1 //===- ScheduleDAGRRList.cpp - Reg pressure reduction list scheduler ------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This implements bottom-up and top-down register pressure reduction list
10 // schedulers, using standard algorithms.  The basic approach uses a priority
11 // queue of available nodes to schedule.  One at a time, nodes are taken from
12 // the priority queue (thus in priority order), checked for legality to
13 // schedule, and emitted if legal.
14 //
15 //===----------------------------------------------------------------------===//
16 
17 #include "ScheduleDAGSDNodes.h"
18 #include "llvm/ADT/ArrayRef.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/CodeGen/ISDOpcodes.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineOperand.h"
27 #include "llvm/CodeGen/Register.h"
28 #include "llvm/CodeGen/ScheduleDAG.h"
29 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
30 #include "llvm/CodeGen/SchedulerRegistry.h"
31 #include "llvm/CodeGen/SelectionDAGISel.h"
32 #include "llvm/CodeGen/SelectionDAGNodes.h"
33 #include "llvm/CodeGen/TargetInstrInfo.h"
34 #include "llvm/CodeGen/TargetLowering.h"
35 #include "llvm/CodeGen/TargetOpcodes.h"
36 #include "llvm/CodeGen/TargetRegisterInfo.h"
37 #include "llvm/CodeGen/TargetSubtargetInfo.h"
38 #include "llvm/Config/llvm-config.h"
39 #include "llvm/IR/InlineAsm.h"
40 #include "llvm/MC/MCInstrDesc.h"
41 #include "llvm/MC/MCRegisterInfo.h"
42 #include "llvm/Support/Casting.h"
43 #include "llvm/Support/CodeGen.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Compiler.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/MachineValueType.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include <algorithm>
51 #include <cassert>
52 #include <cstdint>
53 #include <cstdlib>
54 #include <iterator>
55 #include <limits>
56 #include <memory>
57 #include <utility>
58 #include <vector>
59 
60 using namespace llvm;
61 
62 #define DEBUG_TYPE "pre-RA-sched"
63 
64 STATISTIC(NumBacktracks, "Number of times scheduler backtracked");
65 STATISTIC(NumUnfolds,    "Number of nodes unfolded");
66 STATISTIC(NumDups,       "Number of duplicated nodes");
67 STATISTIC(NumPRCopies,   "Number of physical register copies");
68 
69 static RegisterScheduler
70   burrListDAGScheduler("list-burr",
71                        "Bottom-up register reduction list scheduling",
72                        createBURRListDAGScheduler);
73 
74 static RegisterScheduler
75   sourceListDAGScheduler("source",
76                          "Similar to list-burr but schedules in source "
77                          "order when possible",
78                          createSourceListDAGScheduler);
79 
80 static RegisterScheduler
81   hybridListDAGScheduler("list-hybrid",
82                          "Bottom-up register pressure aware list scheduling "
83                          "which tries to balance latency and register pressure",
84                          createHybridListDAGScheduler);
85 
86 static RegisterScheduler
87   ILPListDAGScheduler("list-ilp",
88                       "Bottom-up register pressure aware list scheduling "
89                       "which tries to balance ILP and register pressure",
90                       createILPListDAGScheduler);
91 
92 static cl::opt<bool> DisableSchedCycles(
93   "disable-sched-cycles", cl::Hidden, cl::init(false),
94   cl::desc("Disable cycle-level precision during preRA scheduling"));
95 
96 // Temporary sched=list-ilp flags until the heuristics are robust.
97 // Some options are also available under sched=list-hybrid.
98 static cl::opt<bool> DisableSchedRegPressure(
99   "disable-sched-reg-pressure", cl::Hidden, cl::init(false),
100   cl::desc("Disable regpressure priority in sched=list-ilp"));
101 static cl::opt<bool> DisableSchedLiveUses(
102   "disable-sched-live-uses", cl::Hidden, cl::init(true),
103   cl::desc("Disable live use priority in sched=list-ilp"));
104 static cl::opt<bool> DisableSchedVRegCycle(
105   "disable-sched-vrcycle", cl::Hidden, cl::init(false),
106   cl::desc("Disable virtual register cycle interference checks"));
107 static cl::opt<bool> DisableSchedPhysRegJoin(
108   "disable-sched-physreg-join", cl::Hidden, cl::init(false),
109   cl::desc("Disable physreg def-use affinity"));
110 static cl::opt<bool> DisableSchedStalls(
111   "disable-sched-stalls", cl::Hidden, cl::init(true),
112   cl::desc("Disable no-stall priority in sched=list-ilp"));
113 static cl::opt<bool> DisableSchedCriticalPath(
114   "disable-sched-critical-path", cl::Hidden, cl::init(false),
115   cl::desc("Disable critical path priority in sched=list-ilp"));
116 static cl::opt<bool> DisableSchedHeight(
117   "disable-sched-height", cl::Hidden, cl::init(false),
118   cl::desc("Disable scheduled-height priority in sched=list-ilp"));
119 static cl::opt<bool> Disable2AddrHack(
120   "disable-2addr-hack", cl::Hidden, cl::init(true),
121   cl::desc("Disable scheduler's two-address hack"));
122 
123 static cl::opt<int> MaxReorderWindow(
124   "max-sched-reorder", cl::Hidden, cl::init(6),
125   cl::desc("Number of instructions to allow ahead of the critical path "
126            "in sched=list-ilp"));
127 
128 static cl::opt<unsigned> AvgIPC(
129   "sched-avg-ipc", cl::Hidden, cl::init(1),
130   cl::desc("Average inst/cycle whan no target itinerary exists."));
131 
132 namespace {
133 
134 //===----------------------------------------------------------------------===//
135 /// ScheduleDAGRRList - The actual register reduction list scheduler
136 /// implementation.  This supports both top-down and bottom-up scheduling.
137 ///
138 class ScheduleDAGRRList : public ScheduleDAGSDNodes {
139 private:
140   /// NeedLatency - True if the scheduler will make use of latency information.
141   bool NeedLatency;
142 
143   /// AvailableQueue - The priority queue to use for the available SUnits.
144   SchedulingPriorityQueue *AvailableQueue;
145 
146   /// PendingQueue - This contains all of the instructions whose operands have
147   /// been issued, but their results are not ready yet (due to the latency of
148   /// the operation).  Once the operands becomes available, the instruction is
149   /// added to the AvailableQueue.
150   std::vector<SUnit *> PendingQueue;
151 
152   /// HazardRec - The hazard recognizer to use.
153   ScheduleHazardRecognizer *HazardRec;
154 
155   /// CurCycle - The current scheduler state corresponds to this cycle.
156   unsigned CurCycle = 0;
157 
158   /// MinAvailableCycle - Cycle of the soonest available instruction.
159   unsigned MinAvailableCycle;
160 
161   /// IssueCount - Count instructions issued in this cycle
162   /// Currently valid only for bottom-up scheduling.
163   unsigned IssueCount;
164 
165   /// LiveRegDefs - A set of physical registers and their definition
166   /// that are "live". These nodes must be scheduled before any other nodes that
167   /// modifies the registers can be scheduled.
168   unsigned NumLiveRegs;
169   std::unique_ptr<SUnit*[]> LiveRegDefs;
170   std::unique_ptr<SUnit*[]> LiveRegGens;
171 
172   // Collect interferences between physical register use/defs.
173   // Each interference is an SUnit and set of physical registers.
174   SmallVector<SUnit*, 4> Interferences;
175 
176   using LRegsMapT = DenseMap<SUnit *, SmallVector<unsigned, 4>>;
177 
178   LRegsMapT LRegsMap;
179 
180   /// Topo - A topological ordering for SUnits which permits fast IsReachable
181   /// and similar queries.
182   ScheduleDAGTopologicalSort Topo;
183 
184   // Hack to keep track of the inverse of FindCallSeqStart without more crazy
185   // DAG crawling.
186   DenseMap<SUnit*, SUnit*> CallSeqEndForStart;
187 
188 public:
189   ScheduleDAGRRList(MachineFunction &mf, bool needlatency,
190                     SchedulingPriorityQueue *availqueue,
191                     CodeGenOpt::Level OptLevel)
192     : ScheduleDAGSDNodes(mf),
193       NeedLatency(needlatency), AvailableQueue(availqueue),
194       Topo(SUnits, nullptr) {
195     const TargetSubtargetInfo &STI = mf.getSubtarget();
196     if (DisableSchedCycles || !NeedLatency)
197       HazardRec = new ScheduleHazardRecognizer();
198     else
199       HazardRec = STI.getInstrInfo()->CreateTargetHazardRecognizer(&STI, this);
200   }
201 
202   ~ScheduleDAGRRList() override {
203     delete HazardRec;
204     delete AvailableQueue;
205   }
206 
207   void Schedule() override;
208 
209   ScheduleHazardRecognizer *getHazardRec() { return HazardRec; }
210 
211   /// IsReachable - Checks if SU is reachable from TargetSU.
212   bool IsReachable(const SUnit *SU, const SUnit *TargetSU) {
213     return Topo.IsReachable(SU, TargetSU);
214   }
215 
216   /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
217   /// create a cycle.
218   bool WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
219     return Topo.WillCreateCycle(SU, TargetSU);
220   }
221 
222   /// AddPredQueued - Queues and update to add a predecessor edge to SUnit SU.
223   /// This returns true if this is a new predecessor.
224   /// Does *NOT* update the topological ordering! It just queues an update.
225   void AddPredQueued(SUnit *SU, const SDep &D) {
226     Topo.AddPredQueued(SU, D.getSUnit());
227     SU->addPred(D);
228   }
229 
230   /// AddPred - adds a predecessor edge to SUnit SU.
231   /// This returns true if this is a new predecessor.
232   /// Updates the topological ordering if required.
233   void AddPred(SUnit *SU, const SDep &D) {
234     Topo.AddPred(SU, D.getSUnit());
235     SU->addPred(D);
236   }
237 
238   /// RemovePred - removes a predecessor edge from SUnit SU.
239   /// This returns true if an edge was removed.
240   /// Updates the topological ordering if required.
241   void RemovePred(SUnit *SU, const SDep &D) {
242     Topo.RemovePred(SU, D.getSUnit());
243     SU->removePred(D);
244   }
245 
246 private:
247   bool isReady(SUnit *SU) {
248     return DisableSchedCycles || !AvailableQueue->hasReadyFilter() ||
249       AvailableQueue->isReady(SU);
250   }
251 
252   void ReleasePred(SUnit *SU, const SDep *PredEdge);
253   void ReleasePredecessors(SUnit *SU);
254   void ReleasePending();
255   void AdvanceToCycle(unsigned NextCycle);
256   void AdvancePastStalls(SUnit *SU);
257   void EmitNode(SUnit *SU);
258   void ScheduleNodeBottomUp(SUnit*);
259   void CapturePred(SDep *PredEdge);
260   void UnscheduleNodeBottomUp(SUnit*);
261   void RestoreHazardCheckerBottomUp();
262   void BacktrackBottomUp(SUnit*, SUnit*);
263   SUnit *TryUnfoldSU(SUnit *);
264   SUnit *CopyAndMoveSuccessors(SUnit*);
265   void InsertCopiesAndMoveSuccs(SUnit*, unsigned,
266                                 const TargetRegisterClass*,
267                                 const TargetRegisterClass*,
268                                 SmallVectorImpl<SUnit*>&);
269   bool DelayForLiveRegsBottomUp(SUnit*, SmallVectorImpl<unsigned>&);
270 
271   void releaseInterferences(unsigned Reg = 0);
272 
273   SUnit *PickNodeToScheduleBottomUp();
274   void ListScheduleBottomUp();
275 
276   /// CreateNewSUnit - Creates a new SUnit and returns a pointer to it.
277   SUnit *CreateNewSUnit(SDNode *N) {
278     unsigned NumSUnits = SUnits.size();
279     SUnit *NewNode = newSUnit(N);
280     // Update the topological ordering.
281     if (NewNode->NodeNum >= NumSUnits)
282       Topo.AddSUnitWithoutPredecessors(NewNode);
283     return NewNode;
284   }
285 
286   /// CreateClone - Creates a new SUnit from an existing one.
287   SUnit *CreateClone(SUnit *N) {
288     unsigned NumSUnits = SUnits.size();
289     SUnit *NewNode = Clone(N);
290     // Update the topological ordering.
291     if (NewNode->NodeNum >= NumSUnits)
292       Topo.AddSUnitWithoutPredecessors(NewNode);
293     return NewNode;
294   }
295 
296   /// forceUnitLatencies - Register-pressure-reducing scheduling doesn't
297   /// need actual latency information but the hybrid scheduler does.
298   bool forceUnitLatencies() const override {
299     return !NeedLatency;
300   }
301 };
302 
303 }  // end anonymous namespace
304 
305 static constexpr unsigned RegSequenceCost = 1;
306 
307 /// GetCostForDef - Looks up the register class and cost for a given definition.
308 /// Typically this just means looking up the representative register class,
309 /// but for untyped values (MVT::Untyped) it means inspecting the node's
310 /// opcode to determine what register class is being generated.
311 static void GetCostForDef(const ScheduleDAGSDNodes::RegDefIter &RegDefPos,
312                           const TargetLowering *TLI,
313                           const TargetInstrInfo *TII,
314                           const TargetRegisterInfo *TRI,
315                           unsigned &RegClass, unsigned &Cost,
316                           const MachineFunction &MF) {
317   MVT VT = RegDefPos.GetValue();
318 
319   // Special handling for untyped values.  These values can only come from
320   // the expansion of custom DAG-to-DAG patterns.
321   if (VT == MVT::Untyped) {
322     const SDNode *Node = RegDefPos.GetNode();
323 
324     // Special handling for CopyFromReg of untyped values.
325     if (!Node->isMachineOpcode() && Node->getOpcode() == ISD::CopyFromReg) {
326       Register Reg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
327       const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(Reg);
328       RegClass = RC->getID();
329       Cost = 1;
330       return;
331     }
332 
333     unsigned Opcode = Node->getMachineOpcode();
334     if (Opcode == TargetOpcode::REG_SEQUENCE) {
335       unsigned DstRCIdx = cast<ConstantSDNode>(Node->getOperand(0))->getZExtValue();
336       const TargetRegisterClass *RC = TRI->getRegClass(DstRCIdx);
337       RegClass = RC->getID();
338       Cost = RegSequenceCost;
339       return;
340     }
341 
342     unsigned Idx = RegDefPos.GetIdx();
343     const MCInstrDesc &Desc = TII->get(Opcode);
344     const TargetRegisterClass *RC = TII->getRegClass(Desc, Idx, TRI, MF);
345     assert(RC && "Not a valid register class");
346     RegClass = RC->getID();
347     // FIXME: Cost arbitrarily set to 1 because there doesn't seem to be a
348     // better way to determine it.
349     Cost = 1;
350   } else {
351     RegClass = TLI->getRepRegClassFor(VT)->getID();
352     Cost = TLI->getRepRegClassCostFor(VT);
353   }
354 }
355 
356 /// Schedule - Schedule the DAG using list scheduling.
357 void ScheduleDAGRRList::Schedule() {
358   LLVM_DEBUG(dbgs() << "********** List Scheduling " << printMBBReference(*BB)
359                     << " '" << BB->getName() << "' **********\n");
360 
361   CurCycle = 0;
362   IssueCount = 0;
363   MinAvailableCycle =
364       DisableSchedCycles ? 0 : std::numeric_limits<unsigned>::max();
365   NumLiveRegs = 0;
366   // Allocate slots for each physical register, plus one for a special register
367   // to track the virtual resource of a calling sequence.
368   LiveRegDefs.reset(new SUnit*[TRI->getNumRegs() + 1]());
369   LiveRegGens.reset(new SUnit*[TRI->getNumRegs() + 1]());
370   CallSeqEndForStart.clear();
371   assert(Interferences.empty() && LRegsMap.empty() && "stale Interferences");
372 
373   // Build the scheduling graph.
374   BuildSchedGraph(nullptr);
375 
376   LLVM_DEBUG(dump());
377   Topo.MarkDirty();
378 
379   AvailableQueue->initNodes(SUnits);
380 
381   HazardRec->Reset();
382 
383   // Execute the actual scheduling loop.
384   ListScheduleBottomUp();
385 
386   AvailableQueue->releaseState();
387 
388   LLVM_DEBUG({
389     dbgs() << "*** Final schedule ***\n";
390     dumpSchedule();
391     dbgs() << '\n';
392   });
393 }
394 
395 //===----------------------------------------------------------------------===//
396 //  Bottom-Up Scheduling
397 //===----------------------------------------------------------------------===//
398 
399 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
400 /// the AvailableQueue if the count reaches zero. Also update its cycle bound.
401 void ScheduleDAGRRList::ReleasePred(SUnit *SU, const SDep *PredEdge) {
402   SUnit *PredSU = PredEdge->getSUnit();
403 
404 #ifndef NDEBUG
405   if (PredSU->NumSuccsLeft == 0) {
406     dbgs() << "*** Scheduling failed! ***\n";
407     dumpNode(*PredSU);
408     dbgs() << " has been released too many times!\n";
409     llvm_unreachable(nullptr);
410   }
411 #endif
412   --PredSU->NumSuccsLeft;
413 
414   if (!forceUnitLatencies()) {
415     // Updating predecessor's height. This is now the cycle when the
416     // predecessor can be scheduled without causing a pipeline stall.
417     PredSU->setHeightToAtLeast(SU->getHeight() + PredEdge->getLatency());
418   }
419 
420   // If all the node's successors are scheduled, this node is ready
421   // to be scheduled. Ignore the special EntrySU node.
422   if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU) {
423     PredSU->isAvailable = true;
424 
425     unsigned Height = PredSU->getHeight();
426     if (Height < MinAvailableCycle)
427       MinAvailableCycle = Height;
428 
429     if (isReady(PredSU)) {
430       AvailableQueue->push(PredSU);
431     }
432     // CapturePred and others may have left the node in the pending queue, avoid
433     // adding it twice.
434     else if (!PredSU->isPending) {
435       PredSU->isPending = true;
436       PendingQueue.push_back(PredSU);
437     }
438   }
439 }
440 
441 /// IsChainDependent - Test if Outer is reachable from Inner through
442 /// chain dependencies.
443 static bool IsChainDependent(SDNode *Outer, SDNode *Inner,
444                              unsigned NestLevel,
445                              const TargetInstrInfo *TII) {
446   SDNode *N = Outer;
447   while (true) {
448     if (N == Inner)
449       return true;
450     // For a TokenFactor, examine each operand. There may be multiple ways
451     // to get to the CALLSEQ_BEGIN, but we need to find the path with the
452     // most nesting in order to ensure that we find the corresponding match.
453     if (N->getOpcode() == ISD::TokenFactor) {
454       for (const SDValue &Op : N->op_values())
455         if (IsChainDependent(Op.getNode(), Inner, NestLevel, TII))
456           return true;
457       return false;
458     }
459     // Check for a lowered CALLSEQ_BEGIN or CALLSEQ_END.
460     if (N->isMachineOpcode()) {
461       if (N->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) {
462         ++NestLevel;
463       } else if (N->getMachineOpcode() == TII->getCallFrameSetupOpcode()) {
464         if (NestLevel == 0)
465           return false;
466         --NestLevel;
467       }
468     }
469     // Otherwise, find the chain and continue climbing.
470     for (const SDValue &Op : N->op_values())
471       if (Op.getValueType() == MVT::Other) {
472         N = Op.getNode();
473         goto found_chain_operand;
474       }
475     return false;
476   found_chain_operand:;
477     if (N->getOpcode() == ISD::EntryToken)
478       return false;
479   }
480 }
481 
482 /// FindCallSeqStart - Starting from the (lowered) CALLSEQ_END node, locate
483 /// the corresponding (lowered) CALLSEQ_BEGIN node.
484 ///
485 /// NestLevel and MaxNested are used in recursion to indcate the current level
486 /// of nesting of CALLSEQ_BEGIN and CALLSEQ_END pairs, as well as the maximum
487 /// level seen so far.
488 ///
489 /// TODO: It would be better to give CALLSEQ_END an explicit operand to point
490 /// to the corresponding CALLSEQ_BEGIN to avoid needing to search for it.
491 static SDNode *
492 FindCallSeqStart(SDNode *N, unsigned &NestLevel, unsigned &MaxNest,
493                  const TargetInstrInfo *TII) {
494   while (true) {
495     // For a TokenFactor, examine each operand. There may be multiple ways
496     // to get to the CALLSEQ_BEGIN, but we need to find the path with the
497     // most nesting in order to ensure that we find the corresponding match.
498     if (N->getOpcode() == ISD::TokenFactor) {
499       SDNode *Best = nullptr;
500       unsigned BestMaxNest = MaxNest;
501       for (const SDValue &Op : N->op_values()) {
502         unsigned MyNestLevel = NestLevel;
503         unsigned MyMaxNest = MaxNest;
504         if (SDNode *New = FindCallSeqStart(Op.getNode(),
505                                            MyNestLevel, MyMaxNest, TII))
506           if (!Best || (MyMaxNest > BestMaxNest)) {
507             Best = New;
508             BestMaxNest = MyMaxNest;
509           }
510       }
511       assert(Best);
512       MaxNest = BestMaxNest;
513       return Best;
514     }
515     // Check for a lowered CALLSEQ_BEGIN or CALLSEQ_END.
516     if (N->isMachineOpcode()) {
517       if (N->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) {
518         ++NestLevel;
519         MaxNest = std::max(MaxNest, NestLevel);
520       } else if (N->getMachineOpcode() == TII->getCallFrameSetupOpcode()) {
521         assert(NestLevel != 0);
522         --NestLevel;
523         if (NestLevel == 0)
524           return N;
525       }
526     }
527     // Otherwise, find the chain and continue climbing.
528     for (const SDValue &Op : N->op_values())
529       if (Op.getValueType() == MVT::Other) {
530         N = Op.getNode();
531         goto found_chain_operand;
532       }
533     return nullptr;
534   found_chain_operand:;
535     if (N->getOpcode() == ISD::EntryToken)
536       return nullptr;
537   }
538 }
539 
540 /// Call ReleasePred for each predecessor, then update register live def/gen.
541 /// Always update LiveRegDefs for a register dependence even if the current SU
542 /// also defines the register. This effectively create one large live range
543 /// across a sequence of two-address node. This is important because the
544 /// entire chain must be scheduled together. Example:
545 ///
546 /// flags = (3) add
547 /// flags = (2) addc flags
548 /// flags = (1) addc flags
549 ///
550 /// results in
551 ///
552 /// LiveRegDefs[flags] = 3
553 /// LiveRegGens[flags] = 1
554 ///
555 /// If (2) addc is unscheduled, then (1) addc must also be unscheduled to avoid
556 /// interference on flags.
557 void ScheduleDAGRRList::ReleasePredecessors(SUnit *SU) {
558   // Bottom up: release predecessors
559   for (SDep &Pred : SU->Preds) {
560     ReleasePred(SU, &Pred);
561     if (Pred.isAssignedRegDep()) {
562       // This is a physical register dependency and it's impossible or
563       // expensive to copy the register. Make sure nothing that can
564       // clobber the register is scheduled between the predecessor and
565       // this node.
566       SUnit *RegDef = LiveRegDefs[Pred.getReg()]; (void)RegDef;
567       assert((!RegDef || RegDef == SU || RegDef == Pred.getSUnit()) &&
568              "interference on register dependence");
569       LiveRegDefs[Pred.getReg()] = Pred.getSUnit();
570       if (!LiveRegGens[Pred.getReg()]) {
571         ++NumLiveRegs;
572         LiveRegGens[Pred.getReg()] = SU;
573       }
574     }
575   }
576 
577   // If we're scheduling a lowered CALLSEQ_END, find the corresponding
578   // CALLSEQ_BEGIN. Inject an artificial physical register dependence between
579   // these nodes, to prevent other calls from being interscheduled with them.
580   unsigned CallResource = TRI->getNumRegs();
581   if (!LiveRegDefs[CallResource])
582     for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode())
583       if (Node->isMachineOpcode() &&
584           Node->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) {
585         unsigned NestLevel = 0;
586         unsigned MaxNest = 0;
587         SDNode *N = FindCallSeqStart(Node, NestLevel, MaxNest, TII);
588         assert(N && "Must find call sequence start");
589 
590         SUnit *Def = &SUnits[N->getNodeId()];
591         CallSeqEndForStart[Def] = SU;
592 
593         ++NumLiveRegs;
594         LiveRegDefs[CallResource] = Def;
595         LiveRegGens[CallResource] = SU;
596         break;
597       }
598 }
599 
600 /// Check to see if any of the pending instructions are ready to issue.  If
601 /// so, add them to the available queue.
602 void ScheduleDAGRRList::ReleasePending() {
603   if (DisableSchedCycles) {
604     assert(PendingQueue.empty() && "pending instrs not allowed in this mode");
605     return;
606   }
607 
608   // If the available queue is empty, it is safe to reset MinAvailableCycle.
609   if (AvailableQueue->empty())
610     MinAvailableCycle = std::numeric_limits<unsigned>::max();
611 
612   // Check to see if any of the pending instructions are ready to issue.  If
613   // so, add them to the available queue.
614   for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
615     unsigned ReadyCycle = PendingQueue[i]->getHeight();
616     if (ReadyCycle < MinAvailableCycle)
617       MinAvailableCycle = ReadyCycle;
618 
619     if (PendingQueue[i]->isAvailable) {
620       if (!isReady(PendingQueue[i]))
621           continue;
622       AvailableQueue->push(PendingQueue[i]);
623     }
624     PendingQueue[i]->isPending = false;
625     PendingQueue[i] = PendingQueue.back();
626     PendingQueue.pop_back();
627     --i; --e;
628   }
629 }
630 
631 /// Move the scheduler state forward by the specified number of Cycles.
632 void ScheduleDAGRRList::AdvanceToCycle(unsigned NextCycle) {
633   if (NextCycle <= CurCycle)
634     return;
635 
636   IssueCount = 0;
637   AvailableQueue->setCurCycle(NextCycle);
638   if (!HazardRec->isEnabled()) {
639     // Bypass lots of virtual calls in case of long latency.
640     CurCycle = NextCycle;
641   }
642   else {
643     for (; CurCycle != NextCycle; ++CurCycle) {
644       HazardRec->RecedeCycle();
645     }
646   }
647   // FIXME: Instead of visiting the pending Q each time, set a dirty flag on the
648   // available Q to release pending nodes at least once before popping.
649   ReleasePending();
650 }
651 
652 /// Move the scheduler state forward until the specified node's dependents are
653 /// ready and can be scheduled with no resource conflicts.
654 void ScheduleDAGRRList::AdvancePastStalls(SUnit *SU) {
655   if (DisableSchedCycles)
656     return;
657 
658   // FIXME: Nodes such as CopyFromReg probably should not advance the current
659   // cycle. Otherwise, we can wrongly mask real stalls. If the non-machine node
660   // has predecessors the cycle will be advanced when they are scheduled.
661   // But given the crude nature of modeling latency though such nodes, we
662   // currently need to treat these nodes like real instructions.
663   // if (!SU->getNode() || !SU->getNode()->isMachineOpcode()) return;
664 
665   unsigned ReadyCycle = SU->getHeight();
666 
667   // Bump CurCycle to account for latency. We assume the latency of other
668   // available instructions may be hidden by the stall (not a full pipe stall).
669   // This updates the hazard recognizer's cycle before reserving resources for
670   // this instruction.
671   AdvanceToCycle(ReadyCycle);
672 
673   // Calls are scheduled in their preceding cycle, so don't conflict with
674   // hazards from instructions after the call. EmitNode will reset the
675   // scoreboard state before emitting the call.
676   if (SU->isCall)
677     return;
678 
679   // FIXME: For resource conflicts in very long non-pipelined stages, we
680   // should probably skip ahead here to avoid useless scoreboard checks.
681   int Stalls = 0;
682   while (true) {
683     ScheduleHazardRecognizer::HazardType HT =
684       HazardRec->getHazardType(SU, -Stalls);
685 
686     if (HT == ScheduleHazardRecognizer::NoHazard)
687       break;
688 
689     ++Stalls;
690   }
691   AdvanceToCycle(CurCycle + Stalls);
692 }
693 
694 /// Record this SUnit in the HazardRecognizer.
695 /// Does not update CurCycle.
696 void ScheduleDAGRRList::EmitNode(SUnit *SU) {
697   if (!HazardRec->isEnabled())
698     return;
699 
700   // Check for phys reg copy.
701   if (!SU->getNode())
702     return;
703 
704   switch (SU->getNode()->getOpcode()) {
705   default:
706     assert(SU->getNode()->isMachineOpcode() &&
707            "This target-independent node should not be scheduled.");
708     break;
709   case ISD::MERGE_VALUES:
710   case ISD::TokenFactor:
711   case ISD::LIFETIME_START:
712   case ISD::LIFETIME_END:
713   case ISD::CopyToReg:
714   case ISD::CopyFromReg:
715   case ISD::EH_LABEL:
716     // Noops don't affect the scoreboard state. Copies are likely to be
717     // removed.
718     return;
719   case ISD::INLINEASM:
720   case ISD::INLINEASM_BR:
721     // For inline asm, clear the pipeline state.
722     HazardRec->Reset();
723     return;
724   }
725   if (SU->isCall) {
726     // Calls are scheduled with their preceding instructions. For bottom-up
727     // scheduling, clear the pipeline state before emitting.
728     HazardRec->Reset();
729   }
730 
731   HazardRec->EmitInstruction(SU);
732 }
733 
734 static void resetVRegCycle(SUnit *SU);
735 
736 /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
737 /// count of its predecessors. If a predecessor pending count is zero, add it to
738 /// the Available queue.
739 void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU) {
740   LLVM_DEBUG(dbgs() << "\n*** Scheduling [" << CurCycle << "]: ");
741   LLVM_DEBUG(dumpNode(*SU));
742 
743 #ifndef NDEBUG
744   if (CurCycle < SU->getHeight())
745     LLVM_DEBUG(dbgs() << "   Height [" << SU->getHeight()
746                       << "] pipeline stall!\n");
747 #endif
748 
749   // FIXME: Do not modify node height. It may interfere with
750   // backtracking. Instead add a "ready cycle" to SUnit. Before scheduling the
751   // node its ready cycle can aid heuristics, and after scheduling it can
752   // indicate the scheduled cycle.
753   SU->setHeightToAtLeast(CurCycle);
754 
755   // Reserve resources for the scheduled instruction.
756   EmitNode(SU);
757 
758   Sequence.push_back(SU);
759 
760   AvailableQueue->scheduledNode(SU);
761 
762   // If HazardRec is disabled, and each inst counts as one cycle, then
763   // advance CurCycle before ReleasePredecessors to avoid useless pushes to
764   // PendingQueue for schedulers that implement HasReadyFilter.
765   if (!HazardRec->isEnabled() && AvgIPC < 2)
766     AdvanceToCycle(CurCycle + 1);
767 
768   // Update liveness of predecessors before successors to avoid treating a
769   // two-address node as a live range def.
770   ReleasePredecessors(SU);
771 
772   // Release all the implicit physical register defs that are live.
773   for (SDep &Succ : SU->Succs) {
774     // LiveRegDegs[Succ.getReg()] != SU when SU is a two-address node.
775     if (Succ.isAssignedRegDep() && LiveRegDefs[Succ.getReg()] == SU) {
776       assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
777       --NumLiveRegs;
778       LiveRegDefs[Succ.getReg()] = nullptr;
779       LiveRegGens[Succ.getReg()] = nullptr;
780       releaseInterferences(Succ.getReg());
781     }
782   }
783   // Release the special call resource dependence, if this is the beginning
784   // of a call.
785   unsigned CallResource = TRI->getNumRegs();
786   if (LiveRegDefs[CallResource] == SU)
787     for (const SDNode *SUNode = SU->getNode(); SUNode;
788          SUNode = SUNode->getGluedNode()) {
789       if (SUNode->isMachineOpcode() &&
790           SUNode->getMachineOpcode() == TII->getCallFrameSetupOpcode()) {
791         assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
792         --NumLiveRegs;
793         LiveRegDefs[CallResource] = nullptr;
794         LiveRegGens[CallResource] = nullptr;
795         releaseInterferences(CallResource);
796       }
797     }
798 
799   resetVRegCycle(SU);
800 
801   SU->isScheduled = true;
802 
803   // Conditions under which the scheduler should eagerly advance the cycle:
804   // (1) No available instructions
805   // (2) All pipelines full, so available instructions must have hazards.
806   //
807   // If HazardRec is disabled, the cycle was pre-advanced before calling
808   // ReleasePredecessors. In that case, IssueCount should remain 0.
809   //
810   // Check AvailableQueue after ReleasePredecessors in case of zero latency.
811   if (HazardRec->isEnabled() || AvgIPC > 1) {
812     if (SU->getNode() && SU->getNode()->isMachineOpcode())
813       ++IssueCount;
814     if ((HazardRec->isEnabled() && HazardRec->atIssueLimit())
815         || (!HazardRec->isEnabled() && IssueCount == AvgIPC))
816       AdvanceToCycle(CurCycle + 1);
817   }
818 }
819 
820 /// CapturePred - This does the opposite of ReleasePred. Since SU is being
821 /// unscheduled, increase the succ left count of its predecessors. Remove
822 /// them from AvailableQueue if necessary.
823 void ScheduleDAGRRList::CapturePred(SDep *PredEdge) {
824   SUnit *PredSU = PredEdge->getSUnit();
825   if (PredSU->isAvailable) {
826     PredSU->isAvailable = false;
827     if (!PredSU->isPending)
828       AvailableQueue->remove(PredSU);
829   }
830 
831   assert(PredSU->NumSuccsLeft < std::numeric_limits<unsigned>::max() &&
832          "NumSuccsLeft will overflow!");
833   ++PredSU->NumSuccsLeft;
834 }
835 
836 /// UnscheduleNodeBottomUp - Remove the node from the schedule, update its and
837 /// its predecessor states to reflect the change.
838 void ScheduleDAGRRList::UnscheduleNodeBottomUp(SUnit *SU) {
839   LLVM_DEBUG(dbgs() << "*** Unscheduling [" << SU->getHeight() << "]: ");
840   LLVM_DEBUG(dumpNode(*SU));
841 
842   for (SDep &Pred : SU->Preds) {
843     CapturePred(&Pred);
844     if (Pred.isAssignedRegDep() && SU == LiveRegGens[Pred.getReg()]){
845       assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
846       assert(LiveRegDefs[Pred.getReg()] == Pred.getSUnit() &&
847              "Physical register dependency violated?");
848       --NumLiveRegs;
849       LiveRegDefs[Pred.getReg()] = nullptr;
850       LiveRegGens[Pred.getReg()] = nullptr;
851       releaseInterferences(Pred.getReg());
852     }
853   }
854 
855   // Reclaim the special call resource dependence, if this is the beginning
856   // of a call.
857   unsigned CallResource = TRI->getNumRegs();
858   for (const SDNode *SUNode = SU->getNode(); SUNode;
859        SUNode = SUNode->getGluedNode()) {
860     if (SUNode->isMachineOpcode() &&
861         SUNode->getMachineOpcode() == TII->getCallFrameSetupOpcode()) {
862       SUnit *SeqEnd = CallSeqEndForStart[SU];
863       assert(SeqEnd && "Call sequence start/end must be known");
864       assert(!LiveRegDefs[CallResource]);
865       assert(!LiveRegGens[CallResource]);
866       ++NumLiveRegs;
867       LiveRegDefs[CallResource] = SU;
868       LiveRegGens[CallResource] = SeqEnd;
869     }
870   }
871 
872   // Release the special call resource dependence, if this is the end
873   // of a call.
874   if (LiveRegGens[CallResource] == SU)
875     for (const SDNode *SUNode = SU->getNode(); SUNode;
876          SUNode = SUNode->getGluedNode()) {
877       if (SUNode->isMachineOpcode() &&
878           SUNode->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) {
879         assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
880         assert(LiveRegDefs[CallResource]);
881         assert(LiveRegGens[CallResource]);
882         --NumLiveRegs;
883         LiveRegDefs[CallResource] = nullptr;
884         LiveRegGens[CallResource] = nullptr;
885         releaseInterferences(CallResource);
886       }
887     }
888 
889   for (auto &Succ : SU->Succs) {
890     if (Succ.isAssignedRegDep()) {
891       auto Reg = Succ.getReg();
892       if (!LiveRegDefs[Reg])
893         ++NumLiveRegs;
894       // This becomes the nearest def. Note that an earlier def may still be
895       // pending if this is a two-address node.
896       LiveRegDefs[Reg] = SU;
897 
898       // Update LiveRegGen only if was empty before this unscheduling.
899       // This is to avoid incorrect updating LiveRegGen set in previous run.
900       if (!LiveRegGens[Reg]) {
901         // Find the successor with the lowest height.
902         LiveRegGens[Reg] = Succ.getSUnit();
903         for (auto &Succ2 : SU->Succs) {
904           if (Succ2.isAssignedRegDep() && Succ2.getReg() == Reg &&
905               Succ2.getSUnit()->getHeight() < LiveRegGens[Reg]->getHeight())
906             LiveRegGens[Reg] = Succ2.getSUnit();
907         }
908       }
909     }
910   }
911   if (SU->getHeight() < MinAvailableCycle)
912     MinAvailableCycle = SU->getHeight();
913 
914   SU->setHeightDirty();
915   SU->isScheduled = false;
916   SU->isAvailable = true;
917   if (!DisableSchedCycles && AvailableQueue->hasReadyFilter()) {
918     // Don't make available until backtracking is complete.
919     SU->isPending = true;
920     PendingQueue.push_back(SU);
921   }
922   else {
923     AvailableQueue->push(SU);
924   }
925   AvailableQueue->unscheduledNode(SU);
926 }
927 
928 /// After backtracking, the hazard checker needs to be restored to a state
929 /// corresponding the current cycle.
930 void ScheduleDAGRRList::RestoreHazardCheckerBottomUp() {
931   HazardRec->Reset();
932 
933   unsigned LookAhead = std::min((unsigned)Sequence.size(),
934                                 HazardRec->getMaxLookAhead());
935   if (LookAhead == 0)
936     return;
937 
938   std::vector<SUnit *>::const_iterator I = (Sequence.end() - LookAhead);
939   unsigned HazardCycle = (*I)->getHeight();
940   for (auto E = Sequence.end(); I != E; ++I) {
941     SUnit *SU = *I;
942     for (; SU->getHeight() > HazardCycle; ++HazardCycle) {
943       HazardRec->RecedeCycle();
944     }
945     EmitNode(SU);
946   }
947 }
948 
949 /// BacktrackBottomUp - Backtrack scheduling to a previous cycle specified in
950 /// BTCycle in order to schedule a specific node.
951 void ScheduleDAGRRList::BacktrackBottomUp(SUnit *SU, SUnit *BtSU) {
952   SUnit *OldSU = Sequence.back();
953   while (true) {
954     Sequence.pop_back();
955     // FIXME: use ready cycle instead of height
956     CurCycle = OldSU->getHeight();
957     UnscheduleNodeBottomUp(OldSU);
958     AvailableQueue->setCurCycle(CurCycle);
959     if (OldSU == BtSU)
960       break;
961     OldSU = Sequence.back();
962   }
963 
964   assert(!SU->isSucc(OldSU) && "Something is wrong!");
965 
966   RestoreHazardCheckerBottomUp();
967 
968   ReleasePending();
969 
970   ++NumBacktracks;
971 }
972 
973 static bool isOperandOf(const SUnit *SU, SDNode *N) {
974   for (const SDNode *SUNode = SU->getNode(); SUNode;
975        SUNode = SUNode->getGluedNode()) {
976     if (SUNode->isOperandOf(N))
977       return true;
978   }
979   return false;
980 }
981 
982 /// TryUnfold - Attempt to unfold
983 SUnit *ScheduleDAGRRList::TryUnfoldSU(SUnit *SU) {
984   SDNode *N = SU->getNode();
985   // Use while over if to ease fall through.
986   SmallVector<SDNode *, 2> NewNodes;
987   if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
988     return nullptr;
989 
990   // unfolding an x86 DEC64m operation results in store, dec, load which
991   // can't be handled here so quit
992   if (NewNodes.size() == 3)
993     return nullptr;
994 
995   assert(NewNodes.size() == 2 && "Expected a load folding node!");
996 
997   N = NewNodes[1];
998   SDNode *LoadNode = NewNodes[0];
999   unsigned NumVals = N->getNumValues();
1000   unsigned OldNumVals = SU->getNode()->getNumValues();
1001 
1002   // LoadNode may already exist. This can happen when there is another
1003   // load from the same location and producing the same type of value
1004   // but it has different alignment or volatileness.
1005   bool isNewLoad = true;
1006   SUnit *LoadSU;
1007   if (LoadNode->getNodeId() != -1) {
1008     LoadSU = &SUnits[LoadNode->getNodeId()];
1009     // If LoadSU has already been scheduled, we should clone it but
1010     // this would negate the benefit to unfolding so just return SU.
1011     if (LoadSU->isScheduled)
1012       return SU;
1013     isNewLoad = false;
1014   } else {
1015     LoadSU = CreateNewSUnit(LoadNode);
1016     LoadNode->setNodeId(LoadSU->NodeNum);
1017 
1018     InitNumRegDefsLeft(LoadSU);
1019     computeLatency(LoadSU);
1020   }
1021 
1022   bool isNewN = true;
1023   SUnit *NewSU;
1024   // This can only happen when isNewLoad is false.
1025   if (N->getNodeId() != -1) {
1026     NewSU = &SUnits[N->getNodeId()];
1027     // If NewSU has already been scheduled, we need to clone it, but this
1028     // negates the benefit to unfolding so just return SU.
1029     if (NewSU->isScheduled) {
1030       return SU;
1031     }
1032     isNewN = false;
1033   } else {
1034     NewSU = CreateNewSUnit(N);
1035     N->setNodeId(NewSU->NodeNum);
1036 
1037     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1038     for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
1039       if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
1040         NewSU->isTwoAddress = true;
1041         break;
1042       }
1043     }
1044     if (MCID.isCommutable())
1045       NewSU->isCommutable = true;
1046 
1047     InitNumRegDefsLeft(NewSU);
1048     computeLatency(NewSU);
1049   }
1050 
1051   LLVM_DEBUG(dbgs() << "Unfolding SU #" << SU->NodeNum << "\n");
1052 
1053   // Now that we are committed to unfolding replace DAG Uses.
1054   for (unsigned i = 0; i != NumVals; ++i)
1055     DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i));
1056   DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals - 1),
1057                                  SDValue(LoadNode, 1));
1058 
1059   // Record all the edges to and from the old SU, by category.
1060   SmallVector<SDep, 4> ChainPreds;
1061   SmallVector<SDep, 4> ChainSuccs;
1062   SmallVector<SDep, 4> LoadPreds;
1063   SmallVector<SDep, 4> NodePreds;
1064   SmallVector<SDep, 4> NodeSuccs;
1065   for (SDep &Pred : SU->Preds) {
1066     if (Pred.isCtrl())
1067       ChainPreds.push_back(Pred);
1068     else if (isOperandOf(Pred.getSUnit(), LoadNode))
1069       LoadPreds.push_back(Pred);
1070     else
1071       NodePreds.push_back(Pred);
1072   }
1073   for (SDep &Succ : SU->Succs) {
1074     if (Succ.isCtrl())
1075       ChainSuccs.push_back(Succ);
1076     else
1077       NodeSuccs.push_back(Succ);
1078   }
1079 
1080   // Now assign edges to the newly-created nodes.
1081   for (const SDep &Pred : ChainPreds) {
1082     RemovePred(SU, Pred);
1083     if (isNewLoad)
1084       AddPredQueued(LoadSU, Pred);
1085   }
1086   for (const SDep &Pred : LoadPreds) {
1087     RemovePred(SU, Pred);
1088     if (isNewLoad)
1089       AddPredQueued(LoadSU, Pred);
1090   }
1091   for (const SDep &Pred : NodePreds) {
1092     RemovePred(SU, Pred);
1093     AddPredQueued(NewSU, Pred);
1094   }
1095   for (SDep &D : NodeSuccs) {
1096     SUnit *SuccDep = D.getSUnit();
1097     D.setSUnit(SU);
1098     RemovePred(SuccDep, D);
1099     D.setSUnit(NewSU);
1100     AddPredQueued(SuccDep, D);
1101     // Balance register pressure.
1102     if (AvailableQueue->tracksRegPressure() && SuccDep->isScheduled &&
1103         !D.isCtrl() && NewSU->NumRegDefsLeft > 0)
1104       --NewSU->NumRegDefsLeft;
1105   }
1106   for (SDep &D : ChainSuccs) {
1107     SUnit *SuccDep = D.getSUnit();
1108     D.setSUnit(SU);
1109     RemovePred(SuccDep, D);
1110     if (isNewLoad) {
1111       D.setSUnit(LoadSU);
1112       AddPredQueued(SuccDep, D);
1113     }
1114   }
1115 
1116   // Add a data dependency to reflect that NewSU reads the value defined
1117   // by LoadSU.
1118   SDep D(LoadSU, SDep::Data, 0);
1119   D.setLatency(LoadSU->Latency);
1120   AddPredQueued(NewSU, D);
1121 
1122   if (isNewLoad)
1123     AvailableQueue->addNode(LoadSU);
1124   if (isNewN)
1125     AvailableQueue->addNode(NewSU);
1126 
1127   ++NumUnfolds;
1128 
1129   if (NewSU->NumSuccsLeft == 0)
1130     NewSU->isAvailable = true;
1131 
1132   return NewSU;
1133 }
1134 
1135 /// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
1136 /// successors to the newly created node.
1137 SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) {
1138   SDNode *N = SU->getNode();
1139   if (!N)
1140     return nullptr;
1141 
1142   LLVM_DEBUG(dbgs() << "Considering duplicating the SU\n");
1143   LLVM_DEBUG(dumpNode(*SU));
1144 
1145   if (N->getGluedNode() &&
1146       !TII->canCopyGluedNodeDuringSchedule(N)) {
1147     LLVM_DEBUG(
1148         dbgs()
1149         << "Giving up because it has incoming glue and the target does not "
1150            "want to copy it\n");
1151     return nullptr;
1152   }
1153 
1154   SUnit *NewSU;
1155   bool TryUnfold = false;
1156   for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
1157     MVT VT = N->getSimpleValueType(i);
1158     if (VT == MVT::Glue) {
1159       LLVM_DEBUG(dbgs() << "Giving up because it has outgoing glue\n");
1160       return nullptr;
1161     } else if (VT == MVT::Other)
1162       TryUnfold = true;
1163   }
1164   for (const SDValue &Op : N->op_values()) {
1165     MVT VT = Op.getNode()->getSimpleValueType(Op.getResNo());
1166     if (VT == MVT::Glue && !TII->canCopyGluedNodeDuringSchedule(N)) {
1167       LLVM_DEBUG(
1168           dbgs() << "Giving up because it one of the operands is glue and "
1169                     "the target does not want to copy it\n");
1170       return nullptr;
1171     }
1172   }
1173 
1174   // If possible unfold instruction.
1175   if (TryUnfold) {
1176     SUnit *UnfoldSU = TryUnfoldSU(SU);
1177     if (!UnfoldSU)
1178       return nullptr;
1179     SU = UnfoldSU;
1180     N = SU->getNode();
1181     // If this can be scheduled don't bother duplicating and just return
1182     if (SU->NumSuccsLeft == 0)
1183       return SU;
1184   }
1185 
1186   LLVM_DEBUG(dbgs() << "    Duplicating SU #" << SU->NodeNum << "\n");
1187   NewSU = CreateClone(SU);
1188 
1189   // New SUnit has the exact same predecessors.
1190   for (SDep &Pred : SU->Preds)
1191     if (!Pred.isArtificial())
1192       AddPredQueued(NewSU, Pred);
1193 
1194   // Make sure the clone comes after the original. (InstrEmitter assumes
1195   // this ordering.)
1196   AddPredQueued(NewSU, SDep(SU, SDep::Artificial));
1197 
1198   // Only copy scheduled successors. Cut them from old node's successor
1199   // list and move them over.
1200   SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
1201   for (SDep &Succ : SU->Succs) {
1202     if (Succ.isArtificial())
1203       continue;
1204     SUnit *SuccSU = Succ.getSUnit();
1205     if (SuccSU->isScheduled) {
1206       SDep D = Succ;
1207       D.setSUnit(NewSU);
1208       AddPredQueued(SuccSU, D);
1209       D.setSUnit(SU);
1210       DelDeps.emplace_back(SuccSU, D);
1211     }
1212   }
1213   for (const auto &[DelSU, DelD] : DelDeps)
1214     RemovePred(DelSU, DelD);
1215 
1216   AvailableQueue->updateNode(SU);
1217   AvailableQueue->addNode(NewSU);
1218 
1219   ++NumDups;
1220   return NewSU;
1221 }
1222 
1223 /// InsertCopiesAndMoveSuccs - Insert register copies and move all
1224 /// scheduled successors of the given SUnit to the last copy.
1225 void ScheduleDAGRRList::InsertCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
1226                                               const TargetRegisterClass *DestRC,
1227                                               const TargetRegisterClass *SrcRC,
1228                                               SmallVectorImpl<SUnit*> &Copies) {
1229   SUnit *CopyFromSU = CreateNewSUnit(nullptr);
1230   CopyFromSU->CopySrcRC = SrcRC;
1231   CopyFromSU->CopyDstRC = DestRC;
1232 
1233   SUnit *CopyToSU = CreateNewSUnit(nullptr);
1234   CopyToSU->CopySrcRC = DestRC;
1235   CopyToSU->CopyDstRC = SrcRC;
1236 
1237   // Only copy scheduled successors. Cut them from old node's successor
1238   // list and move them over.
1239   SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
1240   for (SDep &Succ : SU->Succs) {
1241     if (Succ.isArtificial())
1242       continue;
1243     SUnit *SuccSU = Succ.getSUnit();
1244     if (SuccSU->isScheduled) {
1245       SDep D = Succ;
1246       D.setSUnit(CopyToSU);
1247       AddPredQueued(SuccSU, D);
1248       DelDeps.emplace_back(SuccSU, Succ);
1249     }
1250     else {
1251       // Avoid scheduling the def-side copy before other successors. Otherwise,
1252       // we could introduce another physreg interference on the copy and
1253       // continue inserting copies indefinitely.
1254       AddPredQueued(SuccSU, SDep(CopyFromSU, SDep::Artificial));
1255     }
1256   }
1257   for (const auto &[DelSU, DelD] : DelDeps)
1258     RemovePred(DelSU, DelD);
1259 
1260   SDep FromDep(SU, SDep::Data, Reg);
1261   FromDep.setLatency(SU->Latency);
1262   AddPredQueued(CopyFromSU, FromDep);
1263   SDep ToDep(CopyFromSU, SDep::Data, 0);
1264   ToDep.setLatency(CopyFromSU->Latency);
1265   AddPredQueued(CopyToSU, ToDep);
1266 
1267   AvailableQueue->updateNode(SU);
1268   AvailableQueue->addNode(CopyFromSU);
1269   AvailableQueue->addNode(CopyToSU);
1270   Copies.push_back(CopyFromSU);
1271   Copies.push_back(CopyToSU);
1272 
1273   ++NumPRCopies;
1274 }
1275 
1276 /// getPhysicalRegisterVT - Returns the ValueType of the physical register
1277 /// definition of the specified node.
1278 /// FIXME: Move to SelectionDAG?
1279 static MVT getPhysicalRegisterVT(SDNode *N, unsigned Reg,
1280                                  const TargetInstrInfo *TII) {
1281   unsigned NumRes;
1282   if (N->getOpcode() == ISD::CopyFromReg) {
1283     // CopyFromReg has: "chain, Val, glue" so operand 1 gives the type.
1284     NumRes = 1;
1285   } else {
1286     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1287     assert(!MCID.implicit_defs().empty() &&
1288            "Physical reg def must be in implicit def list!");
1289     NumRes = MCID.getNumDefs();
1290     for (MCPhysReg ImpDef : MCID.implicit_defs()) {
1291       if (Reg == ImpDef)
1292         break;
1293       ++NumRes;
1294     }
1295   }
1296   return N->getSimpleValueType(NumRes);
1297 }
1298 
1299 /// CheckForLiveRegDef - Return true and update live register vector if the
1300 /// specified register def of the specified SUnit clobbers any "live" registers.
1301 static void CheckForLiveRegDef(SUnit *SU, unsigned Reg, SUnit **LiveRegDefs,
1302                                SmallSet<unsigned, 4> &RegAdded,
1303                                SmallVectorImpl<unsigned> &LRegs,
1304                                const TargetRegisterInfo *TRI,
1305                                const SDNode *Node = nullptr) {
1306   for (MCRegAliasIterator AliasI(Reg, TRI, true); AliasI.isValid(); ++AliasI) {
1307 
1308     // Check if Ref is live.
1309     if (!LiveRegDefs[*AliasI]) continue;
1310 
1311     // Allow multiple uses of the same def.
1312     if (LiveRegDefs[*AliasI] == SU) continue;
1313 
1314     // Allow multiple uses of same def
1315     if (Node && LiveRegDefs[*AliasI]->getNode() == Node)
1316       continue;
1317 
1318     // Add Reg to the set of interfering live regs.
1319     if (RegAdded.insert(*AliasI).second) {
1320       LRegs.push_back(*AliasI);
1321     }
1322   }
1323 }
1324 
1325 /// CheckForLiveRegDefMasked - Check for any live physregs that are clobbered
1326 /// by RegMask, and add them to LRegs.
1327 static void CheckForLiveRegDefMasked(SUnit *SU, const uint32_t *RegMask,
1328                                      ArrayRef<SUnit*> LiveRegDefs,
1329                                      SmallSet<unsigned, 4> &RegAdded,
1330                                      SmallVectorImpl<unsigned> &LRegs) {
1331   // Look at all live registers. Skip Reg0 and the special CallResource.
1332   for (unsigned i = 1, e = LiveRegDefs.size()-1; i != e; ++i) {
1333     if (!LiveRegDefs[i]) continue;
1334     if (LiveRegDefs[i] == SU) continue;
1335     if (!MachineOperand::clobbersPhysReg(RegMask, i)) continue;
1336     if (RegAdded.insert(i).second)
1337       LRegs.push_back(i);
1338   }
1339 }
1340 
1341 /// getNodeRegMask - Returns the register mask attached to an SDNode, if any.
1342 static const uint32_t *getNodeRegMask(const SDNode *N) {
1343   for (const SDValue &Op : N->op_values())
1344     if (const auto *RegOp = dyn_cast<RegisterMaskSDNode>(Op.getNode()))
1345       return RegOp->getRegMask();
1346   return nullptr;
1347 }
1348 
1349 /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
1350 /// scheduling of the given node to satisfy live physical register dependencies.
1351 /// If the specific node is the last one that's available to schedule, do
1352 /// whatever is necessary (i.e. backtracking or cloning) to make it possible.
1353 bool ScheduleDAGRRList::
1354 DelayForLiveRegsBottomUp(SUnit *SU, SmallVectorImpl<unsigned> &LRegs) {
1355   if (NumLiveRegs == 0)
1356     return false;
1357 
1358   SmallSet<unsigned, 4> RegAdded;
1359   // If this node would clobber any "live" register, then it's not ready.
1360   //
1361   // If SU is the currently live definition of the same register that it uses,
1362   // then we are free to schedule it.
1363   for (SDep &Pred : SU->Preds) {
1364     if (Pred.isAssignedRegDep() && LiveRegDefs[Pred.getReg()] != SU)
1365       CheckForLiveRegDef(Pred.getSUnit(), Pred.getReg(), LiveRegDefs.get(),
1366                          RegAdded, LRegs, TRI);
1367   }
1368 
1369   for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) {
1370     if (Node->getOpcode() == ISD::INLINEASM ||
1371         Node->getOpcode() == ISD::INLINEASM_BR) {
1372       // Inline asm can clobber physical defs.
1373       unsigned NumOps = Node->getNumOperands();
1374       if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
1375         --NumOps;  // Ignore the glue operand.
1376 
1377       for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
1378         unsigned Flags =
1379           cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
1380         unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
1381 
1382         ++i; // Skip the ID value.
1383         if (InlineAsm::isRegDefKind(Flags) ||
1384             InlineAsm::isRegDefEarlyClobberKind(Flags) ||
1385             InlineAsm::isClobberKind(Flags)) {
1386           // Check for def of register or earlyclobber register.
1387           for (; NumVals; --NumVals, ++i) {
1388             Register Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
1389             if (Reg.isPhysical())
1390               CheckForLiveRegDef(SU, Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI);
1391           }
1392         } else
1393           i += NumVals;
1394       }
1395       continue;
1396     }
1397 
1398     if (Node->getOpcode() == ISD::CopyToReg) {
1399       Register Reg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
1400       if (Reg.isPhysical()) {
1401         SDNode *SrcNode = Node->getOperand(2).getNode();
1402         CheckForLiveRegDef(SU, Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI,
1403                            SrcNode);
1404       }
1405     }
1406 
1407     if (!Node->isMachineOpcode())
1408       continue;
1409     // If we're in the middle of scheduling a call, don't begin scheduling
1410     // another call. Also, don't allow any physical registers to be live across
1411     // the call.
1412     if (Node->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) {
1413       // Check the special calling-sequence resource.
1414       unsigned CallResource = TRI->getNumRegs();
1415       if (LiveRegDefs[CallResource]) {
1416         SDNode *Gen = LiveRegGens[CallResource]->getNode();
1417         while (SDNode *Glued = Gen->getGluedNode())
1418           Gen = Glued;
1419         if (!IsChainDependent(Gen, Node, 0, TII) &&
1420             RegAdded.insert(CallResource).second)
1421           LRegs.push_back(CallResource);
1422       }
1423     }
1424     if (const uint32_t *RegMask = getNodeRegMask(Node))
1425       CheckForLiveRegDefMasked(SU, RegMask,
1426                                ArrayRef(LiveRegDefs.get(), TRI->getNumRegs()),
1427                                RegAdded, LRegs);
1428 
1429     const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode());
1430     if (MCID.hasOptionalDef()) {
1431       // Most ARM instructions have an OptionalDef for CPSR, to model the S-bit.
1432       // This operand can be either a def of CPSR, if the S bit is set; or a use
1433       // of %noreg.  When the OptionalDef is set to a valid register, we need to
1434       // handle it in the same way as an ImplicitDef.
1435       for (unsigned i = 0; i < MCID.getNumDefs(); ++i)
1436         if (MCID.operands()[i].isOptionalDef()) {
1437           const SDValue &OptionalDef = Node->getOperand(i - Node->getNumValues());
1438           Register Reg = cast<RegisterSDNode>(OptionalDef)->getReg();
1439           CheckForLiveRegDef(SU, Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI);
1440         }
1441     }
1442     for (MCPhysReg Reg : MCID.implicit_defs())
1443       CheckForLiveRegDef(SU, Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI);
1444   }
1445 
1446   return !LRegs.empty();
1447 }
1448 
1449 void ScheduleDAGRRList::releaseInterferences(unsigned Reg) {
1450   // Add the nodes that aren't ready back onto the available list.
1451   for (unsigned i = Interferences.size(); i > 0; --i) {
1452     SUnit *SU = Interferences[i-1];
1453     LRegsMapT::iterator LRegsPos = LRegsMap.find(SU);
1454     if (Reg) {
1455       SmallVectorImpl<unsigned> &LRegs = LRegsPos->second;
1456       if (!is_contained(LRegs, Reg))
1457         continue;
1458     }
1459     SU->isPending = false;
1460     // The interfering node may no longer be available due to backtracking.
1461     // Furthermore, it may have been made available again, in which case it is
1462     // now already in the AvailableQueue.
1463     if (SU->isAvailable && !SU->NodeQueueId) {
1464       LLVM_DEBUG(dbgs() << "    Repushing SU #" << SU->NodeNum << '\n');
1465       AvailableQueue->push(SU);
1466     }
1467     if (i < Interferences.size())
1468       Interferences[i-1] = Interferences.back();
1469     Interferences.pop_back();
1470     LRegsMap.erase(LRegsPos);
1471   }
1472 }
1473 
1474 /// Return a node that can be scheduled in this cycle. Requirements:
1475 /// (1) Ready: latency has been satisfied
1476 /// (2) No Hazards: resources are available
1477 /// (3) No Interferences: may unschedule to break register interferences.
1478 SUnit *ScheduleDAGRRList::PickNodeToScheduleBottomUp() {
1479   SUnit *CurSU = AvailableQueue->empty() ? nullptr : AvailableQueue->pop();
1480   auto FindAvailableNode = [&]() {
1481     while (CurSU) {
1482       SmallVector<unsigned, 4> LRegs;
1483       if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
1484         break;
1485       LLVM_DEBUG(dbgs() << "    Interfering reg ";
1486                  if (LRegs[0] == TRI->getNumRegs()) dbgs() << "CallResource";
1487                  else dbgs() << printReg(LRegs[0], TRI);
1488                  dbgs() << " SU #" << CurSU->NodeNum << '\n');
1489       auto [LRegsIter, LRegsInserted] = LRegsMap.try_emplace(CurSU, LRegs);
1490       if (LRegsInserted) {
1491         CurSU->isPending = true;  // This SU is not in AvailableQueue right now.
1492         Interferences.push_back(CurSU);
1493       }
1494       else {
1495         assert(CurSU->isPending && "Interferences are pending");
1496         // Update the interference with current live regs.
1497         LRegsIter->second = LRegs;
1498       }
1499       CurSU = AvailableQueue->pop();
1500     }
1501   };
1502   FindAvailableNode();
1503   if (CurSU)
1504     return CurSU;
1505 
1506   // We query the topological order in the loop body, so make sure outstanding
1507   // updates are applied before entering it (we only enter the loop if there
1508   // are some interferences). If we make changes to the ordering, we exit
1509   // the loop.
1510 
1511   // All candidates are delayed due to live physical reg dependencies.
1512   // Try backtracking, code duplication, or inserting cross class copies
1513   // to resolve it.
1514   for (SUnit *TrySU : Interferences) {
1515     SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
1516 
1517     // Try unscheduling up to the point where it's safe to schedule
1518     // this node.
1519     SUnit *BtSU = nullptr;
1520     unsigned LiveCycle = std::numeric_limits<unsigned>::max();
1521     for (unsigned Reg : LRegs) {
1522       if (LiveRegGens[Reg]->getHeight() < LiveCycle) {
1523         BtSU = LiveRegGens[Reg];
1524         LiveCycle = BtSU->getHeight();
1525       }
1526     }
1527     if (!WillCreateCycle(TrySU, BtSU))  {
1528       // BacktrackBottomUp mutates Interferences!
1529       BacktrackBottomUp(TrySU, BtSU);
1530 
1531       // Force the current node to be scheduled before the node that
1532       // requires the physical reg dep.
1533       if (BtSU->isAvailable) {
1534         BtSU->isAvailable = false;
1535         if (!BtSU->isPending)
1536           AvailableQueue->remove(BtSU);
1537       }
1538       LLVM_DEBUG(dbgs() << "ARTIFICIAL edge from SU(" << BtSU->NodeNum
1539                         << ") to SU(" << TrySU->NodeNum << ")\n");
1540       AddPredQueued(TrySU, SDep(BtSU, SDep::Artificial));
1541 
1542       // If one or more successors has been unscheduled, then the current
1543       // node is no longer available.
1544       if (!TrySU->isAvailable || !TrySU->NodeQueueId) {
1545         LLVM_DEBUG(dbgs() << "TrySU not available; choosing node from queue\n");
1546         CurSU = AvailableQueue->pop();
1547       } else {
1548         LLVM_DEBUG(dbgs() << "TrySU available\n");
1549         // Available and in AvailableQueue
1550         AvailableQueue->remove(TrySU);
1551         CurSU = TrySU;
1552       }
1553       FindAvailableNode();
1554       // Interferences has been mutated. We must break.
1555       break;
1556     }
1557   }
1558 
1559   if (!CurSU) {
1560     // Can't backtrack. If it's too expensive to copy the value, then try
1561     // duplicate the nodes that produces these "too expensive to copy"
1562     // values to break the dependency. In case even that doesn't work,
1563     // insert cross class copies.
1564     // If it's not too expensive, i.e. cost != -1, issue copies.
1565     SUnit *TrySU = Interferences[0];
1566     SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
1567     assert(LRegs.size() == 1 && "Can't handle this yet!");
1568     unsigned Reg = LRegs[0];
1569     SUnit *LRDef = LiveRegDefs[Reg];
1570     MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
1571     const TargetRegisterClass *RC =
1572       TRI->getMinimalPhysRegClass(Reg, VT);
1573     const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
1574 
1575     // If cross copy register class is the same as RC, then it must be possible
1576     // copy the value directly. Do not try duplicate the def.
1577     // If cross copy register class is not the same as RC, then it's possible to
1578     // copy the value but it require cross register class copies and it is
1579     // expensive.
1580     // If cross copy register class is null, then it's not possible to copy
1581     // the value at all.
1582     SUnit *NewDef = nullptr;
1583     if (DestRC != RC) {
1584       NewDef = CopyAndMoveSuccessors(LRDef);
1585       if (!DestRC && !NewDef)
1586         report_fatal_error("Can't handle live physical register dependency!");
1587     }
1588     if (!NewDef) {
1589       // Issue copies, these can be expensive cross register class copies.
1590       SmallVector<SUnit*, 2> Copies;
1591       InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
1592       LLVM_DEBUG(dbgs() << "    Adding an edge from SU #" << TrySU->NodeNum
1593                         << " to SU #" << Copies.front()->NodeNum << "\n");
1594       AddPredQueued(TrySU, SDep(Copies.front(), SDep::Artificial));
1595       NewDef = Copies.back();
1596     }
1597 
1598     LLVM_DEBUG(dbgs() << "    Adding an edge from SU #" << NewDef->NodeNum
1599                       << " to SU #" << TrySU->NodeNum << "\n");
1600     LiveRegDefs[Reg] = NewDef;
1601     AddPredQueued(NewDef, SDep(TrySU, SDep::Artificial));
1602     TrySU->isAvailable = false;
1603     CurSU = NewDef;
1604   }
1605   assert(CurSU && "Unable to resolve live physical register dependencies!");
1606   return CurSU;
1607 }
1608 
1609 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
1610 /// schedulers.
1611 void ScheduleDAGRRList::ListScheduleBottomUp() {
1612   // Release any predecessors of the special Exit node.
1613   ReleasePredecessors(&ExitSU);
1614 
1615   // Add root to Available queue.
1616   if (!SUnits.empty()) {
1617     SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()];
1618     assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!");
1619     RootSU->isAvailable = true;
1620     AvailableQueue->push(RootSU);
1621   }
1622 
1623   // While Available queue is not empty, grab the node with the highest
1624   // priority. If it is not ready put it back.  Schedule the node.
1625   Sequence.reserve(SUnits.size());
1626   while (!AvailableQueue->empty() || !Interferences.empty()) {
1627     LLVM_DEBUG(dbgs() << "\nExamining Available:\n";
1628                AvailableQueue->dump(this));
1629 
1630     // Pick the best node to schedule taking all constraints into
1631     // consideration.
1632     SUnit *SU = PickNodeToScheduleBottomUp();
1633 
1634     AdvancePastStalls(SU);
1635 
1636     ScheduleNodeBottomUp(SU);
1637 
1638     while (AvailableQueue->empty() && !PendingQueue.empty()) {
1639       // Advance the cycle to free resources. Skip ahead to the next ready SU.
1640       assert(MinAvailableCycle < std::numeric_limits<unsigned>::max() &&
1641              "MinAvailableCycle uninitialized");
1642       AdvanceToCycle(std::max(CurCycle + 1, MinAvailableCycle));
1643     }
1644   }
1645 
1646   // Reverse the order if it is bottom up.
1647   std::reverse(Sequence.begin(), Sequence.end());
1648 
1649 #ifndef NDEBUG
1650   VerifyScheduledSequence(/*isBottomUp=*/true);
1651 #endif
1652 }
1653 
1654 namespace {
1655 
1656 class RegReductionPQBase;
1657 
1658 struct queue_sort {
1659   bool isReady(SUnit* SU, unsigned CurCycle) const { return true; }
1660 };
1661 
1662 #ifndef NDEBUG
1663 template<class SF>
1664 struct reverse_sort : public queue_sort {
1665   SF &SortFunc;
1666 
1667   reverse_sort(SF &sf) : SortFunc(sf) {}
1668 
1669   bool operator()(SUnit* left, SUnit* right) const {
1670     // reverse left/right rather than simply !SortFunc(left, right)
1671     // to expose different paths in the comparison logic.
1672     return SortFunc(right, left);
1673   }
1674 };
1675 #endif // NDEBUG
1676 
1677 /// bu_ls_rr_sort - Priority function for bottom up register pressure
1678 // reduction scheduler.
1679 struct bu_ls_rr_sort : public queue_sort {
1680   enum {
1681     IsBottomUp = true,
1682     HasReadyFilter = false
1683   };
1684 
1685   RegReductionPQBase *SPQ;
1686 
1687   bu_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
1688 
1689   bool operator()(SUnit* left, SUnit* right) const;
1690 };
1691 
1692 // src_ls_rr_sort - Priority function for source order scheduler.
1693 struct src_ls_rr_sort : public queue_sort {
1694   enum {
1695     IsBottomUp = true,
1696     HasReadyFilter = false
1697   };
1698 
1699   RegReductionPQBase *SPQ;
1700 
1701   src_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
1702 
1703   bool operator()(SUnit* left, SUnit* right) const;
1704 };
1705 
1706 // hybrid_ls_rr_sort - Priority function for hybrid scheduler.
1707 struct hybrid_ls_rr_sort : public queue_sort {
1708   enum {
1709     IsBottomUp = true,
1710     HasReadyFilter = false
1711   };
1712 
1713   RegReductionPQBase *SPQ;
1714 
1715   hybrid_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
1716 
1717   bool isReady(SUnit *SU, unsigned CurCycle) const;
1718 
1719   bool operator()(SUnit* left, SUnit* right) const;
1720 };
1721 
1722 // ilp_ls_rr_sort - Priority function for ILP (instruction level parallelism)
1723 // scheduler.
1724 struct ilp_ls_rr_sort : public queue_sort {
1725   enum {
1726     IsBottomUp = true,
1727     HasReadyFilter = false
1728   };
1729 
1730   RegReductionPQBase *SPQ;
1731 
1732   ilp_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
1733 
1734   bool isReady(SUnit *SU, unsigned CurCycle) const;
1735 
1736   bool operator()(SUnit* left, SUnit* right) const;
1737 };
1738 
1739 class RegReductionPQBase : public SchedulingPriorityQueue {
1740 protected:
1741   std::vector<SUnit *> Queue;
1742   unsigned CurQueueId = 0;
1743   bool TracksRegPressure;
1744   bool SrcOrder;
1745 
1746   // SUnits - The SUnits for the current graph.
1747   std::vector<SUnit> *SUnits;
1748 
1749   MachineFunction &MF;
1750   const TargetInstrInfo *TII;
1751   const TargetRegisterInfo *TRI;
1752   const TargetLowering *TLI;
1753   ScheduleDAGRRList *scheduleDAG = nullptr;
1754 
1755   // SethiUllmanNumbers - The SethiUllman number for each node.
1756   std::vector<unsigned> SethiUllmanNumbers;
1757 
1758   /// RegPressure - Tracking current reg pressure per register class.
1759   std::vector<unsigned> RegPressure;
1760 
1761   /// RegLimit - Tracking the number of allocatable registers per register
1762   /// class.
1763   std::vector<unsigned> RegLimit;
1764 
1765 public:
1766   RegReductionPQBase(MachineFunction &mf,
1767                      bool hasReadyFilter,
1768                      bool tracksrp,
1769                      bool srcorder,
1770                      const TargetInstrInfo *tii,
1771                      const TargetRegisterInfo *tri,
1772                      const TargetLowering *tli)
1773     : SchedulingPriorityQueue(hasReadyFilter), TracksRegPressure(tracksrp),
1774       SrcOrder(srcorder), MF(mf), TII(tii), TRI(tri), TLI(tli) {
1775     if (TracksRegPressure) {
1776       unsigned NumRC = TRI->getNumRegClasses();
1777       RegLimit.resize(NumRC);
1778       RegPressure.resize(NumRC);
1779       std::fill(RegLimit.begin(), RegLimit.end(), 0);
1780       std::fill(RegPressure.begin(), RegPressure.end(), 0);
1781       for (const TargetRegisterClass *RC : TRI->regclasses())
1782         RegLimit[RC->getID()] = tri->getRegPressureLimit(RC, MF);
1783     }
1784   }
1785 
1786   void setScheduleDAG(ScheduleDAGRRList *scheduleDag) {
1787     scheduleDAG = scheduleDag;
1788   }
1789 
1790   ScheduleHazardRecognizer* getHazardRec() {
1791     return scheduleDAG->getHazardRec();
1792   }
1793 
1794   void initNodes(std::vector<SUnit> &sunits) override;
1795 
1796   void addNode(const SUnit *SU) override;
1797 
1798   void updateNode(const SUnit *SU) override;
1799 
1800   void releaseState() override {
1801     SUnits = nullptr;
1802     SethiUllmanNumbers.clear();
1803     std::fill(RegPressure.begin(), RegPressure.end(), 0);
1804   }
1805 
1806   unsigned getNodePriority(const SUnit *SU) const;
1807 
1808   unsigned getNodeOrdering(const SUnit *SU) const {
1809     if (!SU->getNode()) return 0;
1810 
1811     return SU->getNode()->getIROrder();
1812   }
1813 
1814   bool empty() const override { return Queue.empty(); }
1815 
1816   void push(SUnit *U) override {
1817     assert(!U->NodeQueueId && "Node in the queue already");
1818     U->NodeQueueId = ++CurQueueId;
1819     Queue.push_back(U);
1820   }
1821 
1822   void remove(SUnit *SU) override {
1823     assert(!Queue.empty() && "Queue is empty!");
1824     assert(SU->NodeQueueId != 0 && "Not in queue!");
1825     std::vector<SUnit *>::iterator I = llvm::find(Queue, SU);
1826     if (I != std::prev(Queue.end()))
1827       std::swap(*I, Queue.back());
1828     Queue.pop_back();
1829     SU->NodeQueueId = 0;
1830   }
1831 
1832   bool tracksRegPressure() const override { return TracksRegPressure; }
1833 
1834   void dumpRegPressure() const;
1835 
1836   bool HighRegPressure(const SUnit *SU) const;
1837 
1838   bool MayReduceRegPressure(SUnit *SU) const;
1839 
1840   int RegPressureDiff(SUnit *SU, unsigned &LiveUses) const;
1841 
1842   void scheduledNode(SUnit *SU) override;
1843 
1844   void unscheduledNode(SUnit *SU) override;
1845 
1846 protected:
1847   bool canClobber(const SUnit *SU, const SUnit *Op);
1848   void AddPseudoTwoAddrDeps();
1849   void PrescheduleNodesWithMultipleUses();
1850   void CalculateSethiUllmanNumbers();
1851 };
1852 
1853 template<class SF>
1854 static SUnit *popFromQueueImpl(std::vector<SUnit *> &Q, SF &Picker) {
1855   unsigned BestIdx = 0;
1856   // Only compute the cost for the first 1000 items in the queue, to avoid
1857   // excessive compile-times for very large queues.
1858   for (unsigned I = 1, E = std::min(Q.size(), (decltype(Q.size()))1000); I != E;
1859        I++)
1860     if (Picker(Q[BestIdx], Q[I]))
1861       BestIdx = I;
1862   SUnit *V = Q[BestIdx];
1863   if (BestIdx + 1 != Q.size())
1864     std::swap(Q[BestIdx], Q.back());
1865   Q.pop_back();
1866   return V;
1867 }
1868 
1869 template<class SF>
1870 SUnit *popFromQueue(std::vector<SUnit *> &Q, SF &Picker, ScheduleDAG *DAG) {
1871 #ifndef NDEBUG
1872   if (DAG->StressSched) {
1873     reverse_sort<SF> RPicker(Picker);
1874     return popFromQueueImpl(Q, RPicker);
1875   }
1876 #endif
1877   (void)DAG;
1878   return popFromQueueImpl(Q, Picker);
1879 }
1880 
1881 //===----------------------------------------------------------------------===//
1882 //                RegReductionPriorityQueue Definition
1883 //===----------------------------------------------------------------------===//
1884 //
1885 // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers
1886 // to reduce register pressure.
1887 //
1888 template<class SF>
1889 class RegReductionPriorityQueue : public RegReductionPQBase {
1890   SF Picker;
1891 
1892 public:
1893   RegReductionPriorityQueue(MachineFunction &mf,
1894                             bool tracksrp,
1895                             bool srcorder,
1896                             const TargetInstrInfo *tii,
1897                             const TargetRegisterInfo *tri,
1898                             const TargetLowering *tli)
1899     : RegReductionPQBase(mf, SF::HasReadyFilter, tracksrp, srcorder,
1900                          tii, tri, tli),
1901       Picker(this) {}
1902 
1903   bool isBottomUp() const override { return SF::IsBottomUp; }
1904 
1905   bool isReady(SUnit *U) const override {
1906     return Picker.HasReadyFilter && Picker.isReady(U, getCurCycle());
1907   }
1908 
1909   SUnit *pop() override {
1910     if (Queue.empty()) return nullptr;
1911 
1912     SUnit *V = popFromQueue(Queue, Picker, scheduleDAG);
1913     V->NodeQueueId = 0;
1914     return V;
1915   }
1916 
1917 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1918   LLVM_DUMP_METHOD void dump(ScheduleDAG *DAG) const override {
1919     // Emulate pop() without clobbering NodeQueueIds.
1920     std::vector<SUnit *> DumpQueue = Queue;
1921     SF DumpPicker = Picker;
1922     while (!DumpQueue.empty()) {
1923       SUnit *SU = popFromQueue(DumpQueue, DumpPicker, scheduleDAG);
1924       dbgs() << "Height " << SU->getHeight() << ": ";
1925       DAG->dumpNode(*SU);
1926     }
1927   }
1928 #endif
1929 };
1930 
1931 using BURegReductionPriorityQueue = RegReductionPriorityQueue<bu_ls_rr_sort>;
1932 using SrcRegReductionPriorityQueue = RegReductionPriorityQueue<src_ls_rr_sort>;
1933 using HybridBURRPriorityQueue = RegReductionPriorityQueue<hybrid_ls_rr_sort>;
1934 using ILPBURRPriorityQueue = RegReductionPriorityQueue<ilp_ls_rr_sort>;
1935 
1936 } // end anonymous namespace
1937 
1938 //===----------------------------------------------------------------------===//
1939 //           Static Node Priority for Register Pressure Reduction
1940 //===----------------------------------------------------------------------===//
1941 
1942 // Check for special nodes that bypass scheduling heuristics.
1943 // Currently this pushes TokenFactor nodes down, but may be used for other
1944 // pseudo-ops as well.
1945 //
1946 // Return -1 to schedule right above left, 1 for left above right.
1947 // Return 0 if no bias exists.
1948 static int checkSpecialNodes(const SUnit *left, const SUnit *right) {
1949   bool LSchedLow = left->isScheduleLow;
1950   bool RSchedLow = right->isScheduleLow;
1951   if (LSchedLow != RSchedLow)
1952     return LSchedLow < RSchedLow ? 1 : -1;
1953   return 0;
1954 }
1955 
1956 /// CalcNodeSethiUllmanNumber - Compute Sethi Ullman number.
1957 /// Smaller number is the higher priority.
1958 static unsigned
1959 CalcNodeSethiUllmanNumber(const SUnit *SU, std::vector<unsigned> &SUNumbers) {
1960   if (SUNumbers[SU->NodeNum] != 0)
1961     return SUNumbers[SU->NodeNum];
1962 
1963   // Use WorkList to avoid stack overflow on excessively large IRs.
1964   struct WorkState {
1965     WorkState(const SUnit *SU) : SU(SU) {}
1966     const SUnit *SU;
1967     unsigned PredsProcessed = 0;
1968   };
1969 
1970   SmallVector<WorkState, 16> WorkList;
1971   WorkList.push_back(SU);
1972   while (!WorkList.empty()) {
1973     auto &Temp = WorkList.back();
1974     auto *TempSU = Temp.SU;
1975     bool AllPredsKnown = true;
1976     // Try to find a non-evaluated pred and push it into the processing stack.
1977     for (unsigned P = Temp.PredsProcessed; P < TempSU->Preds.size(); ++P) {
1978       auto &Pred = TempSU->Preds[P];
1979       if (Pred.isCtrl()) continue;  // ignore chain preds
1980       SUnit *PredSU = Pred.getSUnit();
1981       if (SUNumbers[PredSU->NodeNum] == 0) {
1982 #ifndef NDEBUG
1983         // In debug mode, check that we don't have such element in the stack.
1984         for (auto It : WorkList)
1985           assert(It.SU != PredSU && "Trying to push an element twice?");
1986 #endif
1987         // Next time start processing this one starting from the next pred.
1988         Temp.PredsProcessed = P + 1;
1989         WorkList.push_back(PredSU);
1990         AllPredsKnown = false;
1991         break;
1992       }
1993     }
1994 
1995     if (!AllPredsKnown)
1996       continue;
1997 
1998     // Once all preds are known, we can calculate the answer for this one.
1999     unsigned SethiUllmanNumber = 0;
2000     unsigned Extra = 0;
2001     for (const SDep &Pred : TempSU->Preds) {
2002       if (Pred.isCtrl()) continue;  // ignore chain preds
2003       SUnit *PredSU = Pred.getSUnit();
2004       unsigned PredSethiUllman = SUNumbers[PredSU->NodeNum];
2005       assert(PredSethiUllman > 0 && "We should have evaluated this pred!");
2006       if (PredSethiUllman > SethiUllmanNumber) {
2007         SethiUllmanNumber = PredSethiUllman;
2008         Extra = 0;
2009       } else if (PredSethiUllman == SethiUllmanNumber)
2010         ++Extra;
2011     }
2012 
2013     SethiUllmanNumber += Extra;
2014     if (SethiUllmanNumber == 0)
2015       SethiUllmanNumber = 1;
2016     SUNumbers[TempSU->NodeNum] = SethiUllmanNumber;
2017     WorkList.pop_back();
2018   }
2019 
2020   assert(SUNumbers[SU->NodeNum] > 0 && "SethiUllman should never be zero!");
2021   return SUNumbers[SU->NodeNum];
2022 }
2023 
2024 /// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all
2025 /// scheduling units.
2026 void RegReductionPQBase::CalculateSethiUllmanNumbers() {
2027   SethiUllmanNumbers.assign(SUnits->size(), 0);
2028 
2029   for (const SUnit &SU : *SUnits)
2030     CalcNodeSethiUllmanNumber(&SU, SethiUllmanNumbers);
2031 }
2032 
2033 void RegReductionPQBase::addNode(const SUnit *SU) {
2034   unsigned SUSize = SethiUllmanNumbers.size();
2035   if (SUnits->size() > SUSize)
2036     SethiUllmanNumbers.resize(SUSize*2, 0);
2037   CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
2038 }
2039 
2040 void RegReductionPQBase::updateNode(const SUnit *SU) {
2041   SethiUllmanNumbers[SU->NodeNum] = 0;
2042   CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
2043 }
2044 
2045 // Lower priority means schedule further down. For bottom-up scheduling, lower
2046 // priority SUs are scheduled before higher priority SUs.
2047 unsigned RegReductionPQBase::getNodePriority(const SUnit *SU) const {
2048   assert(SU->NodeNum < SethiUllmanNumbers.size());
2049   unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0;
2050   if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
2051     // CopyToReg should be close to its uses to facilitate coalescing and
2052     // avoid spilling.
2053     return 0;
2054   if (Opc == TargetOpcode::EXTRACT_SUBREG ||
2055       Opc == TargetOpcode::SUBREG_TO_REG ||
2056       Opc == TargetOpcode::INSERT_SUBREG)
2057     // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be
2058     // close to their uses to facilitate coalescing.
2059     return 0;
2060   if (SU->NumSuccs == 0 && SU->NumPreds != 0)
2061     // If SU does not have a register use, i.e. it doesn't produce a value
2062     // that would be consumed (e.g. store), then it terminates a chain of
2063     // computation.  Give it a large SethiUllman number so it will be
2064     // scheduled right before its predecessors that it doesn't lengthen
2065     // their live ranges.
2066     return 0xffff;
2067   if (SU->NumPreds == 0 && SU->NumSuccs != 0)
2068     // If SU does not have a register def, schedule it close to its uses
2069     // because it does not lengthen any live ranges.
2070     return 0;
2071 #if 1
2072   return SethiUllmanNumbers[SU->NodeNum];
2073 #else
2074   unsigned Priority = SethiUllmanNumbers[SU->NodeNum];
2075   if (SU->isCallOp) {
2076     // FIXME: This assumes all of the defs are used as call operands.
2077     int NP = (int)Priority - SU->getNode()->getNumValues();
2078     return (NP > 0) ? NP : 0;
2079   }
2080   return Priority;
2081 #endif
2082 }
2083 
2084 //===----------------------------------------------------------------------===//
2085 //                     Register Pressure Tracking
2086 //===----------------------------------------------------------------------===//
2087 
2088 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2089 LLVM_DUMP_METHOD void RegReductionPQBase::dumpRegPressure() const {
2090   for (const TargetRegisterClass *RC : TRI->regclasses()) {
2091     unsigned Id = RC->getID();
2092     unsigned RP = RegPressure[Id];
2093     if (!RP) continue;
2094     LLVM_DEBUG(dbgs() << TRI->getRegClassName(RC) << ": " << RP << " / "
2095                       << RegLimit[Id] << '\n');
2096   }
2097 }
2098 #endif
2099 
2100 bool RegReductionPQBase::HighRegPressure(const SUnit *SU) const {
2101   if (!TLI)
2102     return false;
2103 
2104   for (const SDep &Pred : SU->Preds) {
2105     if (Pred.isCtrl())
2106       continue;
2107     SUnit *PredSU = Pred.getSUnit();
2108     // NumRegDefsLeft is zero when enough uses of this node have been scheduled
2109     // to cover the number of registers defined (they are all live).
2110     if (PredSU->NumRegDefsLeft == 0) {
2111       continue;
2112     }
2113     for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
2114          RegDefPos.IsValid(); RegDefPos.Advance()) {
2115       unsigned RCId, Cost;
2116       GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
2117 
2118       if ((RegPressure[RCId] + Cost) >= RegLimit[RCId])
2119         return true;
2120     }
2121   }
2122   return false;
2123 }
2124 
2125 bool RegReductionPQBase::MayReduceRegPressure(SUnit *SU) const {
2126   const SDNode *N = SU->getNode();
2127 
2128   if (!N->isMachineOpcode() || !SU->NumSuccs)
2129     return false;
2130 
2131   unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2132   for (unsigned i = 0; i != NumDefs; ++i) {
2133     MVT VT = N->getSimpleValueType(i);
2134     if (!N->hasAnyUseOfValue(i))
2135       continue;
2136     unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2137     if (RegPressure[RCId] >= RegLimit[RCId])
2138       return true;
2139   }
2140   return false;
2141 }
2142 
2143 // Compute the register pressure contribution by this instruction by count up
2144 // for uses that are not live and down for defs. Only count register classes
2145 // that are already under high pressure. As a side effect, compute the number of
2146 // uses of registers that are already live.
2147 //
2148 // FIXME: This encompasses the logic in HighRegPressure and MayReduceRegPressure
2149 // so could probably be factored.
2150 int RegReductionPQBase::RegPressureDiff(SUnit *SU, unsigned &LiveUses) const {
2151   LiveUses = 0;
2152   int PDiff = 0;
2153   for (const SDep &Pred : SU->Preds) {
2154     if (Pred.isCtrl())
2155       continue;
2156     SUnit *PredSU = Pred.getSUnit();
2157     // NumRegDefsLeft is zero when enough uses of this node have been scheduled
2158     // to cover the number of registers defined (they are all live).
2159     if (PredSU->NumRegDefsLeft == 0) {
2160       if (PredSU->getNode()->isMachineOpcode())
2161         ++LiveUses;
2162       continue;
2163     }
2164     for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
2165          RegDefPos.IsValid(); RegDefPos.Advance()) {
2166       MVT VT = RegDefPos.GetValue();
2167       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2168       if (RegPressure[RCId] >= RegLimit[RCId])
2169         ++PDiff;
2170     }
2171   }
2172   const SDNode *N = SU->getNode();
2173 
2174   if (!N || !N->isMachineOpcode() || !SU->NumSuccs)
2175     return PDiff;
2176 
2177   unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2178   for (unsigned i = 0; i != NumDefs; ++i) {
2179     MVT VT = N->getSimpleValueType(i);
2180     if (!N->hasAnyUseOfValue(i))
2181       continue;
2182     unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2183     if (RegPressure[RCId] >= RegLimit[RCId])
2184       --PDiff;
2185   }
2186   return PDiff;
2187 }
2188 
2189 void RegReductionPQBase::scheduledNode(SUnit *SU) {
2190   if (!TracksRegPressure)
2191     return;
2192 
2193   if (!SU->getNode())
2194     return;
2195 
2196   for (const SDep &Pred : SU->Preds) {
2197     if (Pred.isCtrl())
2198       continue;
2199     SUnit *PredSU = Pred.getSUnit();
2200     // NumRegDefsLeft is zero when enough uses of this node have been scheduled
2201     // to cover the number of registers defined (they are all live).
2202     if (PredSU->NumRegDefsLeft == 0) {
2203       continue;
2204     }
2205     // FIXME: The ScheduleDAG currently loses information about which of a
2206     // node's values is consumed by each dependence. Consequently, if the node
2207     // defines multiple register classes, we don't know which to pressurize
2208     // here. Instead the following loop consumes the register defs in an
2209     // arbitrary order. At least it handles the common case of clustered loads
2210     // to the same class. For precise liveness, each SDep needs to indicate the
2211     // result number. But that tightly couples the ScheduleDAG with the
2212     // SelectionDAG making updates tricky. A simpler hack would be to attach a
2213     // value type or register class to SDep.
2214     //
2215     // The most important aspect of register tracking is balancing the increase
2216     // here with the reduction further below. Note that this SU may use multiple
2217     // defs in PredSU. The can't be determined here, but we've already
2218     // compensated by reducing NumRegDefsLeft in PredSU during
2219     // ScheduleDAGSDNodes::AddSchedEdges.
2220     --PredSU->NumRegDefsLeft;
2221     unsigned SkipRegDefs = PredSU->NumRegDefsLeft;
2222     for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
2223          RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
2224       if (SkipRegDefs)
2225         continue;
2226 
2227       unsigned RCId, Cost;
2228       GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
2229       RegPressure[RCId] += Cost;
2230       break;
2231     }
2232   }
2233 
2234   // We should have this assert, but there may be dead SDNodes that never
2235   // materialize as SUnits, so they don't appear to generate liveness.
2236   //assert(SU->NumRegDefsLeft == 0 && "not all regdefs have scheduled uses");
2237   int SkipRegDefs = (int)SU->NumRegDefsLeft;
2238   for (ScheduleDAGSDNodes::RegDefIter RegDefPos(SU, scheduleDAG);
2239        RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
2240     if (SkipRegDefs > 0)
2241       continue;
2242     unsigned RCId, Cost;
2243     GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
2244     if (RegPressure[RCId] < Cost) {
2245       // Register pressure tracking is imprecise. This can happen. But we try
2246       // hard not to let it happen because it likely results in poor scheduling.
2247       LLVM_DEBUG(dbgs() << "  SU(" << SU->NodeNum
2248                         << ") has too many regdefs\n");
2249       RegPressure[RCId] = 0;
2250     }
2251     else {
2252       RegPressure[RCId] -= Cost;
2253     }
2254   }
2255   LLVM_DEBUG(dumpRegPressure());
2256 }
2257 
2258 void RegReductionPQBase::unscheduledNode(SUnit *SU) {
2259   if (!TracksRegPressure)
2260     return;
2261 
2262   const SDNode *N = SU->getNode();
2263   if (!N) return;
2264 
2265   if (!N->isMachineOpcode()) {
2266     if (N->getOpcode() != ISD::CopyToReg)
2267       return;
2268   } else {
2269     unsigned Opc = N->getMachineOpcode();
2270     if (Opc == TargetOpcode::EXTRACT_SUBREG ||
2271         Opc == TargetOpcode::INSERT_SUBREG ||
2272         Opc == TargetOpcode::SUBREG_TO_REG ||
2273         Opc == TargetOpcode::REG_SEQUENCE ||
2274         Opc == TargetOpcode::IMPLICIT_DEF)
2275       return;
2276   }
2277 
2278   for (const SDep &Pred : SU->Preds) {
2279     if (Pred.isCtrl())
2280       continue;
2281     SUnit *PredSU = Pred.getSUnit();
2282     // NumSuccsLeft counts all deps. Don't compare it with NumSuccs which only
2283     // counts data deps.
2284     if (PredSU->NumSuccsLeft != PredSU->Succs.size())
2285       continue;
2286     const SDNode *PN = PredSU->getNode();
2287     if (!PN->isMachineOpcode()) {
2288       if (PN->getOpcode() == ISD::CopyFromReg) {
2289         MVT VT = PN->getSimpleValueType(0);
2290         unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2291         RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2292       }
2293       continue;
2294     }
2295     unsigned POpc = PN->getMachineOpcode();
2296     if (POpc == TargetOpcode::IMPLICIT_DEF)
2297       continue;
2298     if (POpc == TargetOpcode::EXTRACT_SUBREG ||
2299         POpc == TargetOpcode::INSERT_SUBREG ||
2300         POpc == TargetOpcode::SUBREG_TO_REG) {
2301       MVT VT = PN->getSimpleValueType(0);
2302       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2303       RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2304       continue;
2305     }
2306     if (POpc == TargetOpcode::REG_SEQUENCE) {
2307       unsigned DstRCIdx =
2308           cast<ConstantSDNode>(PN->getOperand(0))->getZExtValue();
2309       const TargetRegisterClass *RC = TRI->getRegClass(DstRCIdx);
2310       unsigned RCId = RC->getID();
2311       // REG_SEQUENCE is untyped, so getRepRegClassCostFor could not be used
2312       // here. Instead use the same constant as in GetCostForDef.
2313       RegPressure[RCId] += RegSequenceCost;
2314       continue;
2315     }
2316     unsigned NumDefs = TII->get(PN->getMachineOpcode()).getNumDefs();
2317     for (unsigned i = 0; i != NumDefs; ++i) {
2318       MVT VT = PN->getSimpleValueType(i);
2319       if (!PN->hasAnyUseOfValue(i))
2320         continue;
2321       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2322       if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT))
2323         // Register pressure tracking is imprecise. This can happen.
2324         RegPressure[RCId] = 0;
2325       else
2326         RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
2327     }
2328   }
2329 
2330   // Check for isMachineOpcode() as PrescheduleNodesWithMultipleUses()
2331   // may transfer data dependencies to CopyToReg.
2332   if (SU->NumSuccs && N->isMachineOpcode()) {
2333     unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2334     for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
2335       MVT VT = N->getSimpleValueType(i);
2336       if (VT == MVT::Glue || VT == MVT::Other)
2337         continue;
2338       if (!N->hasAnyUseOfValue(i))
2339         continue;
2340       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2341       RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2342     }
2343   }
2344 
2345   LLVM_DEBUG(dumpRegPressure());
2346 }
2347 
2348 //===----------------------------------------------------------------------===//
2349 //           Dynamic Node Priority for Register Pressure Reduction
2350 //===----------------------------------------------------------------------===//
2351 
2352 /// closestSucc - Returns the scheduled cycle of the successor which is
2353 /// closest to the current cycle.
2354 static unsigned closestSucc(const SUnit *SU) {
2355   unsigned MaxHeight = 0;
2356   for (const SDep &Succ : SU->Succs) {
2357     if (Succ.isCtrl()) continue;  // ignore chain succs
2358     unsigned Height = Succ.getSUnit()->getHeight();
2359     // If there are bunch of CopyToRegs stacked up, they should be considered
2360     // to be at the same position.
2361     if (Succ.getSUnit()->getNode() &&
2362         Succ.getSUnit()->getNode()->getOpcode() == ISD::CopyToReg)
2363       Height = closestSucc(Succ.getSUnit())+1;
2364     if (Height > MaxHeight)
2365       MaxHeight = Height;
2366   }
2367   return MaxHeight;
2368 }
2369 
2370 /// calcMaxScratches - Returns an cost estimate of the worse case requirement
2371 /// for scratch registers, i.e. number of data dependencies.
2372 static unsigned calcMaxScratches(const SUnit *SU) {
2373   unsigned Scratches = 0;
2374   for (const SDep &Pred : SU->Preds) {
2375     if (Pred.isCtrl()) continue;  // ignore chain preds
2376     Scratches++;
2377   }
2378   return Scratches;
2379 }
2380 
2381 /// hasOnlyLiveInOpers - Return true if SU has only value predecessors that are
2382 /// CopyFromReg from a virtual register.
2383 static bool hasOnlyLiveInOpers(const SUnit *SU) {
2384   bool RetVal = false;
2385   for (const SDep &Pred : SU->Preds) {
2386     if (Pred.isCtrl()) continue;
2387     const SUnit *PredSU = Pred.getSUnit();
2388     if (PredSU->getNode() &&
2389         PredSU->getNode()->getOpcode() == ISD::CopyFromReg) {
2390       Register Reg =
2391           cast<RegisterSDNode>(PredSU->getNode()->getOperand(1))->getReg();
2392       if (Reg.isVirtual()) {
2393         RetVal = true;
2394         continue;
2395       }
2396     }
2397     return false;
2398   }
2399   return RetVal;
2400 }
2401 
2402 /// hasOnlyLiveOutUses - Return true if SU has only value successors that are
2403 /// CopyToReg to a virtual register. This SU def is probably a liveout and
2404 /// it has no other use. It should be scheduled closer to the terminator.
2405 static bool hasOnlyLiveOutUses(const SUnit *SU) {
2406   bool RetVal = false;
2407   for (const SDep &Succ : SU->Succs) {
2408     if (Succ.isCtrl()) continue;
2409     const SUnit *SuccSU = Succ.getSUnit();
2410     if (SuccSU->getNode() && SuccSU->getNode()->getOpcode() == ISD::CopyToReg) {
2411       Register Reg =
2412           cast<RegisterSDNode>(SuccSU->getNode()->getOperand(1))->getReg();
2413       if (Reg.isVirtual()) {
2414         RetVal = true;
2415         continue;
2416       }
2417     }
2418     return false;
2419   }
2420   return RetVal;
2421 }
2422 
2423 // Set isVRegCycle for a node with only live in opers and live out uses. Also
2424 // set isVRegCycle for its CopyFromReg operands.
2425 //
2426 // This is only relevant for single-block loops, in which case the VRegCycle
2427 // node is likely an induction variable in which the operand and target virtual
2428 // registers should be coalesced (e.g. pre/post increment values). Setting the
2429 // isVRegCycle flag helps the scheduler prioritize other uses of the same
2430 // CopyFromReg so that this node becomes the virtual register "kill". This
2431 // avoids interference between the values live in and out of the block and
2432 // eliminates a copy inside the loop.
2433 static void initVRegCycle(SUnit *SU) {
2434   if (DisableSchedVRegCycle)
2435     return;
2436 
2437   if (!hasOnlyLiveInOpers(SU) || !hasOnlyLiveOutUses(SU))
2438     return;
2439 
2440   LLVM_DEBUG(dbgs() << "VRegCycle: SU(" << SU->NodeNum << ")\n");
2441 
2442   SU->isVRegCycle = true;
2443 
2444   for (const SDep &Pred : SU->Preds) {
2445     if (Pred.isCtrl()) continue;
2446     Pred.getSUnit()->isVRegCycle = true;
2447   }
2448 }
2449 
2450 // After scheduling the definition of a VRegCycle, clear the isVRegCycle flag of
2451 // CopyFromReg operands. We should no longer penalize other uses of this VReg.
2452 static void resetVRegCycle(SUnit *SU) {
2453   if (!SU->isVRegCycle)
2454     return;
2455 
2456   for (const SDep &Pred : SU->Preds) {
2457     if (Pred.isCtrl()) continue;  // ignore chain preds
2458     SUnit *PredSU = Pred.getSUnit();
2459     if (PredSU->isVRegCycle) {
2460       assert(PredSU->getNode()->getOpcode() == ISD::CopyFromReg &&
2461              "VRegCycle def must be CopyFromReg");
2462       Pred.getSUnit()->isVRegCycle = false;
2463     }
2464   }
2465 }
2466 
2467 // Return true if this SUnit uses a CopyFromReg node marked as a VRegCycle. This
2468 // means a node that defines the VRegCycle has not been scheduled yet.
2469 static bool hasVRegCycleUse(const SUnit *SU) {
2470   // If this SU also defines the VReg, don't hoist it as a "use".
2471   if (SU->isVRegCycle)
2472     return false;
2473 
2474   for (const SDep &Pred : SU->Preds) {
2475     if (Pred.isCtrl()) continue;  // ignore chain preds
2476     if (Pred.getSUnit()->isVRegCycle &&
2477         Pred.getSUnit()->getNode()->getOpcode() == ISD::CopyFromReg) {
2478       LLVM_DEBUG(dbgs() << "  VReg cycle use: SU (" << SU->NodeNum << ")\n");
2479       return true;
2480     }
2481   }
2482   return false;
2483 }
2484 
2485 // Check for either a dependence (latency) or resource (hazard) stall.
2486 //
2487 // Note: The ScheduleHazardRecognizer interface requires a non-const SU.
2488 static bool BUHasStall(SUnit *SU, int Height, RegReductionPQBase *SPQ) {
2489   if ((int)SPQ->getCurCycle() < Height) return true;
2490   if (SPQ->getHazardRec()->getHazardType(SU, 0)
2491       != ScheduleHazardRecognizer::NoHazard)
2492     return true;
2493   return false;
2494 }
2495 
2496 // Return -1 if left has higher priority, 1 if right has higher priority.
2497 // Return 0 if latency-based priority is equivalent.
2498 static int BUCompareLatency(SUnit *left, SUnit *right, bool checkPref,
2499                             RegReductionPQBase *SPQ) {
2500   // Scheduling an instruction that uses a VReg whose postincrement has not yet
2501   // been scheduled will induce a copy. Model this as an extra cycle of latency.
2502   int LPenalty = hasVRegCycleUse(left) ? 1 : 0;
2503   int RPenalty = hasVRegCycleUse(right) ? 1 : 0;
2504   int LHeight = (int)left->getHeight() + LPenalty;
2505   int RHeight = (int)right->getHeight() + RPenalty;
2506 
2507   bool LStall = (!checkPref || left->SchedulingPref == Sched::ILP) &&
2508     BUHasStall(left, LHeight, SPQ);
2509   bool RStall = (!checkPref || right->SchedulingPref == Sched::ILP) &&
2510     BUHasStall(right, RHeight, SPQ);
2511 
2512   // If scheduling one of the node will cause a pipeline stall, delay it.
2513   // If scheduling either one of the node will cause a pipeline stall, sort
2514   // them according to their height.
2515   if (LStall) {
2516     if (!RStall)
2517       return 1;
2518     if (LHeight != RHeight)
2519       return LHeight > RHeight ? 1 : -1;
2520   } else if (RStall)
2521     return -1;
2522 
2523   // If either node is scheduling for latency, sort them by height/depth
2524   // and latency.
2525   if (!checkPref || (left->SchedulingPref == Sched::ILP ||
2526                      right->SchedulingPref == Sched::ILP)) {
2527     // If neither instruction stalls (!LStall && !RStall) and HazardRecognizer
2528     // is enabled, grouping instructions by cycle, then its height is already
2529     // covered so only its depth matters. We also reach this point if both stall
2530     // but have the same height.
2531     if (!SPQ->getHazardRec()->isEnabled()) {
2532       if (LHeight != RHeight)
2533         return LHeight > RHeight ? 1 : -1;
2534     }
2535     int LDepth = left->getDepth() - LPenalty;
2536     int RDepth = right->getDepth() - RPenalty;
2537     if (LDepth != RDepth) {
2538       LLVM_DEBUG(dbgs() << "  Comparing latency of SU (" << left->NodeNum
2539                         << ") depth " << LDepth << " vs SU (" << right->NodeNum
2540                         << ") depth " << RDepth << "\n");
2541       return LDepth < RDepth ? 1 : -1;
2542     }
2543     if (left->Latency != right->Latency)
2544       return left->Latency > right->Latency ? 1 : -1;
2545   }
2546   return 0;
2547 }
2548 
2549 static bool BURRSort(SUnit *left, SUnit *right, RegReductionPQBase *SPQ) {
2550   // Schedule physical register definitions close to their use. This is
2551   // motivated by microarchitectures that can fuse cmp+jump macro-ops. But as
2552   // long as shortening physreg live ranges is generally good, we can defer
2553   // creating a subtarget hook.
2554   if (!DisableSchedPhysRegJoin) {
2555     bool LHasPhysReg = left->hasPhysRegDefs;
2556     bool RHasPhysReg = right->hasPhysRegDefs;
2557     if (LHasPhysReg != RHasPhysReg) {
2558       #ifndef NDEBUG
2559       static const char *const PhysRegMsg[] = { " has no physreg",
2560                                                 " defines a physreg" };
2561       #endif
2562       LLVM_DEBUG(dbgs() << "  SU (" << left->NodeNum << ") "
2563                         << PhysRegMsg[LHasPhysReg] << " SU(" << right->NodeNum
2564                         << ") " << PhysRegMsg[RHasPhysReg] << "\n");
2565       return LHasPhysReg < RHasPhysReg;
2566     }
2567   }
2568 
2569   // Prioritize by Sethi-Ulmann number and push CopyToReg nodes down.
2570   unsigned LPriority = SPQ->getNodePriority(left);
2571   unsigned RPriority = SPQ->getNodePriority(right);
2572 
2573   // Be really careful about hoisting call operands above previous calls.
2574   // Only allows it if it would reduce register pressure.
2575   if (left->isCall && right->isCallOp) {
2576     unsigned RNumVals = right->getNode()->getNumValues();
2577     RPriority = (RPriority > RNumVals) ? (RPriority - RNumVals) : 0;
2578   }
2579   if (right->isCall && left->isCallOp) {
2580     unsigned LNumVals = left->getNode()->getNumValues();
2581     LPriority = (LPriority > LNumVals) ? (LPriority - LNumVals) : 0;
2582   }
2583 
2584   if (LPriority != RPriority)
2585     return LPriority > RPriority;
2586 
2587   // One or both of the nodes are calls and their sethi-ullman numbers are the
2588   // same, then keep source order.
2589   if (left->isCall || right->isCall) {
2590     unsigned LOrder = SPQ->getNodeOrdering(left);
2591     unsigned ROrder = SPQ->getNodeOrdering(right);
2592 
2593     // Prefer an ordering where the lower the non-zero order number, the higher
2594     // the preference.
2595     if ((LOrder || ROrder) && LOrder != ROrder)
2596       return LOrder != 0 && (LOrder < ROrder || ROrder == 0);
2597   }
2598 
2599   // Try schedule def + use closer when Sethi-Ullman numbers are the same.
2600   // e.g.
2601   // t1 = op t2, c1
2602   // t3 = op t4, c2
2603   //
2604   // and the following instructions are both ready.
2605   // t2 = op c3
2606   // t4 = op c4
2607   //
2608   // Then schedule t2 = op first.
2609   // i.e.
2610   // t4 = op c4
2611   // t2 = op c3
2612   // t1 = op t2, c1
2613   // t3 = op t4, c2
2614   //
2615   // This creates more short live intervals.
2616   unsigned LDist = closestSucc(left);
2617   unsigned RDist = closestSucc(right);
2618   if (LDist != RDist)
2619     return LDist < RDist;
2620 
2621   // How many registers becomes live when the node is scheduled.
2622   unsigned LScratch = calcMaxScratches(left);
2623   unsigned RScratch = calcMaxScratches(right);
2624   if (LScratch != RScratch)
2625     return LScratch > RScratch;
2626 
2627   // Comparing latency against a call makes little sense unless the node
2628   // is register pressure-neutral.
2629   if ((left->isCall && RPriority > 0) || (right->isCall && LPriority > 0))
2630     return (left->NodeQueueId > right->NodeQueueId);
2631 
2632   // Do not compare latencies when one or both of the nodes are calls.
2633   if (!DisableSchedCycles &&
2634       !(left->isCall || right->isCall)) {
2635     int result = BUCompareLatency(left, right, false /*checkPref*/, SPQ);
2636     if (result != 0)
2637       return result > 0;
2638   }
2639   else {
2640     if (left->getHeight() != right->getHeight())
2641       return left->getHeight() > right->getHeight();
2642 
2643     if (left->getDepth() != right->getDepth())
2644       return left->getDepth() < right->getDepth();
2645   }
2646 
2647   assert(left->NodeQueueId && right->NodeQueueId &&
2648          "NodeQueueId cannot be zero");
2649   return (left->NodeQueueId > right->NodeQueueId);
2650 }
2651 
2652 // Bottom up
2653 bool bu_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2654   if (int res = checkSpecialNodes(left, right))
2655     return res > 0;
2656 
2657   return BURRSort(left, right, SPQ);
2658 }
2659 
2660 // Source order, otherwise bottom up.
2661 bool src_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2662   if (int res = checkSpecialNodes(left, right))
2663     return res > 0;
2664 
2665   unsigned LOrder = SPQ->getNodeOrdering(left);
2666   unsigned ROrder = SPQ->getNodeOrdering(right);
2667 
2668   // Prefer an ordering where the lower the non-zero order number, the higher
2669   // the preference.
2670   if ((LOrder || ROrder) && LOrder != ROrder)
2671     return LOrder != 0 && (LOrder < ROrder || ROrder == 0);
2672 
2673   return BURRSort(left, right, SPQ);
2674 }
2675 
2676 // If the time between now and when the instruction will be ready can cover
2677 // the spill code, then avoid adding it to the ready queue. This gives long
2678 // stalls highest priority and allows hoisting across calls. It should also
2679 // speed up processing the available queue.
2680 bool hybrid_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
2681   static const unsigned ReadyDelay = 3;
2682 
2683   if (SPQ->MayReduceRegPressure(SU)) return true;
2684 
2685   if (SU->getHeight() > (CurCycle + ReadyDelay)) return false;
2686 
2687   if (SPQ->getHazardRec()->getHazardType(SU, -ReadyDelay)
2688       != ScheduleHazardRecognizer::NoHazard)
2689     return false;
2690 
2691   return true;
2692 }
2693 
2694 // Return true if right should be scheduled with higher priority than left.
2695 bool hybrid_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2696   if (int res = checkSpecialNodes(left, right))
2697     return res > 0;
2698 
2699   if (left->isCall || right->isCall)
2700     // No way to compute latency of calls.
2701     return BURRSort(left, right, SPQ);
2702 
2703   bool LHigh = SPQ->HighRegPressure(left);
2704   bool RHigh = SPQ->HighRegPressure(right);
2705   // Avoid causing spills. If register pressure is high, schedule for
2706   // register pressure reduction.
2707   if (LHigh && !RHigh) {
2708     LLVM_DEBUG(dbgs() << "  pressure SU(" << left->NodeNum << ") > SU("
2709                       << right->NodeNum << ")\n");
2710     return true;
2711   }
2712   else if (!LHigh && RHigh) {
2713     LLVM_DEBUG(dbgs() << "  pressure SU(" << right->NodeNum << ") > SU("
2714                       << left->NodeNum << ")\n");
2715     return false;
2716   }
2717   if (!LHigh && !RHigh) {
2718     int result = BUCompareLatency(left, right, true /*checkPref*/, SPQ);
2719     if (result != 0)
2720       return result > 0;
2721   }
2722   return BURRSort(left, right, SPQ);
2723 }
2724 
2725 // Schedule as many instructions in each cycle as possible. So don't make an
2726 // instruction available unless it is ready in the current cycle.
2727 bool ilp_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
2728   if (SU->getHeight() > CurCycle) return false;
2729 
2730   if (SPQ->getHazardRec()->getHazardType(SU, 0)
2731       != ScheduleHazardRecognizer::NoHazard)
2732     return false;
2733 
2734   return true;
2735 }
2736 
2737 static bool canEnableCoalescing(SUnit *SU) {
2738   unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0;
2739   if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
2740     // CopyToReg should be close to its uses to facilitate coalescing and
2741     // avoid spilling.
2742     return true;
2743 
2744   if (Opc == TargetOpcode::EXTRACT_SUBREG ||
2745       Opc == TargetOpcode::SUBREG_TO_REG ||
2746       Opc == TargetOpcode::INSERT_SUBREG)
2747     // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be
2748     // close to their uses to facilitate coalescing.
2749     return true;
2750 
2751   if (SU->NumPreds == 0 && SU->NumSuccs != 0)
2752     // If SU does not have a register def, schedule it close to its uses
2753     // because it does not lengthen any live ranges.
2754     return true;
2755 
2756   return false;
2757 }
2758 
2759 // list-ilp is currently an experimental scheduler that allows various
2760 // heuristics to be enabled prior to the normal register reduction logic.
2761 bool ilp_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2762   if (int res = checkSpecialNodes(left, right))
2763     return res > 0;
2764 
2765   if (left->isCall || right->isCall)
2766     // No way to compute latency of calls.
2767     return BURRSort(left, right, SPQ);
2768 
2769   unsigned LLiveUses = 0, RLiveUses = 0;
2770   int LPDiff = 0, RPDiff = 0;
2771   if (!DisableSchedRegPressure || !DisableSchedLiveUses) {
2772     LPDiff = SPQ->RegPressureDiff(left, LLiveUses);
2773     RPDiff = SPQ->RegPressureDiff(right, RLiveUses);
2774   }
2775   if (!DisableSchedRegPressure && LPDiff != RPDiff) {
2776     LLVM_DEBUG(dbgs() << "RegPressureDiff SU(" << left->NodeNum
2777                       << "): " << LPDiff << " != SU(" << right->NodeNum
2778                       << "): " << RPDiff << "\n");
2779     return LPDiff > RPDiff;
2780   }
2781 
2782   if (!DisableSchedRegPressure && (LPDiff > 0 || RPDiff > 0)) {
2783     bool LReduce = canEnableCoalescing(left);
2784     bool RReduce = canEnableCoalescing(right);
2785     if (LReduce && !RReduce) return false;
2786     if (RReduce && !LReduce) return true;
2787   }
2788 
2789   if (!DisableSchedLiveUses && (LLiveUses != RLiveUses)) {
2790     LLVM_DEBUG(dbgs() << "Live uses SU(" << left->NodeNum << "): " << LLiveUses
2791                       << " != SU(" << right->NodeNum << "): " << RLiveUses
2792                       << "\n");
2793     return LLiveUses < RLiveUses;
2794   }
2795 
2796   if (!DisableSchedStalls) {
2797     bool LStall = BUHasStall(left, left->getHeight(), SPQ);
2798     bool RStall = BUHasStall(right, right->getHeight(), SPQ);
2799     if (LStall != RStall)
2800       return left->getHeight() > right->getHeight();
2801   }
2802 
2803   if (!DisableSchedCriticalPath) {
2804     int spread = (int)left->getDepth() - (int)right->getDepth();
2805     if (std::abs(spread) > MaxReorderWindow) {
2806       LLVM_DEBUG(dbgs() << "Depth of SU(" << left->NodeNum << "): "
2807                         << left->getDepth() << " != SU(" << right->NodeNum
2808                         << "): " << right->getDepth() << "\n");
2809       return left->getDepth() < right->getDepth();
2810     }
2811   }
2812 
2813   if (!DisableSchedHeight && left->getHeight() != right->getHeight()) {
2814     int spread = (int)left->getHeight() - (int)right->getHeight();
2815     if (std::abs(spread) > MaxReorderWindow)
2816       return left->getHeight() > right->getHeight();
2817   }
2818 
2819   return BURRSort(left, right, SPQ);
2820 }
2821 
2822 void RegReductionPQBase::initNodes(std::vector<SUnit> &sunits) {
2823   SUnits = &sunits;
2824   // Add pseudo dependency edges for two-address nodes.
2825   if (!Disable2AddrHack)
2826     AddPseudoTwoAddrDeps();
2827   // Reroute edges to nodes with multiple uses.
2828   if (!TracksRegPressure && !SrcOrder)
2829     PrescheduleNodesWithMultipleUses();
2830   // Calculate node priorities.
2831   CalculateSethiUllmanNumbers();
2832 
2833   // For single block loops, mark nodes that look like canonical IV increments.
2834   if (scheduleDAG->BB->isSuccessor(scheduleDAG->BB))
2835     for (SUnit &SU : sunits)
2836       initVRegCycle(&SU);
2837 }
2838 
2839 //===----------------------------------------------------------------------===//
2840 //                    Preschedule for Register Pressure
2841 //===----------------------------------------------------------------------===//
2842 
2843 bool RegReductionPQBase::canClobber(const SUnit *SU, const SUnit *Op) {
2844   if (SU->isTwoAddress) {
2845     unsigned Opc = SU->getNode()->getMachineOpcode();
2846     const MCInstrDesc &MCID = TII->get(Opc);
2847     unsigned NumRes = MCID.getNumDefs();
2848     unsigned NumOps = MCID.getNumOperands() - NumRes;
2849     for (unsigned i = 0; i != NumOps; ++i) {
2850       if (MCID.getOperandConstraint(i+NumRes, MCOI::TIED_TO) != -1) {
2851         SDNode *DU = SU->getNode()->getOperand(i).getNode();
2852         if (DU->getNodeId() != -1 &&
2853             Op->OrigNode == &(*SUnits)[DU->getNodeId()])
2854           return true;
2855       }
2856     }
2857   }
2858   return false;
2859 }
2860 
2861 /// canClobberReachingPhysRegUse - True if SU would clobber one of it's
2862 /// successor's explicit physregs whose definition can reach DepSU.
2863 /// i.e. DepSU should not be scheduled above SU.
2864 static bool canClobberReachingPhysRegUse(const SUnit *DepSU, const SUnit *SU,
2865                                          ScheduleDAGRRList *scheduleDAG,
2866                                          const TargetInstrInfo *TII,
2867                                          const TargetRegisterInfo *TRI) {
2868   ArrayRef<MCPhysReg> ImpDefs =
2869       TII->get(SU->getNode()->getMachineOpcode()).implicit_defs();
2870   const uint32_t *RegMask = getNodeRegMask(SU->getNode());
2871   if (ImpDefs.empty() && !RegMask)
2872     return false;
2873 
2874   for (const SDep &Succ : SU->Succs) {
2875     SUnit *SuccSU = Succ.getSUnit();
2876     for (const SDep &SuccPred : SuccSU->Preds) {
2877       if (!SuccPred.isAssignedRegDep())
2878         continue;
2879 
2880       if (RegMask &&
2881           MachineOperand::clobbersPhysReg(RegMask, SuccPred.getReg()) &&
2882           scheduleDAG->IsReachable(DepSU, SuccPred.getSUnit()))
2883         return true;
2884 
2885       for (MCPhysReg ImpDef : ImpDefs) {
2886         // Return true if SU clobbers this physical register use and the
2887         // definition of the register reaches from DepSU. IsReachable queries
2888         // a topological forward sort of the DAG (following the successors).
2889         if (TRI->regsOverlap(ImpDef, SuccPred.getReg()) &&
2890             scheduleDAG->IsReachable(DepSU, SuccPred.getSUnit()))
2891           return true;
2892       }
2893     }
2894   }
2895   return false;
2896 }
2897 
2898 /// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's
2899 /// physical register defs.
2900 static bool canClobberPhysRegDefs(const SUnit *SuccSU, const SUnit *SU,
2901                                   const TargetInstrInfo *TII,
2902                                   const TargetRegisterInfo *TRI) {
2903   SDNode *N = SuccSU->getNode();
2904   unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2905   ArrayRef<MCPhysReg> ImpDefs = TII->get(N->getMachineOpcode()).implicit_defs();
2906   assert(!ImpDefs.empty() && "Caller should check hasPhysRegDefs");
2907   for (const SDNode *SUNode = SU->getNode(); SUNode;
2908        SUNode = SUNode->getGluedNode()) {
2909     if (!SUNode->isMachineOpcode())
2910       continue;
2911     ArrayRef<MCPhysReg> SUImpDefs =
2912         TII->get(SUNode->getMachineOpcode()).implicit_defs();
2913     const uint32_t *SURegMask = getNodeRegMask(SUNode);
2914     if (SUImpDefs.empty() && !SURegMask)
2915       continue;
2916     for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
2917       MVT VT = N->getSimpleValueType(i);
2918       if (VT == MVT::Glue || VT == MVT::Other)
2919         continue;
2920       if (!N->hasAnyUseOfValue(i))
2921         continue;
2922       MCPhysReg Reg = ImpDefs[i - NumDefs];
2923       if (SURegMask && MachineOperand::clobbersPhysReg(SURegMask, Reg))
2924         return true;
2925       for (MCPhysReg SUReg : SUImpDefs) {
2926         if (TRI->regsOverlap(Reg, SUReg))
2927           return true;
2928       }
2929     }
2930   }
2931   return false;
2932 }
2933 
2934 /// PrescheduleNodesWithMultipleUses - Nodes with multiple uses
2935 /// are not handled well by the general register pressure reduction
2936 /// heuristics. When presented with code like this:
2937 ///
2938 ///      N
2939 ///    / |
2940 ///   /  |
2941 ///  U  store
2942 ///  |
2943 /// ...
2944 ///
2945 /// the heuristics tend to push the store up, but since the
2946 /// operand of the store has another use (U), this would increase
2947 /// the length of that other use (the U->N edge).
2948 ///
2949 /// This function transforms code like the above to route U's
2950 /// dependence through the store when possible, like this:
2951 ///
2952 ///      N
2953 ///      ||
2954 ///      ||
2955 ///     store
2956 ///       |
2957 ///       U
2958 ///       |
2959 ///      ...
2960 ///
2961 /// This results in the store being scheduled immediately
2962 /// after N, which shortens the U->N live range, reducing
2963 /// register pressure.
2964 void RegReductionPQBase::PrescheduleNodesWithMultipleUses() {
2965   // Visit all the nodes in topological order, working top-down.
2966   for (SUnit &SU : *SUnits) {
2967     // For now, only look at nodes with no data successors, such as stores.
2968     // These are especially important, due to the heuristics in
2969     // getNodePriority for nodes with no data successors.
2970     if (SU.NumSuccs != 0)
2971       continue;
2972     // For now, only look at nodes with exactly one data predecessor.
2973     if (SU.NumPreds != 1)
2974       continue;
2975     // Avoid prescheduling copies to virtual registers, which don't behave
2976     // like other nodes from the perspective of scheduling heuristics.
2977     if (SDNode *N = SU.getNode())
2978       if (N->getOpcode() == ISD::CopyToReg &&
2979           cast<RegisterSDNode>(N->getOperand(1))->getReg().isVirtual())
2980         continue;
2981 
2982     SDNode *PredFrameSetup = nullptr;
2983     for (const SDep &Pred : SU.Preds)
2984       if (Pred.isCtrl() && Pred.getSUnit()) {
2985         // Find the predecessor which is not data dependence.
2986         SDNode *PredND = Pred.getSUnit()->getNode();
2987 
2988         // If PredND is FrameSetup, we should not pre-scheduled the node,
2989         // or else, when bottom up scheduling, ADJCALLSTACKDOWN and
2990         // ADJCALLSTACKUP may hold CallResource too long and make other
2991         // calls can't be scheduled. If there's no other available node
2992         // to schedule, the schedular will try to rename the register by
2993         // creating copy to avoid the conflict which will fail because
2994         // CallResource is not a real physical register.
2995         if (PredND && PredND->isMachineOpcode() &&
2996             (PredND->getMachineOpcode() == TII->getCallFrameSetupOpcode())) {
2997           PredFrameSetup = PredND;
2998           break;
2999         }
3000       }
3001     // Skip the node has FrameSetup parent.
3002     if (PredFrameSetup != nullptr)
3003       continue;
3004 
3005     // Locate the single data predecessor.
3006     SUnit *PredSU = nullptr;
3007     for (const SDep &Pred : SU.Preds)
3008       if (!Pred.isCtrl()) {
3009         PredSU = Pred.getSUnit();
3010         break;
3011       }
3012     assert(PredSU);
3013 
3014     // Don't rewrite edges that carry physregs, because that requires additional
3015     // support infrastructure.
3016     if (PredSU->hasPhysRegDefs)
3017       continue;
3018     // Short-circuit the case where SU is PredSU's only data successor.
3019     if (PredSU->NumSuccs == 1)
3020       continue;
3021     // Avoid prescheduling to copies from virtual registers, which don't behave
3022     // like other nodes from the perspective of scheduling heuristics.
3023     if (SDNode *N = SU.getNode())
3024       if (N->getOpcode() == ISD::CopyFromReg &&
3025           cast<RegisterSDNode>(N->getOperand(1))->getReg().isVirtual())
3026         continue;
3027 
3028     // Perform checks on the successors of PredSU.
3029     for (const SDep &PredSucc : PredSU->Succs) {
3030       SUnit *PredSuccSU = PredSucc.getSUnit();
3031       if (PredSuccSU == &SU) continue;
3032       // If PredSU has another successor with no data successors, for
3033       // now don't attempt to choose either over the other.
3034       if (PredSuccSU->NumSuccs == 0)
3035         goto outer_loop_continue;
3036       // Don't break physical register dependencies.
3037       if (SU.hasPhysRegClobbers && PredSuccSU->hasPhysRegDefs)
3038         if (canClobberPhysRegDefs(PredSuccSU, &SU, TII, TRI))
3039           goto outer_loop_continue;
3040       // Don't introduce graph cycles.
3041       if (scheduleDAG->IsReachable(&SU, PredSuccSU))
3042         goto outer_loop_continue;
3043     }
3044 
3045     // Ok, the transformation is safe and the heuristics suggest it is
3046     // profitable. Update the graph.
3047     LLVM_DEBUG(
3048         dbgs() << "    Prescheduling SU #" << SU.NodeNum << " next to PredSU #"
3049                << PredSU->NodeNum
3050                << " to guide scheduling in the presence of multiple uses\n");
3051     for (unsigned i = 0; i != PredSU->Succs.size(); ++i) {
3052       SDep Edge = PredSU->Succs[i];
3053       assert(!Edge.isAssignedRegDep());
3054       SUnit *SuccSU = Edge.getSUnit();
3055       if (SuccSU != &SU) {
3056         Edge.setSUnit(PredSU);
3057         scheduleDAG->RemovePred(SuccSU, Edge);
3058         scheduleDAG->AddPredQueued(&SU, Edge);
3059         Edge.setSUnit(&SU);
3060         scheduleDAG->AddPredQueued(SuccSU, Edge);
3061         --i;
3062       }
3063     }
3064   outer_loop_continue:;
3065   }
3066 }
3067 
3068 /// AddPseudoTwoAddrDeps - If two nodes share an operand and one of them uses
3069 /// it as a def&use operand. Add a pseudo control edge from it to the other
3070 /// node (if it won't create a cycle) so the two-address one will be scheduled
3071 /// first (lower in the schedule). If both nodes are two-address, favor the
3072 /// one that has a CopyToReg use (more likely to be a loop induction update).
3073 /// If both are two-address, but one is commutable while the other is not
3074 /// commutable, favor the one that's not commutable.
3075 void RegReductionPQBase::AddPseudoTwoAddrDeps() {
3076   for (SUnit &SU : *SUnits) {
3077     if (!SU.isTwoAddress)
3078       continue;
3079 
3080     SDNode *Node = SU.getNode();
3081     if (!Node || !Node->isMachineOpcode() || SU.getNode()->getGluedNode())
3082       continue;
3083 
3084     bool isLiveOut = hasOnlyLiveOutUses(&SU);
3085     unsigned Opc = Node->getMachineOpcode();
3086     const MCInstrDesc &MCID = TII->get(Opc);
3087     unsigned NumRes = MCID.getNumDefs();
3088     unsigned NumOps = MCID.getNumOperands() - NumRes;
3089     for (unsigned j = 0; j != NumOps; ++j) {
3090       if (MCID.getOperandConstraint(j+NumRes, MCOI::TIED_TO) == -1)
3091         continue;
3092       SDNode *DU = SU.getNode()->getOperand(j).getNode();
3093       if (DU->getNodeId() == -1)
3094         continue;
3095       const SUnit *DUSU = &(*SUnits)[DU->getNodeId()];
3096       if (!DUSU)
3097         continue;
3098       for (const SDep &Succ : DUSU->Succs) {
3099         if (Succ.isCtrl())
3100           continue;
3101         SUnit *SuccSU = Succ.getSUnit();
3102         if (SuccSU == &SU)
3103           continue;
3104         // Be conservative. Ignore if nodes aren't at roughly the same
3105         // depth and height.
3106         if (SuccSU->getHeight() < SU.getHeight() &&
3107             (SU.getHeight() - SuccSU->getHeight()) > 1)
3108           continue;
3109         // Skip past COPY_TO_REGCLASS nodes, so that the pseudo edge
3110         // constrains whatever is using the copy, instead of the copy
3111         // itself. In the case that the copy is coalesced, this
3112         // preserves the intent of the pseudo two-address heurietics.
3113         while (SuccSU->Succs.size() == 1 &&
3114                SuccSU->getNode()->isMachineOpcode() &&
3115                SuccSU->getNode()->getMachineOpcode() ==
3116                  TargetOpcode::COPY_TO_REGCLASS)
3117           SuccSU = SuccSU->Succs.front().getSUnit();
3118         // Don't constrain non-instruction nodes.
3119         if (!SuccSU->getNode() || !SuccSU->getNode()->isMachineOpcode())
3120           continue;
3121         // Don't constrain nodes with physical register defs if the
3122         // predecessor can clobber them.
3123         if (SuccSU->hasPhysRegDefs && SU.hasPhysRegClobbers) {
3124           if (canClobberPhysRegDefs(SuccSU, &SU, TII, TRI))
3125             continue;
3126         }
3127         // Don't constrain EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG;
3128         // these may be coalesced away. We want them close to their uses.
3129         unsigned SuccOpc = SuccSU->getNode()->getMachineOpcode();
3130         if (SuccOpc == TargetOpcode::EXTRACT_SUBREG ||
3131             SuccOpc == TargetOpcode::INSERT_SUBREG ||
3132             SuccOpc == TargetOpcode::SUBREG_TO_REG)
3133           continue;
3134         if (!canClobberReachingPhysRegUse(SuccSU, &SU, scheduleDAG, TII, TRI) &&
3135             (!canClobber(SuccSU, DUSU) ||
3136              (isLiveOut && !hasOnlyLiveOutUses(SuccSU)) ||
3137              (!SU.isCommutable && SuccSU->isCommutable)) &&
3138             !scheduleDAG->IsReachable(SuccSU, &SU)) {
3139           LLVM_DEBUG(dbgs()
3140                      << "    Adding a pseudo-two-addr edge from SU #"
3141                      << SU.NodeNum << " to SU #" << SuccSU->NodeNum << "\n");
3142           scheduleDAG->AddPredQueued(&SU, SDep(SuccSU, SDep::Artificial));
3143         }
3144       }
3145     }
3146   }
3147 }
3148 
3149 //===----------------------------------------------------------------------===//
3150 //                         Public Constructor Functions
3151 //===----------------------------------------------------------------------===//
3152 
3153 ScheduleDAGSDNodes *
3154 llvm::createBURRListDAGScheduler(SelectionDAGISel *IS,
3155                                  CodeGenOpt::Level OptLevel) {
3156   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3157   const TargetInstrInfo *TII = STI.getInstrInfo();
3158   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3159 
3160   BURegReductionPriorityQueue *PQ =
3161     new BURegReductionPriorityQueue(*IS->MF, false, false, TII, TRI, nullptr);
3162   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
3163   PQ->setScheduleDAG(SD);
3164   return SD;
3165 }
3166 
3167 ScheduleDAGSDNodes *
3168 llvm::createSourceListDAGScheduler(SelectionDAGISel *IS,
3169                                    CodeGenOpt::Level OptLevel) {
3170   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3171   const TargetInstrInfo *TII = STI.getInstrInfo();
3172   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3173 
3174   SrcRegReductionPriorityQueue *PQ =
3175     new SrcRegReductionPriorityQueue(*IS->MF, false, true, TII, TRI, nullptr);
3176   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
3177   PQ->setScheduleDAG(SD);
3178   return SD;
3179 }
3180 
3181 ScheduleDAGSDNodes *
3182 llvm::createHybridListDAGScheduler(SelectionDAGISel *IS,
3183                                    CodeGenOpt::Level OptLevel) {
3184   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3185   const TargetInstrInfo *TII = STI.getInstrInfo();
3186   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3187   const TargetLowering *TLI = IS->TLI;
3188 
3189   HybridBURRPriorityQueue *PQ =
3190     new HybridBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI);
3191 
3192   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
3193   PQ->setScheduleDAG(SD);
3194   return SD;
3195 }
3196 
3197 ScheduleDAGSDNodes *
3198 llvm::createILPListDAGScheduler(SelectionDAGISel *IS,
3199                                 CodeGenOpt::Level OptLevel) {
3200   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3201   const TargetInstrInfo *TII = STI.getInstrInfo();
3202   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3203   const TargetLowering *TLI = IS->TLI;
3204 
3205   ILPBURRPriorityQueue *PQ =
3206     new ILPBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI);
3207   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
3208   PQ->setScheduleDAG(SD);
3209   return SD;
3210 }
3211