G); dbgs() << ": ";
383 Dep.print(dbgs(), G->TRI); dbgs() << '\n';
386 if (Succs.size() != 0) {
387 dbgs() << " Successors:\n";
388 for (const SDep &Dep : Succs) {
390 Dep.getSUnit()->print(dbgs(), G); dbgs() << ": ";
391 Dep.print(dbgs(), G->TRI); dbgs() << '\n';
398 unsigned ScheduleDAG::VerifyScheduledDAG(bool isBottomUp) {
399 bool AnyNotSched = false;
400 unsigned DeadNodes = 0;
401 for (const SUnit &SUnit : SUnits) {
402 if (!SUnit.isScheduled) {
403 if (SUnit.NumPreds == 0 && SUnit.NumSuccs == 0) {
408 dbgs() << "*** Scheduling failed! ***\n";
410 dbgs() << "has not been scheduled!\n";
413 if (SUnit.isScheduled &&
414 (isBottomUp ? SUnit.getHeight() : SUnit.getDepth()) >
415 unsigned(std::numeric_limits<int>::max())) {
417 dbgs() << "*** Scheduling failed! ***\n";
419 dbgs() << "has an unexpected "
420 << (isBottomUp ? "Height" : "Depth") << " value!\n";
424 if (SUnit.NumSuccsLeft != 0) {
426 dbgs() << "*** Scheduling failed! ***\n";
428 dbgs() << "has successors left!\n";
432 if (SUnit.NumPredsLeft != 0) {
434 dbgs() << "*** Scheduling failed! ***\n";
436 dbgs() << "has predecessors left!\n";
441 assert(!AnyNotSched);
442 return SUnits.size() - DeadNodes;
446 void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() {
447 // The idea of the algorithm is taken from
448 // "Online algorithms for managing the topological order of
449 // a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly
450 // This is the MNR algorithm, which was first introduced by
451 // A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in
452 // "Maintaining a topological order under edge insertions".
454 // Short description of the algorithm:
456 // Topological ordering, ord, of a DAG maps each node to a topological
457 // index so that for all edges X->Y it is the case that ord(X) < ord(Y).
459 // This means that if there is a path from the node X to the node Z,
460 // then ord(X) < ord(Z).
462 // This property can be used to check for reachability of nodes:
463 // if Z is reachable from X, then an insertion of the edge Z->X would
466 // The algorithm first computes a topological ordering for the DAG by
467 // initializing the Index2Node and Node2Index arrays and then tries to keep
468 // the ordering up-to-date after edge insertions by reordering the DAG.
470 // On insertion of the edge X->Y, the algorithm first marks by calling DFS
471 // the nodes reachable from Y, and then shifts them using Shift to lie
472 // immediately after X in Index2Node.
473 unsigned DAGSize = SUnits.size();
474 std::vector<SUnit*> WorkList;
475 WorkList.reserve(DAGSize);
477 Index2Node.resize(DAGSize);
478 Node2Index.resize(DAGSize);
480 // Initialize the data structures.
482 WorkList.push_back(ExitSU);
483 for (SUnit &SU : SUnits) {
484 int NodeNum = SU.NodeNum;
485 unsigned Degree = SU.Succs.size();
486 // Temporarily use the Node2Index array as scratch space for degree counts.
487 Node2Index[NodeNum] = Degree;
489 // Is it a node without dependencies?
491 assert(SU.Succs.empty() && "SUnit should have no successors");
492 // Collect leaf nodes.
493 WorkList.push_back(&SU);
498 while (!WorkList.empty()) {
499 SUnit *SU = WorkList.back();
501 if (SU->NodeNum < DAGSize)
502 Allocate(SU->NodeNum, --Id);
503 for (const SDep &PredDep : SU->Preds) {
504 SUnit *SU = PredDep.getSUnit();
505 if (SU->NodeNum < DAGSize && !--Node2Index[SU->NodeNum])
506 // If all dependencies of the node are processed already,
507 // then the node can be computed now.
508 WorkList.push_back(SU);
512 Visited.resize(DAGSize);
515 // Check correctness of the ordering
516 for (SUnit &SU : SUnits) {
517 for (const SDep &PD : SU.Preds) {
518 assert(Node2Index[SU.NodeNum] > Node2Index[PD.getSUnit()->NodeNum] &&
519 "Wrong topological sorting");
525 void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) {
526 int UpperBound, LowerBound;
527 LowerBound = Node2Index[Y->NodeNum];
528 UpperBound = Node2Index[X->NodeNum];
529 bool HasLoop = false;
530 // Is Ord(X) < Ord(Y) ?
531 if (LowerBound < UpperBound) {
532 // Update the topological order.
534 DFS(Y, UpperBound, HasLoop);
535 assert(!HasLoop && "Inserted edge creates a loop!");
536 // Recompute topological indexes.
537 Shift(Visited, LowerBound, UpperBound);
541 void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) {
542 // InitDAGTopologicalSorting();
545 void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound,
547 std::vector<const SUnit*> WorkList;
548 WorkList.reserve(SUnits.size());
550 WorkList.push_back(SU);
552 SU = WorkList.back();
554 Visited.set(SU->NodeNum);
555 for (const SDep &SuccDep
556 : make_range(SU->Succs.rbegin(), SU->Succs.rend())) {
557 unsigned s = SuccDep.getSUnit()->NodeNum;
558 // Edges to non-SUnits are allowed but ignored (e.g. ExitSU).
559 if (s >= Node2Index.size())
561 if (Node2Index[s] == UpperBound) {
565 // Visit successors if not already and in affected region.
566 if (!Visited.test(s) && Node2Index[s] < UpperBound) {
567 WorkList.push_back(SuccDep.getSUnit());
570 } while (!WorkList.empty());
573 std::vector<int> ScheduleDAGTopologicalSort::GetSubGraph(const SUnit &StartSU,
574 const SUnit &TargetSU,
576 std::vector<const SUnit*> WorkList;
577 int LowerBound = Node2Index[StartSU.NodeNum];
578 int UpperBound = Node2Index[TargetSU.NodeNum];
580 BitVector VisitedBack;
581 std::vector<int> Nodes;
583 if (LowerBound > UpperBound) {
588 WorkList.reserve(SUnits.size());
591 // Starting from StartSU, visit all successors up
593 WorkList.push_back(&StartSU);
595 const SUnit *SU = WorkList.back();
597 for (int I = SU->Succs.size()-1; I >= 0; --I) {
598 const SUnit *Succ = SU->Succs[I].getSUnit();
599 unsigned s = Succ->NodeNum;
600 // Edges to non-SUnits are allowed but ignored (e.g. ExitSU).
601 if (Succ->isBoundaryNode())
603 if (Node2Index[s] == UpperBound) {
607 // Visit successors if not already and in affected region.
608 if (!Visited.test(s) && Node2Index[s] < UpperBound) {
610 WorkList.push_back(Succ);
613 } while (!WorkList.empty());
621 VisitedBack.resize(SUnits.size());
624 // Starting from TargetSU, visit all predecessors up
625 // to LowerBound. SUs that are visited by the two
626 // passes are added to Nodes.
627 WorkList.push_back(&TargetSU);
629 const SUnit *SU = WorkList.back();
631 for (int I = SU->Preds.size()-1; I >= 0; --I) {
632 const SUnit *Pred = SU->Preds[I].getSUnit();
633 unsigned s = Pred->NodeNum;
634 // Edges to non-SUnits are allowed but ignored (e.g. EntrySU).
635 if (Pred->isBoundaryNode())
637 if (Node2Index[s] == LowerBound) {
641 if (!VisitedBack.test(s) && Visited.test(s)) {
643 WorkList.push_back(Pred);
647 } while (!WorkList.empty());
649 assert(Found && "Error in SUnit Graph!");
654 void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound,
660 for (i = LowerBound; i <= UpperBound; ++i) {
661 // w is node at topological index i.
662 int w = Index2Node[i];
663 if (Visited.test(w)) {
669 Allocate(w, i - shift);
673 for (unsigned LI : L) {
674 Allocate(LI, i - shift);
679 bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *TargetSU, SUnit *SU) {
680 // Is SU reachable from TargetSU via successor edges?
681 if (IsReachable(SU, TargetSU))
683 for (const SDep &PredDep : TargetSU->Preds)
684 if (PredDep.isAssignedRegDep() &&
685 IsReachable(SU, PredDep.getSUnit()))
690 bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU,
691 const SUnit *TargetSU) {
692 // If insertion of the edge SU->TargetSU would create a cycle
693 // then there is a path from TargetSU to SU.
694 int UpperBound, LowerBound;
695 LowerBound = Node2Index[TargetSU->NodeNum];
696 UpperBound = Node2Index[SU->NodeNum];
697 bool HasLoop = false;
698 // Is Ord(TargetSU) < Ord(SU) ?
699 if (LowerBound < UpperBound) {
701 // There may be a path from TargetSU to SU. Check for it.
702 DFS(TargetSU, UpperBound, HasLoop);
707 void ScheduleDAGTopologicalSort::Allocate(int n, int index) {
708 Node2Index[n] = index;
709 Index2Node[index] = n;
712 ScheduleDAGTopologicalSort::
713 ScheduleDAGTopologicalSort(std::vector<SUnit> &sunits, SUnit *exitsu)
714 : SUnits(sunits), ExitSU(exitsu) {}
716 ScheduleHazardRecognizer::~ScheduleHazardRecognizer() = default;