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 isStore : 1; // Is a store.
55 bool isTwoAddress : 1; // Is a two-address instruction.
56 bool isDefNUseOperand : 1; // Is a def&use operand.
57 bool isPending : 1; // True once pending.
58 bool isAvailable : 1; // True once available.
59 bool isScheduled : 1; // True once scheduled.
60 unsigned short Latency; // Node latency.
61 unsigned CycleBound; // Upper/lower cycle to be scheduled at.
62 unsigned Cycle; // Once scheduled, the cycle of the op.
63 unsigned NodeNum; // Entry # of node in the node vector.
65 SUnit(SDNode *node, unsigned nodenum)
66 : Node(node), NumPredsLeft(0), NumSuccsLeft(0),
67 NumChainPredsLeft(0), NumChainSuccsLeft(0), isStore(false),
68 isTwoAddress(false), isDefNUseOperand(false),
69 isPending(false), isAvailable(false), isScheduled(false),
70 Latency(0), CycleBound(0), Cycle(0), NodeNum(nodenum) {}
72 void dump(const SelectionDAG *G) const;
73 void dumpAll(const SelectionDAG *G) const;
77 void SUnit::dump(const SelectionDAG *G) const {
81 if (FlaggedNodes.size() != 0) {
82 for (unsigned i = 0, e = FlaggedNodes.size(); i != e; i++) {
84 FlaggedNodes[i]->dump(G);
90 void SUnit::dumpAll(const SelectionDAG *G) const {
93 std::cerr << " # preds left : " << NumPredsLeft << "\n";
94 std::cerr << " # succs left : " << NumSuccsLeft << "\n";
95 std::cerr << " # chain preds left : " << NumChainPredsLeft << "\n";
96 std::cerr << " # chain succs left : " << NumChainSuccsLeft << "\n";
97 std::cerr << " Latency : " << Latency << "\n";
99 if (Preds.size() != 0) {
100 std::cerr << " Predecessors:\n";
101 for (std::set<std::pair<SUnit*,bool> >::const_iterator I = Preds.begin(),
102 E = Preds.end(); I != E; ++I) {
106 std::cerr << " val ";
110 if (Succs.size() != 0) {
111 std::cerr << " Successors:\n";
112 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = Succs.begin(),
113 E = Succs.end(); I != E; ++I) {
117 std::cerr << " val ";
124 //===----------------------------------------------------------------------===//
125 /// SchedulingPriorityQueue - This interface is used to plug different
126 /// priorities computation algorithms into the list scheduler. It implements the
127 /// interface of a standard priority queue, where nodes are inserted in
128 /// arbitrary order and returned in priority order. The computation of the
129 /// priority and the representation of the queue are totally up to the
130 /// implementation to decide.
133 class SchedulingPriorityQueue {
135 virtual ~SchedulingPriorityQueue() {}
137 virtual void initNodes(const std::vector<SUnit> &SUnits) = 0;
138 virtual void releaseState() = 0;
140 virtual bool empty() const = 0;
141 virtual void push(SUnit *U) = 0;
143 virtual void push_all(const std::vector<SUnit *> &Nodes) = 0;
144 virtual SUnit *pop() = 0;
146 /// ScheduledNode - As each node is scheduled, this method is invoked. This
147 /// allows the priority function to adjust the priority of node that have
148 /// already been emitted.
149 virtual void ScheduledNode(SUnit *Node) {}
156 //===----------------------------------------------------------------------===//
157 /// ScheduleDAGList - The actual list scheduler implementation. This supports
158 /// both top-down and bottom-up scheduling.
160 class ScheduleDAGList : public ScheduleDAG {
162 // SDNode to SUnit mapping (many to one).
163 std::map<SDNode*, SUnit*> SUnitMap;
164 // The schedule. Null SUnit*'s represent noop instructions.
165 std::vector<SUnit*> Sequence;
167 // The scheduling units.
168 std::vector<SUnit> SUnits;
170 /// isBottomUp - This is true if the scheduling problem is bottom-up, false if
174 /// AvailableQueue - The priority queue to use for the available SUnits.
176 SchedulingPriorityQueue *AvailableQueue;
178 /// PendingQueue - This contains all of the instructions whose operands have
179 /// been issued, but their results are not ready yet (due to the latency of
180 /// the operation). Once the operands becomes available, the instruction is
181 /// added to the AvailableQueue. This keeps track of each SUnit and the
182 /// number of cycles left to execute before the operation is available.
183 std::vector<std::pair<unsigned, SUnit*> > PendingQueue;
185 /// HazardRec - The hazard recognizer to use.
186 HazardRecognizer *HazardRec;
189 ScheduleDAGList(SelectionDAG &dag, MachineBasicBlock *bb,
190 const TargetMachine &tm, bool isbottomup,
191 SchedulingPriorityQueue *availqueue,
192 HazardRecognizer *HR)
193 : ScheduleDAG(dag, bb, tm), isBottomUp(isbottomup),
194 AvailableQueue(availqueue), HazardRec(HR) {
199 delete AvailableQueue;
204 void dumpSchedule() const;
207 SUnit *NewSUnit(SDNode *N);
208 void ReleasePred(SUnit *PredSU, bool isChain, unsigned CurCycle);
209 void ReleaseSucc(SUnit *SuccSU, bool isChain);
210 void ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle);
211 void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle);
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 /// BuildSchedUnits - Build SUnits from the selection dag that we are input.
229 /// This SUnit graph is similar to the SelectionDAG, but represents flagged
230 /// together nodes with a single SUnit.
231 void ScheduleDAGList::BuildSchedUnits() {
232 // Reserve entries in the vector for each of the SUnits we are creating. This
233 // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
235 SUnits.reserve(std::distance(DAG.allnodes_begin(), DAG.allnodes_end()));
237 const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
239 for (SelectionDAG::allnodes_iterator NI = DAG.allnodes_begin(),
240 E = DAG.allnodes_end(); NI != E; ++NI) {
241 if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
244 // If this node has already been processed, stop now.
245 if (SUnitMap[NI]) continue;
247 SUnit *NodeSUnit = NewSUnit(NI);
249 // See if anything is flagged to this node, if so, add them to flagged
250 // nodes. Nodes can have at most one flag input and one flag output. Flags
251 // are required the be the last operand and result of a node.
253 // Scan up, adding flagged preds to FlaggedNodes.
255 while (N->getNumOperands() &&
256 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
257 N = N->getOperand(N->getNumOperands()-1).Val;
258 NodeSUnit->FlaggedNodes.push_back(N);
259 SUnitMap[N] = NodeSUnit;
262 // Scan down, adding this node and any flagged succs to FlaggedNodes if they
263 // have a user of the flag operand.
265 while (N->getValueType(N->getNumValues()-1) == MVT::Flag) {
266 SDOperand FlagVal(N, N->getNumValues()-1);
268 // There are either zero or one users of the Flag result.
269 bool HasFlagUse = false;
270 for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
272 if (FlagVal.isOperand(*UI)) {
274 NodeSUnit->FlaggedNodes.push_back(N);
275 SUnitMap[N] = NodeSUnit;
279 if (!HasFlagUse) break;
282 // Now all flagged nodes are in FlaggedNodes and N is the bottom-most node.
285 SUnitMap[N] = NodeSUnit;
287 // Compute the latency for the node. We use the sum of the latencies for
288 // all nodes flagged together into this SUnit.
289 if (InstrItins.isEmpty()) {
290 // No latency information.
291 NodeSUnit->Latency = 1;
293 NodeSUnit->Latency = 0;
294 if (N->isTargetOpcode()) {
295 unsigned SchedClass = TII->getSchedClass(N->getTargetOpcode());
296 InstrStage *S = InstrItins.begin(SchedClass);
297 InstrStage *E = InstrItins.end(SchedClass);
299 NodeSUnit->Latency += S->Cycles;
301 for (unsigned i = 0, e = NodeSUnit->FlaggedNodes.size(); i != e; ++i) {
302 SDNode *FNode = NodeSUnit->FlaggedNodes[i];
303 if (FNode->isTargetOpcode()) {
304 unsigned SchedClass = TII->getSchedClass(FNode->getTargetOpcode());
305 InstrStage *S = InstrItins.begin(SchedClass);
306 InstrStage *E = InstrItins.end(SchedClass);
308 NodeSUnit->Latency += S->Cycles;
314 // Pass 2: add the preds, succs, etc.
315 for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
316 SUnit *SU = &SUnits[su];
317 SDNode *MainNode = SU->Node;
319 if (MainNode->isTargetOpcode()) {
320 unsigned Opc = MainNode->getTargetOpcode();
321 if (TII->isTwoAddrInstr(Opc))
322 SU->isTwoAddress = true;
323 if (TII->isStore(Opc))
327 // Find all predecessors and successors of the group.
328 // Temporarily add N to make code simpler.
329 SU->FlaggedNodes.push_back(MainNode);
331 for (unsigned n = 0, e = SU->FlaggedNodes.size(); n != e; ++n) {
332 SDNode *N = SU->FlaggedNodes[n];
334 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
335 SDNode *OpN = N->getOperand(i).Val;
336 if (isPassiveNode(OpN)) continue; // Not scheduled.
337 SUnit *OpSU = SUnitMap[OpN];
338 assert(OpSU && "Node has no SUnit!");
339 if (OpSU == SU) continue; // In the same group.
341 MVT::ValueType OpVT = N->getOperand(i).getValueType();
342 assert(OpVT != MVT::Flag && "Flagged nodes should be in same sunit!");
343 bool isChain = OpVT == MVT::Other;
345 if (SU->Preds.insert(std::make_pair(OpSU, isChain)).second) {
349 SU->NumChainPredsLeft++;
352 if (OpSU->Succs.insert(std::make_pair(SU, isChain)).second) {
354 OpSU->NumSuccsLeft++;
356 OpSU->NumChainSuccsLeft++;
362 // Remove MainNode from FlaggedNodes again.
363 SU->FlaggedNodes.pop_back();
366 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
367 SUnits[su].dumpAll(&DAG));
371 /// EmitSchedule - Emit the machine code in scheduled order.
372 void ScheduleDAGList::EmitSchedule() {
373 std::map<SDNode*, unsigned> VRBaseMap;
374 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
375 if (SUnit *SU = Sequence[i]) {
376 for (unsigned j = 0, ee = SU->FlaggedNodes.size(); j != ee; j++)
377 EmitNode(SU->FlaggedNodes[j], VRBaseMap);
378 EmitNode(SU->Node, VRBaseMap);
380 // Null SUnit* is a noop.
386 /// dump - dump the schedule.
387 void ScheduleDAGList::dumpSchedule() const {
388 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
389 if (SUnit *SU = Sequence[i])
392 std::cerr << "**** NOOP ****\n";
396 /// Schedule - Schedule the DAG using list scheduling.
397 void ScheduleDAGList::Schedule() {
398 DEBUG(std::cerr << "********** List Scheduling **********\n");
400 // Build scheduling units.
403 AvailableQueue->initNodes(SUnits);
405 // Execute the actual scheduling loop Top-Down or Bottom-Up as appropriate.
407 ListScheduleBottomUp();
409 ListScheduleTopDown();
411 AvailableQueue->releaseState();
413 DEBUG(std::cerr << "*** Final schedule ***\n");
414 DEBUG(dumpSchedule());
415 DEBUG(std::cerr << "\n");
417 // Emit in scheduled order
421 //===----------------------------------------------------------------------===//
422 // Bottom-Up Scheduling
423 //===----------------------------------------------------------------------===//
425 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
426 /// the Available queue is the count reaches zero. Also update its cycle bound.
427 void ScheduleDAGList::ReleasePred(SUnit *PredSU, bool isChain,
429 // FIXME: the distance between two nodes is not always == the predecessor's
430 // latency. For example, the reader can very well read the register written
431 // by the predecessor later than the issue cycle. It also depends on the
432 // interrupt model (drain vs. freeze).
433 PredSU->CycleBound = std::max(PredSU->CycleBound, CurCycle + PredSU->Latency);
436 PredSU->NumSuccsLeft--;
438 PredSU->NumChainSuccsLeft--;
441 if (PredSU->NumSuccsLeft < 0 || PredSU->NumChainSuccsLeft < 0) {
442 std::cerr << "*** List scheduling failed! ***\n";
444 std::cerr << " has been released too many times!\n";
449 if ((PredSU->NumSuccsLeft + PredSU->NumChainSuccsLeft) == 0) {
450 // EntryToken has to go last! Special case it here.
451 if (PredSU->Node->getOpcode() != ISD::EntryToken) {
452 PredSU->isAvailable = true;
453 AvailableQueue->push(PredSU);
457 /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
458 /// count of its predecessors. If a predecessor pending count is zero, add it to
459 /// the Available queue.
460 void ScheduleDAGList::ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle) {
461 DEBUG(std::cerr << "*** Scheduling [" << CurCycle << "]: ");
462 DEBUG(SU->dump(&DAG));
463 SU->Cycle = CurCycle;
465 Sequence.push_back(SU);
467 // Bottom up: release predecessors
468 for (std::set<std::pair<SUnit*, bool> >::iterator I = SU->Preds.begin(),
469 E = SU->Preds.end(); I != E; ++I) {
470 ReleasePred(I->first, I->second, CurCycle);
471 // FIXME: This is something used by the priority function that it should
472 // calculate directly.
478 /// isReady - True if node's lower cycle bound is less or equal to the current
479 /// scheduling cycle. Always true if all nodes have uniform latency 1.
480 static inline bool isReady(SUnit *SU, unsigned CurrCycle) {
481 return SU->CycleBound <= CurrCycle;
484 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
486 void ScheduleDAGList::ListScheduleBottomUp() {
487 unsigned CurrCycle = 0;
488 // Add root to Available queue.
489 AvailableQueue->push(SUnitMap[DAG.getRoot().Val]);
491 // While Available queue is not empty, grab the node with the highest
492 // priority. If it is not ready put it back. Schedule the node.
493 std::vector<SUnit*> NotReady;
494 while (!AvailableQueue->empty()) {
495 SUnit *CurrNode = AvailableQueue->pop();
497 while (!isReady(CurrNode, CurrCycle)) {
498 NotReady.push_back(CurrNode);
499 CurrNode = AvailableQueue->pop();
502 // Add the nodes that aren't ready back onto the available list.
503 AvailableQueue->push_all(NotReady);
506 ScheduleNodeBottomUp(CurrNode, CurrCycle);
508 CurrNode->isScheduled = true;
509 AvailableQueue->ScheduledNode(CurrNode);
512 // Add entry node last
513 if (DAG.getEntryNode().Val != DAG.getRoot().Val) {
514 SUnit *Entry = SUnitMap[DAG.getEntryNode().Val];
515 Sequence.push_back(Entry);
518 // Reverse the order if it is bottom up.
519 std::reverse(Sequence.begin(), Sequence.end());
523 // Verify that all SUnits were scheduled.
524 bool AnyNotSched = false;
525 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
526 if (SUnits[i].NumSuccsLeft != 0 || SUnits[i].NumChainSuccsLeft != 0) {
528 std::cerr << "*** List scheduling failed! ***\n";
529 SUnits[i].dump(&DAG);
530 std::cerr << "has not been scheduled!\n";
534 assert(!AnyNotSched);
538 //===----------------------------------------------------------------------===//
539 // Top-Down Scheduling
540 //===----------------------------------------------------------------------===//
542 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
543 /// the PendingQueue if the count reaches zero.
544 void ScheduleDAGList::ReleaseSucc(SUnit *SuccSU, bool isChain) {
546 SuccSU->NumPredsLeft--;
548 SuccSU->NumChainPredsLeft--;
550 assert(SuccSU->NumPredsLeft >= 0 && SuccSU->NumChainPredsLeft >= 0 &&
551 "List scheduling internal error");
553 if ((SuccSU->NumPredsLeft + SuccSU->NumChainPredsLeft) == 0) {
554 // Compute how many cycles it will be before this actually becomes
555 // available. This is the max of the start time of all predecessors plus
557 unsigned AvailableCycle = 0;
558 for (std::set<std::pair<SUnit*, bool> >::iterator I = SuccSU->Preds.begin(),
559 E = SuccSU->Preds.end(); I != E; ++I) {
560 // If this is a token edge, we don't need to wait for the latency of the
561 // preceeding instruction (e.g. a long-latency load) unless there is also
562 // some other data dependence.
563 unsigned PredDoneCycle = I->first->Cycle;
565 PredDoneCycle += I->first->Latency;
566 else if (I->first->Latency)
569 AvailableCycle = std::max(AvailableCycle, PredDoneCycle);
572 PendingQueue.push_back(std::make_pair(AvailableCycle, SuccSU));
573 SuccSU->isPending = true;
577 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
578 /// count of its successors. If a successor pending count is zero, add it to
579 /// the Available queue.
580 void ScheduleDAGList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
581 DEBUG(std::cerr << "*** Scheduling [" << CurCycle << "]: ");
582 DEBUG(SU->dump(&DAG));
584 Sequence.push_back(SU);
585 SU->Cycle = CurCycle;
587 // Bottom up: release successors.
588 for (std::set<std::pair<SUnit*, bool> >::iterator I = SU->Succs.begin(),
589 E = SU->Succs.end(); I != E; ++I)
590 ReleaseSucc(I->first, I->second);
593 /// ListScheduleTopDown - The main loop of list scheduling for top-down
595 void ScheduleDAGList::ListScheduleTopDown() {
596 unsigned CurCycle = 0;
597 SUnit *Entry = SUnitMap[DAG.getEntryNode().Val];
599 // All leaves to Available queue.
600 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
601 // It is available if it has no predecessors.
602 if (SUnits[i].Preds.size() == 0 && &SUnits[i] != Entry) {
603 AvailableQueue->push(&SUnits[i]);
604 SUnits[i].isAvailable = SUnits[i].isPending = true;
608 // Emit the entry node first.
609 ScheduleNodeTopDown(Entry, CurCycle);
610 HazardRec->EmitInstruction(Entry->Node);
612 // While Available queue is not empty, grab the node with the highest
613 // priority. If it is not ready put it back. Schedule the node.
614 std::vector<SUnit*> NotReady;
615 while (!AvailableQueue->empty() || !PendingQueue.empty()) {
616 // Check to see if any of the pending instructions are ready to issue. If
617 // so, add them to the available queue.
618 for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
619 if (PendingQueue[i].first == CurCycle) {
620 AvailableQueue->push(PendingQueue[i].second);
621 PendingQueue[i].second->isAvailable = true;
622 PendingQueue[i] = PendingQueue.back();
623 PendingQueue.pop_back();
626 assert(PendingQueue[i].first > CurCycle && "Negative latency?");
630 // If there are no instructions available, don't try to issue anything, and
631 // don't advance the hazard recognizer.
632 if (AvailableQueue->empty()) {
637 SUnit *FoundSUnit = 0;
638 SDNode *FoundNode = 0;
640 bool HasNoopHazards = false;
641 while (!AvailableQueue->empty()) {
642 SUnit *CurSUnit = AvailableQueue->pop();
644 // Get the node represented by this SUnit.
645 FoundNode = CurSUnit->Node;
647 // If this is a pseudo op, like copyfromreg, look to see if there is a
648 // real target node flagged to it. If so, use the target node.
649 for (unsigned i = 0, e = CurSUnit->FlaggedNodes.size();
650 FoundNode->getOpcode() < ISD::BUILTIN_OP_END && i != e; ++i)
651 FoundNode = CurSUnit->FlaggedNodes[i];
653 HazardRecognizer::HazardType HT = HazardRec->getHazardType(FoundNode);
654 if (HT == HazardRecognizer::NoHazard) {
655 FoundSUnit = CurSUnit;
659 // Remember if this is a noop hazard.
660 HasNoopHazards |= HT == HazardRecognizer::NoopHazard;
662 NotReady.push_back(CurSUnit);
665 // Add the nodes that aren't ready back onto the available list.
666 if (!NotReady.empty()) {
667 AvailableQueue->push_all(NotReady);
671 // If we found a node to schedule, do it now.
673 ScheduleNodeTopDown(FoundSUnit, CurCycle);
674 HazardRec->EmitInstruction(FoundNode);
675 FoundSUnit->isScheduled = true;
676 AvailableQueue->ScheduledNode(FoundSUnit);
678 // If this is a pseudo-op node, we don't want to increment the current
680 if (FoundSUnit->Latency) // Don't increment CurCycle for pseudo-ops!
682 } else if (!HasNoopHazards) {
683 // Otherwise, we have a pipeline stall, but no other problem, just advance
684 // the current cycle and try again.
685 DEBUG(std::cerr << "*** Advancing cycle, no work to do\n");
686 HazardRec->AdvanceCycle();
690 // Otherwise, we have no instructions to issue and we have instructions
691 // that will fault if we don't do this right. This is the case for
692 // processors without pipeline interlocks and other cases.
693 DEBUG(std::cerr << "*** Emitting noop\n");
694 HazardRec->EmitNoop();
695 Sequence.push_back(0); // NULL SUnit* -> noop
702 // Verify that all SUnits were scheduled.
703 bool AnyNotSched = false;
704 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
705 if (SUnits[i].NumPredsLeft != 0 || SUnits[i].NumChainPredsLeft != 0) {
707 std::cerr << "*** List scheduling failed! ***\n";
708 SUnits[i].dump(&DAG);
709 std::cerr << "has not been scheduled!\n";
713 assert(!AnyNotSched);
717 //===----------------------------------------------------------------------===//
718 // RegReductionPriorityQueue Implementation
719 //===----------------------------------------------------------------------===//
721 // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers
722 // to reduce register pressure.
725 class RegReductionPriorityQueue;
727 /// Sorting functions for the Available queue.
728 struct ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
729 RegReductionPriorityQueue *SPQ;
730 ls_rr_sort(RegReductionPriorityQueue *spq) : SPQ(spq) {}
731 ls_rr_sort(const ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
733 bool operator()(const SUnit* left, const SUnit* right) const;
735 } // end anonymous namespace
738 class RegReductionPriorityQueue : public SchedulingPriorityQueue {
739 // SUnits - The SUnits for the current graph.
740 const std::vector<SUnit> *SUnits;
742 // SethiUllmanNumbers - The SethiUllman number for each node.
743 std::vector<unsigned> SethiUllmanNumbers;
745 std::priority_queue<SUnit*, std::vector<SUnit*>, ls_rr_sort> Queue;
747 RegReductionPriorityQueue() : Queue(ls_rr_sort(this)) {
750 void initNodes(const std::vector<SUnit> &sunits) {
752 // Calculate node priorities.
753 CalculatePriorities();
755 void releaseState() {
757 SethiUllmanNumbers.clear();
760 unsigned getSethiUllmanNumber(unsigned NodeNum) const {
761 assert(NodeNum < SethiUllmanNumbers.size());
762 return SethiUllmanNumbers[NodeNum];
765 bool empty() const { return Queue.empty(); }
767 void push(SUnit *U) {
770 void push_all(const std::vector<SUnit *> &Nodes) {
771 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
772 Queue.push(Nodes[i]);
776 SUnit *V = Queue.top();
781 void CalculatePriorities();
782 unsigned CalcNodePriority(const SUnit *SU);
786 bool ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
787 unsigned LeftNum = left->NodeNum;
788 unsigned RightNum = right->NodeNum;
790 int LBonus = (int)left ->isDefNUseOperand;
791 int RBonus = (int)right->isDefNUseOperand;
793 // Special tie breaker: if two nodes share a operand, the one that
794 // use it as a def&use operand is preferred.
795 if (left->isTwoAddress && !right->isTwoAddress) {
796 SDNode *DUNode = left->Node->getOperand(0).Val;
797 if (DUNode->isOperand(right->Node))
800 if (!left->isTwoAddress && right->isTwoAddress) {
801 SDNode *DUNode = right->Node->getOperand(0).Val;
802 if (DUNode->isOperand(left->Node))
806 // Push stores up as much as possible. This really help code like this:
813 // This would make sure the scheduled code completed all computations and
814 // the stores before the next series of computation starts.
815 if (!left->isStore && right->isStore)
817 if (left->isStore && !right->isStore)
820 // Priority1 is just the number of live range genned.
821 int LPriority1 = left ->NumPredsLeft - LBonus;
822 int RPriority1 = right->NumPredsLeft - RBonus;
823 int LPriority2 = SPQ->getSethiUllmanNumber(LeftNum) + LBonus;
824 int RPriority2 = SPQ->getSethiUllmanNumber(RightNum) + RBonus;
826 if (LPriority1 > RPriority1)
828 else if (LPriority1 == RPriority1)
829 if (LPriority2 < RPriority2)
831 else if (LPriority2 == RPriority2)
832 if (left->CycleBound > right->CycleBound)
839 /// CalcNodePriority - Priority is the Sethi Ullman number.
840 /// Smaller number is the higher priority.
841 unsigned RegReductionPriorityQueue::CalcNodePriority(const SUnit *SU) {
842 unsigned &SethiUllmanNumber = SethiUllmanNumbers[SU->NodeNum];
843 if (SethiUllmanNumber != 0)
844 return SethiUllmanNumber;
846 if (SU->Preds.size() == 0) {
847 SethiUllmanNumber = 1;
850 for (std::set<std::pair<SUnit*, bool> >::const_iterator
851 I = SU->Preds.begin(), E = SU->Preds.end(); I != E; ++I) {
852 if (I->second) continue; // ignore chain preds.
853 SUnit *PredSU = I->first;
854 unsigned PredSethiUllman = CalcNodePriority(PredSU);
855 if (PredSethiUllman > SethiUllmanNumber) {
856 SethiUllmanNumber = PredSethiUllman;
858 } else if (PredSethiUllman == SethiUllmanNumber)
862 SethiUllmanNumber += Extra;
865 return SethiUllmanNumber;
868 /// CalculatePriorities - Calculate priorities of all scheduling units.
869 void RegReductionPriorityQueue::CalculatePriorities() {
870 SethiUllmanNumbers.assign(SUnits->size(), 0);
872 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
873 CalcNodePriority(&(*SUnits)[i]);
876 //===----------------------------------------------------------------------===//
877 // LatencyPriorityQueue Implementation
878 //===----------------------------------------------------------------------===//
880 // This is a SchedulingPriorityQueue that schedules using latency information to
881 // reduce the length of the critical path through the basic block.
884 class LatencyPriorityQueue;
886 /// Sorting functions for the Available queue.
887 struct latency_sort : public std::binary_function<SUnit*, SUnit*, bool> {
888 LatencyPriorityQueue *PQ;
889 latency_sort(LatencyPriorityQueue *pq) : PQ(pq) {}
890 latency_sort(const latency_sort &RHS) : PQ(RHS.PQ) {}
892 bool operator()(const SUnit* left, const SUnit* right) const;
894 } // end anonymous namespace
897 class LatencyPriorityQueue : public SchedulingPriorityQueue {
898 // SUnits - The SUnits for the current graph.
899 const std::vector<SUnit> *SUnits;
901 // Latencies - The latency (max of latency from this node to the bb exit)
903 std::vector<int> Latencies;
905 /// NumNodesSolelyBlocking - This vector contains, for every node in the
906 /// Queue, the number of nodes that the node is the sole unscheduled
907 /// predecessor for. This is used as a tie-breaker heuristic for better
909 std::vector<unsigned> NumNodesSolelyBlocking;
911 std::priority_queue<SUnit*, std::vector<SUnit*>, latency_sort> Queue;
913 LatencyPriorityQueue() : Queue(latency_sort(this)) {
916 void initNodes(const std::vector<SUnit> &sunits) {
918 // Calculate node priorities.
919 CalculatePriorities();
921 void releaseState() {
926 unsigned getLatency(unsigned NodeNum) const {
927 assert(NodeNum < Latencies.size());
928 return Latencies[NodeNum];
931 unsigned getNumSolelyBlockNodes(unsigned NodeNum) const {
932 assert(NodeNum < NumNodesSolelyBlocking.size());
933 return NumNodesSolelyBlocking[NodeNum];
936 bool empty() const { return Queue.empty(); }
938 virtual void push(SUnit *U) {
941 void push_impl(SUnit *U);
943 void push_all(const std::vector<SUnit *> &Nodes) {
944 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
949 SUnit *V = Queue.top();
954 // ScheduledNode - As nodes are scheduled, we look to see if there are any
955 // successor nodes that have a single unscheduled predecessor. If so, that
956 // single predecessor has a higher priority, since scheduling it will make
957 // the node available.
958 void ScheduledNode(SUnit *Node);
961 void CalculatePriorities();
962 int CalcLatency(const SUnit &SU);
963 void AdjustPriorityOfUnscheduledPreds(SUnit *SU);
965 /// RemoveFromPriorityQueue - This is a really inefficient way to remove a
966 /// node from a priority queue. We should roll our own heap to make this
967 /// better or something.
968 void RemoveFromPriorityQueue(SUnit *SU) {
969 std::vector<SUnit*> Temp;
971 assert(!Queue.empty() && "Not in queue!");
972 while (Queue.top() != SU) {
973 Temp.push_back(Queue.top());
975 assert(!Queue.empty() && "Not in queue!");
978 // Remove the node from the PQ.
981 // Add all the other nodes back.
982 for (unsigned i = 0, e = Temp.size(); i != e; ++i)
988 bool latency_sort::operator()(const SUnit *LHS, const SUnit *RHS) const {
989 unsigned LHSNum = LHS->NodeNum;
990 unsigned RHSNum = RHS->NodeNum;
992 // The most important heuristic is scheduling the critical path.
993 unsigned LHSLatency = PQ->getLatency(LHSNum);
994 unsigned RHSLatency = PQ->getLatency(RHSNum);
995 if (LHSLatency < RHSLatency) return true;
996 if (LHSLatency > RHSLatency) return false;
998 // After that, if two nodes have identical latencies, look to see if one will
999 // unblock more other nodes than the other.
1000 unsigned LHSBlocked = PQ->getNumSolelyBlockNodes(LHSNum);
1001 unsigned RHSBlocked = PQ->getNumSolelyBlockNodes(RHSNum);
1002 if (LHSBlocked < RHSBlocked) return true;
1003 if (LHSBlocked > RHSBlocked) return false;
1005 // Finally, just to provide a stable ordering, use the node number as a
1007 return LHSNum < RHSNum;
1011 /// CalcNodePriority - Calculate the maximal path from the node to the exit.
1013 int LatencyPriorityQueue::CalcLatency(const SUnit &SU) {
1014 int &Latency = Latencies[SU.NodeNum];
1018 int MaxSuccLatency = 0;
1019 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = SU.Succs.begin(),
1020 E = SU.Succs.end(); I != E; ++I)
1021 MaxSuccLatency = std::max(MaxSuccLatency, CalcLatency(*I->first));
1023 return Latency = MaxSuccLatency + SU.Latency;
1026 /// CalculatePriorities - Calculate priorities of all scheduling units.
1027 void LatencyPriorityQueue::CalculatePriorities() {
1028 Latencies.assign(SUnits->size(), -1);
1029 NumNodesSolelyBlocking.assign(SUnits->size(), 0);
1031 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
1032 CalcLatency((*SUnits)[i]);
1035 /// getSingleUnscheduledPred - If there is exactly one unscheduled predecessor
1036 /// of SU, return it, otherwise return null.
1037 static SUnit *getSingleUnscheduledPred(SUnit *SU) {
1038 SUnit *OnlyAvailablePred = 0;
1039 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = SU->Preds.begin(),
1040 E = SU->Preds.end(); I != E; ++I)
1041 if (!I->first->isScheduled) {
1042 // We found an available, but not scheduled, predecessor. If it's the
1043 // only one we have found, keep track of it... otherwise give up.
1044 if (OnlyAvailablePred && OnlyAvailablePred != I->first)
1046 OnlyAvailablePred = I->first;
1049 return OnlyAvailablePred;
1052 void LatencyPriorityQueue::push_impl(SUnit *SU) {
1053 // Look at all of the successors of this node. Count the number of nodes that
1054 // this node is the sole unscheduled node for.
1055 unsigned NumNodesBlocking = 0;
1056 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = SU->Succs.begin(),
1057 E = SU->Succs.end(); I != E; ++I)
1058 if (getSingleUnscheduledPred(I->first) == SU)
1060 NumNodesSolelyBlocking[SU->NodeNum] = NumNodesBlocking;
1066 // ScheduledNode - As nodes are scheduled, we look to see if there are any
1067 // successor nodes that have a single unscheduled predecessor. If so, that
1068 // single predecessor has a higher priority, since scheduling it will make
1069 // the node available.
1070 void LatencyPriorityQueue::ScheduledNode(SUnit *SU) {
1071 for (std::set<std::pair<SUnit*, bool> >::const_iterator I = SU->Succs.begin(),
1072 E = SU->Succs.end(); I != E; ++I)
1073 AdjustPriorityOfUnscheduledPreds(I->first);
1076 /// AdjustPriorityOfUnscheduledPreds - One of the predecessors of SU was just
1077 /// scheduled. If SU is not itself available, then there is at least one
1078 /// predecessor node that has not been scheduled yet. If SU has exactly ONE
1079 /// unscheduled predecessor, we want to increase its priority: it getting
1080 /// scheduled will make this node available, so it is better than some other
1081 /// node of the same priority that will not make a node available.
1082 void LatencyPriorityQueue::AdjustPriorityOfUnscheduledPreds(SUnit *SU) {
1083 if (SU->isPending) return; // All preds scheduled.
1085 SUnit *OnlyAvailablePred = getSingleUnscheduledPred(SU);
1086 if (OnlyAvailablePred == 0 || !OnlyAvailablePred->isAvailable) return;
1088 // Okay, we found a single predecessor that is available, but not scheduled.
1089 // Since it is available, it must be in the priority queue. First remove it.
1090 RemoveFromPriorityQueue(OnlyAvailablePred);
1092 // Reinsert the node into the priority queue, which recomputes its
1093 // NumNodesSolelyBlocking value.
1094 push(OnlyAvailablePred);
1098 //===----------------------------------------------------------------------===//
1099 // Public Constructor Functions
1100 //===----------------------------------------------------------------------===//
1102 llvm::ScheduleDAG* llvm::createBURRListDAGScheduler(SelectionDAG &DAG,
1103 MachineBasicBlock *BB) {
1104 return new ScheduleDAGList(DAG, BB, DAG.getTarget(), true,
1105 new RegReductionPriorityQueue(),
1106 new HazardRecognizer());
1109 /// createTDListDAGScheduler - This creates a top-down list scheduler with the
1110 /// specified hazard recognizer.
1111 ScheduleDAG* llvm::createTDListDAGScheduler(SelectionDAG &DAG,
1112 MachineBasicBlock *BB,
1113 HazardRecognizer *HR) {
1114 return new ScheduleDAGList(DAG, BB, DAG.getTarget(), false,
1115 new LatencyPriorityQueue(),