1 //===---- ScheduleDAGList.cpp - Implement a list scheduler for isel DAG ---===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements a top-down list scheduler, using standard algorithms.
11 // The basic approach uses a priority queue of available nodes to schedule.
12 // One at a time, nodes are taken from the priority queue (thus in priority
13 // 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 "pre-RA-sched"
22 #include "llvm/CodeGen/ScheduleDAG.h"
23 #include "llvm/CodeGen/SchedulerRegistry.h"
24 #include "llvm/CodeGen/SelectionDAGISel.h"
25 #include "llvm/Target/TargetRegisterInfo.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Target/TargetMachine.h"
28 #include "llvm/Target/TargetInstrInfo.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/Compiler.h"
31 #include "llvm/ADT/PriorityQueue.h"
32 #include "llvm/ADT/Statistic.h"
36 STATISTIC(NumNoops , "Number of noops inserted");
37 STATISTIC(NumStalls, "Number of pipeline stalls");
39 static RegisterScheduler
40 tdListDAGScheduler("list-td", " Top-down list scheduler",
41 createTDListDAGScheduler);
44 //===----------------------------------------------------------------------===//
45 /// ScheduleDAGList - The actual list scheduler implementation. This supports
46 /// top-down scheduling.
48 class VISIBILITY_HIDDEN ScheduleDAGList : public ScheduleDAG {
50 /// AvailableQueue - The priority queue to use for the available SUnits.
52 SchedulingPriorityQueue *AvailableQueue;
54 /// PendingQueue - This contains all of the instructions whose operands have
55 /// been issued, but their results are not ready yet (due to the latency of
56 /// the operation). Once the operands becomes available, the instruction is
57 /// added to the AvailableQueue. This keeps track of each SUnit and the
58 /// number of cycles left to execute before the operation is available.
59 std::vector<std::pair<unsigned, SUnit*> > PendingQueue;
61 /// HazardRec - The hazard recognizer to use.
62 HazardRecognizer *HazardRec;
65 ScheduleDAGList(SelectionDAG &dag, MachineBasicBlock *bb,
66 const TargetMachine &tm,
67 SchedulingPriorityQueue *availqueue,
69 : ScheduleDAG(dag, bb, tm),
70 AvailableQueue(availqueue), HazardRec(HR) {
75 delete AvailableQueue;
81 void ReleaseSucc(SUnit *SuccSU, bool isChain);
82 void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle);
83 void ListScheduleTopDown();
85 } // end anonymous namespace
87 HazardRecognizer::~HazardRecognizer() {}
90 /// Schedule - Schedule the DAG using list scheduling.
91 void ScheduleDAGList::Schedule() {
92 DOUT << "********** List Scheduling **********\n";
94 // Build scheduling units.
97 AvailableQueue->initNodes(SUnits);
99 ListScheduleTopDown();
101 AvailableQueue->releaseState();
103 DOUT << "*** Final schedule ***\n";
104 DEBUG(dumpSchedule());
108 //===----------------------------------------------------------------------===//
109 // Top-Down Scheduling
110 //===----------------------------------------------------------------------===//
112 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
113 /// the PendingQueue if the count reaches zero.
114 void ScheduleDAGList::ReleaseSucc(SUnit *SuccSU, bool isChain) {
115 SuccSU->NumPredsLeft--;
117 assert(SuccSU->NumPredsLeft >= 0 &&
118 "List scheduling internal error");
120 if (SuccSU->NumPredsLeft == 0) {
121 // Compute how many cycles it will be before this actually becomes
122 // available. This is the max of the start time of all predecessors plus
124 unsigned AvailableCycle = 0;
125 for (SUnit::pred_iterator I = SuccSU->Preds.begin(),
126 E = SuccSU->Preds.end(); I != E; ++I) {
127 // If this is a token edge, we don't need to wait for the latency of the
128 // preceeding instruction (e.g. a long-latency load) unless there is also
129 // some other data dependence.
130 SUnit &Pred = *I->Dep;
131 unsigned PredDoneCycle = Pred.Cycle;
133 PredDoneCycle += Pred.Latency;
134 else if (Pred.Latency)
137 AvailableCycle = std::max(AvailableCycle, PredDoneCycle);
140 PendingQueue.push_back(std::make_pair(AvailableCycle, SuccSU));
144 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
145 /// count of its successors. If a successor pending count is zero, add it to
146 /// the Available queue.
147 void ScheduleDAGList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
148 DOUT << "*** Scheduling [" << CurCycle << "]: ";
149 DEBUG(SU->dump(&DAG));
151 Sequence.push_back(SU);
152 SU->Cycle = CurCycle;
154 // Bottom up: release successors.
155 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
157 ReleaseSucc(I->Dep, I->isCtrl);
160 /// ListScheduleTopDown - The main loop of list scheduling for top-down
162 void ScheduleDAGList::ListScheduleTopDown() {
163 unsigned CurCycle = 0;
165 // All leaves to Available queue.
166 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
167 // It is available if it has no predecessors.
168 if (SUnits[i].Preds.empty()) {
169 AvailableQueue->push(&SUnits[i]);
170 SUnits[i].isAvailable = SUnits[i].isPending = true;
174 // While Available queue is not empty, grab the node with the highest
175 // priority. If it is not ready put it back. Schedule the node.
176 std::vector<SUnit*> NotReady;
177 Sequence.reserve(SUnits.size());
178 while (!AvailableQueue->empty() || !PendingQueue.empty()) {
179 // Check to see if any of the pending instructions are ready to issue. If
180 // so, add them to the available queue.
181 for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
182 if (PendingQueue[i].first == CurCycle) {
183 AvailableQueue->push(PendingQueue[i].second);
184 PendingQueue[i].second->isAvailable = true;
185 PendingQueue[i] = PendingQueue.back();
186 PendingQueue.pop_back();
189 assert(PendingQueue[i].first > CurCycle && "Negative latency?");
193 // If there are no instructions available, don't try to issue anything, and
194 // don't advance the hazard recognizer.
195 if (AvailableQueue->empty()) {
200 SUnit *FoundSUnit = 0;
201 SDNode *FoundNode = 0;
203 bool HasNoopHazards = false;
204 while (!AvailableQueue->empty()) {
205 SUnit *CurSUnit = AvailableQueue->pop();
207 // Get the node represented by this SUnit.
208 FoundNode = CurSUnit->Node;
210 // If this is a pseudo op, like copyfromreg, look to see if there is a
211 // real target node flagged to it. If so, use the target node.
212 for (unsigned i = 0, e = CurSUnit->FlaggedNodes.size();
213 FoundNode->getOpcode() < ISD::BUILTIN_OP_END && i != e; ++i)
214 FoundNode = CurSUnit->FlaggedNodes[i];
216 HazardRecognizer::HazardType HT = HazardRec->getHazardType(FoundNode);
217 if (HT == HazardRecognizer::NoHazard) {
218 FoundSUnit = CurSUnit;
222 // Remember if this is a noop hazard.
223 HasNoopHazards |= HT == HazardRecognizer::NoopHazard;
225 NotReady.push_back(CurSUnit);
228 // Add the nodes that aren't ready back onto the available list.
229 if (!NotReady.empty()) {
230 AvailableQueue->push_all(NotReady);
234 // If we found a node to schedule, do it now.
236 ScheduleNodeTopDown(FoundSUnit, CurCycle);
237 HazardRec->EmitInstruction(FoundNode);
238 FoundSUnit->isScheduled = true;
239 AvailableQueue->ScheduledNode(FoundSUnit);
241 // If this is a pseudo-op node, we don't want to increment the current
243 if (FoundSUnit->Latency) // Don't increment CurCycle for pseudo-ops!
245 } else if (!HasNoopHazards) {
246 // Otherwise, we have a pipeline stall, but no other problem, just advance
247 // the current cycle and try again.
248 DOUT << "*** Advancing cycle, no work to do\n";
249 HazardRec->AdvanceCycle();
253 // Otherwise, we have no instructions to issue and we have instructions
254 // that will fault if we don't do this right. This is the case for
255 // processors without pipeline interlocks and other cases.
256 DOUT << "*** Emitting noop\n";
257 HazardRec->EmitNoop();
258 Sequence.push_back(0); // NULL SUnit* -> noop
265 // Verify that all SUnits were scheduled.
266 bool AnyNotSched = false;
267 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
268 if (SUnits[i].NumPredsLeft != 0) {
270 cerr << "*** List scheduling failed! ***\n";
271 SUnits[i].dump(&DAG);
272 cerr << "has not been scheduled!\n";
276 assert(!AnyNotSched);
280 //===----------------------------------------------------------------------===//
281 // LatencyPriorityQueue Implementation
282 //===----------------------------------------------------------------------===//
284 // This is a SchedulingPriorityQueue that schedules using latency information to
285 // reduce the length of the critical path through the basic block.
288 class LatencyPriorityQueue;
290 /// Sorting functions for the Available queue.
291 struct latency_sort : public std::binary_function<SUnit*, SUnit*, bool> {
292 LatencyPriorityQueue *PQ;
293 latency_sort(LatencyPriorityQueue *pq) : PQ(pq) {}
294 latency_sort(const latency_sort &RHS) : PQ(RHS.PQ) {}
296 bool operator()(const SUnit* left, const SUnit* right) const;
298 } // end anonymous namespace
301 class LatencyPriorityQueue : public SchedulingPriorityQueue {
302 // SUnits - The SUnits for the current graph.
303 std::vector<SUnit> *SUnits;
305 // Latencies - The latency (max of latency from this node to the bb exit)
307 std::vector<int> Latencies;
309 /// NumNodesSolelyBlocking - This vector contains, for every node in the
310 /// Queue, the number of nodes that the node is the sole unscheduled
311 /// predecessor for. This is used as a tie-breaker heuristic for better
313 std::vector<unsigned> NumNodesSolelyBlocking;
315 PriorityQueue<SUnit*, std::vector<SUnit*>, latency_sort> Queue;
317 LatencyPriorityQueue() : Queue(latency_sort(this)) {
320 void initNodes(std::vector<SUnit> &sunits) {
322 // Calculate node priorities.
323 CalculatePriorities();
326 void addNode(const SUnit *SU) {
327 Latencies.resize(SUnits->size(), -1);
328 NumNodesSolelyBlocking.resize(SUnits->size(), 0);
332 void updateNode(const SUnit *SU) {
333 Latencies[SU->NodeNum] = -1;
337 void releaseState() {
342 unsigned getLatency(unsigned NodeNum) const {
343 assert(NodeNum < Latencies.size());
344 return Latencies[NodeNum];
347 unsigned getNumSolelyBlockNodes(unsigned NodeNum) const {
348 assert(NodeNum < NumNodesSolelyBlocking.size());
349 return NumNodesSolelyBlocking[NodeNum];
352 unsigned size() const { return Queue.size(); }
354 bool empty() const { return Queue.empty(); }
356 virtual void push(SUnit *U) {
359 void push_impl(SUnit *U);
361 void push_all(const std::vector<SUnit *> &Nodes) {
362 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
367 if (empty()) return NULL;
368 SUnit *V = Queue.top();
373 void remove(SUnit *SU) {
374 assert(!Queue.empty() && "Not in queue!");
378 // ScheduledNode - As nodes are scheduled, we look to see if there are any
379 // successor nodes that have a single unscheduled predecessor. If so, that
380 // single predecessor has a higher priority, since scheduling it will make
381 // the node available.
382 void ScheduledNode(SUnit *Node);
385 void CalculatePriorities();
386 int CalcLatency(const SUnit &SU);
387 void AdjustPriorityOfUnscheduledPreds(SUnit *SU);
388 SUnit *getSingleUnscheduledPred(SUnit *SU);
392 bool latency_sort::operator()(const SUnit *LHS, const SUnit *RHS) const {
393 unsigned LHSNum = LHS->NodeNum;
394 unsigned RHSNum = RHS->NodeNum;
396 // The most important heuristic is scheduling the critical path.
397 unsigned LHSLatency = PQ->getLatency(LHSNum);
398 unsigned RHSLatency = PQ->getLatency(RHSNum);
399 if (LHSLatency < RHSLatency) return true;
400 if (LHSLatency > RHSLatency) return false;
402 // After that, if two nodes have identical latencies, look to see if one will
403 // unblock more other nodes than the other.
404 unsigned LHSBlocked = PQ->getNumSolelyBlockNodes(LHSNum);
405 unsigned RHSBlocked = PQ->getNumSolelyBlockNodes(RHSNum);
406 if (LHSBlocked < RHSBlocked) return true;
407 if (LHSBlocked > RHSBlocked) return false;
409 // Finally, just to provide a stable ordering, use the node number as a
411 return LHSNum < RHSNum;
415 /// CalcNodePriority - Calculate the maximal path from the node to the exit.
417 int LatencyPriorityQueue::CalcLatency(const SUnit &SU) {
418 int &Latency = Latencies[SU.NodeNum];
422 std::vector<const SUnit*> WorkList;
423 WorkList.push_back(&SU);
424 while (!WorkList.empty()) {
425 const SUnit *Cur = WorkList.back();
427 int MaxSuccLatency = 0;
428 for (SUnit::const_succ_iterator I = Cur->Succs.begin(),E = Cur->Succs.end();
430 int SuccLatency = Latencies[I->Dep->NodeNum];
431 if (SuccLatency == -1) {
433 WorkList.push_back(I->Dep);
435 MaxSuccLatency = std::max(MaxSuccLatency, SuccLatency);
439 Latencies[Cur->NodeNum] = MaxSuccLatency + Cur->Latency;
447 /// CalculatePriorities - Calculate priorities of all scheduling units.
448 void LatencyPriorityQueue::CalculatePriorities() {
449 Latencies.assign(SUnits->size(), -1);
450 NumNodesSolelyBlocking.assign(SUnits->size(), 0);
452 // For each node, calculate the maximal path from the node to the exit.
453 std::vector<std::pair<const SUnit*, unsigned> > WorkList;
454 for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
455 const SUnit *SU = &(*SUnits)[i];
456 if (SU->Succs.empty())
457 WorkList.push_back(std::make_pair(SU, 0U));
460 while (!WorkList.empty()) {
461 const SUnit *SU = WorkList.back().first;
462 unsigned SuccLat = WorkList.back().second;
464 int &Latency = Latencies[SU->NodeNum];
465 if (Latency == -1 || (SU->Latency + SuccLat) > (unsigned)Latency) {
466 Latency = SU->Latency + SuccLat;
467 for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
469 WorkList.push_back(std::make_pair(I->Dep, Latency));
474 /// getSingleUnscheduledPred - If there is exactly one unscheduled predecessor
475 /// of SU, return it, otherwise return null.
476 SUnit *LatencyPriorityQueue::getSingleUnscheduledPred(SUnit *SU) {
477 SUnit *OnlyAvailablePred = 0;
478 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
480 SUnit &Pred = *I->Dep;
481 if (!Pred.isScheduled) {
482 // We found an available, but not scheduled, predecessor. If it's the
483 // only one we have found, keep track of it... otherwise give up.
484 if (OnlyAvailablePred && OnlyAvailablePred != &Pred)
486 OnlyAvailablePred = &Pred;
490 return OnlyAvailablePred;
493 void LatencyPriorityQueue::push_impl(SUnit *SU) {
494 // Look at all of the successors of this node. Count the number of nodes that
495 // this node is the sole unscheduled node for.
496 unsigned NumNodesBlocking = 0;
497 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
499 if (getSingleUnscheduledPred(I->Dep) == SU)
501 NumNodesSolelyBlocking[SU->NodeNum] = NumNodesBlocking;
507 // ScheduledNode - As nodes are scheduled, we look to see if there are any
508 // successor nodes that have a single unscheduled predecessor. If so, that
509 // single predecessor has a higher priority, since scheduling it will make
510 // the node available.
511 void LatencyPriorityQueue::ScheduledNode(SUnit *SU) {
512 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
514 AdjustPriorityOfUnscheduledPreds(I->Dep);
517 /// AdjustPriorityOfUnscheduledPreds - One of the predecessors of SU was just
518 /// scheduled. If SU is not itself available, then there is at least one
519 /// predecessor node that has not been scheduled yet. If SU has exactly ONE
520 /// unscheduled predecessor, we want to increase its priority: it getting
521 /// scheduled will make this node available, so it is better than some other
522 /// node of the same priority that will not make a node available.
523 void LatencyPriorityQueue::AdjustPriorityOfUnscheduledPreds(SUnit *SU) {
524 if (SU->isPending) return; // All preds scheduled.
526 SUnit *OnlyAvailablePred = getSingleUnscheduledPred(SU);
527 if (OnlyAvailablePred == 0 || !OnlyAvailablePred->isAvailable) return;
529 // Okay, we found a single predecessor that is available, but not scheduled.
530 // Since it is available, it must be in the priority queue. First remove it.
531 remove(OnlyAvailablePred);
533 // Reinsert the node into the priority queue, which recomputes its
534 // NumNodesSolelyBlocking value.
535 push(OnlyAvailablePred);
539 //===----------------------------------------------------------------------===//
540 // Public Constructor Functions
541 //===----------------------------------------------------------------------===//
543 /// createTDListDAGScheduler - This creates a top-down list scheduler with a
544 /// new hazard recognizer. This scheduler takes ownership of the hazard
545 /// recognizer and deletes it when done.
546 ScheduleDAG* llvm::createTDListDAGScheduler(SelectionDAGISel *IS,
548 MachineBasicBlock *BB, bool Fast) {
549 return new ScheduleDAGList(*DAG, BB, DAG->getTarget(),
550 new LatencyPriorityQueue(),
551 IS->CreateTargetHazardRecognizer());