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.
45 // Preds/Succs - The SUnits before/after us in the graph. The boolean value
46 // is true if the edge is a token chain edge, false if it is a value edge.
47 std::set<std::pair<SUnit*,bool> > Preds; // All sunit predecessors.
48 std::set<std::pair<SUnit*,bool> > Succs; // All sunit successors.
50 short NumPredsLeft; // # of preds not scheduled.
51 short NumSuccsLeft; // # of succs not scheduled.
52 short NumChainPredsLeft; // # of chain preds not scheduled.
53 short NumChainSuccsLeft; // # of chain succs not scheduled.
54 bool isTwoAddress : 1; // Is a two-address instruction.
55 bool isDefNUseOperand : 1; // Is a def&use operand.
56 bool isPending : 1; // True once pending.
57 bool isAvailable : 1; // True once available.
58 bool isScheduled : 1; // True once scheduled.
59 unsigned short Latency; // Node latency.
60 unsigned CycleBound; // Upper/lower cycle to be scheduled at.
61 unsigned Cycle; // Once scheduled, the cycle of the op.
62 unsigned NodeNum; // Entry # of node in the node vector.
64 SUnit(SDNode *node, unsigned nodenum)
65 : Node(node), NumPredsLeft(0), NumSuccsLeft(0),
66 NumChainPredsLeft(0), NumChainSuccsLeft(0),
67 isTwoAddress(false), isDefNUseOperand(false), isPending(false),
68 isAvailable(false), isScheduled(false),
69 Latency(0), CycleBound(0), Cycle(0), NodeNum(nodenum) {}
71 void dump(const SelectionDAG *G) const;
72 void dumpAll(const SelectionDAG *G) const;
76 void SUnit::dump(const SelectionDAG *G) const {
77 std::cerr << "SU(" << NodeNum << "): ";
80 if (FlaggedNodes.size() != 0) {
81 for (unsigned i = 0, e = FlaggedNodes.size(); i != e; i++) {
83 FlaggedNodes[i]->dump(G);
89 void SUnit::dumpAll(const SelectionDAG *G) const {
92 std::cerr << " # preds left : " << NumPredsLeft << "\n";
93 std::cerr << " # succs left : " << NumSuccsLeft << "\n";
94 std::cerr << " # chain preds left : " << NumChainPredsLeft << "\n";
95 std::cerr << " # chain succs left : " << NumChainSuccsLeft << "\n";
96 std::cerr << " Latency : " << Latency << "\n";
98 if (Preds.size() != 0) {
99 std::cerr << " Predecessors:\n";
100 for (std::set<std::pair<SUnit*,bool> >::const_iterator I = Preds.begin(),
101 E = Preds.end(); I != E; ++I) {
105 std::cerr << " val ";
109 if (Succs.size() != 0) {
110 std::cerr << " Successors:\n";
111 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = Succs.begin(),
112 E = Succs.end(); I != E; ++I) {
116 std::cerr << " val ";
123 //===----------------------------------------------------------------------===//
124 /// SchedulingPriorityQueue - This interface is used to plug different
125 /// priorities computation algorithms into the list scheduler. It implements the
126 /// interface of a standard priority queue, where nodes are inserted in
127 /// arbitrary order and returned in priority order. The computation of the
128 /// priority and the representation of the queue are totally up to the
129 /// implementation to decide.
132 class SchedulingPriorityQueue {
134 virtual ~SchedulingPriorityQueue() {}
136 virtual void initNodes(const std::vector<SUnit> &SUnits) = 0;
137 virtual void releaseState() = 0;
139 virtual bool empty() const = 0;
140 virtual void push(SUnit *U) = 0;
142 virtual void push_all(const std::vector<SUnit *> &Nodes) = 0;
143 virtual SUnit *pop() = 0;
145 /// ScheduledNode - As each node is scheduled, this method is invoked. This
146 /// allows the priority function to adjust the priority of node that have
147 /// already been emitted.
148 virtual void ScheduledNode(SUnit *Node) {}
155 //===----------------------------------------------------------------------===//
156 /// ScheduleDAGList - The actual list scheduler implementation. This supports
157 /// both top-down and bottom-up scheduling.
159 class ScheduleDAGList : public ScheduleDAG {
161 // SDNode to SUnit mapping (many to one).
162 std::map<SDNode*, SUnit*> SUnitMap;
163 // The schedule. Null SUnit*'s represent noop instructions.
164 std::vector<SUnit*> Sequence;
166 // The scheduling units.
167 std::vector<SUnit> SUnits;
169 /// isBottomUp - This is true if the scheduling problem is bottom-up, false if
173 /// AvailableQueue - The priority queue to use for the available SUnits.
175 SchedulingPriorityQueue *AvailableQueue;
177 /// PendingQueue - This contains all of the instructions whose operands have
178 /// been issued, but their results are not ready yet (due to the latency of
179 /// the operation). Once the operands becomes available, the instruction is
180 /// added to the AvailableQueue. This keeps track of each SUnit and the
181 /// number of cycles left to execute before the operation is available.
182 std::vector<std::pair<unsigned, SUnit*> > PendingQueue;
184 /// HazardRec - The hazard recognizer to use.
185 HazardRecognizer *HazardRec;
188 ScheduleDAGList(SelectionDAG &dag, MachineBasicBlock *bb,
189 const TargetMachine &tm, bool isbottomup,
190 SchedulingPriorityQueue *availqueue,
191 HazardRecognizer *HR)
192 : ScheduleDAG(dag, bb, tm), isBottomUp(isbottomup),
193 AvailableQueue(availqueue), HazardRec(HR) {
198 delete AvailableQueue;
203 void dumpSchedule() const;
206 SUnit *NewSUnit(SDNode *N);
207 void ReleasePred(SUnit *PredSU, bool isChain, unsigned CurCycle);
208 void ReleaseSucc(SUnit *SuccSU, bool isChain);
209 void ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle);
210 void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle);
211 void ListScheduleTopDown();
212 void ListScheduleBottomUp();
213 void BuildSchedUnits();
216 } // end anonymous namespace
218 HazardRecognizer::~HazardRecognizer() {}
221 /// NewSUnit - Creates a new SUnit and return a ptr to it.
222 SUnit *ScheduleDAGList::NewSUnit(SDNode *N) {
223 SUnits.push_back(SUnit(N, SUnits.size()));
224 return &SUnits.back();
227 /// BuildSchedUnits - Build SUnits from the selection dag that we are input.
228 /// This SUnit graph is similar to the SelectionDAG, but represents flagged
229 /// together nodes with a single SUnit.
230 void ScheduleDAGList::BuildSchedUnits() {
231 // Reserve entries in the vector for each of the SUnits we are creating. This
232 // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
234 SUnits.reserve(std::distance(DAG.allnodes_begin(), DAG.allnodes_end()));
236 const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
238 for (SelectionDAG::allnodes_iterator NI = DAG.allnodes_begin(),
239 E = DAG.allnodes_end(); NI != E; ++NI) {
240 if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
243 // If this node has already been processed, stop now.
244 if (SUnitMap[NI]) continue;
246 SUnit *NodeSUnit = NewSUnit(NI);
248 // See if anything is flagged to this node, if so, add them to flagged
249 // nodes. Nodes can have at most one flag input and one flag output. Flags
250 // are required the be the last operand and result of a node.
252 // Scan up, adding flagged preds to FlaggedNodes.
254 while (N->getNumOperands() &&
255 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
256 N = N->getOperand(N->getNumOperands()-1).Val;
257 NodeSUnit->FlaggedNodes.push_back(N);
258 SUnitMap[N] = NodeSUnit;
261 // Scan down, adding this node and any flagged succs to FlaggedNodes if they
262 // have a user of the flag operand.
264 while (N->getValueType(N->getNumValues()-1) == MVT::Flag) {
265 SDOperand FlagVal(N, N->getNumValues()-1);
267 // There are either zero or one users of the Flag result.
268 bool HasFlagUse = false;
269 for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
271 if (FlagVal.isOperand(*UI)) {
273 NodeSUnit->FlaggedNodes.push_back(N);
274 SUnitMap[N] = NodeSUnit;
278 if (!HasFlagUse) break;
281 // Now all flagged nodes are in FlaggedNodes and N is the bottom-most node.
284 SUnitMap[N] = NodeSUnit;
286 // Compute the latency for the node. We use the sum of the latencies for
287 // all nodes flagged together into this SUnit.
288 if (InstrItins.isEmpty()) {
289 // No latency information.
290 NodeSUnit->Latency = 1;
292 NodeSUnit->Latency = 0;
293 if (N->isTargetOpcode()) {
294 unsigned SchedClass = TII->getSchedClass(N->getTargetOpcode());
295 InstrStage *S = InstrItins.begin(SchedClass);
296 InstrStage *E = InstrItins.end(SchedClass);
298 NodeSUnit->Latency += S->Cycles;
300 for (unsigned i = 0, e = NodeSUnit->FlaggedNodes.size(); i != e; ++i) {
301 SDNode *FNode = NodeSUnit->FlaggedNodes[i];
302 if (FNode->isTargetOpcode()) {
303 unsigned SchedClass = TII->getSchedClass(FNode->getTargetOpcode());
304 InstrStage *S = InstrItins.begin(SchedClass);
305 InstrStage *E = InstrItins.end(SchedClass);
307 NodeSUnit->Latency += S->Cycles;
313 // Pass 2: add the preds, succs, etc.
314 for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
315 SUnit *SU = &SUnits[su];
316 SDNode *MainNode = SU->Node;
318 if (MainNode->isTargetOpcode()) {
319 unsigned Opc = MainNode->getTargetOpcode();
320 if (TII->isTwoAddrInstr(Opc)) {
321 SU->isTwoAddress = true;
322 SDNode *OpN = MainNode->getOperand(0).Val;
323 SUnit *OpSU = SUnitMap[OpN];
325 OpSU->isDefNUseOperand = true;
329 // Find all predecessors and successors of the group.
330 // Temporarily add N to make code simpler.
331 SU->FlaggedNodes.push_back(MainNode);
333 for (unsigned n = 0, e = SU->FlaggedNodes.size(); n != e; ++n) {
334 SDNode *N = SU->FlaggedNodes[n];
336 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
337 SDNode *OpN = N->getOperand(i).Val;
338 if (isPassiveNode(OpN)) continue; // Not scheduled.
339 SUnit *OpSU = SUnitMap[OpN];
340 assert(OpSU && "Node has no SUnit!");
341 if (OpSU == SU) continue; // In the same group.
343 MVT::ValueType OpVT = N->getOperand(i).getValueType();
344 assert(OpVT != MVT::Flag && "Flagged nodes should be in same sunit!");
345 bool isChain = OpVT == MVT::Other;
347 if (SU->Preds.insert(std::make_pair(OpSU, isChain)).second) {
351 SU->NumChainPredsLeft++;
354 if (OpSU->Succs.insert(std::make_pair(SU, isChain)).second) {
356 OpSU->NumSuccsLeft++;
358 OpSU->NumChainSuccsLeft++;
364 // Remove MainNode from FlaggedNodes again.
365 SU->FlaggedNodes.pop_back();
368 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
369 SUnits[su].dumpAll(&DAG));
373 /// EmitSchedule - Emit the machine code in scheduled order.
374 void ScheduleDAGList::EmitSchedule() {
375 std::map<SDNode*, unsigned> VRBaseMap;
376 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
377 if (SUnit *SU = Sequence[i]) {
378 for (unsigned j = 0, ee = SU->FlaggedNodes.size(); j != ee; j++)
379 EmitNode(SU->FlaggedNodes[j], VRBaseMap);
380 EmitNode(SU->Node, VRBaseMap);
382 // Null SUnit* is a noop.
388 /// dump - dump the schedule.
389 void ScheduleDAGList::dumpSchedule() const {
390 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
391 if (SUnit *SU = Sequence[i])
394 std::cerr << "**** NOOP ****\n";
398 /// Schedule - Schedule the DAG using list scheduling.
399 void ScheduleDAGList::Schedule() {
400 DEBUG(std::cerr << "********** List Scheduling **********\n");
402 // Build scheduling units.
405 AvailableQueue->initNodes(SUnits);
407 // Execute the actual scheduling loop Top-Down or Bottom-Up as appropriate.
409 ListScheduleBottomUp();
411 ListScheduleTopDown();
413 AvailableQueue->releaseState();
415 DEBUG(std::cerr << "*** Final schedule ***\n");
416 DEBUG(dumpSchedule());
417 DEBUG(std::cerr << "\n");
419 // Emit in scheduled order
423 //===----------------------------------------------------------------------===//
424 // Bottom-Up Scheduling
425 //===----------------------------------------------------------------------===//
427 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
428 /// the Available queue is the count reaches zero. Also update its cycle bound.
429 void ScheduleDAGList::ReleasePred(SUnit *PredSU, bool isChain,
431 // FIXME: the distance between two nodes is not always == the predecessor's
432 // latency. For example, the reader can very well read the register written
433 // by the predecessor later than the issue cycle. It also depends on the
434 // interrupt model (drain vs. freeze).
435 PredSU->CycleBound = std::max(PredSU->CycleBound, CurCycle + PredSU->Latency);
438 PredSU->NumSuccsLeft--;
440 PredSU->NumChainSuccsLeft--;
443 if (PredSU->NumSuccsLeft < 0 || PredSU->NumChainSuccsLeft < 0) {
444 std::cerr << "*** List scheduling failed! ***\n";
446 std::cerr << " has been released too many times!\n";
451 if ((PredSU->NumSuccsLeft + PredSU->NumChainSuccsLeft) == 0) {
452 // EntryToken has to go last! Special case it here.
453 if (PredSU->Node->getOpcode() != ISD::EntryToken) {
454 PredSU->isAvailable = true;
455 AvailableQueue->push(PredSU);
459 /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
460 /// count of its predecessors. If a predecessor pending count is zero, add it to
461 /// the Available queue.
462 void ScheduleDAGList::ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle) {
463 DEBUG(std::cerr << "*** Scheduling [" << CurCycle << "]: ");
464 DEBUG(SU->dump(&DAG));
465 SU->Cycle = CurCycle;
467 AvailableQueue->ScheduledNode(SU);
468 Sequence.push_back(SU);
470 // Bottom up: release predecessors
471 for (std::set<std::pair<SUnit*, bool> >::iterator I = SU->Preds.begin(),
472 E = SU->Preds.end(); I != E; ++I) {
473 ReleasePred(I->first, I->second, CurCycle);
474 // FIXME: This is something used by the priority function that it should
475 // calculate directly.
479 SU->NumChainPredsLeft--;
483 /// isReady - True if node's lower cycle bound is less or equal to the current
484 /// scheduling cycle. Always true if all nodes have uniform latency 1.
485 static inline bool isReady(SUnit *SU, unsigned CurrCycle) {
486 return SU->CycleBound <= CurrCycle;
489 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
491 void ScheduleDAGList::ListScheduleBottomUp() {
492 unsigned CurrCycle = 0;
493 // Add root to Available queue.
494 AvailableQueue->push(SUnitMap[DAG.getRoot().Val]);
496 // While Available queue is not empty, grab the node with the highest
497 // priority. If it is not ready put it back. Schedule the node.
498 std::vector<SUnit*> NotReady;
499 SUnit *CurrNode = NULL;
500 while (!AvailableQueue->empty()) {
501 SUnit *CurrNode = AvailableQueue->pop();
502 while (!isReady(CurrNode, CurrCycle)) {
503 NotReady.push_back(CurrNode);
504 CurrNode = AvailableQueue->pop();
507 // Add the nodes that aren't ready back onto the available list.
508 AvailableQueue->push_all(NotReady);
511 ScheduleNodeBottomUp(CurrNode, CurrCycle);
513 CurrNode->isScheduled = true;
516 // Add entry node last
517 if (DAG.getEntryNode().Val != DAG.getRoot().Val) {
518 SUnit *Entry = SUnitMap[DAG.getEntryNode().Val];
519 Sequence.push_back(Entry);
522 // Reverse the order if it is bottom up.
523 std::reverse(Sequence.begin(), Sequence.end());
527 // Verify that all SUnits were scheduled.
528 bool AnyNotSched = false;
529 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
530 if (SUnits[i].NumSuccsLeft != 0 || SUnits[i].NumChainSuccsLeft != 0) {
532 std::cerr << "*** List scheduling failed! ***\n";
533 SUnits[i].dump(&DAG);
534 std::cerr << "has not been scheduled!\n";
538 assert(!AnyNotSched);
542 //===----------------------------------------------------------------------===//
543 // Top-Down Scheduling
544 //===----------------------------------------------------------------------===//
546 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
547 /// the PendingQueue if the count reaches zero.
548 void ScheduleDAGList::ReleaseSucc(SUnit *SuccSU, bool isChain) {
550 SuccSU->NumPredsLeft--;
552 SuccSU->NumChainPredsLeft--;
554 assert(SuccSU->NumPredsLeft >= 0 && SuccSU->NumChainPredsLeft >= 0 &&
555 "List scheduling internal error");
557 if ((SuccSU->NumPredsLeft + SuccSU->NumChainPredsLeft) == 0) {
558 // Compute how many cycles it will be before this actually becomes
559 // available. This is the max of the start time of all predecessors plus
561 unsigned AvailableCycle = 0;
562 for (std::set<std::pair<SUnit*, bool> >::iterator I = SuccSU->Preds.begin(),
563 E = SuccSU->Preds.end(); I != E; ++I) {
564 // If this is a token edge, we don't need to wait for the latency of the
565 // preceeding instruction (e.g. a long-latency load) unless there is also
566 // some other data dependence.
567 unsigned PredDoneCycle = I->first->Cycle;
569 PredDoneCycle += I->first->Latency;
570 else if (I->first->Latency)
573 AvailableCycle = std::max(AvailableCycle, PredDoneCycle);
576 PendingQueue.push_back(std::make_pair(AvailableCycle, SuccSU));
577 SuccSU->isPending = true;
581 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
582 /// count of its successors. If a successor pending count is zero, add it to
583 /// the Available queue.
584 void ScheduleDAGList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
585 DEBUG(std::cerr << "*** Scheduling [" << CurCycle << "]: ");
586 DEBUG(SU->dump(&DAG));
588 Sequence.push_back(SU);
589 SU->Cycle = CurCycle;
591 // Bottom up: release successors.
592 for (std::set<std::pair<SUnit*, bool> >::iterator I = SU->Succs.begin(),
593 E = SU->Succs.end(); I != E; ++I)
594 ReleaseSucc(I->first, I->second);
597 /// ListScheduleTopDown - The main loop of list scheduling for top-down
599 void ScheduleDAGList::ListScheduleTopDown() {
600 unsigned CurCycle = 0;
601 SUnit *Entry = SUnitMap[DAG.getEntryNode().Val];
603 // All leaves to Available queue.
604 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
605 // It is available if it has no predecessors.
606 if (SUnits[i].Preds.size() == 0 && &SUnits[i] != Entry) {
607 AvailableQueue->push(&SUnits[i]);
608 SUnits[i].isAvailable = SUnits[i].isPending = true;
612 // Emit the entry node first.
613 ScheduleNodeTopDown(Entry, CurCycle);
614 HazardRec->EmitInstruction(Entry->Node);
616 // While Available queue is not empty, grab the node with the highest
617 // priority. If it is not ready put it back. Schedule the node.
618 std::vector<SUnit*> NotReady;
619 while (!AvailableQueue->empty() || !PendingQueue.empty()) {
620 // Check to see if any of the pending instructions are ready to issue. If
621 // so, add them to the available queue.
622 for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
623 if (PendingQueue[i].first == CurCycle) {
624 AvailableQueue->push(PendingQueue[i].second);
625 PendingQueue[i].second->isAvailable = true;
626 PendingQueue[i] = PendingQueue.back();
627 PendingQueue.pop_back();
630 assert(PendingQueue[i].first > CurCycle && "Negative latency?");
634 // If there are no instructions available, don't try to issue anything, and
635 // don't advance the hazard recognizer.
636 if (AvailableQueue->empty()) {
641 SUnit *FoundSUnit = 0;
642 SDNode *FoundNode = 0;
644 bool HasNoopHazards = false;
645 while (!AvailableQueue->empty()) {
646 SUnit *CurSUnit = AvailableQueue->pop();
648 // Get the node represented by this SUnit.
649 FoundNode = CurSUnit->Node;
651 // If this is a pseudo op, like copyfromreg, look to see if there is a
652 // real target node flagged to it. If so, use the target node.
653 for (unsigned i = 0, e = CurSUnit->FlaggedNodes.size();
654 FoundNode->getOpcode() < ISD::BUILTIN_OP_END && i != e; ++i)
655 FoundNode = CurSUnit->FlaggedNodes[i];
657 HazardRecognizer::HazardType HT = HazardRec->getHazardType(FoundNode);
658 if (HT == HazardRecognizer::NoHazard) {
659 FoundSUnit = CurSUnit;
663 // Remember if this is a noop hazard.
664 HasNoopHazards |= HT == HazardRecognizer::NoopHazard;
666 NotReady.push_back(CurSUnit);
669 // Add the nodes that aren't ready back onto the available list.
670 if (!NotReady.empty()) {
671 AvailableQueue->push_all(NotReady);
675 // If we found a node to schedule, do it now.
677 ScheduleNodeTopDown(FoundSUnit, CurCycle);
678 HazardRec->EmitInstruction(FoundNode);
679 FoundSUnit->isScheduled = true;
680 AvailableQueue->ScheduledNode(FoundSUnit);
682 // If this is a pseudo-op node, we don't want to increment the current
684 if (FoundSUnit->Latency) // Don't increment CurCycle for pseudo-ops!
686 } else if (!HasNoopHazards) {
687 // Otherwise, we have a pipeline stall, but no other problem, just advance
688 // the current cycle and try again.
689 DEBUG(std::cerr << "*** Advancing cycle, no work to do\n");
690 HazardRec->AdvanceCycle();
694 // Otherwise, we have no instructions to issue and we have instructions
695 // that will fault if we don't do this right. This is the case for
696 // processors without pipeline interlocks and other cases.
697 DEBUG(std::cerr << "*** Emitting noop\n");
698 HazardRec->EmitNoop();
699 Sequence.push_back(0); // NULL SUnit* -> noop
706 // Verify that all SUnits were scheduled.
707 bool AnyNotSched = false;
708 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
709 if (SUnits[i].NumPredsLeft != 0 || SUnits[i].NumChainPredsLeft != 0) {
711 std::cerr << "*** List scheduling failed! ***\n";
712 SUnits[i].dump(&DAG);
713 std::cerr << "has not been scheduled!\n";
717 assert(!AnyNotSched);
721 //===----------------------------------------------------------------------===//
722 // RegReductionPriorityQueue Implementation
723 //===----------------------------------------------------------------------===//
725 // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers
726 // to reduce register pressure.
729 class RegReductionPriorityQueue;
731 /// Sorting functions for the Available queue.
732 struct ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
733 RegReductionPriorityQueue *SPQ;
734 ls_rr_sort(RegReductionPriorityQueue *spq) : SPQ(spq) {}
735 ls_rr_sort(const ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
737 bool operator()(const SUnit* left, const SUnit* right) const;
739 } // end anonymous namespace
742 class RegReductionPriorityQueue : public SchedulingPriorityQueue {
743 // SUnits - The SUnits for the current graph.
744 const std::vector<SUnit> *SUnits;
746 // SethiUllmanNumbers - The SethiUllman number for each node.
747 std::vector<int> SethiUllmanNumbers;
749 std::priority_queue<SUnit*, std::vector<SUnit*>, ls_rr_sort> Queue;
751 RegReductionPriorityQueue() :
752 Queue(ls_rr_sort(this)) {}
754 void initNodes(const std::vector<SUnit> &sunits) {
756 // Calculate node priorities.
757 CalculatePriorities();
759 void releaseState() {
761 SethiUllmanNumbers.clear();
764 int getSethiUllmanNumber(unsigned NodeNum) const {
765 assert(NodeNum < SethiUllmanNumbers.size());
766 return SethiUllmanNumbers[NodeNum];
769 bool empty() const { return Queue.empty(); }
771 void push(SUnit *U) {
774 void push_all(const std::vector<SUnit *> &Nodes) {
775 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
776 Queue.push(Nodes[i]);
780 SUnit *V = Queue.top();
786 void CalculatePriorities();
787 int CalcNodePriority(const SUnit *SU);
791 bool ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
792 unsigned LeftNum = left->NodeNum;
793 unsigned RightNum = right->NodeNum;
794 bool LIsTarget = left->Node->isTargetOpcode();
795 bool RIsTarget = right->Node->isTargetOpcode();
796 int LPriority = SPQ->getSethiUllmanNumber(LeftNum);
797 int RPriority = SPQ->getSethiUllmanNumber(RightNum);
798 bool LIsFloater = LIsTarget && (LPriority == 1 || LPriority == 0);
799 bool RIsFloater = RIsTarget && (RPriority == 1 || RPriority == 0);
801 // Schedule floaters (e.g. load from some constant address) and immediate use
802 // of floaters (with no other operands) just before the use.
803 if (LIsFloater && !RIsFloater)
805 else if (!LIsFloater && RIsFloater)
808 // Special tie breaker: if two nodes share a operand, the one that use it
809 // as a def&use operand is preferred.
810 if (LIsTarget && RIsTarget) {
811 if (left->isTwoAddress && !right->isTwoAddress) {
812 SDNode *DUNode = left->Node->getOperand(0).Val;
813 if (DUNode->isOperand(right->Node))
816 if (!left->isTwoAddress && right->isTwoAddress) {
817 SDNode *DUNode = right->Node->getOperand(0).Val;
818 if (DUNode->isOperand(left->Node))
823 if (LPriority < RPriority)
825 else if (LPriority == RPriority)
826 if (left->NumPredsLeft > right->NumPredsLeft)
828 else if (left->NumPredsLeft == right->NumPredsLeft)
829 if (left->CycleBound > right->CycleBound)
835 /// CalcNodePriority - Priority is the Sethi Ullman number.
836 /// Smaller number is the higher priority.
837 int RegReductionPriorityQueue::CalcNodePriority(const SUnit *SU) {
838 int &SethiUllmanNumber = SethiUllmanNumbers[SU->NodeNum];
839 if (SethiUllmanNumber != 0)
840 return SethiUllmanNumber;
842 unsigned Opc = SU->Node->getOpcode();
843 if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
844 SethiUllmanNumber = INT_MAX - 10;
845 else if (SU->NumSuccsLeft == 0)
846 // If SU does not have a use, i.e. it doesn't produce a value that would
847 // be consumed (e.g. store), then it terminates a chain of computation.
848 // Give it a small SethiUllman number so it will be scheduled right before its
849 // predecessors that it doesn't lengthen their live ranges.
850 SethiUllmanNumber = INT_MIN + 10;
851 else if (SU->NumPredsLeft == 0 && Opc != ISD::CopyFromReg)
852 SethiUllmanNumber = 1;
855 for (std::set<std::pair<SUnit*, bool> >::const_iterator
856 I = SU->Preds.begin(), E = SU->Preds.end(); I != E; ++I) {
857 if (I->second) continue; // ignore chain preds
858 SUnit *PredSU = I->first;
859 int PredSethiUllman = CalcNodePriority(PredSU);
860 if (PredSethiUllman > SethiUllmanNumber) {
861 SethiUllmanNumber = PredSethiUllman;
863 } else if (PredSethiUllman == SethiUllmanNumber && !I->second)
867 SethiUllmanNumber += Extra;
870 return SethiUllmanNumber;
873 /// CalculatePriorities - Calculate priorities of all scheduling units.
874 void RegReductionPriorityQueue::CalculatePriorities() {
875 SethiUllmanNumbers.assign(SUnits->size(), 0);
877 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
878 CalcNodePriority(&(*SUnits)[i]);
881 //===----------------------------------------------------------------------===//
882 // LatencyPriorityQueue Implementation
883 //===----------------------------------------------------------------------===//
885 // This is a SchedulingPriorityQueue that schedules using latency information to
886 // reduce the length of the critical path through the basic block.
889 class LatencyPriorityQueue;
891 /// Sorting functions for the Available queue.
892 struct latency_sort : public std::binary_function<SUnit*, SUnit*, bool> {
893 LatencyPriorityQueue *PQ;
894 latency_sort(LatencyPriorityQueue *pq) : PQ(pq) {}
895 latency_sort(const latency_sort &RHS) : PQ(RHS.PQ) {}
897 bool operator()(const SUnit* left, const SUnit* right) const;
899 } // end anonymous namespace
902 class LatencyPriorityQueue : public SchedulingPriorityQueue {
903 // SUnits - The SUnits for the current graph.
904 const std::vector<SUnit> *SUnits;
906 // Latencies - The latency (max of latency from this node to the bb exit)
908 std::vector<int> Latencies;
910 /// NumNodesSolelyBlocking - This vector contains, for every node in the
911 /// Queue, the number of nodes that the node is the sole unscheduled
912 /// predecessor for. This is used as a tie-breaker heuristic for better
914 std::vector<unsigned> NumNodesSolelyBlocking;
916 std::priority_queue<SUnit*, std::vector<SUnit*>, latency_sort> Queue;
918 LatencyPriorityQueue() : Queue(latency_sort(this)) {
921 void initNodes(const std::vector<SUnit> &sunits) {
923 // Calculate node priorities.
924 CalculatePriorities();
926 void releaseState() {
931 unsigned getLatency(unsigned NodeNum) const {
932 assert(NodeNum < Latencies.size());
933 return Latencies[NodeNum];
936 unsigned getNumSolelyBlockNodes(unsigned NodeNum) const {
937 assert(NodeNum < NumNodesSolelyBlocking.size());
938 return NumNodesSolelyBlocking[NodeNum];
941 bool empty() const { return Queue.empty(); }
943 virtual void push(SUnit *U) {
946 void push_impl(SUnit *U);
948 void push_all(const std::vector<SUnit *> &Nodes) {
949 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
954 SUnit *V = Queue.top();
959 // ScheduledNode - As nodes are scheduled, we look to see if there are any
960 // successor nodes that have a single unscheduled predecessor. If so, that
961 // single predecessor has a higher priority, since scheduling it will make
962 // the node available.
963 void ScheduledNode(SUnit *Node);
966 void CalculatePriorities();
967 int CalcLatency(const SUnit &SU);
968 void AdjustPriorityOfUnscheduledPreds(SUnit *SU);
970 /// RemoveFromPriorityQueue - This is a really inefficient way to remove a
971 /// node from a priority queue. We should roll our own heap to make this
972 /// better or something.
973 void RemoveFromPriorityQueue(SUnit *SU) {
974 std::vector<SUnit*> Temp;
976 assert(!Queue.empty() && "Not in queue!");
977 while (Queue.top() != SU) {
978 Temp.push_back(Queue.top());
980 assert(!Queue.empty() && "Not in queue!");
983 // Remove the node from the PQ.
986 // Add all the other nodes back.
987 for (unsigned i = 0, e = Temp.size(); i != e; ++i)
993 bool latency_sort::operator()(const SUnit *LHS, const SUnit *RHS) const {
994 unsigned LHSNum = LHS->NodeNum;
995 unsigned RHSNum = RHS->NodeNum;
997 // The most important heuristic is scheduling the critical path.
998 unsigned LHSLatency = PQ->getLatency(LHSNum);
999 unsigned RHSLatency = PQ->getLatency(RHSNum);
1000 if (LHSLatency < RHSLatency) return true;
1001 if (LHSLatency > RHSLatency) return false;
1003 // After that, if two nodes have identical latencies, look to see if one will
1004 // unblock more other nodes than the other.
1005 unsigned LHSBlocked = PQ->getNumSolelyBlockNodes(LHSNum);
1006 unsigned RHSBlocked = PQ->getNumSolelyBlockNodes(RHSNum);
1007 if (LHSBlocked < RHSBlocked) return true;
1008 if (LHSBlocked > RHSBlocked) return false;
1010 // Finally, just to provide a stable ordering, use the node number as a
1012 return LHSNum < RHSNum;
1016 /// CalcNodePriority - Calculate the maximal path from the node to the exit.
1018 int LatencyPriorityQueue::CalcLatency(const SUnit &SU) {
1019 int &Latency = Latencies[SU.NodeNum];
1023 int MaxSuccLatency = 0;
1024 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = SU.Succs.begin(),
1025 E = SU.Succs.end(); I != E; ++I)
1026 MaxSuccLatency = std::max(MaxSuccLatency, CalcLatency(*I->first));
1028 return Latency = MaxSuccLatency + SU.Latency;
1031 /// CalculatePriorities - Calculate priorities of all scheduling units.
1032 void LatencyPriorityQueue::CalculatePriorities() {
1033 Latencies.assign(SUnits->size(), -1);
1034 NumNodesSolelyBlocking.assign(SUnits->size(), 0);
1036 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
1037 CalcLatency((*SUnits)[i]);
1040 /// getSingleUnscheduledPred - If there is exactly one unscheduled predecessor
1041 /// of SU, return it, otherwise return null.
1042 static SUnit *getSingleUnscheduledPred(SUnit *SU) {
1043 SUnit *OnlyAvailablePred = 0;
1044 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = SU->Preds.begin(),
1045 E = SU->Preds.end(); I != E; ++I)
1046 if (!I->first->isScheduled) {
1047 // We found an available, but not scheduled, predecessor. If it's the
1048 // only one we have found, keep track of it... otherwise give up.
1049 if (OnlyAvailablePred && OnlyAvailablePred != I->first)
1051 OnlyAvailablePred = I->first;
1054 return OnlyAvailablePred;
1057 void LatencyPriorityQueue::push_impl(SUnit *SU) {
1058 // Look at all of the successors of this node. Count the number of nodes that
1059 // this node is the sole unscheduled node for.
1060 unsigned NumNodesBlocking = 0;
1061 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = SU->Succs.begin(),
1062 E = SU->Succs.end(); I != E; ++I)
1063 if (getSingleUnscheduledPred(I->first) == SU)
1065 NumNodesSolelyBlocking[SU->NodeNum] = NumNodesBlocking;
1071 // ScheduledNode - As nodes are scheduled, we look to see if there are any
1072 // successor nodes that have a single unscheduled predecessor. If so, that
1073 // single predecessor has a higher priority, since scheduling it will make
1074 // the node available.
1075 void LatencyPriorityQueue::ScheduledNode(SUnit *SU) {
1076 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = SU->Succs.begin(),
1077 E = SU->Succs.end(); I != E; ++I)
1078 AdjustPriorityOfUnscheduledPreds(I->first);
1081 /// AdjustPriorityOfUnscheduledPreds - One of the predecessors of SU was just
1082 /// scheduled. If SU is not itself available, then there is at least one
1083 /// predecessor node that has not been scheduled yet. If SU has exactly ONE
1084 /// unscheduled predecessor, we want to increase its priority: it getting
1085 /// scheduled will make this node available, so it is better than some other
1086 /// node of the same priority that will not make a node available.
1087 void LatencyPriorityQueue::AdjustPriorityOfUnscheduledPreds(SUnit *SU) {
1088 if (SU->isPending) return; // All preds scheduled.
1090 SUnit *OnlyAvailablePred = getSingleUnscheduledPred(SU);
1091 if (OnlyAvailablePred == 0 || !OnlyAvailablePred->isAvailable) return;
1093 // Okay, we found a single predecessor that is available, but not scheduled.
1094 // Since it is available, it must be in the priority queue. First remove it.
1095 RemoveFromPriorityQueue(OnlyAvailablePred);
1097 // Reinsert the node into the priority queue, which recomputes its
1098 // NumNodesSolelyBlocking value.
1099 push(OnlyAvailablePred);
1103 //===----------------------------------------------------------------------===//
1104 // Public Constructor Functions
1105 //===----------------------------------------------------------------------===//
1107 llvm::ScheduleDAG* llvm::createBURRListDAGScheduler(SelectionDAG &DAG,
1108 MachineBasicBlock *BB) {
1109 return new ScheduleDAGList(DAG, BB, DAG.getTarget(), true,
1110 new RegReductionPriorityQueue(),
1111 new HazardRecognizer());
1114 /// createTDListDAGScheduler - This creates a top-down list scheduler with the
1115 /// specified hazard recognizer.
1116 ScheduleDAG* llvm::createTDListDAGScheduler(SelectionDAG &DAG,
1117 MachineBasicBlock *BB,
1118 HazardRecognizer *HR) {
1119 return new ScheduleDAGList(DAG, BB, DAG.getTarget(), false,
1120 new LatencyPriorityQueue(),