1 //===---- ScheduleDAGList.cpp - Implement a list scheduler for isel DAG ---===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Evan Cheng and is distributed under the
6 // University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements bottom-up and top-down list schedulers, using standard
11 // algorithms. The basic approach uses a priority queue of available nodes to
12 // schedule. One at a time, nodes are taken from the priority queue (thus in
13 // priority order), checked for legality to schedule, and emitted if legal.
15 // Nodes may not be legal to schedule either due to structural hazards (e.g.
16 // pipeline or resource constraints) or because an input to the instruction has
17 // not completed execution.
19 //===----------------------------------------------------------------------===//
21 #define DEBUG_TYPE "sched"
22 #include "llvm/CodeGen/ScheduleDAG.h"
23 #include "llvm/Target/TargetMachine.h"
24 #include "llvm/Target/TargetInstrInfo.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Support/CommandLine.h"
36 Statistic<> NumNoops ("scheduler", "Number of noops inserted");
37 Statistic<> NumStalls("scheduler", "Number of pipeline stalls");
39 /// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
40 /// a group of nodes flagged together.
42 SDNode *Node; // Representative node.
43 std::vector<SDNode*> FlaggedNodes; // All nodes flagged to Node.
44 std::set<SUnit*> Preds; // All real predecessors.
45 std::set<SUnit*> ChainPreds; // All chain predecessors.
46 std::set<SUnit*> Succs; // All real successors.
47 std::set<SUnit*> ChainSuccs; // All chain successors.
48 short NumPredsLeft; // # of preds not scheduled.
49 short NumSuccsLeft; // # of succs not scheduled.
50 short NumChainPredsLeft; // # of chain preds not scheduled.
51 short NumChainSuccsLeft; // # of chain succs not scheduled.
52 bool isTwoAddress : 1; // Is a two-address instruction.
53 bool isDefNUseOperand : 1; // Is a def&use operand.
54 bool isAvailable : 1; // True once available.
55 bool isScheduled : 1; // True once scheduled.
56 unsigned short Latency; // Node latency.
57 unsigned CycleBound; // Upper/lower cycle to be scheduled at.
58 unsigned NodeNum; // Entry # of node in the node vector.
60 SUnit(SDNode *node, unsigned nodenum)
61 : Node(node), NumPredsLeft(0), NumSuccsLeft(0),
62 NumChainPredsLeft(0), NumChainSuccsLeft(0),
63 isTwoAddress(false), isDefNUseOperand(false),
64 isAvailable(false), isScheduled(false),
65 Latency(0), CycleBound(0), NodeNum(nodenum) {}
67 void dump(const SelectionDAG *G) const;
68 void dumpAll(const SelectionDAG *G) const;
72 void SUnit::dump(const SelectionDAG *G) const {
76 if (FlaggedNodes.size() != 0) {
77 for (unsigned i = 0, e = FlaggedNodes.size(); i != e; i++) {
79 FlaggedNodes[i]->dump(G);
85 void SUnit::dumpAll(const SelectionDAG *G) const {
88 std::cerr << " # preds left : " << NumPredsLeft << "\n";
89 std::cerr << " # succs left : " << NumSuccsLeft << "\n";
90 std::cerr << " # chain preds left : " << NumChainPredsLeft << "\n";
91 std::cerr << " # chain succs left : " << NumChainSuccsLeft << "\n";
92 std::cerr << " Latency : " << Latency << "\n";
94 if (Preds.size() != 0) {
95 std::cerr << " Predecessors:\n";
96 for (std::set<SUnit*>::const_iterator I = Preds.begin(),
97 E = Preds.end(); I != E; ++I) {
102 if (ChainPreds.size() != 0) {
103 std::cerr << " Chained Preds:\n";
104 for (std::set<SUnit*>::const_iterator I = ChainPreds.begin(),
105 E = ChainPreds.end(); I != E; ++I) {
110 if (Succs.size() != 0) {
111 std::cerr << " Successors:\n";
112 for (std::set<SUnit*>::const_iterator I = Succs.begin(),
113 E = Succs.end(); I != E; ++I) {
118 if (ChainSuccs.size() != 0) {
119 std::cerr << " Chained succs:\n";
120 for (std::set<SUnit*>::const_iterator I = ChainSuccs.begin(),
121 E = ChainSuccs.end(); I != E; ++I) {
129 //===----------------------------------------------------------------------===//
130 /// SchedulingPriorityQueue - This interface is used to plug different
131 /// priorities computation algorithms into the list scheduler. It implements the
132 /// interface of a standard priority queue, where nodes are inserted in
133 /// arbitrary order and returned in priority order. The computation of the
134 /// priority and the representation of the queue are totally up to the
135 /// implementation to decide.
138 class SchedulingPriorityQueue {
140 virtual ~SchedulingPriorityQueue() {}
142 virtual void initNodes(const std::vector<SUnit> &SUnits) = 0;
143 virtual void releaseState() = 0;
145 virtual bool empty() const = 0;
146 virtual void push(SUnit *U) = 0;
148 virtual void push_all(const std::vector<SUnit *> &Nodes) = 0;
149 virtual SUnit *pop() = 0;
151 /// ScheduledNode - As each node is scheduled, this method is invoked. This
152 /// allows the priority function to adjust the priority of node that have
153 /// already been emitted.
154 virtual void ScheduledNode(SUnit *Node) {}
161 //===----------------------------------------------------------------------===//
162 /// ScheduleDAGList - The actual list scheduler implementation. This supports
163 /// both top-down and bottom-up scheduling.
165 class ScheduleDAGList : public ScheduleDAG {
167 // SDNode to SUnit mapping (many to one).
168 std::map<SDNode*, SUnit*> SUnitMap;
169 // The schedule. Null SUnit*'s represent noop instructions.
170 std::vector<SUnit*> Sequence;
171 // Current scheduling cycle.
174 // The scheduling units.
175 std::vector<SUnit> SUnits;
177 /// isBottomUp - This is true if the scheduling problem is bottom-up, false if
181 /// PriorityQueue - The priority queue to use.
182 SchedulingPriorityQueue *PriorityQueue;
184 /// HazardRec - The hazard recognizer to use.
185 HazardRecognizer *HazardRec;
188 ScheduleDAGList(SelectionDAG &dag, MachineBasicBlock *bb,
189 const TargetMachine &tm, bool isbottomup,
190 SchedulingPriorityQueue *priorityqueue,
191 HazardRecognizer *HR)
192 : ScheduleDAG(listSchedulingBURR, dag, bb, tm),
193 CurrCycle(0), isBottomUp(isbottomup),
194 PriorityQueue(priorityqueue), HazardRec(HR) {
199 delete PriorityQueue;
204 void dumpSchedule() const;
207 SUnit *NewSUnit(SDNode *N);
208 void ReleasePred(SUnit *PredSU, bool isChain = false);
209 void ReleaseSucc(SUnit *SuccSU, bool isChain = false);
210 void ScheduleNodeBottomUp(SUnit *SU);
211 void ScheduleNodeTopDown(SUnit *SU);
212 void ListScheduleTopDown();
213 void ListScheduleBottomUp();
214 void BuildSchedUnits();
217 } // end anonymous namespace
219 HazardRecognizer::~HazardRecognizer() {}
222 /// NewSUnit - Creates a new SUnit and return a ptr to it.
223 SUnit *ScheduleDAGList::NewSUnit(SDNode *N) {
224 SUnits.push_back(SUnit(N, SUnits.size()));
225 return &SUnits.back();
228 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
229 /// the Available queue is the count reaches zero. Also update its cycle bound.
230 void ScheduleDAGList::ReleasePred(SUnit *PredSU, bool isChain) {
231 // FIXME: the distance between two nodes is not always == the predecessor's
232 // latency. For example, the reader can very well read the register written
233 // by the predecessor later than the issue cycle. It also depends on the
234 // interrupt model (drain vs. freeze).
235 PredSU->CycleBound = std::max(PredSU->CycleBound,CurrCycle + PredSU->Latency);
238 PredSU->NumSuccsLeft--;
240 PredSU->NumChainSuccsLeft--;
243 if (PredSU->NumSuccsLeft < 0 || PredSU->NumChainSuccsLeft < 0) {
244 std::cerr << "*** List scheduling failed! ***\n";
246 std::cerr << " has been released too many times!\n";
251 if ((PredSU->NumSuccsLeft + PredSU->NumChainSuccsLeft) == 0) {
252 // EntryToken has to go last! Special case it here.
253 if (PredSU->Node->getOpcode() != ISD::EntryToken) {
254 PredSU->isAvailable = true;
255 PriorityQueue->push(PredSU);
260 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
261 /// the Available queue is the count reaches zero. Also update its cycle bound.
262 void ScheduleDAGList::ReleaseSucc(SUnit *SuccSU, bool isChain) {
263 // FIXME: the distance between two nodes is not always == the predecessor's
264 // latency. For example, the reader can very well read the register written
265 // by the predecessor later than the issue cycle. It also depends on the
266 // interrupt model (drain vs. freeze).
267 SuccSU->CycleBound = std::max(SuccSU->CycleBound,CurrCycle + SuccSU->Latency);
270 SuccSU->NumPredsLeft--;
272 SuccSU->NumChainPredsLeft--;
275 if (SuccSU->NumPredsLeft < 0 || SuccSU->NumChainPredsLeft < 0) {
276 std::cerr << "*** List scheduling failed! ***\n";
278 std::cerr << " has been released too many times!\n";
283 if ((SuccSU->NumPredsLeft + SuccSU->NumChainPredsLeft) == 0) {
284 SuccSU->isAvailable = true;
285 PriorityQueue->push(SuccSU);
289 /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
290 /// count of its predecessors. If a predecessor pending count is zero, add it to
291 /// the Available queue.
292 void ScheduleDAGList::ScheduleNodeBottomUp(SUnit *SU) {
293 DEBUG(std::cerr << "*** Scheduling: ");
294 DEBUG(SU->dump(&DAG));
296 Sequence.push_back(SU);
298 // Bottom up: release predecessors
299 for (std::set<SUnit*>::iterator I1 = SU->Preds.begin(),
300 E1 = SU->Preds.end(); I1 != E1; ++I1) {
304 for (std::set<SUnit*>::iterator I2 = SU->ChainPreds.begin(),
305 E2 = SU->ChainPreds.end(); I2 != E2; ++I2)
306 ReleasePred(*I2, true);
311 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
312 /// count of its successors. If a successor pending count is zero, add it to
313 /// the Available queue.
314 void ScheduleDAGList::ScheduleNodeTopDown(SUnit *SU) {
315 DEBUG(std::cerr << "*** Scheduling: ");
316 DEBUG(SU->dump(&DAG));
318 Sequence.push_back(SU);
320 // Bottom up: release successors.
321 for (std::set<SUnit*>::iterator I1 = SU->Succs.begin(),
322 E1 = SU->Succs.end(); I1 != E1; ++I1) {
326 for (std::set<SUnit*>::iterator I2 = SU->ChainSuccs.begin(),
327 E2 = SU->ChainSuccs.end(); I2 != E2; ++I2)
328 ReleaseSucc(*I2, true);
333 /// isReady - True if node's lower cycle bound is less or equal to the current
334 /// scheduling cycle. Always true if all nodes have uniform latency 1.
335 static inline bool isReady(SUnit *SU, unsigned CurrCycle) {
336 return SU->CycleBound <= CurrCycle;
339 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
341 void ScheduleDAGList::ListScheduleBottomUp() {
342 // Add root to Available queue.
343 PriorityQueue->push(SUnitMap[DAG.getRoot().Val]);
345 // While Available queue is not empty, grab the node with the highest
346 // priority. If it is not ready put it back. Schedule the node.
347 std::vector<SUnit*> NotReady;
348 while (!PriorityQueue->empty()) {
349 SUnit *CurrNode = PriorityQueue->pop();
351 while (!isReady(CurrNode, CurrCycle)) {
352 NotReady.push_back(CurrNode);
353 CurrNode = PriorityQueue->pop();
356 // Add the nodes that aren't ready back onto the available list.
357 PriorityQueue->push_all(NotReady);
360 ScheduleNodeBottomUp(CurrNode);
361 CurrNode->isScheduled = true;
362 PriorityQueue->ScheduledNode(CurrNode);
365 // Add entry node last
366 if (DAG.getEntryNode().Val != DAG.getRoot().Val) {
367 SUnit *Entry = SUnitMap[DAG.getEntryNode().Val];
368 Sequence.push_back(Entry);
371 // Reverse the order if it is bottom up.
372 std::reverse(Sequence.begin(), Sequence.end());
376 // Verify that all SUnits were scheduled.
377 bool AnyNotSched = false;
378 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
379 if (SUnits[i].NumSuccsLeft != 0 || SUnits[i].NumChainSuccsLeft != 0) {
381 std::cerr << "*** List scheduling failed! ***\n";
382 SUnits[i].dump(&DAG);
383 std::cerr << "has not been scheduled!\n";
387 assert(!AnyNotSched);
391 /// ListScheduleTopDown - The main loop of list scheduling for top-down
393 void ScheduleDAGList::ListScheduleTopDown() {
394 // Emit the entry node first.
395 SUnit *Entry = SUnitMap[DAG.getEntryNode().Val];
396 ScheduleNodeTopDown(Entry);
397 HazardRec->EmitInstruction(Entry->Node);
399 // All leaves to Available queue.
400 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
401 // It is available if it has no predecessors.
402 if ((SUnits[i].Preds.size() + SUnits[i].ChainPreds.size()) == 0 &&
404 PriorityQueue->push(&SUnits[i]);
407 // While Available queue is not empty, grab the node with the highest
408 // priority. If it is not ready put it back. Schedule the node.
409 std::vector<SUnit*> NotReady;
410 while (!PriorityQueue->empty()) {
411 SUnit *FoundNode = 0;
413 bool HasNoopHazards = false;
415 SUnit *CurNode = PriorityQueue->pop();
417 // Get the node represented by this SUnit.
418 SDNode *N = CurNode->Node;
419 // If this is a pseudo op, like copyfromreg, look to see if there is a
420 // real target node flagged to it. If so, use the target node.
421 for (unsigned i = 0, e = CurNode->FlaggedNodes.size();
422 N->getOpcode() < ISD::BUILTIN_OP_END && i != e; ++i)
423 N = CurNode->FlaggedNodes[i];
425 HazardRecognizer::HazardType HT = HazardRec->getHazardType(N);
426 if (HT == HazardRecognizer::NoHazard) {
431 // Remember if this is a noop hazard.
432 HasNoopHazards |= HT == HazardRecognizer::NoopHazard;
434 NotReady.push_back(CurNode);
435 } while (!PriorityQueue->empty());
437 // Add the nodes that aren't ready back onto the available list.
438 PriorityQueue->push_all(NotReady);
441 // If we found a node to schedule, do it now.
443 ScheduleNodeTopDown(FoundNode);
444 HazardRec->EmitInstruction(FoundNode->Node);
445 FoundNode->isScheduled = true;
446 PriorityQueue->ScheduledNode(FoundNode);
447 } else if (!HasNoopHazards) {
448 // Otherwise, we have a pipeline stall, but no other problem, just advance
449 // the current cycle and try again.
450 DEBUG(std::cerr << "*** Advancing cycle, no work to do\n");
451 HazardRec->AdvanceCycle();
454 // Otherwise, we have no instructions to issue and we have instructions
455 // that will fault if we don't do this right. This is the case for
456 // processors without pipeline interlocks and other cases.
457 DEBUG(std::cerr << "*** Emitting noop\n");
458 HazardRec->EmitNoop();
459 Sequence.push_back(0); // NULL SUnit* -> noop
465 // Verify that all SUnits were scheduled.
466 bool AnyNotSched = false;
467 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
468 if (SUnits[i].NumPredsLeft != 0 || SUnits[i].NumChainPredsLeft != 0) {
470 std::cerr << "*** List scheduling failed! ***\n";
471 SUnits[i].dump(&DAG);
472 std::cerr << "has not been scheduled!\n";
476 assert(!AnyNotSched);
481 void ScheduleDAGList::BuildSchedUnits() {
482 // Reserve entries in the vector for each of the SUnits we are creating. This
483 // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
485 SUnits.reserve(NodeCount);
487 const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
489 // Pass 1: create the SUnit's.
490 for (unsigned i = 0, NC = NodeCount; i < NC; i++) {
491 NodeInfo *NI = &Info[i];
492 SDNode *N = NI->Node;
493 if (isPassiveNode(N))
497 if (NI->isInGroup()) {
498 if (NI != NI->Group->getBottom()) // Bottom up, so only look at bottom
499 continue; // node of the NodeGroup
502 // Find the flagged nodes.
503 SDOperand FlagOp = N->getOperand(N->getNumOperands() - 1);
504 SDNode *Flag = FlagOp.Val;
505 unsigned ResNo = FlagOp.ResNo;
506 while (Flag->getValueType(ResNo) == MVT::Flag) {
507 NodeInfo *FNI = getNI(Flag);
508 assert(FNI->Group == NI->Group);
509 SU->FlaggedNodes.insert(SU->FlaggedNodes.begin(), Flag);
512 FlagOp = Flag->getOperand(Flag->getNumOperands() - 1);
514 ResNo = FlagOp.ResNo;
521 // Compute the latency for the node. We use the sum of the latencies for
522 // all nodes flagged together into this SUnit.
523 if (InstrItins.isEmpty()) {
524 // No latency information.
528 if (N->isTargetOpcode()) {
529 unsigned SchedClass = TII->getSchedClass(N->getTargetOpcode());
530 InstrStage *S = InstrItins.begin(SchedClass);
531 InstrStage *E = InstrItins.end(SchedClass);
533 SU->Latency += S->Cycles;
535 for (unsigned i = 0, e = SU->FlaggedNodes.size(); i != e; ++i) {
536 SDNode *FNode = SU->FlaggedNodes[i];
537 if (FNode->isTargetOpcode()) {
538 unsigned SchedClass = TII->getSchedClass(FNode->getTargetOpcode());
539 InstrStage *S = InstrItins.begin(SchedClass);
540 InstrStage *E = InstrItins.end(SchedClass);
542 SU->Latency += S->Cycles;
548 // Pass 2: add the preds, succs, etc.
549 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
550 SUnit *SU = &SUnits[i];
551 SDNode *N = SU->Node;
552 NodeInfo *NI = getNI(N);
554 if (N->isTargetOpcode() && TII->isTwoAddrInstr(N->getTargetOpcode()))
555 SU->isTwoAddress = true;
557 if (NI->isInGroup()) {
558 // Find all predecessors (of the group).
559 NodeGroupOpIterator NGOI(NI);
560 while (!NGOI.isEnd()) {
561 SDOperand Op = NGOI.next();
562 SDNode *OpN = Op.Val;
563 MVT::ValueType VT = OpN->getValueType(Op.ResNo);
564 NodeInfo *OpNI = getNI(OpN);
565 if (OpNI->Group != NI->Group && !isPassiveNode(OpN)) {
566 assert(VT != MVT::Flag);
567 SUnit *OpSU = SUnitMap[OpN];
568 if (VT == MVT::Other) {
569 if (SU->ChainPreds.insert(OpSU).second)
570 SU->NumChainPredsLeft++;
571 if (OpSU->ChainSuccs.insert(SU).second)
572 OpSU->NumChainSuccsLeft++;
574 if (SU->Preds.insert(OpSU).second)
576 if (OpSU->Succs.insert(SU).second)
577 OpSU->NumSuccsLeft++;
582 // Find node predecessors.
583 for (unsigned j = 0, e = N->getNumOperands(); j != e; j++) {
584 SDOperand Op = N->getOperand(j);
585 SDNode *OpN = Op.Val;
586 MVT::ValueType VT = OpN->getValueType(Op.ResNo);
587 if (!isPassiveNode(OpN)) {
588 assert(VT != MVT::Flag);
589 SUnit *OpSU = SUnitMap[OpN];
590 if (VT == MVT::Other) {
591 if (SU->ChainPreds.insert(OpSU).second)
592 SU->NumChainPredsLeft++;
593 if (OpSU->ChainSuccs.insert(SU).second)
594 OpSU->NumChainSuccsLeft++;
596 if (SU->Preds.insert(OpSU).second)
598 if (OpSU->Succs.insert(SU).second)
599 OpSU->NumSuccsLeft++;
600 if (j == 0 && SU->isTwoAddress)
601 OpSU->isDefNUseOperand = true;
607 DEBUG(SU->dumpAll(&DAG));
611 /// EmitSchedule - Emit the machine code in scheduled order.
612 void ScheduleDAGList::EmitSchedule() {
613 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
614 if (SUnit *SU = Sequence[i]) {
615 for (unsigned j = 0, ee = SU->FlaggedNodes.size(); j != ee; j++) {
616 SDNode *N = SU->FlaggedNodes[j];
619 EmitNode(getNI(SU->Node));
621 // Null SUnit* is a noop.
627 /// dump - dump the schedule.
628 void ScheduleDAGList::dumpSchedule() const {
629 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
630 if (SUnit *SU = Sequence[i])
633 std::cerr << "**** NOOP ****\n";
637 /// Schedule - Schedule the DAG using list scheduling.
638 /// FIXME: Right now it only supports the burr (bottom up register reducing)
640 void ScheduleDAGList::Schedule() {
641 DEBUG(std::cerr << "********** List Scheduling **********\n");
643 // Build scheduling units.
646 PriorityQueue->initNodes(SUnits);
648 // Execute the actual scheduling loop Top-Down or Bottom-Up as appropriate.
650 ListScheduleBottomUp();
652 ListScheduleTopDown();
654 PriorityQueue->releaseState();
656 DEBUG(std::cerr << "*** Final schedule ***\n");
657 DEBUG(dumpSchedule());
658 DEBUG(std::cerr << "\n");
660 // Emit in scheduled order
664 //===----------------------------------------------------------------------===//
665 // RegReductionPriorityQueue Implementation
666 //===----------------------------------------------------------------------===//
668 // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers
669 // to reduce register pressure.
672 class RegReductionPriorityQueue;
674 /// Sorting functions for the Available queue.
675 struct ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
676 RegReductionPriorityQueue *SPQ;
677 ls_rr_sort(RegReductionPriorityQueue *spq) : SPQ(spq) {}
678 ls_rr_sort(const ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
680 bool operator()(const SUnit* left, const SUnit* right) const;
682 } // end anonymous namespace
685 class RegReductionPriorityQueue : public SchedulingPriorityQueue {
686 // SUnits - The SUnits for the current graph.
687 const std::vector<SUnit> *SUnits;
689 // SethiUllmanNumbers - The SethiUllman number for each node.
690 std::vector<int> SethiUllmanNumbers;
692 std::priority_queue<SUnit*, std::vector<SUnit*>, ls_rr_sort> Queue;
694 RegReductionPriorityQueue() : Queue(ls_rr_sort(this)) {
697 void initNodes(const std::vector<SUnit> &sunits) {
699 // Calculate node priorities.
700 CalculatePriorities();
702 void releaseState() {
704 SethiUllmanNumbers.clear();
707 unsigned getSethiUllmanNumber(unsigned NodeNum) const {
708 assert(NodeNum < SethiUllmanNumbers.size());
709 return SethiUllmanNumbers[NodeNum];
712 bool empty() const { return Queue.empty(); }
714 void push(SUnit *U) {
717 void push_all(const std::vector<SUnit *> &Nodes) {
718 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
719 Queue.push(Nodes[i]);
723 SUnit *V = Queue.top();
728 void CalculatePriorities();
729 int CalcNodePriority(const SUnit *SU);
733 bool ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
734 unsigned LeftNum = left->NodeNum;
735 unsigned RightNum = right->NodeNum;
737 int LBonus = (int)left ->isDefNUseOperand;
738 int RBonus = (int)right->isDefNUseOperand;
740 // Special tie breaker: if two nodes share a operand, the one that
741 // use it as a def&use operand is preferred.
742 if (left->isTwoAddress && !right->isTwoAddress) {
743 SDNode *DUNode = left->Node->getOperand(0).Val;
744 if (DUNode->isOperand(right->Node))
747 if (!left->isTwoAddress && right->isTwoAddress) {
748 SDNode *DUNode = right->Node->getOperand(0).Val;
749 if (DUNode->isOperand(left->Node))
753 // Priority1 is just the number of live range genned.
754 int LPriority1 = left ->NumPredsLeft - LBonus;
755 int RPriority1 = right->NumPredsLeft - RBonus;
756 int LPriority2 = SPQ->getSethiUllmanNumber(LeftNum) + LBonus;
757 int RPriority2 = SPQ->getSethiUllmanNumber(RightNum) + RBonus;
759 if (LPriority1 > RPriority1)
761 else if (LPriority1 == RPriority1)
762 if (LPriority2 < RPriority2)
764 else if (LPriority2 == RPriority2)
765 if (left->CycleBound > right->CycleBound)
772 /// CalcNodePriority - Priority is the Sethi Ullman number.
773 /// Smaller number is the higher priority.
774 int RegReductionPriorityQueue::CalcNodePriority(const SUnit *SU) {
775 int &SethiUllmanNumber = SethiUllmanNumbers[SU->NodeNum];
776 if (SethiUllmanNumber != INT_MIN)
777 return SethiUllmanNumber;
779 if (SU->Preds.size() == 0) {
780 SethiUllmanNumber = 1;
783 for (std::set<SUnit*>::const_iterator I = SU->Preds.begin(),
784 E = SU->Preds.end(); I != E; ++I) {
786 int PredSethiUllman = CalcNodePriority(PredSU);
787 if (PredSethiUllman > SethiUllmanNumber) {
788 SethiUllmanNumber = PredSethiUllman;
790 } else if (PredSethiUllman == SethiUllmanNumber)
794 if (SU->Node->getOpcode() != ISD::TokenFactor)
795 SethiUllmanNumber += Extra;
797 SethiUllmanNumber = (Extra == 1) ? 0 : Extra-1;
800 return SethiUllmanNumber;
803 /// CalculatePriorities - Calculate priorities of all scheduling units.
804 void RegReductionPriorityQueue::CalculatePriorities() {
805 SethiUllmanNumbers.assign(SUnits->size(), INT_MIN);
807 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
808 CalcNodePriority(&(*SUnits)[i]);
811 //===----------------------------------------------------------------------===//
812 // LatencyPriorityQueue Implementation
813 //===----------------------------------------------------------------------===//
815 // This is a SchedulingPriorityQueue that schedules using latency information to
816 // reduce the length of the critical path through the basic block.
819 class LatencyPriorityQueue;
821 /// Sorting functions for the Available queue.
822 struct latency_sort : public std::binary_function<SUnit*, SUnit*, bool> {
823 LatencyPriorityQueue *PQ;
824 latency_sort(LatencyPriorityQueue *pq) : PQ(pq) {}
825 latency_sort(const latency_sort &RHS) : PQ(RHS.PQ) {}
827 bool operator()(const SUnit* left, const SUnit* right) const;
829 } // end anonymous namespace
832 class LatencyPriorityQueue : public SchedulingPriorityQueue {
833 // SUnits - The SUnits for the current graph.
834 const std::vector<SUnit> *SUnits;
836 // Latencies - The latency (max of latency from this node to the bb exit)
838 std::vector<int> Latencies;
840 /// NumNodesSolelyBlocking - This vector contains, for every node in the
841 /// Queue, the number of nodes that the node is the sole unscheduled
842 /// predecessor for. This is used as a tie-breaker heuristic for better
844 std::vector<unsigned> NumNodesSolelyBlocking;
846 std::priority_queue<SUnit*, std::vector<SUnit*>, latency_sort> Queue;
848 LatencyPriorityQueue() : Queue(latency_sort(this)) {
851 void initNodes(const std::vector<SUnit> &sunits) {
853 // Calculate node priorities.
854 CalculatePriorities();
856 void releaseState() {
861 unsigned getLatency(unsigned NodeNum) const {
862 assert(NodeNum < Latencies.size());
863 return Latencies[NodeNum];
866 unsigned getNumSolelyBlockNodes(unsigned NodeNum) const {
867 assert(NodeNum < NumNodesSolelyBlocking.size());
868 return NumNodesSolelyBlocking[NodeNum];
871 bool empty() const { return Queue.empty(); }
873 virtual void push(SUnit *U) {
876 void push_impl(SUnit *U);
878 void push_all(const std::vector<SUnit *> &Nodes) {
879 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
884 SUnit *V = Queue.top();
889 // ScheduledNode - As nodes are scheduled, we look to see if there are any
890 // successor nodes that have a single unscheduled predecessor. If so, that
891 // single predecessor has a higher priority, since scheduling it will make
892 // the node available.
893 void ScheduledNode(SUnit *Node);
896 void CalculatePriorities();
897 int CalcLatency(const SUnit &SU);
898 void AdjustPriorityOfUnscheduledPreds(SUnit *SU);
900 /// RemoveFromPriorityQueue - This is a really inefficient way to remove a
901 /// node from a priority queue. We should roll our own heap to make this
902 /// better or something.
903 void RemoveFromPriorityQueue(SUnit *SU) {
904 std::vector<SUnit*> Temp;
906 assert(!Queue.empty() && "Not in queue!");
907 while (Queue.top() != SU) {
908 Temp.push_back(Queue.top());
910 assert(!Queue.empty() && "Not in queue!");
913 // Remove the node from the PQ.
916 // Add all the other nodes back.
917 for (unsigned i = 0, e = Temp.size(); i != e; ++i)
923 bool latency_sort::operator()(const SUnit *LHS, const SUnit *RHS) const {
924 unsigned LHSNum = LHS->NodeNum;
925 unsigned RHSNum = RHS->NodeNum;
927 // The most important heuristic is scheduling the critical path.
928 unsigned LHSLatency = PQ->getLatency(LHSNum);
929 unsigned RHSLatency = PQ->getLatency(RHSNum);
930 if (LHSLatency < RHSLatency) return true;
931 if (LHSLatency > RHSLatency) return false;
933 // After that, if two nodes have identical latencies, look to see if one will
934 // unblock more other nodes than the other.
935 unsigned LHSBlocked = PQ->getNumSolelyBlockNodes(LHSNum);
936 unsigned RHSBlocked = PQ->getNumSolelyBlockNodes(RHSNum);
937 if (LHSBlocked < RHSBlocked) return true;
938 if (LHSBlocked > RHSBlocked) return false;
940 // Finally, just to provide a stable ordering, use the node number as a
942 return LHSNum < RHSNum;
946 /// CalcNodePriority - Calculate the maximal path from the node to the exit.
948 int LatencyPriorityQueue::CalcLatency(const SUnit &SU) {
949 int &Latency = Latencies[SU.NodeNum];
953 int MaxSuccLatency = 0;
954 for (std::set<SUnit*>::const_iterator I = SU.Succs.begin(),
955 E = SU.Succs.end(); I != E; ++I)
956 MaxSuccLatency = std::max(MaxSuccLatency, CalcLatency(**I));
958 for (std::set<SUnit*>::const_iterator I = SU.ChainSuccs.begin(),
959 E = SU.ChainSuccs.end(); I != E; ++I)
960 MaxSuccLatency = std::max(MaxSuccLatency, CalcLatency(**I));
962 return Latency = MaxSuccLatency + SU.Latency;
965 /// CalculatePriorities - Calculate priorities of all scheduling units.
966 void LatencyPriorityQueue::CalculatePriorities() {
967 Latencies.assign(SUnits->size(), -1);
968 NumNodesSolelyBlocking.assign(SUnits->size(), 0);
970 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
971 CalcLatency((*SUnits)[i]);
974 /// getSingleUnscheduledPred - If there is exactly one unscheduled predecessor
975 /// of SU, return it, otherwise return null.
976 static SUnit *getSingleUnscheduledPred(SUnit *SU) {
977 SUnit *OnlyAvailablePred = 0;
978 for (std::set<SUnit*>::const_iterator I = SU->Preds.begin(),
979 E = SU->Preds.end(); I != E; ++I)
980 if (!(*I)->isScheduled) {
981 // We found an available, but not scheduled, predecessor. If it's the
982 // only one we have found, keep track of it... otherwise give up.
983 if (OnlyAvailablePred && OnlyAvailablePred != *I)
985 OnlyAvailablePred = *I;
988 for (std::set<SUnit*>::const_iterator I = SU->ChainSuccs.begin(),
989 E = SU->ChainSuccs.end(); I != E; ++I)
990 if (!(*I)->isScheduled) {
991 // We found an available, but not scheduled, predecessor. If it's the
992 // only one we have found, keep track of it... otherwise give up.
993 if (OnlyAvailablePred && OnlyAvailablePred != *I)
995 OnlyAvailablePred = *I;
998 return OnlyAvailablePred;
1001 void LatencyPriorityQueue::push_impl(SUnit *SU) {
1002 // Look at all of the successors of this node. Count the number of nodes that
1003 // this node is the sole unscheduled node for.
1004 unsigned NumNodesBlocking = 0;
1005 for (std::set<SUnit*>::const_iterator I = SU->Succs.begin(),
1006 E = SU->Succs.end(); I != E; ++I)
1007 if (getSingleUnscheduledPred(*I) == SU)
1010 for (std::set<SUnit*>::const_iterator I = SU->ChainSuccs.begin(),
1011 E = SU->ChainSuccs.end(); I != E; ++I)
1012 if (getSingleUnscheduledPred(*I) == SU)
1019 // ScheduledNode - As nodes are scheduled, we look to see if there are any
1020 // successor nodes that have a single unscheduled predecessor. If so, that
1021 // single predecessor has a higher priority, since scheduling it will make
1022 // the node available.
1023 void LatencyPriorityQueue::ScheduledNode(SUnit *SU) {
1024 for (std::set<SUnit*>::const_iterator I = SU->Succs.begin(),
1025 E = SU->Succs.end(); I != E; ++I)
1026 AdjustPriorityOfUnscheduledPreds(*I);
1028 for (std::set<SUnit*>::const_iterator I = SU->ChainSuccs.begin(),
1029 E = SU->ChainSuccs.end(); I != E; ++I)
1030 AdjustPriorityOfUnscheduledPreds(*I);
1033 /// AdjustPriorityOfUnscheduledPreds - One of the predecessors of SU was just
1034 /// scheduled. If SU is not itself available, then there is at least one
1035 /// predecessor node that has not been scheduled yet. If SU has exactly ONE
1036 /// unscheduled predecessor, we want to increase its priority: it getting
1037 /// scheduled will make this node available, so it is better than some other
1038 /// node of the same priority that will not make a node available.
1039 void LatencyPriorityQueue::AdjustPriorityOfUnscheduledPreds(SUnit *SU) {
1040 if (SU->isAvailable) return; // All preds scheduled.
1042 SUnit *OnlyAvailablePred = getSingleUnscheduledPred(SU);
1043 if (OnlyAvailablePred == 0 || !OnlyAvailablePred->isAvailable) return;
1045 // Okay, we found a single predecessor that is available, but not scheduled.
1046 // Since it is available, it must be in the priority queue. First remove it.
1047 RemoveFromPriorityQueue(OnlyAvailablePred);
1049 // Reinsert the node into the priority queue, which recomputes its
1050 // NumNodesSolelyBlocking value.
1051 push(OnlyAvailablePred);
1055 //===----------------------------------------------------------------------===//
1056 // Public Constructor Functions
1057 //===----------------------------------------------------------------------===//
1059 llvm::ScheduleDAG* llvm::createBURRListDAGScheduler(SelectionDAG &DAG,
1060 MachineBasicBlock *BB) {
1061 return new ScheduleDAGList(DAG, BB, DAG.getTarget(), true,
1062 new RegReductionPriorityQueue(),
1063 new HazardRecognizer());
1066 /// createTDListDAGScheduler - This creates a top-down list scheduler with the
1067 /// specified hazard recognizer.
1068 ScheduleDAG* llvm::createTDListDAGScheduler(SelectionDAG &DAG,
1069 MachineBasicBlock *BB,
1070 HazardRecognizer *HR) {
1071 return new ScheduleDAGList(DAG, BB, DAG.getTarget(), false,
1072 new LatencyPriorityQueue(),