1 //===----- ScheduleDAGRRList.cpp - Reg pressure reduction list scheduler --===//
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 bottom-up and top-down register pressure reduction list
11 // schedulers, using standard algorithms. The basic approach uses a priority
12 // queue of available nodes to schedule. One at a time, nodes are taken from
13 // the priority queue (thus in priority order), checked for legality to
14 // schedule, and emitted if legal.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "pre-RA-sched"
19 #include "ScheduleDAGSDNodes.h"
20 #include "llvm/InlineAsm.h"
21 #include "llvm/CodeGen/SchedulerRegistry.h"
22 #include "llvm/CodeGen/SelectionDAGISel.h"
23 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
24 #include "llvm/Target/TargetRegisterInfo.h"
25 #include "llvm/Target/TargetData.h"
26 #include "llvm/Target/TargetMachine.h"
27 #include "llvm/Target/TargetInstrInfo.h"
28 #include "llvm/Target/TargetLowering.h"
29 #include "llvm/ADT/SmallSet.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/ADT/STLExtras.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/raw_ostream.h"
38 STATISTIC(NumBacktracks, "Number of times scheduler backtracked");
39 STATISTIC(NumUnfolds, "Number of nodes unfolded");
40 STATISTIC(NumDups, "Number of duplicated nodes");
41 STATISTIC(NumPRCopies, "Number of physical register copies");
43 static RegisterScheduler
44 burrListDAGScheduler("list-burr",
45 "Bottom-up register reduction list scheduling",
46 createBURRListDAGScheduler);
47 static RegisterScheduler
48 tdrListrDAGScheduler("list-tdrr",
49 "Top-down register reduction list scheduling",
50 createTDRRListDAGScheduler);
51 static RegisterScheduler
52 sourceListDAGScheduler("source",
53 "Similar to list-burr but schedules in source "
54 "order when possible",
55 createSourceListDAGScheduler);
57 static RegisterScheduler
58 hybridListDAGScheduler("list-hybrid",
59 "Bottom-up register pressure aware list scheduling "
60 "which tries to balance latency and register pressure",
61 createHybridListDAGScheduler);
63 static RegisterScheduler
64 ILPListDAGScheduler("list-ilp",
65 "Bottom-up register pressure aware list scheduling "
66 "which tries to balance ILP and register pressure",
67 createILPListDAGScheduler);
69 static cl::opt<bool> DisableSchedCycles(
70 "disable-sched-cycles", cl::Hidden, cl::init(false),
71 cl::desc("Disable cycle-level precision during preRA scheduling"));
73 // Temporary sched=list-ilp flags until the heuristics are robust.
74 static cl::opt<bool> DisableSchedRegPressure(
75 "disable-sched-reg-pressure", cl::Hidden, cl::init(false),
76 cl::desc("Disable regpressure priority in sched=list-ilp"));
77 static cl::opt<bool> DisableSchedLiveUses(
78 "disable-sched-live-uses", cl::Hidden, cl::init(true),
79 cl::desc("Disable live use priority in sched=list-ilp"));
80 static cl::opt<bool> DisableSchedStalls(
81 "disable-sched-stalls", cl::Hidden, cl::init(true),
82 cl::desc("Disable no-stall priority in sched=list-ilp"));
83 static cl::opt<bool> DisableSchedCriticalPath(
84 "disable-sched-critical-path", cl::Hidden, cl::init(false),
85 cl::desc("Disable critical path priority in sched=list-ilp"));
86 static cl::opt<bool> DisableSchedHeight(
87 "disable-sched-height", cl::Hidden, cl::init(false),
88 cl::desc("Disable scheduled-height priority in sched=list-ilp"));
90 static cl::opt<int> MaxReorderWindow(
91 "max-sched-reorder", cl::Hidden, cl::init(6),
92 cl::desc("Number of instructions to allow ahead of the critical path "
93 "in sched=list-ilp"));
95 static cl::opt<unsigned> AvgIPC(
96 "sched-avg-ipc", cl::Hidden, cl::init(1),
97 cl::desc("Average inst/cycle whan no target itinerary exists."));
101 // For sched=list-ilp, Count the number of times each factor comes into play.
102 enum { FactPressureDiff, FactRegUses, FactHeight, FactDepth, FactStatic,
103 FactOther, NumFactors };
105 static const char *FactorName[NumFactors] =
106 {"PressureDiff", "RegUses", "Height", "Depth","Static", "Other"};
107 static int FactorCount[NumFactors];
111 //===----------------------------------------------------------------------===//
112 /// ScheduleDAGRRList - The actual register reduction list scheduler
113 /// implementation. This supports both top-down and bottom-up scheduling.
115 class ScheduleDAGRRList : public ScheduleDAGSDNodes {
117 /// isBottomUp - This is true if the scheduling problem is bottom-up, false if
121 /// NeedLatency - True if the scheduler will make use of latency information.
125 /// AvailableQueue - The priority queue to use for the available SUnits.
126 SchedulingPriorityQueue *AvailableQueue;
128 /// PendingQueue - This contains all of the instructions whose operands have
129 /// been issued, but their results are not ready yet (due to the latency of
130 /// the operation). Once the operands becomes available, the instruction is
131 /// added to the AvailableQueue.
132 std::vector<SUnit*> PendingQueue;
134 /// HazardRec - The hazard recognizer to use.
135 ScheduleHazardRecognizer *HazardRec;
137 /// CurCycle - The current scheduler state corresponds to this cycle.
140 /// MinAvailableCycle - Cycle of the soonest available instruction.
141 unsigned MinAvailableCycle;
143 /// IssueCount - Count instructions issued in this cycle
144 /// Currently valid only for bottom-up scheduling.
147 /// LiveRegDefs - A set of physical registers and their definition
148 /// that are "live". These nodes must be scheduled before any other nodes that
149 /// modifies the registers can be scheduled.
150 unsigned NumLiveRegs;
151 std::vector<SUnit*> LiveRegDefs;
152 std::vector<SUnit*> LiveRegGens;
154 /// Topo - A topological ordering for SUnits which permits fast IsReachable
155 /// and similar queries.
156 ScheduleDAGTopologicalSort Topo;
159 ScheduleDAGRRList(MachineFunction &mf, bool needlatency,
160 SchedulingPriorityQueue *availqueue,
161 CodeGenOpt::Level OptLevel)
162 : ScheduleDAGSDNodes(mf), isBottomUp(availqueue->isBottomUp()),
163 NeedLatency(needlatency), AvailableQueue(availqueue), CurCycle(0),
166 const TargetMachine &tm = mf.getTarget();
167 if (DisableSchedCycles || !NeedLatency)
168 HazardRec = new ScheduleHazardRecognizer();
170 HazardRec = tm.getInstrInfo()->CreateTargetHazardRecognizer(&tm, this);
173 ~ScheduleDAGRRList() {
175 delete AvailableQueue;
180 ScheduleHazardRecognizer *getHazardRec() { return HazardRec; }
182 /// IsReachable - Checks if SU is reachable from TargetSU.
183 bool IsReachable(const SUnit *SU, const SUnit *TargetSU) {
184 return Topo.IsReachable(SU, TargetSU);
187 /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
189 bool WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
190 return Topo.WillCreateCycle(SU, TargetSU);
193 /// AddPred - adds a predecessor edge to SUnit SU.
194 /// This returns true if this is a new predecessor.
195 /// Updates the topological ordering if required.
196 void AddPred(SUnit *SU, const SDep &D) {
197 Topo.AddPred(SU, D.getSUnit());
201 /// RemovePred - removes a predecessor edge from SUnit SU.
202 /// This returns true if an edge was removed.
203 /// Updates the topological ordering if required.
204 void RemovePred(SUnit *SU, const SDep &D) {
205 Topo.RemovePred(SU, D.getSUnit());
210 bool isReady(SUnit *SU) {
211 return DisableSchedCycles || !AvailableQueue->hasReadyFilter() ||
212 AvailableQueue->isReady(SU);
215 void ReleasePred(SUnit *SU, const SDep *PredEdge);
216 void ReleasePredecessors(SUnit *SU);
217 void ReleaseSucc(SUnit *SU, const SDep *SuccEdge);
218 void ReleaseSuccessors(SUnit *SU);
219 void ReleasePending();
220 void AdvanceToCycle(unsigned NextCycle);
221 void AdvancePastStalls(SUnit *SU);
222 void EmitNode(SUnit *SU);
223 void ScheduleNodeBottomUp(SUnit*);
224 void CapturePred(SDep *PredEdge);
225 void UnscheduleNodeBottomUp(SUnit*);
226 void RestoreHazardCheckerBottomUp();
227 void BacktrackBottomUp(SUnit*, SUnit*);
228 SUnit *CopyAndMoveSuccessors(SUnit*);
229 void InsertCopiesAndMoveSuccs(SUnit*, unsigned,
230 const TargetRegisterClass*,
231 const TargetRegisterClass*,
232 SmallVector<SUnit*, 2>&);
233 bool DelayForLiveRegsBottomUp(SUnit*, SmallVector<unsigned, 4>&);
235 SUnit *PickNodeToScheduleBottomUp();
236 void ListScheduleBottomUp();
238 void ScheduleNodeTopDown(SUnit*);
239 void ListScheduleTopDown();
242 /// CreateNewSUnit - Creates a new SUnit and returns a pointer to it.
243 /// Updates the topological ordering if required.
244 SUnit *CreateNewSUnit(SDNode *N) {
245 unsigned NumSUnits = SUnits.size();
246 SUnit *NewNode = NewSUnit(N);
247 // Update the topological ordering.
248 if (NewNode->NodeNum >= NumSUnits)
249 Topo.InitDAGTopologicalSorting();
253 /// CreateClone - Creates a new SUnit from an existing one.
254 /// Updates the topological ordering if required.
255 SUnit *CreateClone(SUnit *N) {
256 unsigned NumSUnits = SUnits.size();
257 SUnit *NewNode = Clone(N);
258 // Update the topological ordering.
259 if (NewNode->NodeNum >= NumSUnits)
260 Topo.InitDAGTopologicalSorting();
264 /// ForceUnitLatencies - Register-pressure-reducing scheduling doesn't
265 /// need actual latency information but the hybrid scheduler does.
266 bool ForceUnitLatencies() const {
270 } // end anonymous namespace
273 /// Schedule - Schedule the DAG using list scheduling.
274 void ScheduleDAGRRList::Schedule() {
276 << "********** List Scheduling BB#" << BB->getNumber()
277 << " '" << BB->getName() << "' **********\n");
279 for (int i = 0; i < NumFactors; ++i) {
286 MinAvailableCycle = DisableSchedCycles ? 0 : UINT_MAX;
288 LiveRegDefs.resize(TRI->getNumRegs(), NULL);
289 LiveRegGens.resize(TRI->getNumRegs(), NULL);
291 // Build the scheduling graph.
292 BuildSchedGraph(NULL);
294 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
295 SUnits[su].dumpAll(this));
296 Topo.InitDAGTopologicalSorting();
298 AvailableQueue->initNodes(SUnits);
302 // Execute the actual scheduling loop Top-Down or Bottom-Up as appropriate.
304 ListScheduleBottomUp();
306 ListScheduleTopDown();
309 for (int i = 0; i < NumFactors; ++i) {
310 DEBUG(dbgs() << FactorName[i] << "\t" << FactorCount[i] << "\n");
313 AvailableQueue->releaseState();
316 //===----------------------------------------------------------------------===//
317 // Bottom-Up Scheduling
318 //===----------------------------------------------------------------------===//
320 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
321 /// the AvailableQueue if the count reaches zero. Also update its cycle bound.
322 void ScheduleDAGRRList::ReleasePred(SUnit *SU, const SDep *PredEdge) {
323 SUnit *PredSU = PredEdge->getSUnit();
326 if (PredSU->NumSuccsLeft == 0) {
327 dbgs() << "*** Scheduling failed! ***\n";
329 dbgs() << " has been released too many times!\n";
333 --PredSU->NumSuccsLeft;
335 if (!ForceUnitLatencies()) {
336 // Updating predecessor's height. This is now the cycle when the
337 // predecessor can be scheduled without causing a pipeline stall.
338 PredSU->setHeightToAtLeast(SU->getHeight() + PredEdge->getLatency());
341 // If all the node's successors are scheduled, this node is ready
342 // to be scheduled. Ignore the special EntrySU node.
343 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU) {
344 PredSU->isAvailable = true;
346 unsigned Height = PredSU->getHeight();
347 if (Height < MinAvailableCycle)
348 MinAvailableCycle = Height;
350 if (isReady(PredSU)) {
351 AvailableQueue->push(PredSU);
353 // CapturePred and others may have left the node in the pending queue, avoid
355 else if (!PredSU->isPending) {
356 PredSU->isPending = true;
357 PendingQueue.push_back(PredSU);
362 /// Call ReleasePred for each predecessor, then update register live def/gen.
363 /// Always update LiveRegDefs for a register dependence even if the current SU
364 /// also defines the register. This effectively create one large live range
365 /// across a sequence of two-address node. This is important because the
366 /// entire chain must be scheduled together. Example:
369 /// flags = (2) addc flags
370 /// flags = (1) addc flags
374 /// LiveRegDefs[flags] = 3
375 /// LiveRegGens[flags] = 1
377 /// If (2) addc is unscheduled, then (1) addc must also be unscheduled to avoid
378 /// interference on flags.
379 void ScheduleDAGRRList::ReleasePredecessors(SUnit *SU) {
380 // Bottom up: release predecessors
381 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
383 ReleasePred(SU, &*I);
384 if (I->isAssignedRegDep()) {
385 // This is a physical register dependency and it's impossible or
386 // expensive to copy the register. Make sure nothing that can
387 // clobber the register is scheduled between the predecessor and
389 SUnit *RegDef = LiveRegDefs[I->getReg()]; (void)RegDef;
390 assert((!RegDef || RegDef == SU || RegDef == I->getSUnit()) &&
391 "interference on register dependence");
392 LiveRegDefs[I->getReg()] = I->getSUnit();
393 if (!LiveRegGens[I->getReg()]) {
395 LiveRegGens[I->getReg()] = SU;
401 /// Check to see if any of the pending instructions are ready to issue. If
402 /// so, add them to the available queue.
403 void ScheduleDAGRRList::ReleasePending() {
404 if (DisableSchedCycles) {
405 assert(PendingQueue.empty() && "pending instrs not allowed in this mode");
409 // If the available queue is empty, it is safe to reset MinAvailableCycle.
410 if (AvailableQueue->empty())
411 MinAvailableCycle = UINT_MAX;
413 // Check to see if any of the pending instructions are ready to issue. If
414 // so, add them to the available queue.
415 for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
416 unsigned ReadyCycle =
417 isBottomUp ? PendingQueue[i]->getHeight() : PendingQueue[i]->getDepth();
418 if (ReadyCycle < MinAvailableCycle)
419 MinAvailableCycle = ReadyCycle;
421 if (PendingQueue[i]->isAvailable) {
422 if (!isReady(PendingQueue[i]))
424 AvailableQueue->push(PendingQueue[i]);
426 PendingQueue[i]->isPending = false;
427 PendingQueue[i] = PendingQueue.back();
428 PendingQueue.pop_back();
433 /// Move the scheduler state forward by the specified number of Cycles.
434 void ScheduleDAGRRList::AdvanceToCycle(unsigned NextCycle) {
435 if (NextCycle <= CurCycle)
439 AvailableQueue->setCurCycle(NextCycle);
440 if (!HazardRec->isEnabled()) {
441 // Bypass lots of virtual calls in case of long latency.
442 CurCycle = NextCycle;
445 for (; CurCycle != NextCycle; ++CurCycle) {
447 HazardRec->RecedeCycle();
449 HazardRec->AdvanceCycle();
452 // FIXME: Instead of visiting the pending Q each time, set a dirty flag on the
453 // available Q to release pending nodes at least once before popping.
457 /// Move the scheduler state forward until the specified node's dependents are
458 /// ready and can be scheduled with no resource conflicts.
459 void ScheduleDAGRRList::AdvancePastStalls(SUnit *SU) {
460 if (DisableSchedCycles)
463 unsigned ReadyCycle = isBottomUp ? SU->getHeight() : SU->getDepth();
465 // Bump CurCycle to account for latency. We assume the latency of other
466 // available instructions may be hidden by the stall (not a full pipe stall).
467 // This updates the hazard recognizer's cycle before reserving resources for
469 AdvanceToCycle(ReadyCycle);
471 // Calls are scheduled in their preceding cycle, so don't conflict with
472 // hazards from instructions after the call. EmitNode will reset the
473 // scoreboard state before emitting the call.
474 if (isBottomUp && SU->isCall)
477 // FIXME: For resource conflicts in very long non-pipelined stages, we
478 // should probably skip ahead here to avoid useless scoreboard checks.
481 ScheduleHazardRecognizer::HazardType HT =
482 HazardRec->getHazardType(SU, isBottomUp ? -Stalls : Stalls);
484 if (HT == ScheduleHazardRecognizer::NoHazard)
489 AdvanceToCycle(CurCycle + Stalls);
492 /// Record this SUnit in the HazardRecognizer.
493 /// Does not update CurCycle.
494 void ScheduleDAGRRList::EmitNode(SUnit *SU) {
495 if (!HazardRec->isEnabled())
498 // Check for phys reg copy.
502 switch (SU->getNode()->getOpcode()) {
504 assert(SU->getNode()->isMachineOpcode() &&
505 "This target-independent node should not be scheduled.");
507 case ISD::MERGE_VALUES:
508 case ISD::TokenFactor:
510 case ISD::CopyFromReg:
512 // Noops don't affect the scoreboard state. Copies are likely to be
516 // For inline asm, clear the pipeline state.
520 if (isBottomUp && SU->isCall) {
521 // Calls are scheduled with their preceding instructions. For bottom-up
522 // scheduling, clear the pipeline state before emitting.
526 HazardRec->EmitInstruction(SU);
528 if (!isBottomUp && SU->isCall) {
533 /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
534 /// count of its predecessors. If a predecessor pending count is zero, add it to
535 /// the Available queue.
536 void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU) {
537 DEBUG(dbgs() << "\n*** Scheduling [" << CurCycle << "]: ");
538 DEBUG(SU->dump(this));
541 if (CurCycle < SU->getHeight())
542 DEBUG(dbgs() << " Height [" << SU->getHeight() << "] pipeline stall!\n");
545 // FIXME: Do not modify node height. It may interfere with
546 // backtracking. Instead add a "ready cycle" to SUnit. Before scheduling the
547 // node it's ready cycle can aid heuristics, and after scheduling it can
548 // indicate the scheduled cycle.
549 SU->setHeightToAtLeast(CurCycle);
551 // Reserve resources for the scheduled intruction.
554 Sequence.push_back(SU);
556 AvailableQueue->ScheduledNode(SU);
558 // If HazardRec is disabled, and each inst counts as one cycle, then
559 // advance CurCycle before ReleasePredecessors to avoid useles pushed to
560 // PendingQueue for schedulers that implement HasReadyFilter.
561 if (!HazardRec->isEnabled() && AvgIPC < 2)
562 AdvanceToCycle(CurCycle + 1);
564 // Update liveness of predecessors before successors to avoid treating a
565 // two-address node as a live range def.
566 ReleasePredecessors(SU);
568 // Release all the implicit physical register defs that are live.
569 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
571 // LiveRegDegs[I->getReg()] != SU when SU is a two-address node.
572 if (I->isAssignedRegDep() && LiveRegDefs[I->getReg()] == SU) {
573 assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
575 LiveRegDefs[I->getReg()] = NULL;
576 LiveRegGens[I->getReg()] = NULL;
580 SU->isScheduled = true;
582 // Conditions under which the scheduler should eagerly advance the cycle:
583 // (1) No available instructions
584 // (2) All pipelines full, so available instructions must have hazards.
586 // If HazardRec is disabled, the cycle was advanced earlier.
588 // Check AvailableQueue after ReleasePredecessors in case of zero latency.
590 if ((HazardRec->isEnabled() && HazardRec->atIssueLimit())
591 || (!HazardRec->isEnabled() && AvgIPC > 1 && IssueCount == AvgIPC)
592 || AvailableQueue->empty())
593 AdvanceToCycle(CurCycle + 1);
596 /// CapturePred - This does the opposite of ReleasePred. Since SU is being
597 /// unscheduled, incrcease the succ left count of its predecessors. Remove
598 /// them from AvailableQueue if necessary.
599 void ScheduleDAGRRList::CapturePred(SDep *PredEdge) {
600 SUnit *PredSU = PredEdge->getSUnit();
601 if (PredSU->isAvailable) {
602 PredSU->isAvailable = false;
603 if (!PredSU->isPending)
604 AvailableQueue->remove(PredSU);
607 assert(PredSU->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!");
608 ++PredSU->NumSuccsLeft;
611 /// UnscheduleNodeBottomUp - Remove the node from the schedule, update its and
612 /// its predecessor states to reflect the change.
613 void ScheduleDAGRRList::UnscheduleNodeBottomUp(SUnit *SU) {
614 DEBUG(dbgs() << "*** Unscheduling [" << SU->getHeight() << "]: ");
615 DEBUG(SU->dump(this));
617 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
620 if (I->isAssignedRegDep() && SU == LiveRegGens[I->getReg()]){
621 assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
622 assert(LiveRegDefs[I->getReg()] == I->getSUnit() &&
623 "Physical register dependency violated?");
625 LiveRegDefs[I->getReg()] = NULL;
626 LiveRegGens[I->getReg()] = NULL;
630 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
632 if (I->isAssignedRegDep()) {
633 // This becomes the nearest def. Note that an earlier def may still be
634 // pending if this is a two-address node.
635 LiveRegDefs[I->getReg()] = SU;
636 if (!LiveRegDefs[I->getReg()]) {
639 if (LiveRegGens[I->getReg()] == NULL ||
640 I->getSUnit()->getHeight() < LiveRegGens[I->getReg()]->getHeight())
641 LiveRegGens[I->getReg()] = I->getSUnit();
644 if (SU->getHeight() < MinAvailableCycle)
645 MinAvailableCycle = SU->getHeight();
647 SU->setHeightDirty();
648 SU->isScheduled = false;
649 SU->isAvailable = true;
650 if (!DisableSchedCycles && AvailableQueue->hasReadyFilter()) {
651 // Don't make available until backtracking is complete.
652 SU->isPending = true;
653 PendingQueue.push_back(SU);
656 AvailableQueue->push(SU);
658 AvailableQueue->UnscheduledNode(SU);
661 /// After backtracking, the hazard checker needs to be restored to a state
662 /// corresponding the the current cycle.
663 void ScheduleDAGRRList::RestoreHazardCheckerBottomUp() {
666 unsigned LookAhead = std::min((unsigned)Sequence.size(),
667 HazardRec->getMaxLookAhead());
671 std::vector<SUnit*>::const_iterator I = (Sequence.end() - LookAhead);
672 unsigned HazardCycle = (*I)->getHeight();
673 for (std::vector<SUnit*>::const_iterator E = Sequence.end(); I != E; ++I) {
675 for (; SU->getHeight() > HazardCycle; ++HazardCycle) {
676 HazardRec->RecedeCycle();
682 /// BacktrackBottomUp - Backtrack scheduling to a previous cycle specified in
683 /// BTCycle in order to schedule a specific node.
684 void ScheduleDAGRRList::BacktrackBottomUp(SUnit *SU, SUnit *BtSU) {
685 SUnit *OldSU = Sequence.back();
688 if (SU->isSucc(OldSU))
689 // Don't try to remove SU from AvailableQueue.
690 SU->isAvailable = false;
691 // FIXME: use ready cycle instead of height
692 CurCycle = OldSU->getHeight();
693 UnscheduleNodeBottomUp(OldSU);
694 AvailableQueue->setCurCycle(CurCycle);
697 OldSU = Sequence.back();
700 assert(!SU->isSucc(OldSU) && "Something is wrong!");
702 RestoreHazardCheckerBottomUp();
709 static bool isOperandOf(const SUnit *SU, SDNode *N) {
710 for (const SDNode *SUNode = SU->getNode(); SUNode;
711 SUNode = SUNode->getGluedNode()) {
712 if (SUNode->isOperandOf(N))
718 /// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
719 /// successors to the newly created node.
720 SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) {
721 SDNode *N = SU->getNode();
725 if (SU->getNode()->getGluedNode())
729 bool TryUnfold = false;
730 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
731 EVT VT = N->getValueType(i);
734 else if (VT == MVT::Other)
737 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
738 const SDValue &Op = N->getOperand(i);
739 EVT VT = Op.getNode()->getValueType(Op.getResNo());
745 SmallVector<SDNode*, 2> NewNodes;
746 if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
749 DEBUG(dbgs() << "Unfolding SU #" << SU->NodeNum << "\n");
750 assert(NewNodes.size() == 2 && "Expected a load folding node!");
753 SDNode *LoadNode = NewNodes[0];
754 unsigned NumVals = N->getNumValues();
755 unsigned OldNumVals = SU->getNode()->getNumValues();
756 for (unsigned i = 0; i != NumVals; ++i)
757 DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i));
758 DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals-1),
759 SDValue(LoadNode, 1));
761 // LoadNode may already exist. This can happen when there is another
762 // load from the same location and producing the same type of value
763 // but it has different alignment or volatileness.
764 bool isNewLoad = true;
766 if (LoadNode->getNodeId() != -1) {
767 LoadSU = &SUnits[LoadNode->getNodeId()];
770 LoadSU = CreateNewSUnit(LoadNode);
771 LoadNode->setNodeId(LoadSU->NodeNum);
773 InitNumRegDefsLeft(LoadSU);
774 ComputeLatency(LoadSU);
777 SUnit *NewSU = CreateNewSUnit(N);
778 assert(N->getNodeId() == -1 && "Node already inserted!");
779 N->setNodeId(NewSU->NodeNum);
781 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
782 for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
783 if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
784 NewSU->isTwoAddress = true;
788 if (TID.isCommutable())
789 NewSU->isCommutable = true;
791 InitNumRegDefsLeft(NewSU);
792 ComputeLatency(NewSU);
794 // Record all the edges to and from the old SU, by category.
795 SmallVector<SDep, 4> ChainPreds;
796 SmallVector<SDep, 4> ChainSuccs;
797 SmallVector<SDep, 4> LoadPreds;
798 SmallVector<SDep, 4> NodePreds;
799 SmallVector<SDep, 4> NodeSuccs;
800 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
803 ChainPreds.push_back(*I);
804 else if (isOperandOf(I->getSUnit(), LoadNode))
805 LoadPreds.push_back(*I);
807 NodePreds.push_back(*I);
809 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
812 ChainSuccs.push_back(*I);
814 NodeSuccs.push_back(*I);
817 // Now assign edges to the newly-created nodes.
818 for (unsigned i = 0, e = ChainPreds.size(); i != e; ++i) {
819 const SDep &Pred = ChainPreds[i];
820 RemovePred(SU, Pred);
822 AddPred(LoadSU, Pred);
824 for (unsigned i = 0, e = LoadPreds.size(); i != e; ++i) {
825 const SDep &Pred = LoadPreds[i];
826 RemovePred(SU, Pred);
828 AddPred(LoadSU, Pred);
830 for (unsigned i = 0, e = NodePreds.size(); i != e; ++i) {
831 const SDep &Pred = NodePreds[i];
832 RemovePred(SU, Pred);
833 AddPred(NewSU, Pred);
835 for (unsigned i = 0, e = NodeSuccs.size(); i != e; ++i) {
836 SDep D = NodeSuccs[i];
837 SUnit *SuccDep = D.getSUnit();
839 RemovePred(SuccDep, D);
842 // Balance register pressure.
843 if (AvailableQueue->tracksRegPressure() && SuccDep->isScheduled
844 && !D.isCtrl() && NewSU->NumRegDefsLeft > 0)
845 --NewSU->NumRegDefsLeft;
847 for (unsigned i = 0, e = ChainSuccs.size(); i != e; ++i) {
848 SDep D = ChainSuccs[i];
849 SUnit *SuccDep = D.getSUnit();
851 RemovePred(SuccDep, D);
858 // Add a data dependency to reflect that NewSU reads the value defined
860 AddPred(NewSU, SDep(LoadSU, SDep::Data, LoadSU->Latency));
863 AvailableQueue->addNode(LoadSU);
864 AvailableQueue->addNode(NewSU);
868 if (NewSU->NumSuccsLeft == 0) {
869 NewSU->isAvailable = true;
875 DEBUG(dbgs() << " Duplicating SU #" << SU->NodeNum << "\n");
876 NewSU = CreateClone(SU);
878 // New SUnit has the exact same predecessors.
879 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
881 if (!I->isArtificial())
884 // Only copy scheduled successors. Cut them from old node's successor
885 // list and move them over.
886 SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
887 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
889 if (I->isArtificial())
891 SUnit *SuccSU = I->getSUnit();
892 if (SuccSU->isScheduled) {
897 DelDeps.push_back(std::make_pair(SuccSU, D));
900 for (unsigned i = 0, e = DelDeps.size(); i != e; ++i)
901 RemovePred(DelDeps[i].first, DelDeps[i].second);
903 AvailableQueue->updateNode(SU);
904 AvailableQueue->addNode(NewSU);
910 /// InsertCopiesAndMoveSuccs - Insert register copies and move all
911 /// scheduled successors of the given SUnit to the last copy.
912 void ScheduleDAGRRList::InsertCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
913 const TargetRegisterClass *DestRC,
914 const TargetRegisterClass *SrcRC,
915 SmallVector<SUnit*, 2> &Copies) {
916 SUnit *CopyFromSU = CreateNewSUnit(NULL);
917 CopyFromSU->CopySrcRC = SrcRC;
918 CopyFromSU->CopyDstRC = DestRC;
920 SUnit *CopyToSU = CreateNewSUnit(NULL);
921 CopyToSU->CopySrcRC = DestRC;
922 CopyToSU->CopyDstRC = SrcRC;
924 // Only copy scheduled successors. Cut them from old node's successor
925 // list and move them over.
926 SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
927 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
929 if (I->isArtificial())
931 SUnit *SuccSU = I->getSUnit();
932 if (SuccSU->isScheduled) {
934 D.setSUnit(CopyToSU);
936 DelDeps.push_back(std::make_pair(SuccSU, *I));
939 for (unsigned i = 0, e = DelDeps.size(); i != e; ++i)
940 RemovePred(DelDeps[i].first, DelDeps[i].second);
942 AddPred(CopyFromSU, SDep(SU, SDep::Data, SU->Latency, Reg));
943 AddPred(CopyToSU, SDep(CopyFromSU, SDep::Data, CopyFromSU->Latency, 0));
945 AvailableQueue->updateNode(SU);
946 AvailableQueue->addNode(CopyFromSU);
947 AvailableQueue->addNode(CopyToSU);
948 Copies.push_back(CopyFromSU);
949 Copies.push_back(CopyToSU);
954 /// getPhysicalRegisterVT - Returns the ValueType of the physical register
955 /// definition of the specified node.
956 /// FIXME: Move to SelectionDAG?
957 static EVT getPhysicalRegisterVT(SDNode *N, unsigned Reg,
958 const TargetInstrInfo *TII) {
959 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
960 assert(TID.ImplicitDefs && "Physical reg def must be in implicit def list!");
961 unsigned NumRes = TID.getNumDefs();
962 for (const unsigned *ImpDef = TID.getImplicitDefs(); *ImpDef; ++ImpDef) {
967 return N->getValueType(NumRes);
970 /// CheckForLiveRegDef - Return true and update live register vector if the
971 /// specified register def of the specified SUnit clobbers any "live" registers.
972 static void CheckForLiveRegDef(SUnit *SU, unsigned Reg,
973 std::vector<SUnit*> &LiveRegDefs,
974 SmallSet<unsigned, 4> &RegAdded,
975 SmallVector<unsigned, 4> &LRegs,
976 const TargetRegisterInfo *TRI) {
977 for (const unsigned *AliasI = TRI->getOverlaps(Reg); *AliasI; ++AliasI) {
979 // Check if Ref is live.
980 if (!LiveRegDefs[Reg]) continue;
982 // Allow multiple uses of the same def.
983 if (LiveRegDefs[Reg] == SU) continue;
985 // Add Reg to the set of interfering live regs.
986 if (RegAdded.insert(Reg))
987 LRegs.push_back(Reg);
991 /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
992 /// scheduling of the given node to satisfy live physical register dependencies.
993 /// If the specific node is the last one that's available to schedule, do
994 /// whatever is necessary (i.e. backtracking or cloning) to make it possible.
995 bool ScheduleDAGRRList::
996 DelayForLiveRegsBottomUp(SUnit *SU, SmallVector<unsigned, 4> &LRegs) {
997 if (NumLiveRegs == 0)
1000 SmallSet<unsigned, 4> RegAdded;
1001 // If this node would clobber any "live" register, then it's not ready.
1003 // If SU is the currently live definition of the same register that it uses,
1004 // then we are free to schedule it.
1005 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1007 if (I->isAssignedRegDep() && LiveRegDefs[I->getReg()] != SU)
1008 CheckForLiveRegDef(I->getSUnit(), I->getReg(), LiveRegDefs,
1009 RegAdded, LRegs, TRI);
1012 for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) {
1013 if (Node->getOpcode() == ISD::INLINEASM) {
1014 // Inline asm can clobber physical defs.
1015 unsigned NumOps = Node->getNumOperands();
1016 if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
1017 --NumOps; // Ignore the glue operand.
1019 for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
1021 cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
1022 unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
1024 ++i; // Skip the ID value.
1025 if (InlineAsm::isRegDefKind(Flags) ||
1026 InlineAsm::isRegDefEarlyClobberKind(Flags)) {
1027 // Check for def of register or earlyclobber register.
1028 for (; NumVals; --NumVals, ++i) {
1029 unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
1030 if (TargetRegisterInfo::isPhysicalRegister(Reg))
1031 CheckForLiveRegDef(SU, Reg, LiveRegDefs, RegAdded, LRegs, TRI);
1039 if (!Node->isMachineOpcode())
1041 const TargetInstrDesc &TID = TII->get(Node->getMachineOpcode());
1042 if (!TID.ImplicitDefs)
1044 for (const unsigned *Reg = TID.ImplicitDefs; *Reg; ++Reg)
1045 CheckForLiveRegDef(SU, *Reg, LiveRegDefs, RegAdded, LRegs, TRI);
1048 return !LRegs.empty();
1051 /// Return a node that can be scheduled in this cycle. Requirements:
1052 /// (1) Ready: latency has been satisfied
1053 /// (2) No Hazards: resources are available
1054 /// (3) No Interferences: may unschedule to break register interferences.
1055 SUnit *ScheduleDAGRRList::PickNodeToScheduleBottomUp() {
1056 SmallVector<SUnit*, 4> Interferences;
1057 DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMap;
1059 SUnit *CurSU = AvailableQueue->pop();
1061 SmallVector<unsigned, 4> LRegs;
1062 if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
1064 LRegsMap.insert(std::make_pair(CurSU, LRegs));
1066 CurSU->isPending = true; // This SU is not in AvailableQueue right now.
1067 Interferences.push_back(CurSU);
1068 CurSU = AvailableQueue->pop();
1071 // Add the nodes that aren't ready back onto the available list.
1072 for (unsigned i = 0, e = Interferences.size(); i != e; ++i) {
1073 Interferences[i]->isPending = false;
1074 assert(Interferences[i]->isAvailable && "must still be available");
1075 AvailableQueue->push(Interferences[i]);
1080 // All candidates are delayed due to live physical reg dependencies.
1081 // Try backtracking, code duplication, or inserting cross class copies
1083 for (unsigned i = 0, e = Interferences.size(); i != e; ++i) {
1084 SUnit *TrySU = Interferences[i];
1085 SmallVector<unsigned, 4> &LRegs = LRegsMap[TrySU];
1087 // Try unscheduling up to the point where it's safe to schedule
1090 unsigned LiveCycle = UINT_MAX;
1091 for (unsigned j = 0, ee = LRegs.size(); j != ee; ++j) {
1092 unsigned Reg = LRegs[j];
1093 if (LiveRegGens[Reg]->getHeight() < LiveCycle) {
1094 BtSU = LiveRegGens[Reg];
1095 LiveCycle = BtSU->getHeight();
1098 if (!WillCreateCycle(TrySU, BtSU)) {
1099 BacktrackBottomUp(TrySU, BtSU);
1101 // Force the current node to be scheduled before the node that
1102 // requires the physical reg dep.
1103 if (BtSU->isAvailable) {
1104 BtSU->isAvailable = false;
1105 if (!BtSU->isPending)
1106 AvailableQueue->remove(BtSU);
1108 AddPred(TrySU, SDep(BtSU, SDep::Order, /*Latency=*/1,
1109 /*Reg=*/0, /*isNormalMemory=*/false,
1110 /*isMustAlias=*/false, /*isArtificial=*/true));
1112 // If one or more successors has been unscheduled, then the current
1113 // node is no longer avaialable. Schedule a successor that's now
1114 // available instead.
1115 if (!TrySU->isAvailable) {
1116 CurSU = AvailableQueue->pop();
1120 TrySU->isPending = false;
1121 Interferences.erase(Interferences.begin()+i);
1128 // Can't backtrack. If it's too expensive to copy the value, then try
1129 // duplicate the nodes that produces these "too expensive to copy"
1130 // values to break the dependency. In case even that doesn't work,
1131 // insert cross class copies.
1132 // If it's not too expensive, i.e. cost != -1, issue copies.
1133 SUnit *TrySU = Interferences[0];
1134 SmallVector<unsigned, 4> &LRegs = LRegsMap[TrySU];
1135 assert(LRegs.size() == 1 && "Can't handle this yet!");
1136 unsigned Reg = LRegs[0];
1137 SUnit *LRDef = LiveRegDefs[Reg];
1138 EVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
1139 const TargetRegisterClass *RC =
1140 TRI->getMinimalPhysRegClass(Reg, VT);
1141 const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
1143 // If cross copy register class is null, then it must be possible copy
1144 // the value directly. Do not try duplicate the def.
1147 NewDef = CopyAndMoveSuccessors(LRDef);
1151 // Issue copies, these can be expensive cross register class copies.
1152 SmallVector<SUnit*, 2> Copies;
1153 InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
1154 DEBUG(dbgs() << " Adding an edge from SU #" << TrySU->NodeNum
1155 << " to SU #" << Copies.front()->NodeNum << "\n");
1156 AddPred(TrySU, SDep(Copies.front(), SDep::Order, /*Latency=*/1,
1157 /*Reg=*/0, /*isNormalMemory=*/false,
1158 /*isMustAlias=*/false,
1159 /*isArtificial=*/true));
1160 NewDef = Copies.back();
1163 DEBUG(dbgs() << " Adding an edge from SU #" << NewDef->NodeNum
1164 << " to SU #" << TrySU->NodeNum << "\n");
1165 LiveRegDefs[Reg] = NewDef;
1166 AddPred(NewDef, SDep(TrySU, SDep::Order, /*Latency=*/1,
1167 /*Reg=*/0, /*isNormalMemory=*/false,
1168 /*isMustAlias=*/false,
1169 /*isArtificial=*/true));
1170 TrySU->isAvailable = false;
1174 assert(CurSU && "Unable to resolve live physical register dependencies!");
1176 // Add the nodes that aren't ready back onto the available list.
1177 for (unsigned i = 0, e = Interferences.size(); i != e; ++i) {
1178 Interferences[i]->isPending = false;
1179 // May no longer be available due to backtracking.
1180 if (Interferences[i]->isAvailable) {
1181 AvailableQueue->push(Interferences[i]);
1187 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
1189 void ScheduleDAGRRList::ListScheduleBottomUp() {
1190 // Release any predecessors of the special Exit node.
1191 ReleasePredecessors(&ExitSU);
1193 // Add root to Available queue.
1194 if (!SUnits.empty()) {
1195 SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()];
1196 assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!");
1197 RootSU->isAvailable = true;
1198 AvailableQueue->push(RootSU);
1201 // While Available queue is not empty, grab the node with the highest
1202 // priority. If it is not ready put it back. Schedule the node.
1203 Sequence.reserve(SUnits.size());
1204 while (!AvailableQueue->empty()) {
1205 DEBUG(dbgs() << "\n*** Examining Available\n";
1206 AvailableQueue->dump(this));
1208 // Pick the best node to schedule taking all constraints into
1210 SUnit *SU = PickNodeToScheduleBottomUp();
1212 AdvancePastStalls(SU);
1214 ScheduleNodeBottomUp(SU);
1216 while (AvailableQueue->empty() && !PendingQueue.empty()) {
1217 // Advance the cycle to free resources. Skip ahead to the next ready SU.
1218 assert(MinAvailableCycle < UINT_MAX && "MinAvailableCycle uninitialized");
1219 AdvanceToCycle(std::max(CurCycle + 1, MinAvailableCycle));
1223 // Reverse the order if it is bottom up.
1224 std::reverse(Sequence.begin(), Sequence.end());
1227 VerifySchedule(isBottomUp);
1231 //===----------------------------------------------------------------------===//
1232 // Top-Down Scheduling
1233 //===----------------------------------------------------------------------===//
1235 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
1236 /// the AvailableQueue if the count reaches zero. Also update its cycle bound.
1237 void ScheduleDAGRRList::ReleaseSucc(SUnit *SU, const SDep *SuccEdge) {
1238 SUnit *SuccSU = SuccEdge->getSUnit();
1241 if (SuccSU->NumPredsLeft == 0) {
1242 dbgs() << "*** Scheduling failed! ***\n";
1244 dbgs() << " has been released too many times!\n";
1245 llvm_unreachable(0);
1248 --SuccSU->NumPredsLeft;
1250 // If all the node's predecessors are scheduled, this node is ready
1251 // to be scheduled. Ignore the special ExitSU node.
1252 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU) {
1253 SuccSU->isAvailable = true;
1254 AvailableQueue->push(SuccSU);
1258 void ScheduleDAGRRList::ReleaseSuccessors(SUnit *SU) {
1259 // Top down: release successors
1260 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1262 assert(!I->isAssignedRegDep() &&
1263 "The list-tdrr scheduler doesn't yet support physreg dependencies!");
1265 ReleaseSucc(SU, &*I);
1269 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
1270 /// count of its successors. If a successor pending count is zero, add it to
1271 /// the Available queue.
1272 void ScheduleDAGRRList::ScheduleNodeTopDown(SUnit *SU) {
1273 DEBUG(dbgs() << "*** Scheduling [" << CurCycle << "]: ");
1274 DEBUG(SU->dump(this));
1276 assert(CurCycle >= SU->getDepth() && "Node scheduled above its depth!");
1277 SU->setDepthToAtLeast(CurCycle);
1278 Sequence.push_back(SU);
1280 ReleaseSuccessors(SU);
1281 SU->isScheduled = true;
1282 AvailableQueue->ScheduledNode(SU);
1285 /// ListScheduleTopDown - The main loop of list scheduling for top-down
1287 void ScheduleDAGRRList::ListScheduleTopDown() {
1288 AvailableQueue->setCurCycle(CurCycle);
1290 // Release any successors of the special Entry node.
1291 ReleaseSuccessors(&EntrySU);
1293 // All leaves to Available queue.
1294 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
1295 // It is available if it has no predecessors.
1296 if (SUnits[i].Preds.empty()) {
1297 AvailableQueue->push(&SUnits[i]);
1298 SUnits[i].isAvailable = true;
1302 // While Available queue is not empty, grab the node with the highest
1303 // priority. If it is not ready put it back. Schedule the node.
1304 Sequence.reserve(SUnits.size());
1305 while (!AvailableQueue->empty()) {
1306 SUnit *CurSU = AvailableQueue->pop();
1309 ScheduleNodeTopDown(CurSU);
1311 AvailableQueue->setCurCycle(CurCycle);
1315 VerifySchedule(isBottomUp);
1320 //===----------------------------------------------------------------------===//
1321 // RegReductionPriorityQueue Definition
1322 //===----------------------------------------------------------------------===//
1324 // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers
1325 // to reduce register pressure.
1328 class RegReductionPQBase;
1330 struct queue_sort : public std::binary_function<SUnit*, SUnit*, bool> {
1331 bool isReady(SUnit* SU, unsigned CurCycle) const { return true; }
1334 /// bu_ls_rr_sort - Priority function for bottom up register pressure
1335 // reduction scheduler.
1336 struct bu_ls_rr_sort : public queue_sort {
1339 HasReadyFilter = false
1342 RegReductionPQBase *SPQ;
1343 bu_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
1344 bu_ls_rr_sort(const bu_ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
1346 bool operator()(SUnit* left, SUnit* right) const;
1349 // td_ls_rr_sort - Priority function for top down register pressure reduction
1351 struct td_ls_rr_sort : public queue_sort {
1354 HasReadyFilter = false
1357 RegReductionPQBase *SPQ;
1358 td_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
1359 td_ls_rr_sort(const td_ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
1361 bool operator()(const SUnit* left, const SUnit* right) const;
1364 // src_ls_rr_sort - Priority function for source order scheduler.
1365 struct src_ls_rr_sort : public queue_sort {
1368 HasReadyFilter = false
1371 RegReductionPQBase *SPQ;
1372 src_ls_rr_sort(RegReductionPQBase *spq)
1374 src_ls_rr_sort(const src_ls_rr_sort &RHS)
1377 bool operator()(SUnit* left, SUnit* right) const;
1380 // hybrid_ls_rr_sort - Priority function for hybrid scheduler.
1381 struct hybrid_ls_rr_sort : public queue_sort {
1384 HasReadyFilter = false
1387 RegReductionPQBase *SPQ;
1388 hybrid_ls_rr_sort(RegReductionPQBase *spq)
1390 hybrid_ls_rr_sort(const hybrid_ls_rr_sort &RHS)
1393 bool isReady(SUnit *SU, unsigned CurCycle) const;
1395 bool operator()(SUnit* left, SUnit* right) const;
1398 // ilp_ls_rr_sort - Priority function for ILP (instruction level parallelism)
1400 struct ilp_ls_rr_sort : public queue_sort {
1403 HasReadyFilter = false
1406 RegReductionPQBase *SPQ;
1407 ilp_ls_rr_sort(RegReductionPQBase *spq)
1409 ilp_ls_rr_sort(const ilp_ls_rr_sort &RHS)
1412 bool isReady(SUnit *SU, unsigned CurCycle) const;
1414 bool operator()(SUnit* left, SUnit* right) const;
1417 class RegReductionPQBase : public SchedulingPriorityQueue {
1419 std::vector<SUnit*> Queue;
1420 unsigned CurQueueId;
1421 bool TracksRegPressure;
1423 // SUnits - The SUnits for the current graph.
1424 std::vector<SUnit> *SUnits;
1426 MachineFunction &MF;
1427 const TargetInstrInfo *TII;
1428 const TargetRegisterInfo *TRI;
1429 const TargetLowering *TLI;
1430 ScheduleDAGRRList *scheduleDAG;
1432 // SethiUllmanNumbers - The SethiUllman number for each node.
1433 std::vector<unsigned> SethiUllmanNumbers;
1435 /// RegPressure - Tracking current reg pressure per register class.
1437 std::vector<unsigned> RegPressure;
1439 /// RegLimit - Tracking the number of allocatable registers per register
1441 std::vector<unsigned> RegLimit;
1444 RegReductionPQBase(MachineFunction &mf,
1445 bool hasReadyFilter,
1447 const TargetInstrInfo *tii,
1448 const TargetRegisterInfo *tri,
1449 const TargetLowering *tli)
1450 : SchedulingPriorityQueue(hasReadyFilter),
1451 CurQueueId(0), TracksRegPressure(tracksrp),
1452 MF(mf), TII(tii), TRI(tri), TLI(tli), scheduleDAG(NULL) {
1453 if (TracksRegPressure) {
1454 unsigned NumRC = TRI->getNumRegClasses();
1455 RegLimit.resize(NumRC);
1456 RegPressure.resize(NumRC);
1457 std::fill(RegLimit.begin(), RegLimit.end(), 0);
1458 std::fill(RegPressure.begin(), RegPressure.end(), 0);
1459 for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
1460 E = TRI->regclass_end(); I != E; ++I)
1461 RegLimit[(*I)->getID()] = tri->getRegPressureLimit(*I, MF);
1465 void setScheduleDAG(ScheduleDAGRRList *scheduleDag) {
1466 scheduleDAG = scheduleDag;
1469 ScheduleHazardRecognizer* getHazardRec() {
1470 return scheduleDAG->getHazardRec();
1473 void initNodes(std::vector<SUnit> &sunits);
1475 void addNode(const SUnit *SU);
1477 void updateNode(const SUnit *SU);
1479 void releaseState() {
1481 SethiUllmanNumbers.clear();
1482 std::fill(RegPressure.begin(), RegPressure.end(), 0);
1485 unsigned getNodePriority(const SUnit *SU) const;
1487 unsigned getNodeOrdering(const SUnit *SU) const {
1488 return scheduleDAG->DAG->GetOrdering(SU->getNode());
1491 bool empty() const { return Queue.empty(); }
1493 void push(SUnit *U) {
1494 assert(!U->NodeQueueId && "Node in the queue already");
1495 U->NodeQueueId = ++CurQueueId;
1499 void remove(SUnit *SU) {
1500 assert(!Queue.empty() && "Queue is empty!");
1501 assert(SU->NodeQueueId != 0 && "Not in queue!");
1502 std::vector<SUnit *>::iterator I = std::find(Queue.begin(), Queue.end(),
1504 if (I != prior(Queue.end()))
1505 std::swap(*I, Queue.back());
1507 SU->NodeQueueId = 0;
1510 bool tracksRegPressure() const { return TracksRegPressure; }
1512 void dumpRegPressure() const;
1514 bool HighRegPressure(const SUnit *SU) const;
1516 bool MayReduceRegPressure(SUnit *SU) const;
1518 int RegPressureDiff(SUnit *SU, unsigned &LiveUses) const;
1520 void ScheduledNode(SUnit *SU);
1522 void UnscheduledNode(SUnit *SU);
1525 bool canClobber(const SUnit *SU, const SUnit *Op);
1526 void AddPseudoTwoAddrDeps();
1527 void PrescheduleNodesWithMultipleUses();
1528 void CalculateSethiUllmanNumbers();
1532 class RegReductionPriorityQueue : public RegReductionPQBase {
1533 static SUnit *popFromQueue(std::vector<SUnit*> &Q, SF &Picker) {
1534 std::vector<SUnit *>::iterator Best = Q.begin();
1535 for (std::vector<SUnit *>::iterator I = llvm::next(Q.begin()),
1536 E = Q.end(); I != E; ++I)
1537 if (Picker(*Best, *I))
1540 if (Best != prior(Q.end()))
1541 std::swap(*Best, Q.back());
1549 RegReductionPriorityQueue(MachineFunction &mf,
1551 const TargetInstrInfo *tii,
1552 const TargetRegisterInfo *tri,
1553 const TargetLowering *tli)
1554 : RegReductionPQBase(mf, SF::HasReadyFilter, tracksrp, tii, tri, tli),
1557 bool isBottomUp() const { return SF::IsBottomUp; }
1559 bool isReady(SUnit *U) const {
1560 return Picker.HasReadyFilter && Picker.isReady(U, getCurCycle());
1564 if (Queue.empty()) return NULL;
1566 SUnit *V = popFromQueue(Queue, Picker);
1571 void dump(ScheduleDAG *DAG) const {
1572 // Emulate pop() without clobbering NodeQueueIds.
1573 std::vector<SUnit*> DumpQueue = Queue;
1574 SF DumpPicker = Picker;
1575 while (!DumpQueue.empty()) {
1576 SUnit *SU = popFromQueue(DumpQueue, DumpPicker);
1578 dbgs() << "Height " << SU->getHeight() << ": ";
1580 dbgs() << "Depth " << SU->getDepth() << ": ";
1586 typedef RegReductionPriorityQueue<bu_ls_rr_sort>
1587 BURegReductionPriorityQueue;
1589 typedef RegReductionPriorityQueue<td_ls_rr_sort>
1590 TDRegReductionPriorityQueue;
1592 typedef RegReductionPriorityQueue<src_ls_rr_sort>
1593 SrcRegReductionPriorityQueue;
1595 typedef RegReductionPriorityQueue<hybrid_ls_rr_sort>
1596 HybridBURRPriorityQueue;
1598 typedef RegReductionPriorityQueue<ilp_ls_rr_sort>
1599 ILPBURRPriorityQueue;
1600 } // end anonymous namespace
1602 //===----------------------------------------------------------------------===//
1603 // Static Node Priority for Register Pressure Reduction
1604 //===----------------------------------------------------------------------===//
1606 /// CalcNodeSethiUllmanNumber - Compute Sethi Ullman number.
1607 /// Smaller number is the higher priority.
1609 CalcNodeSethiUllmanNumber(const SUnit *SU, std::vector<unsigned> &SUNumbers) {
1610 unsigned &SethiUllmanNumber = SUNumbers[SU->NodeNum];
1611 if (SethiUllmanNumber != 0)
1612 return SethiUllmanNumber;
1615 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1617 if (I->isCtrl()) continue; // ignore chain preds
1618 SUnit *PredSU = I->getSUnit();
1619 unsigned PredSethiUllman = CalcNodeSethiUllmanNumber(PredSU, SUNumbers);
1620 if (PredSethiUllman > SethiUllmanNumber) {
1621 SethiUllmanNumber = PredSethiUllman;
1623 } else if (PredSethiUllman == SethiUllmanNumber)
1627 SethiUllmanNumber += Extra;
1629 if (SethiUllmanNumber == 0)
1630 SethiUllmanNumber = 1;
1632 return SethiUllmanNumber;
1635 /// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all
1636 /// scheduling units.
1637 void RegReductionPQBase::CalculateSethiUllmanNumbers() {
1638 SethiUllmanNumbers.assign(SUnits->size(), 0);
1640 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
1641 CalcNodeSethiUllmanNumber(&(*SUnits)[i], SethiUllmanNumbers);
1644 void RegReductionPQBase::initNodes(std::vector<SUnit> &sunits) {
1646 // Add pseudo dependency edges for two-address nodes.
1647 AddPseudoTwoAddrDeps();
1648 // Reroute edges to nodes with multiple uses.
1649 if (!TracksRegPressure)
1650 PrescheduleNodesWithMultipleUses();
1651 // Calculate node priorities.
1652 CalculateSethiUllmanNumbers();
1655 void RegReductionPQBase::addNode(const SUnit *SU) {
1656 unsigned SUSize = SethiUllmanNumbers.size();
1657 if (SUnits->size() > SUSize)
1658 SethiUllmanNumbers.resize(SUSize*2, 0);
1659 CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
1662 void RegReductionPQBase::updateNode(const SUnit *SU) {
1663 SethiUllmanNumbers[SU->NodeNum] = 0;
1664 CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
1667 // Lower priority means schedule further down. For bottom-up scheduling, lower
1668 // priority SUs are scheduled before higher priority SUs.
1669 unsigned RegReductionPQBase::getNodePriority(const SUnit *SU) const {
1670 assert(SU->NodeNum < SethiUllmanNumbers.size());
1671 unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0;
1672 if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
1673 // CopyToReg should be close to its uses to facilitate coalescing and
1676 if (Opc == TargetOpcode::EXTRACT_SUBREG ||
1677 Opc == TargetOpcode::SUBREG_TO_REG ||
1678 Opc == TargetOpcode::INSERT_SUBREG)
1679 // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be
1680 // close to their uses to facilitate coalescing.
1682 if (SU->NumSuccs == 0 && SU->NumPreds != 0)
1683 // If SU does not have a register use, i.e. it doesn't produce a value
1684 // that would be consumed (e.g. store), then it terminates a chain of
1685 // computation. Give it a large SethiUllman number so it will be
1686 // scheduled right before its predecessors that it doesn't lengthen
1687 // their live ranges.
1689 if (SU->NumPreds == 0 && SU->NumSuccs != 0)
1690 // If SU does not have a register def, schedule it close to its uses
1691 // because it does not lengthen any live ranges.
1693 return SethiUllmanNumbers[SU->NodeNum];
1696 //===----------------------------------------------------------------------===//
1697 // Register Pressure Tracking
1698 //===----------------------------------------------------------------------===//
1700 void RegReductionPQBase::dumpRegPressure() const {
1701 for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
1702 E = TRI->regclass_end(); I != E; ++I) {
1703 const TargetRegisterClass *RC = *I;
1704 unsigned Id = RC->getID();
1705 unsigned RP = RegPressure[Id];
1707 DEBUG(dbgs() << RC->getName() << ": " << RP << " / " << RegLimit[Id]
1712 bool RegReductionPQBase::HighRegPressure(const SUnit *SU) const {
1716 for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
1720 SUnit *PredSU = I->getSUnit();
1721 // NumRegDefsLeft is zero when enough uses of this node have been scheduled
1722 // to cover the number of registers defined (they are all live).
1723 if (PredSU->NumRegDefsLeft == 0) {
1726 for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
1727 RegDefPos.IsValid(); RegDefPos.Advance()) {
1728 EVT VT = RegDefPos.GetValue();
1729 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1730 unsigned Cost = TLI->getRepRegClassCostFor(VT);
1731 if ((RegPressure[RCId] + Cost) >= RegLimit[RCId])
1738 bool RegReductionPQBase::MayReduceRegPressure(SUnit *SU) const {
1739 const SDNode *N = SU->getNode();
1741 if (!N->isMachineOpcode() || !SU->NumSuccs)
1744 unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
1745 for (unsigned i = 0; i != NumDefs; ++i) {
1746 EVT VT = N->getValueType(i);
1747 if (!N->hasAnyUseOfValue(i))
1749 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1750 if (RegPressure[RCId] >= RegLimit[RCId])
1756 // Compute the register pressure contribution by this instruction by count up
1757 // for uses that are not live and down for defs. Only count register classes
1758 // that are already under high pressure. As a side effect, compute the number of
1759 // uses of registers that are already live.
1761 // FIXME: This encompasses the logic in HighRegPressure and MayReduceRegPressure
1762 // so could probably be factored.
1763 int RegReductionPQBase::RegPressureDiff(SUnit *SU, unsigned &LiveUses) const {
1766 for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
1770 SUnit *PredSU = I->getSUnit();
1771 // NumRegDefsLeft is zero when enough uses of this node have been scheduled
1772 // to cover the number of registers defined (they are all live).
1773 if (PredSU->NumRegDefsLeft == 0) {
1774 if (PredSU->getNode()->isMachineOpcode())
1778 for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
1779 RegDefPos.IsValid(); RegDefPos.Advance()) {
1780 EVT VT = RegDefPos.GetValue();
1781 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1782 if (RegPressure[RCId] >= RegLimit[RCId])
1786 const SDNode *N = SU->getNode();
1788 if (!N || !N->isMachineOpcode() || !SU->NumSuccs)
1791 unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
1792 for (unsigned i = 0; i != NumDefs; ++i) {
1793 EVT VT = N->getValueType(i);
1794 if (!N->hasAnyUseOfValue(i))
1796 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1797 if (RegPressure[RCId] >= RegLimit[RCId])
1803 void RegReductionPQBase::ScheduledNode(SUnit *SU) {
1804 if (!TracksRegPressure)
1810 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1814 SUnit *PredSU = I->getSUnit();
1815 // NumRegDefsLeft is zero when enough uses of this node have been scheduled
1816 // to cover the number of registers defined (they are all live).
1817 if (PredSU->NumRegDefsLeft == 0) {
1820 // FIXME: The ScheduleDAG currently loses information about which of a
1821 // node's values is consumed by each dependence. Consequently, if the node
1822 // defines multiple register classes, we don't know which to pressurize
1823 // here. Instead the following loop consumes the register defs in an
1824 // arbitrary order. At least it handles the common case of clustered loads
1825 // to the same class. For precise liveness, each SDep needs to indicate the
1826 // result number. But that tightly couples the ScheduleDAG with the
1827 // SelectionDAG making updates tricky. A simpler hack would be to attach a
1828 // value type or register class to SDep.
1830 // The most important aspect of register tracking is balancing the increase
1831 // here with the reduction further below. Note that this SU may use multiple
1832 // defs in PredSU. The can't be determined here, but we've already
1833 // compensated by reducing NumRegDefsLeft in PredSU during
1834 // ScheduleDAGSDNodes::AddSchedEdges.
1835 --PredSU->NumRegDefsLeft;
1836 unsigned SkipRegDefs = PredSU->NumRegDefsLeft;
1837 for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
1838 RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
1841 EVT VT = RegDefPos.GetValue();
1842 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1843 RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
1848 // We should have this assert, but there may be dead SDNodes that never
1849 // materialize as SUnits, so they don't appear to generate liveness.
1850 //assert(SU->NumRegDefsLeft == 0 && "not all regdefs have scheduled uses");
1851 int SkipRegDefs = (int)SU->NumRegDefsLeft;
1852 for (ScheduleDAGSDNodes::RegDefIter RegDefPos(SU, scheduleDAG);
1853 RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
1854 if (SkipRegDefs > 0)
1856 EVT VT = RegDefPos.GetValue();
1857 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1858 if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT)) {
1859 // Register pressure tracking is imprecise. This can happen. But we try
1860 // hard not to let it happen because it likely results in poor scheduling.
1861 DEBUG(dbgs() << " SU(" << SU->NodeNum << ") has too many regdefs\n");
1862 RegPressure[RCId] = 0;
1865 RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
1871 void RegReductionPQBase::UnscheduledNode(SUnit *SU) {
1872 if (!TracksRegPressure)
1875 const SDNode *N = SU->getNode();
1878 if (!N->isMachineOpcode()) {
1879 if (N->getOpcode() != ISD::CopyToReg)
1882 unsigned Opc = N->getMachineOpcode();
1883 if (Opc == TargetOpcode::EXTRACT_SUBREG ||
1884 Opc == TargetOpcode::INSERT_SUBREG ||
1885 Opc == TargetOpcode::SUBREG_TO_REG ||
1886 Opc == TargetOpcode::REG_SEQUENCE ||
1887 Opc == TargetOpcode::IMPLICIT_DEF)
1891 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1895 SUnit *PredSU = I->getSUnit();
1896 // NumSuccsLeft counts all deps. Don't compare it with NumSuccs which only
1897 // counts data deps.
1898 if (PredSU->NumSuccsLeft != PredSU->Succs.size())
1900 const SDNode *PN = PredSU->getNode();
1901 if (!PN->isMachineOpcode()) {
1902 if (PN->getOpcode() == ISD::CopyFromReg) {
1903 EVT VT = PN->getValueType(0);
1904 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1905 RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
1909 unsigned POpc = PN->getMachineOpcode();
1910 if (POpc == TargetOpcode::IMPLICIT_DEF)
1912 if (POpc == TargetOpcode::EXTRACT_SUBREG) {
1913 EVT VT = PN->getOperand(0).getValueType();
1914 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1915 RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
1917 } else if (POpc == TargetOpcode::INSERT_SUBREG ||
1918 POpc == TargetOpcode::SUBREG_TO_REG) {
1919 EVT VT = PN->getValueType(0);
1920 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1921 RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
1924 unsigned NumDefs = TII->get(PN->getMachineOpcode()).getNumDefs();
1925 for (unsigned i = 0; i != NumDefs; ++i) {
1926 EVT VT = PN->getValueType(i);
1927 if (!PN->hasAnyUseOfValue(i))
1929 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1930 if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT))
1931 // Register pressure tracking is imprecise. This can happen.
1932 RegPressure[RCId] = 0;
1934 RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
1938 // Check for isMachineOpcode() as PrescheduleNodesWithMultipleUses()
1939 // may transfer data dependencies to CopyToReg.
1940 if (SU->NumSuccs && N->isMachineOpcode()) {
1941 unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
1942 for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
1943 EVT VT = N->getValueType(i);
1944 if (VT == MVT::Glue || VT == MVT::Other)
1946 if (!N->hasAnyUseOfValue(i))
1948 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1949 RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
1956 //===----------------------------------------------------------------------===//
1957 // Dynamic Node Priority for Register Pressure Reduction
1958 //===----------------------------------------------------------------------===//
1960 /// closestSucc - Returns the scheduled cycle of the successor which is
1961 /// closest to the current cycle.
1962 static unsigned closestSucc(const SUnit *SU) {
1963 unsigned MaxHeight = 0;
1964 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1966 if (I->isCtrl()) continue; // ignore chain succs
1967 unsigned Height = I->getSUnit()->getHeight();
1968 // If there are bunch of CopyToRegs stacked up, they should be considered
1969 // to be at the same position.
1970 if (I->getSUnit()->getNode() &&
1971 I->getSUnit()->getNode()->getOpcode() == ISD::CopyToReg)
1972 Height = closestSucc(I->getSUnit())+1;
1973 if (Height > MaxHeight)
1979 /// calcMaxScratches - Returns an cost estimate of the worse case requirement
1980 /// for scratch registers, i.e. number of data dependencies.
1981 static unsigned calcMaxScratches(const SUnit *SU) {
1982 unsigned Scratches = 0;
1983 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1985 if (I->isCtrl()) continue; // ignore chain preds
1991 /// hasOnlyLiveOutUse - Return true if SU has a single value successor that is a
1992 /// CopyToReg to a virtual register. This SU def is probably a liveout and
1993 /// it has no other use. It should be scheduled closer to the terminator.
1994 static bool hasOnlyLiveOutUses(const SUnit *SU) {
1995 bool RetVal = false;
1996 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1998 if (I->isCtrl()) continue;
1999 const SUnit *SuccSU = I->getSUnit();
2000 if (SuccSU->getNode() && SuccSU->getNode()->getOpcode() == ISD::CopyToReg) {
2002 cast<RegisterSDNode>(SuccSU->getNode()->getOperand(1))->getReg();
2003 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
2013 /// UnitsSharePred - Return true if the two scheduling units share a common
2014 /// data predecessor.
2015 static bool UnitsSharePred(const SUnit *left, const SUnit *right) {
2016 SmallSet<const SUnit*, 4> Preds;
2017 for (SUnit::const_pred_iterator I = left->Preds.begin(),E = left->Preds.end();
2019 if (I->isCtrl()) continue; // ignore chain preds
2020 Preds.insert(I->getSUnit());
2022 for (SUnit::const_pred_iterator I = right->Preds.begin(),E = right->Preds.end();
2024 if (I->isCtrl()) continue; // ignore chain preds
2025 if (Preds.count(I->getSUnit()))
2031 // Check for either a dependence (latency) or resource (hazard) stall.
2033 // Note: The ScheduleHazardRecognizer interface requires a non-const SU.
2034 static bool BUHasStall(SUnit *SU, int Height, RegReductionPQBase *SPQ) {
2035 if ((int)SPQ->getCurCycle() < Height) return true;
2036 if (SPQ->getHazardRec()->getHazardType(SU, 0)
2037 != ScheduleHazardRecognizer::NoHazard)
2042 // Return -1 if left has higher priority, 1 if right has higher priority.
2043 // Return 0 if latency-based priority is equivalent.
2044 static int BUCompareLatency(SUnit *left, SUnit *right, bool checkPref,
2045 RegReductionPQBase *SPQ) {
2046 // If the two nodes share an operand and one of them has a single
2047 // use that is a live out copy, favor the one that is live out. Otherwise
2048 // it will be difficult to eliminate the copy if the instruction is a
2049 // loop induction variable update. e.g.
2056 bool SharePred = UnitsSharePred(left, right);
2057 // FIXME: Only adjust if BB is a loop back edge.
2058 // FIXME: What's the cost of a copy?
2059 int LBonus = (SharePred && hasOnlyLiveOutUses(left)) ? 1 : 0;
2060 int RBonus = (SharePred && hasOnlyLiveOutUses(right)) ? 1 : 0;
2061 int LHeight = (int)left->getHeight() - LBonus;
2062 int RHeight = (int)right->getHeight() - RBonus;
2064 bool LStall = (!checkPref || left->SchedulingPref == Sched::Latency) &&
2065 BUHasStall(left, LHeight, SPQ);
2066 bool RStall = (!checkPref || right->SchedulingPref == Sched::Latency) &&
2067 BUHasStall(right, RHeight, SPQ);
2069 // If scheduling one of the node will cause a pipeline stall, delay it.
2070 // If scheduling either one of the node will cause a pipeline stall, sort
2071 // them according to their height.
2075 if (LHeight != RHeight)
2076 return LHeight > RHeight ? 1 : -1;
2080 // If either node is scheduling for latency, sort them by height/depth
2082 if (!checkPref || (left->SchedulingPref == Sched::Latency ||
2083 right->SchedulingPref == Sched::Latency)) {
2084 if (DisableSchedCycles) {
2085 if (LHeight != RHeight)
2086 return LHeight > RHeight ? 1 : -1;
2089 // If neither instruction stalls (!LStall && !RStall) then
2090 // its height is already covered so only its depth matters. We also reach
2091 // this if both stall but have the same height.
2092 unsigned LDepth = left->getDepth();
2093 unsigned RDepth = right->getDepth();
2094 if (LDepth != RDepth) {
2095 DEBUG(dbgs() << " Comparing latency of SU (" << left->NodeNum
2096 << ") depth " << LDepth << " vs SU (" << right->NodeNum
2097 << ") depth " << RDepth << "\n");
2098 return LDepth < RDepth ? 1 : -1;
2101 if (left->Latency != right->Latency)
2102 return left->Latency > right->Latency ? 1 : -1;
2107 static bool BURRSort(SUnit *left, SUnit *right, RegReductionPQBase *SPQ) {
2108 unsigned LPriority = SPQ->getNodePriority(left);
2109 unsigned RPriority = SPQ->getNodePriority(right);
2110 if (LPriority != RPriority) {
2111 DEBUG(++FactorCount[FactStatic]);
2112 return LPriority > RPriority;
2114 DEBUG(++FactorCount[FactOther]);
2116 // Try schedule def + use closer when Sethi-Ullman numbers are the same.
2121 // and the following instructions are both ready.
2125 // Then schedule t2 = op first.
2132 // This creates more short live intervals.
2133 unsigned LDist = closestSucc(left);
2134 unsigned RDist = closestSucc(right);
2136 return LDist < RDist;
2138 // How many registers becomes live when the node is scheduled.
2139 unsigned LScratch = calcMaxScratches(left);
2140 unsigned RScratch = calcMaxScratches(right);
2141 if (LScratch != RScratch)
2142 return LScratch > RScratch;
2144 if (!DisableSchedCycles) {
2145 int result = BUCompareLatency(left, right, false /*checkPref*/, SPQ);
2150 if (left->getHeight() != right->getHeight())
2151 return left->getHeight() > right->getHeight();
2153 if (left->getDepth() != right->getDepth())
2154 return left->getDepth() < right->getDepth();
2157 assert(left->NodeQueueId && right->NodeQueueId &&
2158 "NodeQueueId cannot be zero");
2159 return (left->NodeQueueId > right->NodeQueueId);
2163 bool bu_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2164 return BURRSort(left, right, SPQ);
2167 // Source order, otherwise bottom up.
2168 bool src_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2169 unsigned LOrder = SPQ->getNodeOrdering(left);
2170 unsigned ROrder = SPQ->getNodeOrdering(right);
2172 // Prefer an ordering where the lower the non-zero order number, the higher
2174 if ((LOrder || ROrder) && LOrder != ROrder)
2175 return LOrder != 0 && (LOrder < ROrder || ROrder == 0);
2177 return BURRSort(left, right, SPQ);
2180 // If the time between now and when the instruction will be ready can cover
2181 // the spill code, then avoid adding it to the ready queue. This gives long
2182 // stalls highest priority and allows hoisting across calls. It should also
2183 // speed up processing the available queue.
2184 bool hybrid_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
2185 static const unsigned ReadyDelay = 3;
2187 if (SPQ->MayReduceRegPressure(SU)) return true;
2189 if (SU->getHeight() > (CurCycle + ReadyDelay)) return false;
2191 if (SPQ->getHazardRec()->getHazardType(SU, -ReadyDelay)
2192 != ScheduleHazardRecognizer::NoHazard)
2198 // Return true if right should be scheduled with higher priority than left.
2199 bool hybrid_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2200 if (left->isCall || right->isCall)
2201 // No way to compute latency of calls.
2202 return BURRSort(left, right, SPQ);
2204 bool LHigh = SPQ->HighRegPressure(left);
2205 bool RHigh = SPQ->HighRegPressure(right);
2206 // Avoid causing spills. If register pressure is high, schedule for
2207 // register pressure reduction.
2208 if (LHigh && !RHigh) {
2209 DEBUG(dbgs() << " pressure SU(" << left->NodeNum << ") > SU("
2210 << right->NodeNum << ")\n");
2213 else if (!LHigh && RHigh) {
2214 DEBUG(dbgs() << " pressure SU(" << right->NodeNum << ") > SU("
2215 << left->NodeNum << ")\n");
2218 else if (!LHigh && !RHigh) {
2219 int result = BUCompareLatency(left, right, true /*checkPref*/, SPQ);
2223 return BURRSort(left, right, SPQ);
2226 // Schedule as many instructions in each cycle as possible. So don't make an
2227 // instruction available unless it is ready in the current cycle.
2228 bool ilp_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
2229 if (SU->getHeight() > CurCycle) return false;
2231 if (SPQ->getHazardRec()->getHazardType(SU, 0)
2232 != ScheduleHazardRecognizer::NoHazard)
2238 static bool canEnableCoalescing(SUnit *SU) {
2239 unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0;
2240 if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
2241 // CopyToReg should be close to its uses to facilitate coalescing and
2245 if (Opc == TargetOpcode::EXTRACT_SUBREG ||
2246 Opc == TargetOpcode::SUBREG_TO_REG ||
2247 Opc == TargetOpcode::INSERT_SUBREG)
2248 // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be
2249 // close to their uses to facilitate coalescing.
2252 if (SU->NumPreds == 0 && SU->NumSuccs != 0)
2253 // If SU does not have a register def, schedule it close to its uses
2254 // because it does not lengthen any live ranges.
2260 // list-ilp is currently an experimental scheduler that allows various
2261 // heuristics to be enabled prior to the normal register reduction logic.
2262 bool ilp_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2263 if (left->isCall || right->isCall)
2264 // No way to compute latency of calls.
2265 return BURRSort(left, right, SPQ);
2267 unsigned LLiveUses = 0, RLiveUses = 0;
2268 int LPDiff = 0, RPDiff = 0;
2269 if (!DisableSchedRegPressure || !DisableSchedLiveUses) {
2270 LPDiff = SPQ->RegPressureDiff(left, LLiveUses);
2271 RPDiff = SPQ->RegPressureDiff(right, RLiveUses);
2273 if (!DisableSchedRegPressure && LPDiff != RPDiff) {
2274 DEBUG(++FactorCount[FactPressureDiff]);
2275 DEBUG(dbgs() << "RegPressureDiff SU(" << left->NodeNum << "): " << LPDiff
2276 << " != SU(" << right->NodeNum << "): " << RPDiff << "\n");
2277 return LPDiff > RPDiff;
2280 if (!DisableSchedRegPressure && (LPDiff > 0 || RPDiff > 0)) {
2281 bool LReduce = canEnableCoalescing(left);
2282 bool RReduce = canEnableCoalescing(right);
2283 DEBUG(if (LReduce != RReduce) ++FactorCount[FactPressureDiff]);
2284 if (LReduce && !RReduce) return false;
2285 if (RReduce && !LReduce) return true;
2288 if (!DisableSchedLiveUses && (LLiveUses != RLiveUses)) {
2289 DEBUG(dbgs() << "Live uses SU(" << left->NodeNum << "): " << LLiveUses
2290 << " != SU(" << right->NodeNum << "): " << RLiveUses << "\n");
2291 DEBUG(++FactorCount[FactRegUses]);
2292 return LLiveUses < RLiveUses;
2295 if (!DisableSchedStalls) {
2296 bool LStall = BUHasStall(left, left->getHeight(), SPQ);
2297 bool RStall = BUHasStall(right, right->getHeight(), SPQ);
2298 if (LStall != RStall) {
2299 DEBUG(++FactorCount[FactHeight]);
2300 return left->getHeight() > right->getHeight();
2304 if (!DisableSchedCriticalPath) {
2305 int spread = (int)left->getDepth() - (int)right->getDepth();
2306 if (std::abs(spread) > MaxReorderWindow) {
2307 DEBUG(dbgs() << "Depth of SU(" << left->NodeNum << "): "
2308 << left->getDepth() << " != SU(" << right->NodeNum << "): "
2309 << right->getDepth() << "\n");
2310 DEBUG(++FactorCount[FactDepth]);
2311 return left->getDepth() < right->getDepth();
2315 if (!DisableSchedHeight && left->getHeight() != right->getHeight()) {
2316 int spread = (int)left->getHeight() - (int)right->getHeight();
2317 if (std::abs(spread) > MaxReorderWindow) {
2318 DEBUG(++FactorCount[FactHeight]);
2319 return left->getHeight() > right->getHeight();
2323 return BURRSort(left, right, SPQ);
2326 //===----------------------------------------------------------------------===//
2327 // Preschedule for Register Pressure
2328 //===----------------------------------------------------------------------===//
2330 bool RegReductionPQBase::canClobber(const SUnit *SU, const SUnit *Op) {
2331 if (SU->isTwoAddress) {
2332 unsigned Opc = SU->getNode()->getMachineOpcode();
2333 const TargetInstrDesc &TID = TII->get(Opc);
2334 unsigned NumRes = TID.getNumDefs();
2335 unsigned NumOps = TID.getNumOperands() - NumRes;
2336 for (unsigned i = 0; i != NumOps; ++i) {
2337 if (TID.getOperandConstraint(i+NumRes, TOI::TIED_TO) != -1) {
2338 SDNode *DU = SU->getNode()->getOperand(i).getNode();
2339 if (DU->getNodeId() != -1 &&
2340 Op->OrigNode == &(*SUnits)[DU->getNodeId()])
2348 /// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's
2349 /// physical register defs.
2350 static bool canClobberPhysRegDefs(const SUnit *SuccSU, const SUnit *SU,
2351 const TargetInstrInfo *TII,
2352 const TargetRegisterInfo *TRI) {
2353 SDNode *N = SuccSU->getNode();
2354 unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2355 const unsigned *ImpDefs = TII->get(N->getMachineOpcode()).getImplicitDefs();
2356 assert(ImpDefs && "Caller should check hasPhysRegDefs");
2357 for (const SDNode *SUNode = SU->getNode(); SUNode;
2358 SUNode = SUNode->getGluedNode()) {
2359 if (!SUNode->isMachineOpcode())
2361 const unsigned *SUImpDefs =
2362 TII->get(SUNode->getMachineOpcode()).getImplicitDefs();
2365 for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
2366 EVT VT = N->getValueType(i);
2367 if (VT == MVT::Glue || VT == MVT::Other)
2369 if (!N->hasAnyUseOfValue(i))
2371 unsigned Reg = ImpDefs[i - NumDefs];
2372 for (;*SUImpDefs; ++SUImpDefs) {
2373 unsigned SUReg = *SUImpDefs;
2374 if (TRI->regsOverlap(Reg, SUReg))
2382 /// PrescheduleNodesWithMultipleUses - Nodes with multiple uses
2383 /// are not handled well by the general register pressure reduction
2384 /// heuristics. When presented with code like this:
2393 /// the heuristics tend to push the store up, but since the
2394 /// operand of the store has another use (U), this would increase
2395 /// the length of that other use (the U->N edge).
2397 /// This function transforms code like the above to route U's
2398 /// dependence through the store when possible, like this:
2409 /// This results in the store being scheduled immediately
2410 /// after N, which shortens the U->N live range, reducing
2411 /// register pressure.
2413 void RegReductionPQBase::PrescheduleNodesWithMultipleUses() {
2414 // Visit all the nodes in topological order, working top-down.
2415 for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
2416 SUnit *SU = &(*SUnits)[i];
2417 // For now, only look at nodes with no data successors, such as stores.
2418 // These are especially important, due to the heuristics in
2419 // getNodePriority for nodes with no data successors.
2420 if (SU->NumSuccs != 0)
2422 // For now, only look at nodes with exactly one data predecessor.
2423 if (SU->NumPreds != 1)
2425 // Avoid prescheduling copies to virtual registers, which don't behave
2426 // like other nodes from the perspective of scheduling heuristics.
2427 if (SDNode *N = SU->getNode())
2428 if (N->getOpcode() == ISD::CopyToReg &&
2429 TargetRegisterInfo::isVirtualRegister
2430 (cast<RegisterSDNode>(N->getOperand(1))->getReg()))
2433 // Locate the single data predecessor.
2435 for (SUnit::const_pred_iterator II = SU->Preds.begin(),
2436 EE = SU->Preds.end(); II != EE; ++II)
2437 if (!II->isCtrl()) {
2438 PredSU = II->getSUnit();
2443 // Don't rewrite edges that carry physregs, because that requires additional
2444 // support infrastructure.
2445 if (PredSU->hasPhysRegDefs)
2447 // Short-circuit the case where SU is PredSU's only data successor.
2448 if (PredSU->NumSuccs == 1)
2450 // Avoid prescheduling to copies from virtual registers, which don't behave
2451 // like other nodes from the perspective of scheduling heuristics.
2452 if (SDNode *N = SU->getNode())
2453 if (N->getOpcode() == ISD::CopyFromReg &&
2454 TargetRegisterInfo::isVirtualRegister
2455 (cast<RegisterSDNode>(N->getOperand(1))->getReg()))
2458 // Perform checks on the successors of PredSU.
2459 for (SUnit::const_succ_iterator II = PredSU->Succs.begin(),
2460 EE = PredSU->Succs.end(); II != EE; ++II) {
2461 SUnit *PredSuccSU = II->getSUnit();
2462 if (PredSuccSU == SU) continue;
2463 // If PredSU has another successor with no data successors, for
2464 // now don't attempt to choose either over the other.
2465 if (PredSuccSU->NumSuccs == 0)
2466 goto outer_loop_continue;
2467 // Don't break physical register dependencies.
2468 if (SU->hasPhysRegClobbers && PredSuccSU->hasPhysRegDefs)
2469 if (canClobberPhysRegDefs(PredSuccSU, SU, TII, TRI))
2470 goto outer_loop_continue;
2471 // Don't introduce graph cycles.
2472 if (scheduleDAG->IsReachable(SU, PredSuccSU))
2473 goto outer_loop_continue;
2476 // Ok, the transformation is safe and the heuristics suggest it is
2477 // profitable. Update the graph.
2478 DEBUG(dbgs() << " Prescheduling SU #" << SU->NodeNum
2479 << " next to PredSU #" << PredSU->NodeNum
2480 << " to guide scheduling in the presence of multiple uses\n");
2481 for (unsigned i = 0; i != PredSU->Succs.size(); ++i) {
2482 SDep Edge = PredSU->Succs[i];
2483 assert(!Edge.isAssignedRegDep());
2484 SUnit *SuccSU = Edge.getSUnit();
2486 Edge.setSUnit(PredSU);
2487 scheduleDAG->RemovePred(SuccSU, Edge);
2488 scheduleDAG->AddPred(SU, Edge);
2490 scheduleDAG->AddPred(SuccSU, Edge);
2494 outer_loop_continue:;
2498 /// AddPseudoTwoAddrDeps - If two nodes share an operand and one of them uses
2499 /// it as a def&use operand. Add a pseudo control edge from it to the other
2500 /// node (if it won't create a cycle) so the two-address one will be scheduled
2501 /// first (lower in the schedule). If both nodes are two-address, favor the
2502 /// one that has a CopyToReg use (more likely to be a loop induction update).
2503 /// If both are two-address, but one is commutable while the other is not
2504 /// commutable, favor the one that's not commutable.
2505 void RegReductionPQBase::AddPseudoTwoAddrDeps() {
2506 for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
2507 SUnit *SU = &(*SUnits)[i];
2508 if (!SU->isTwoAddress)
2511 SDNode *Node = SU->getNode();
2512 if (!Node || !Node->isMachineOpcode() || SU->getNode()->getGluedNode())
2515 bool isLiveOut = hasOnlyLiveOutUses(SU);
2516 unsigned Opc = Node->getMachineOpcode();
2517 const TargetInstrDesc &TID = TII->get(Opc);
2518 unsigned NumRes = TID.getNumDefs();
2519 unsigned NumOps = TID.getNumOperands() - NumRes;
2520 for (unsigned j = 0; j != NumOps; ++j) {
2521 if (TID.getOperandConstraint(j+NumRes, TOI::TIED_TO) == -1)
2523 SDNode *DU = SU->getNode()->getOperand(j).getNode();
2524 if (DU->getNodeId() == -1)
2526 const SUnit *DUSU = &(*SUnits)[DU->getNodeId()];
2527 if (!DUSU) continue;
2528 for (SUnit::const_succ_iterator I = DUSU->Succs.begin(),
2529 E = DUSU->Succs.end(); I != E; ++I) {
2530 if (I->isCtrl()) continue;
2531 SUnit *SuccSU = I->getSUnit();
2534 // Be conservative. Ignore if nodes aren't at roughly the same
2535 // depth and height.
2536 if (SuccSU->getHeight() < SU->getHeight() &&
2537 (SU->getHeight() - SuccSU->getHeight()) > 1)
2539 // Skip past COPY_TO_REGCLASS nodes, so that the pseudo edge
2540 // constrains whatever is using the copy, instead of the copy
2541 // itself. In the case that the copy is coalesced, this
2542 // preserves the intent of the pseudo two-address heurietics.
2543 while (SuccSU->Succs.size() == 1 &&
2544 SuccSU->getNode()->isMachineOpcode() &&
2545 SuccSU->getNode()->getMachineOpcode() ==
2546 TargetOpcode::COPY_TO_REGCLASS)
2547 SuccSU = SuccSU->Succs.front().getSUnit();
2548 // Don't constrain non-instruction nodes.
2549 if (!SuccSU->getNode() || !SuccSU->getNode()->isMachineOpcode())
2551 // Don't constrain nodes with physical register defs if the
2552 // predecessor can clobber them.
2553 if (SuccSU->hasPhysRegDefs && SU->hasPhysRegClobbers) {
2554 if (canClobberPhysRegDefs(SuccSU, SU, TII, TRI))
2557 // Don't constrain EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG;
2558 // these may be coalesced away. We want them close to their uses.
2559 unsigned SuccOpc = SuccSU->getNode()->getMachineOpcode();
2560 if (SuccOpc == TargetOpcode::EXTRACT_SUBREG ||
2561 SuccOpc == TargetOpcode::INSERT_SUBREG ||
2562 SuccOpc == TargetOpcode::SUBREG_TO_REG)
2564 if ((!canClobber(SuccSU, DUSU) ||
2565 (isLiveOut && !hasOnlyLiveOutUses(SuccSU)) ||
2566 (!SU->isCommutable && SuccSU->isCommutable)) &&
2567 !scheduleDAG->IsReachable(SuccSU, SU)) {
2568 DEBUG(dbgs() << " Adding a pseudo-two-addr edge from SU #"
2569 << SU->NodeNum << " to SU #" << SuccSU->NodeNum << "\n");
2570 scheduleDAG->AddPred(SU, SDep(SuccSU, SDep::Order, /*Latency=*/0,
2571 /*Reg=*/0, /*isNormalMemory=*/false,
2572 /*isMustAlias=*/false,
2573 /*isArtificial=*/true));
2580 /// LimitedSumOfUnscheduledPredsOfSuccs - Compute the sum of the unscheduled
2581 /// predecessors of the successors of the SUnit SU. Stop when the provided
2582 /// limit is exceeded.
2583 static unsigned LimitedSumOfUnscheduledPredsOfSuccs(const SUnit *SU,
2586 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
2588 const SUnit *SuccSU = I->getSUnit();
2589 for (SUnit::const_pred_iterator II = SuccSU->Preds.begin(),
2590 EE = SuccSU->Preds.end(); II != EE; ++II) {
2591 SUnit *PredSU = II->getSUnit();
2592 if (!PredSU->isScheduled)
2602 bool td_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
2603 unsigned LPriority = SPQ->getNodePriority(left);
2604 unsigned RPriority = SPQ->getNodePriority(right);
2605 bool LIsTarget = left->getNode() && left->getNode()->isMachineOpcode();
2606 bool RIsTarget = right->getNode() && right->getNode()->isMachineOpcode();
2607 bool LIsFloater = LIsTarget && left->NumPreds == 0;
2608 bool RIsFloater = RIsTarget && right->NumPreds == 0;
2609 unsigned LBonus = (LimitedSumOfUnscheduledPredsOfSuccs(left,1) == 1) ? 2 : 0;
2610 unsigned RBonus = (LimitedSumOfUnscheduledPredsOfSuccs(right,1) == 1) ? 2 : 0;
2612 if (left->NumSuccs == 0 && right->NumSuccs != 0)
2614 else if (left->NumSuccs != 0 && right->NumSuccs == 0)
2621 if (left->NumSuccs == 1)
2623 if (right->NumSuccs == 1)
2626 if (LPriority+LBonus != RPriority+RBonus)
2627 return LPriority+LBonus < RPriority+RBonus;
2629 if (left->getDepth() != right->getDepth())
2630 return left->getDepth() < right->getDepth();
2632 if (left->NumSuccsLeft != right->NumSuccsLeft)
2633 return left->NumSuccsLeft > right->NumSuccsLeft;
2635 assert(left->NodeQueueId && right->NodeQueueId &&
2636 "NodeQueueId cannot be zero");
2637 return (left->NodeQueueId > right->NodeQueueId);
2640 //===----------------------------------------------------------------------===//
2641 // Public Constructor Functions
2642 //===----------------------------------------------------------------------===//
2644 llvm::ScheduleDAGSDNodes *
2645 llvm::createBURRListDAGScheduler(SelectionDAGISel *IS,
2646 CodeGenOpt::Level OptLevel) {
2647 const TargetMachine &TM = IS->TM;
2648 const TargetInstrInfo *TII = TM.getInstrInfo();
2649 const TargetRegisterInfo *TRI = TM.getRegisterInfo();
2651 BURegReductionPriorityQueue *PQ =
2652 new BURegReductionPriorityQueue(*IS->MF, false, TII, TRI, 0);
2653 ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
2654 PQ->setScheduleDAG(SD);
2658 llvm::ScheduleDAGSDNodes *
2659 llvm::createTDRRListDAGScheduler(SelectionDAGISel *IS,
2660 CodeGenOpt::Level OptLevel) {
2661 const TargetMachine &TM = IS->TM;
2662 const TargetInstrInfo *TII = TM.getInstrInfo();
2663 const TargetRegisterInfo *TRI = TM.getRegisterInfo();
2665 TDRegReductionPriorityQueue *PQ =
2666 new TDRegReductionPriorityQueue(*IS->MF, false, TII, TRI, 0);
2667 ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
2668 PQ->setScheduleDAG(SD);
2672 llvm::ScheduleDAGSDNodes *
2673 llvm::createSourceListDAGScheduler(SelectionDAGISel *IS,
2674 CodeGenOpt::Level OptLevel) {
2675 const TargetMachine &TM = IS->TM;
2676 const TargetInstrInfo *TII = TM.getInstrInfo();
2677 const TargetRegisterInfo *TRI = TM.getRegisterInfo();
2679 SrcRegReductionPriorityQueue *PQ =
2680 new SrcRegReductionPriorityQueue(*IS->MF, false, TII, TRI, 0);
2681 ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
2682 PQ->setScheduleDAG(SD);
2686 llvm::ScheduleDAGSDNodes *
2687 llvm::createHybridListDAGScheduler(SelectionDAGISel *IS,
2688 CodeGenOpt::Level OptLevel) {
2689 const TargetMachine &TM = IS->TM;
2690 const TargetInstrInfo *TII = TM.getInstrInfo();
2691 const TargetRegisterInfo *TRI = TM.getRegisterInfo();
2692 const TargetLowering *TLI = &IS->getTargetLowering();
2694 HybridBURRPriorityQueue *PQ =
2695 new HybridBURRPriorityQueue(*IS->MF, true, TII, TRI, TLI);
2697 ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
2698 PQ->setScheduleDAG(SD);
2702 llvm::ScheduleDAGSDNodes *
2703 llvm::createILPListDAGScheduler(SelectionDAGISel *IS,
2704 CodeGenOpt::Level OptLevel) {
2705 const TargetMachine &TM = IS->TM;
2706 const TargetInstrInfo *TII = TM.getInstrInfo();
2707 const TargetRegisterInfo *TRI = TM.getRegisterInfo();
2708 const TargetLowering *TLI = &IS->getTargetLowering();
2710 ILPBURRPriorityQueue *PQ =
2711 new ILPBURRPriorityQueue(*IS->MF, true, TII, TRI, TLI);
2712 ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
2713 PQ->setScheduleDAG(SD);