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(std::distance(DAG.allnodes_begin(), DAG.allnodes_end()));
487 const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
489 for (SelectionDAG::allnodes_iterator NI = DAG.allnodes_begin(),
490 E = DAG.allnodes_end(); NI != E; ++NI) {
491 if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
494 // If this node has already been processed, stop now.
495 if (SUnitMap[NI]) continue;
497 SUnit *NodeSUnit = NewSUnit(NI);
499 // See if anything is flagged to this node, if so, add them to flagged
500 // nodes. Nodes can have at most one flag input and one flag output. Flags
501 // are required the be the last operand and result of a node.
503 // Scan up, adding flagged preds to FlaggedNodes.
505 while (N->getNumOperands() &&
506 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
507 N = N->getOperand(N->getNumOperands()-1).Val;
508 NodeSUnit->FlaggedNodes.push_back(N);
509 SUnitMap[N] = NodeSUnit;
512 // Scan down, adding this node and any flagged succs to FlaggedNodes if they
513 // have a user of the flag operand.
515 while (N->getValueType(N->getNumValues()-1) == MVT::Flag) {
516 SDOperand FlagVal(N, N->getNumValues()-1);
518 // There are either zero or one users of the Flag result.
519 bool HasFlagUse = false;
520 for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
522 if (FlagVal.isOperand(*UI)) {
524 NodeSUnit->FlaggedNodes.push_back(N);
525 SUnitMap[N] = NodeSUnit;
529 if (!HasFlagUse) break;
532 // Now all flagged nodes are in FlaggedNodes and N is the bottom-most node.
535 SUnitMap[N] = NodeSUnit;
537 // Compute the latency for the node. We use the sum of the latencies for
538 // all nodes flagged together into this SUnit.
539 if (InstrItins.isEmpty()) {
540 // No latency information.
541 NodeSUnit->Latency = 1;
543 NodeSUnit->Latency = 0;
544 if (N->isTargetOpcode()) {
545 unsigned SchedClass = TII->getSchedClass(N->getTargetOpcode());
546 InstrStage *S = InstrItins.begin(SchedClass);
547 InstrStage *E = InstrItins.end(SchedClass);
549 NodeSUnit->Latency += S->Cycles;
551 for (unsigned i = 0, e = NodeSUnit->FlaggedNodes.size(); i != e; ++i) {
552 SDNode *FNode = NodeSUnit->FlaggedNodes[i];
553 if (FNode->isTargetOpcode()) {
554 unsigned SchedClass = TII->getSchedClass(FNode->getTargetOpcode());
555 InstrStage *S = InstrItins.begin(SchedClass);
556 InstrStage *E = InstrItins.end(SchedClass);
558 NodeSUnit->Latency += S->Cycles;
564 // Pass 2: add the preds, succs, etc.
565 for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
566 SUnit *SU = &SUnits[su];
567 SDNode *MainNode = SU->Node;
569 if (MainNode->isTargetOpcode() &&
570 TII->isTwoAddrInstr(MainNode->getTargetOpcode()))
571 SU->isTwoAddress = true;
573 // Find all predecessors and successors of the group.
574 // Temporarily add N to make code simpler.
575 SU->FlaggedNodes.push_back(MainNode);
577 for (unsigned n = 0, e = SU->FlaggedNodes.size(); n != e; ++n) {
578 SDNode *N = SU->FlaggedNodes[n];
580 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
581 SDNode *OpN = N->getOperand(i).Val;
582 if (isPassiveNode(OpN)) continue; // Not scheduled.
583 SUnit *OpSU = SUnitMap[OpN];
584 assert(OpSU && "Node has no SUnit!");
585 if (OpSU == SU) continue; // In the same group.
587 MVT::ValueType OpVT = N->getOperand(i).getValueType();
588 assert(OpVT != MVT::Flag && "Flagged nodes should be in same sunit!");
590 if (OpVT == 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++;
604 // Remove MainNode from FlaggedNodes again.
605 SU->FlaggedNodes.pop_back();
607 DEBUG(SU->dumpAll(&DAG));
611 /// EmitSchedule - Emit the machine code in scheduled order.
612 void ScheduleDAGList::EmitSchedule() {
613 std::map<SDNode*, unsigned> VRBaseMap;
614 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
615 if (SUnit *SU = Sequence[i]) {
616 for (unsigned j = 0, ee = SU->FlaggedNodes.size(); j != ee; j++)
617 EmitNode(SU->FlaggedNodes[j], VRBaseMap);
618 EmitNode(SU->Node, VRBaseMap);
620 // Null SUnit* is a noop.
626 /// dump - dump the schedule.
627 void ScheduleDAGList::dumpSchedule() const {
628 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
629 if (SUnit *SU = Sequence[i])
632 std::cerr << "**** NOOP ****\n";
636 /// Schedule - Schedule the DAG using list scheduling.
637 /// FIXME: Right now it only supports the burr (bottom up register reducing)
639 void ScheduleDAGList::Schedule() {
640 DEBUG(std::cerr << "********** List Scheduling **********\n");
642 // Build scheduling units.
645 PriorityQueue->initNodes(SUnits);
647 // Execute the actual scheduling loop Top-Down or Bottom-Up as appropriate.
649 ListScheduleBottomUp();
651 ListScheduleTopDown();
653 PriorityQueue->releaseState();
655 DEBUG(std::cerr << "*** Final schedule ***\n");
656 DEBUG(dumpSchedule());
657 DEBUG(std::cerr << "\n");
659 // Emit in scheduled order
663 //===----------------------------------------------------------------------===//
664 // RegReductionPriorityQueue Implementation
665 //===----------------------------------------------------------------------===//
667 // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers
668 // to reduce register pressure.
671 class RegReductionPriorityQueue;
673 /// Sorting functions for the Available queue.
674 struct ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
675 RegReductionPriorityQueue *SPQ;
676 ls_rr_sort(RegReductionPriorityQueue *spq) : SPQ(spq) {}
677 ls_rr_sort(const ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
679 bool operator()(const SUnit* left, const SUnit* right) const;
681 } // end anonymous namespace
684 class RegReductionPriorityQueue : public SchedulingPriorityQueue {
685 // SUnits - The SUnits for the current graph.
686 const std::vector<SUnit> *SUnits;
688 // SethiUllmanNumbers - The SethiUllman number for each node.
689 std::vector<int> SethiUllmanNumbers;
691 std::priority_queue<SUnit*, std::vector<SUnit*>, ls_rr_sort> Queue;
693 RegReductionPriorityQueue() : Queue(ls_rr_sort(this)) {
696 void initNodes(const std::vector<SUnit> &sunits) {
698 // Calculate node priorities.
699 CalculatePriorities();
701 void releaseState() {
703 SethiUllmanNumbers.clear();
706 unsigned getSethiUllmanNumber(unsigned NodeNum) const {
707 assert(NodeNum < SethiUllmanNumbers.size());
708 return SethiUllmanNumbers[NodeNum];
711 bool empty() const { return Queue.empty(); }
713 void push(SUnit *U) {
716 void push_all(const std::vector<SUnit *> &Nodes) {
717 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
718 Queue.push(Nodes[i]);
722 SUnit *V = Queue.top();
727 void CalculatePriorities();
728 int CalcNodePriority(const SUnit *SU);
732 bool ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
733 unsigned LeftNum = left->NodeNum;
734 unsigned RightNum = right->NodeNum;
736 int LBonus = (int)left ->isDefNUseOperand;
737 int RBonus = (int)right->isDefNUseOperand;
739 // Special tie breaker: if two nodes share a operand, the one that
740 // use it as a def&use operand is preferred.
741 if (left->isTwoAddress && !right->isTwoAddress) {
742 SDNode *DUNode = left->Node->getOperand(0).Val;
743 if (DUNode->isOperand(right->Node))
746 if (!left->isTwoAddress && right->isTwoAddress) {
747 SDNode *DUNode = right->Node->getOperand(0).Val;
748 if (DUNode->isOperand(left->Node))
752 // Priority1 is just the number of live range genned.
753 int LPriority1 = left ->NumPredsLeft - LBonus;
754 int RPriority1 = right->NumPredsLeft - RBonus;
755 int LPriority2 = SPQ->getSethiUllmanNumber(LeftNum) + LBonus;
756 int RPriority2 = SPQ->getSethiUllmanNumber(RightNum) + RBonus;
758 if (LPriority1 > RPriority1)
760 else if (LPriority1 == RPriority1)
761 if (LPriority2 < RPriority2)
763 else if (LPriority2 == RPriority2)
764 if (left->CycleBound > right->CycleBound)
771 /// CalcNodePriority - Priority is the Sethi Ullman number.
772 /// Smaller number is the higher priority.
773 int RegReductionPriorityQueue::CalcNodePriority(const SUnit *SU) {
774 int &SethiUllmanNumber = SethiUllmanNumbers[SU->NodeNum];
775 if (SethiUllmanNumber != INT_MIN)
776 return SethiUllmanNumber;
778 if (SU->Preds.size() == 0) {
779 SethiUllmanNumber = 1;
782 for (std::set<SUnit*>::const_iterator I = SU->Preds.begin(),
783 E = SU->Preds.end(); I != E; ++I) {
785 int PredSethiUllman = CalcNodePriority(PredSU);
786 if (PredSethiUllman > SethiUllmanNumber) {
787 SethiUllmanNumber = PredSethiUllman;
789 } else if (PredSethiUllman == SethiUllmanNumber)
793 if (SU->Node->getOpcode() != ISD::TokenFactor)
794 SethiUllmanNumber += Extra;
796 SethiUllmanNumber = (Extra == 1) ? 0 : Extra-1;
799 return SethiUllmanNumber;
802 /// CalculatePriorities - Calculate priorities of all scheduling units.
803 void RegReductionPriorityQueue::CalculatePriorities() {
804 SethiUllmanNumbers.assign(SUnits->size(), INT_MIN);
806 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
807 CalcNodePriority(&(*SUnits)[i]);
810 //===----------------------------------------------------------------------===//
811 // LatencyPriorityQueue Implementation
812 //===----------------------------------------------------------------------===//
814 // This is a SchedulingPriorityQueue that schedules using latency information to
815 // reduce the length of the critical path through the basic block.
818 class LatencyPriorityQueue;
820 /// Sorting functions for the Available queue.
821 struct latency_sort : public std::binary_function<SUnit*, SUnit*, bool> {
822 LatencyPriorityQueue *PQ;
823 latency_sort(LatencyPriorityQueue *pq) : PQ(pq) {}
824 latency_sort(const latency_sort &RHS) : PQ(RHS.PQ) {}
826 bool operator()(const SUnit* left, const SUnit* right) const;
828 } // end anonymous namespace
831 class LatencyPriorityQueue : public SchedulingPriorityQueue {
832 // SUnits - The SUnits for the current graph.
833 const std::vector<SUnit> *SUnits;
835 // Latencies - The latency (max of latency from this node to the bb exit)
837 std::vector<int> Latencies;
839 /// NumNodesSolelyBlocking - This vector contains, for every node in the
840 /// Queue, the number of nodes that the node is the sole unscheduled
841 /// predecessor for. This is used as a tie-breaker heuristic for better
843 std::vector<unsigned> NumNodesSolelyBlocking;
845 std::priority_queue<SUnit*, std::vector<SUnit*>, latency_sort> Queue;
847 LatencyPriorityQueue() : Queue(latency_sort(this)) {
850 void initNodes(const std::vector<SUnit> &sunits) {
852 // Calculate node priorities.
853 CalculatePriorities();
855 void releaseState() {
860 unsigned getLatency(unsigned NodeNum) const {
861 assert(NodeNum < Latencies.size());
862 return Latencies[NodeNum];
865 unsigned getNumSolelyBlockNodes(unsigned NodeNum) const {
866 assert(NodeNum < NumNodesSolelyBlocking.size());
867 return NumNodesSolelyBlocking[NodeNum];
870 bool empty() const { return Queue.empty(); }
872 virtual void push(SUnit *U) {
875 void push_impl(SUnit *U);
877 void push_all(const std::vector<SUnit *> &Nodes) {
878 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
883 SUnit *V = Queue.top();
888 // ScheduledNode - As nodes are scheduled, we look to see if there are any
889 // successor nodes that have a single unscheduled predecessor. If so, that
890 // single predecessor has a higher priority, since scheduling it will make
891 // the node available.
892 void ScheduledNode(SUnit *Node);
895 void CalculatePriorities();
896 int CalcLatency(const SUnit &SU);
897 void AdjustPriorityOfUnscheduledPreds(SUnit *SU);
899 /// RemoveFromPriorityQueue - This is a really inefficient way to remove a
900 /// node from a priority queue. We should roll our own heap to make this
901 /// better or something.
902 void RemoveFromPriorityQueue(SUnit *SU) {
903 std::vector<SUnit*> Temp;
905 assert(!Queue.empty() && "Not in queue!");
906 while (Queue.top() != SU) {
907 Temp.push_back(Queue.top());
909 assert(!Queue.empty() && "Not in queue!");
912 // Remove the node from the PQ.
915 // Add all the other nodes back.
916 for (unsigned i = 0, e = Temp.size(); i != e; ++i)
922 bool latency_sort::operator()(const SUnit *LHS, const SUnit *RHS) const {
923 unsigned LHSNum = LHS->NodeNum;
924 unsigned RHSNum = RHS->NodeNum;
926 // The most important heuristic is scheduling the critical path.
927 unsigned LHSLatency = PQ->getLatency(LHSNum);
928 unsigned RHSLatency = PQ->getLatency(RHSNum);
929 if (LHSLatency < RHSLatency) return true;
930 if (LHSLatency > RHSLatency) return false;
932 // After that, if two nodes have identical latencies, look to see if one will
933 // unblock more other nodes than the other.
934 unsigned LHSBlocked = PQ->getNumSolelyBlockNodes(LHSNum);
935 unsigned RHSBlocked = PQ->getNumSolelyBlockNodes(RHSNum);
936 if (LHSBlocked < RHSBlocked) return true;
937 if (LHSBlocked > RHSBlocked) return false;
939 // Finally, just to provide a stable ordering, use the node number as a
941 return LHSNum < RHSNum;
945 /// CalcNodePriority - Calculate the maximal path from the node to the exit.
947 int LatencyPriorityQueue::CalcLatency(const SUnit &SU) {
948 int &Latency = Latencies[SU.NodeNum];
952 int MaxSuccLatency = 0;
953 for (std::set<SUnit*>::const_iterator I = SU.Succs.begin(),
954 E = SU.Succs.end(); I != E; ++I)
955 MaxSuccLatency = std::max(MaxSuccLatency, CalcLatency(**I));
957 for (std::set<SUnit*>::const_iterator I = SU.ChainSuccs.begin(),
958 E = SU.ChainSuccs.end(); I != E; ++I)
959 MaxSuccLatency = std::max(MaxSuccLatency, CalcLatency(**I));
961 return Latency = MaxSuccLatency + SU.Latency;
964 /// CalculatePriorities - Calculate priorities of all scheduling units.
965 void LatencyPriorityQueue::CalculatePriorities() {
966 Latencies.assign(SUnits->size(), -1);
967 NumNodesSolelyBlocking.assign(SUnits->size(), 0);
969 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
970 CalcLatency((*SUnits)[i]);
973 /// getSingleUnscheduledPred - If there is exactly one unscheduled predecessor
974 /// of SU, return it, otherwise return null.
975 static SUnit *getSingleUnscheduledPred(SUnit *SU) {
976 SUnit *OnlyAvailablePred = 0;
977 for (std::set<SUnit*>::const_iterator I = SU->Preds.begin(),
978 E = SU->Preds.end(); I != E; ++I)
979 if (!(*I)->isScheduled) {
980 // We found an available, but not scheduled, predecessor. If it's the
981 // only one we have found, keep track of it... otherwise give up.
982 if (OnlyAvailablePred && OnlyAvailablePred != *I)
984 OnlyAvailablePred = *I;
987 for (std::set<SUnit*>::const_iterator I = SU->ChainSuccs.begin(),
988 E = SU->ChainSuccs.end(); I != E; ++I)
989 if (!(*I)->isScheduled) {
990 // We found an available, but not scheduled, predecessor. If it's the
991 // only one we have found, keep track of it... otherwise give up.
992 if (OnlyAvailablePred && OnlyAvailablePred != *I)
994 OnlyAvailablePred = *I;
997 return OnlyAvailablePred;
1000 void LatencyPriorityQueue::push_impl(SUnit *SU) {
1001 // Look at all of the successors of this node. Count the number of nodes that
1002 // this node is the sole unscheduled node for.
1003 unsigned NumNodesBlocking = 0;
1004 for (std::set<SUnit*>::const_iterator I = SU->Succs.begin(),
1005 E = SU->Succs.end(); I != E; ++I)
1006 if (getSingleUnscheduledPred(*I) == SU)
1009 for (std::set<SUnit*>::const_iterator I = SU->ChainSuccs.begin(),
1010 E = SU->ChainSuccs.end(); I != E; ++I)
1011 if (getSingleUnscheduledPred(*I) == SU)
1018 // ScheduledNode - As nodes are scheduled, we look to see if there are any
1019 // successor nodes that have a single unscheduled predecessor. If so, that
1020 // single predecessor has a higher priority, since scheduling it will make
1021 // the node available.
1022 void LatencyPriorityQueue::ScheduledNode(SUnit *SU) {
1023 for (std::set<SUnit*>::const_iterator I = SU->Succs.begin(),
1024 E = SU->Succs.end(); I != E; ++I)
1025 AdjustPriorityOfUnscheduledPreds(*I);
1027 for (std::set<SUnit*>::const_iterator I = SU->ChainSuccs.begin(),
1028 E = SU->ChainSuccs.end(); I != E; ++I)
1029 AdjustPriorityOfUnscheduledPreds(*I);
1032 /// AdjustPriorityOfUnscheduledPreds - One of the predecessors of SU was just
1033 /// scheduled. If SU is not itself available, then there is at least one
1034 /// predecessor node that has not been scheduled yet. If SU has exactly ONE
1035 /// unscheduled predecessor, we want to increase its priority: it getting
1036 /// scheduled will make this node available, so it is better than some other
1037 /// node of the same priority that will not make a node available.
1038 void LatencyPriorityQueue::AdjustPriorityOfUnscheduledPreds(SUnit *SU) {
1039 if (SU->isAvailable) return; // All preds scheduled.
1041 SUnit *OnlyAvailablePred = getSingleUnscheduledPred(SU);
1042 if (OnlyAvailablePred == 0 || !OnlyAvailablePred->isAvailable) return;
1044 // Okay, we found a single predecessor that is available, but not scheduled.
1045 // Since it is available, it must be in the priority queue. First remove it.
1046 RemoveFromPriorityQueue(OnlyAvailablePred);
1048 // Reinsert the node into the priority queue, which recomputes its
1049 // NumNodesSolelyBlocking value.
1050 push(OnlyAvailablePred);
1054 //===----------------------------------------------------------------------===//
1055 // Public Constructor Functions
1056 //===----------------------------------------------------------------------===//
1058 llvm::ScheduleDAG* llvm::createBURRListDAGScheduler(SelectionDAG &DAG,
1059 MachineBasicBlock *BB) {
1060 return new ScheduleDAGList(DAG, BB, DAG.getTarget(), true,
1061 new RegReductionPriorityQueue(),
1062 new HazardRecognizer());
1065 /// createTDListDAGScheduler - This creates a top-down list scheduler with the
1066 /// specified hazard recognizer.
1067 ScheduleDAG* llvm::createTDListDAGScheduler(SelectionDAG &DAG,
1068 MachineBasicBlock *BB,
1069 HazardRecognizer *HR) {
1070 return new ScheduleDAGList(DAG, BB, DAG.getTarget(), false,
1071 new LatencyPriorityQueue(),