xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp (revision 7a6dacaca14b62ca4b74406814becb87a3fefac0)
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/MachineValueType.h"
28 #include "llvm/CodeGen/Register.h"
29 #include "llvm/CodeGen/ScheduleDAG.h"
30 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
31 #include "llvm/CodeGen/SchedulerRegistry.h"
32 #include "llvm/CodeGen/SelectionDAGISel.h"
33 #include "llvm/CodeGen/SelectionDAGNodes.h"
34 #include "llvm/CodeGen/TargetInstrInfo.h"
35 #include "llvm/CodeGen/TargetLowering.h"
36 #include "llvm/CodeGen/TargetOpcodes.h"
37 #include "llvm/CodeGen/TargetRegisterInfo.h"
38 #include "llvm/CodeGen/TargetSubtargetInfo.h"
39 #include "llvm/Config/llvm-config.h"
40 #include "llvm/IR/InlineAsm.h"
41 #include "llvm/MC/MCInstrDesc.h"
42 #include "llvm/MC/MCRegisterInfo.h"
43 #include "llvm/Support/Casting.h"
44 #include "llvm/Support/CodeGen.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Compiler.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.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 = ~0u;
160 
161   /// IssueCount - Count instructions issued in this cycle
162   /// Currently valid only for bottom-up scheduling.
163   unsigned IssueCount = 0u;
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 = 0u;
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                     CodeGenOptLevel OptLevel)
192       : ScheduleDAGSDNodes(mf), NeedLatency(needlatency),
193         AvailableQueue(availqueue), Topo(SUnits, nullptr) {
194     const TargetSubtargetInfo &STI = mf.getSubtarget();
195     if (DisableSchedCycles || !NeedLatency)
196       HazardRec = new ScheduleHazardRecognizer();
197     else
198       HazardRec = STI.getInstrInfo()->CreateTargetHazardRecognizer(&STI, this);
199   }
200 
201   ~ScheduleDAGRRList() override {
202     delete HazardRec;
203     delete AvailableQueue;
204   }
205 
206   void Schedule() override;
207 
208   ScheduleHazardRecognizer *getHazardRec() { return HazardRec; }
209 
210   /// IsReachable - Checks if SU is reachable from TargetSU.
211   bool IsReachable(const SUnit *SU, const SUnit *TargetSU) {
212     return Topo.IsReachable(SU, TargetSU);
213   }
214 
215   /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
216   /// create a cycle.
217   bool WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
218     return Topo.WillCreateCycle(SU, TargetSU);
219   }
220 
221   /// AddPredQueued - Queues and update to add a predecessor edge to SUnit SU.
222   /// This returns true if this is a new predecessor.
223   /// Does *NOT* update the topological ordering! It just queues an update.
224   void AddPredQueued(SUnit *SU, const SDep &D) {
225     Topo.AddPredQueued(SU, D.getSUnit());
226     SU->addPred(D);
227   }
228 
229   /// AddPred - adds a predecessor edge to SUnit SU.
230   /// This returns true if this is a new predecessor.
231   /// Updates the topological ordering if required.
232   void AddPred(SUnit *SU, const SDep &D) {
233     Topo.AddPred(SU, D.getSUnit());
234     SU->addPred(D);
235   }
236 
237   /// RemovePred - removes a predecessor edge from SUnit SU.
238   /// This returns true if an edge was removed.
239   /// Updates the topological ordering if required.
240   void RemovePred(SUnit *SU, const SDep &D) {
241     Topo.RemovePred(SU, D.getSUnit());
242     SU->removePred(D);
243   }
244 
245 private:
246   bool isReady(SUnit *SU) {
247     return DisableSchedCycles || !AvailableQueue->hasReadyFilter() ||
248       AvailableQueue->isReady(SU);
249   }
250 
251   void ReleasePred(SUnit *SU, const SDep *PredEdge);
252   void ReleasePredecessors(SUnit *SU);
253   void ReleasePending();
254   void AdvanceToCycle(unsigned NextCycle);
255   void AdvancePastStalls(SUnit *SU);
256   void EmitNode(SUnit *SU);
257   void ScheduleNodeBottomUp(SUnit*);
258   void CapturePred(SDep *PredEdge);
259   void UnscheduleNodeBottomUp(SUnit*);
260   void RestoreHazardCheckerBottomUp();
261   void BacktrackBottomUp(SUnit*, SUnit*);
262   SUnit *TryUnfoldSU(SUnit *);
263   SUnit *CopyAndMoveSuccessors(SUnit*);
264   void InsertCopiesAndMoveSuccs(SUnit*, unsigned,
265                                 const TargetRegisterClass*,
266                                 const TargetRegisterClass*,
267                                 SmallVectorImpl<SUnit*>&);
268   bool DelayForLiveRegsBottomUp(SUnit*, SmallVectorImpl<unsigned>&);
269 
270   void releaseInterferences(unsigned Reg = 0);
271 
272   SUnit *PickNodeToScheduleBottomUp();
273   void ListScheduleBottomUp();
274 
275   /// CreateNewSUnit - Creates a new SUnit and returns a pointer to it.
276   SUnit *CreateNewSUnit(SDNode *N) {
277     unsigned NumSUnits = SUnits.size();
278     SUnit *NewNode = newSUnit(N);
279     // Update the topological ordering.
280     if (NewNode->NodeNum >= NumSUnits)
281       Topo.AddSUnitWithoutPredecessors(NewNode);
282     return NewNode;
283   }
284 
285   /// CreateClone - Creates a new SUnit from an existing one.
286   SUnit *CreateClone(SUnit *N) {
287     unsigned NumSUnits = SUnits.size();
288     SUnit *NewNode = Clone(N);
289     // Update the topological ordering.
290     if (NewNode->NodeNum >= NumSUnits)
291       Topo.AddSUnitWithoutPredecessors(NewNode);
292     return NewNode;
293   }
294 
295   /// forceUnitLatencies - Register-pressure-reducing scheduling doesn't
296   /// need actual latency information but the hybrid scheduler does.
297   bool forceUnitLatencies() const override {
298     return !NeedLatency;
299   }
300 };
301 
302 }  // end anonymous namespace
303 
304 static constexpr unsigned RegSequenceCost = 1;
305 
306 /// GetCostForDef - Looks up the register class and cost for a given definition.
307 /// Typically this just means looking up the representative register class,
308 /// but for untyped values (MVT::Untyped) it means inspecting the node's
309 /// opcode to determine what register class is being generated.
310 static void GetCostForDef(const ScheduleDAGSDNodes::RegDefIter &RegDefPos,
311                           const TargetLowering *TLI,
312                           const TargetInstrInfo *TII,
313                           const TargetRegisterInfo *TRI,
314                           unsigned &RegClass, unsigned &Cost,
315                           const MachineFunction &MF) {
316   MVT VT = RegDefPos.GetValue();
317 
318   // Special handling for untyped values.  These values can only come from
319   // the expansion of custom DAG-to-DAG patterns.
320   if (VT == MVT::Untyped) {
321     const SDNode *Node = RegDefPos.GetNode();
322 
323     // Special handling for CopyFromReg of untyped values.
324     if (!Node->isMachineOpcode() && Node->getOpcode() == ISD::CopyFromReg) {
325       Register Reg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
326       const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(Reg);
327       RegClass = RC->getID();
328       Cost = 1;
329       return;
330     }
331 
332     unsigned Opcode = Node->getMachineOpcode();
333     if (Opcode == TargetOpcode::REG_SEQUENCE) {
334       unsigned DstRCIdx = Node->getConstantOperandVal(0);
335       const TargetRegisterClass *RC = TRI->getRegClass(DstRCIdx);
336       RegClass = RC->getID();
337       Cost = RegSequenceCost;
338       return;
339     }
340 
341     unsigned Idx = RegDefPos.GetIdx();
342     const MCInstrDesc &Desc = TII->get(Opcode);
343     const TargetRegisterClass *RC = TII->getRegClass(Desc, Idx, TRI, MF);
344     assert(RC && "Not a valid register class");
345     RegClass = RC->getID();
346     // FIXME: Cost arbitrarily set to 1 because there doesn't seem to be a
347     // better way to determine it.
348     Cost = 1;
349   } else {
350     RegClass = TLI->getRepRegClassFor(VT)->getID();
351     Cost = TLI->getRepRegClassCostFor(VT);
352   }
353 }
354 
355 /// Schedule - Schedule the DAG using list scheduling.
356 void ScheduleDAGRRList::Schedule() {
357   LLVM_DEBUG(dbgs() << "********** List Scheduling " << printMBBReference(*BB)
358                     << " '" << BB->getName() << "' **********\n");
359 
360   CurCycle = 0;
361   IssueCount = 0;
362   MinAvailableCycle =
363       DisableSchedCycles ? 0 : std::numeric_limits<unsigned>::max();
364   NumLiveRegs = 0;
365   // Allocate slots for each physical register, plus one for a special register
366   // to track the virtual resource of a calling sequence.
367   LiveRegDefs.reset(new SUnit*[TRI->getNumRegs() + 1]());
368   LiveRegGens.reset(new SUnit*[TRI->getNumRegs() + 1]());
369   CallSeqEndForStart.clear();
370   assert(Interferences.empty() && LRegsMap.empty() && "stale Interferences");
371 
372   // Build the scheduling graph.
373   BuildSchedGraph(nullptr);
374 
375   LLVM_DEBUG(dump());
376   Topo.MarkDirty();
377 
378   AvailableQueue->initNodes(SUnits);
379 
380   HazardRec->Reset();
381 
382   // Execute the actual scheduling loop.
383   ListScheduleBottomUp();
384 
385   AvailableQueue->releaseState();
386 
387   LLVM_DEBUG({
388     dbgs() << "*** Final schedule ***\n";
389     dumpSchedule();
390     dbgs() << '\n';
391   });
392 }
393 
394 //===----------------------------------------------------------------------===//
395 //  Bottom-Up Scheduling
396 //===----------------------------------------------------------------------===//
397 
398 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
399 /// the AvailableQueue if the count reaches zero. Also update its cycle bound.
400 void ScheduleDAGRRList::ReleasePred(SUnit *SU, const SDep *PredEdge) {
401   SUnit *PredSU = PredEdge->getSUnit();
402 
403 #ifndef NDEBUG
404   if (PredSU->NumSuccsLeft == 0) {
405     dbgs() << "*** Scheduling failed! ***\n";
406     dumpNode(*PredSU);
407     dbgs() << " has been released too many times!\n";
408     llvm_unreachable(nullptr);
409   }
410 #endif
411   --PredSU->NumSuccsLeft;
412 
413   if (!forceUnitLatencies()) {
414     // Updating predecessor's height. This is now the cycle when the
415     // predecessor can be scheduled without causing a pipeline stall.
416     PredSU->setHeightToAtLeast(SU->getHeight() + PredEdge->getLatency());
417   }
418 
419   // If all the node's successors are scheduled, this node is ready
420   // to be scheduled. Ignore the special EntrySU node.
421   if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU) {
422     PredSU->isAvailable = true;
423 
424     unsigned Height = PredSU->getHeight();
425     if (Height < MinAvailableCycle)
426       MinAvailableCycle = Height;
427 
428     if (isReady(PredSU)) {
429       AvailableQueue->push(PredSU);
430     }
431     // CapturePred and others may have left the node in the pending queue, avoid
432     // adding it twice.
433     else if (!PredSU->isPending) {
434       PredSU->isPending = true;
435       PendingQueue.push_back(PredSU);
436     }
437   }
438 }
439 
440 /// IsChainDependent - Test if Outer is reachable from Inner through
441 /// chain dependencies.
442 static bool IsChainDependent(SDNode *Outer, SDNode *Inner,
443                              unsigned NestLevel,
444                              const TargetInstrInfo *TII) {
445   SDNode *N = Outer;
446   while (true) {
447     if (N == Inner)
448       return true;
449     // For a TokenFactor, examine each operand. There may be multiple ways
450     // to get to the CALLSEQ_BEGIN, but we need to find the path with the
451     // most nesting in order to ensure that we find the corresponding match.
452     if (N->getOpcode() == ISD::TokenFactor) {
453       for (const SDValue &Op : N->op_values())
454         if (IsChainDependent(Op.getNode(), Inner, NestLevel, TII))
455           return true;
456       return false;
457     }
458     // Check for a lowered CALLSEQ_BEGIN or CALLSEQ_END.
459     if (N->isMachineOpcode()) {
460       if (N->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) {
461         ++NestLevel;
462       } else if (N->getMachineOpcode() == TII->getCallFrameSetupOpcode()) {
463         if (NestLevel == 0)
464           return false;
465         --NestLevel;
466       }
467     }
468     // Otherwise, find the chain and continue climbing.
469     for (const SDValue &Op : N->op_values())
470       if (Op.getValueType() == MVT::Other) {
471         N = Op.getNode();
472         goto found_chain_operand;
473       }
474     return false;
475   found_chain_operand:;
476     if (N->getOpcode() == ISD::EntryToken)
477       return false;
478   }
479 }
480 
481 /// FindCallSeqStart - Starting from the (lowered) CALLSEQ_END node, locate
482 /// the corresponding (lowered) CALLSEQ_BEGIN node.
483 ///
484 /// NestLevel and MaxNested are used in recursion to indcate the current level
485 /// of nesting of CALLSEQ_BEGIN and CALLSEQ_END pairs, as well as the maximum
486 /// level seen so far.
487 ///
488 /// TODO: It would be better to give CALLSEQ_END an explicit operand to point
489 /// to the corresponding CALLSEQ_BEGIN to avoid needing to search for it.
490 static SDNode *
491 FindCallSeqStart(SDNode *N, unsigned &NestLevel, unsigned &MaxNest,
492                  const TargetInstrInfo *TII) {
493   while (true) {
494     // For a TokenFactor, examine each operand. There may be multiple ways
495     // to get to the CALLSEQ_BEGIN, but we need to find the path with the
496     // most nesting in order to ensure that we find the corresponding match.
497     if (N->getOpcode() == ISD::TokenFactor) {
498       SDNode *Best = nullptr;
499       unsigned BestMaxNest = MaxNest;
500       for (const SDValue &Op : N->op_values()) {
501         unsigned MyNestLevel = NestLevel;
502         unsigned MyMaxNest = MaxNest;
503         if (SDNode *New = FindCallSeqStart(Op.getNode(),
504                                            MyNestLevel, MyMaxNest, TII))
505           if (!Best || (MyMaxNest > BestMaxNest)) {
506             Best = New;
507             BestMaxNest = MyMaxNest;
508           }
509       }
510       assert(Best);
511       MaxNest = BestMaxNest;
512       return Best;
513     }
514     // Check for a lowered CALLSEQ_BEGIN or CALLSEQ_END.
515     if (N->isMachineOpcode()) {
516       if (N->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) {
517         ++NestLevel;
518         MaxNest = std::max(MaxNest, NestLevel);
519       } else if (N->getMachineOpcode() == TII->getCallFrameSetupOpcode()) {
520         assert(NestLevel != 0);
521         --NestLevel;
522         if (NestLevel == 0)
523           return N;
524       }
525     }
526     // Otherwise, find the chain and continue climbing.
527     for (const SDValue &Op : N->op_values())
528       if (Op.getValueType() == MVT::Other) {
529         N = Op.getNode();
530         goto found_chain_operand;
531       }
532     return nullptr;
533   found_chain_operand:;
534     if (N->getOpcode() == ISD::EntryToken)
535       return nullptr;
536   }
537 }
538 
539 /// Call ReleasePred for each predecessor, then update register live def/gen.
540 /// Always update LiveRegDefs for a register dependence even if the current SU
541 /// also defines the register. This effectively create one large live range
542 /// across a sequence of two-address node. This is important because the
543 /// entire chain must be scheduled together. Example:
544 ///
545 /// flags = (3) add
546 /// flags = (2) addc flags
547 /// flags = (1) addc flags
548 ///
549 /// results in
550 ///
551 /// LiveRegDefs[flags] = 3
552 /// LiveRegGens[flags] = 1
553 ///
554 /// If (2) addc is unscheduled, then (1) addc must also be unscheduled to avoid
555 /// interference on flags.
556 void ScheduleDAGRRList::ReleasePredecessors(SUnit *SU) {
557   // Bottom up: release predecessors
558   for (SDep &Pred : SU->Preds) {
559     ReleasePred(SU, &Pred);
560     if (Pred.isAssignedRegDep()) {
561       // This is a physical register dependency and it's impossible or
562       // expensive to copy the register. Make sure nothing that can
563       // clobber the register is scheduled between the predecessor and
564       // this node.
565       SUnit *RegDef = LiveRegDefs[Pred.getReg()]; (void)RegDef;
566       assert((!RegDef || RegDef == SU || RegDef == Pred.getSUnit()) &&
567              "interference on register dependence");
568       LiveRegDefs[Pred.getReg()] = Pred.getSUnit();
569       if (!LiveRegGens[Pred.getReg()]) {
570         ++NumLiveRegs;
571         LiveRegGens[Pred.getReg()] = SU;
572       }
573     }
574   }
575 
576   // If we're scheduling a lowered CALLSEQ_END, find the corresponding
577   // CALLSEQ_BEGIN. Inject an artificial physical register dependence between
578   // these nodes, to prevent other calls from being interscheduled with them.
579   unsigned CallResource = TRI->getNumRegs();
580   if (!LiveRegDefs[CallResource])
581     for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode())
582       if (Node->isMachineOpcode() &&
583           Node->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) {
584         unsigned NestLevel = 0;
585         unsigned MaxNest = 0;
586         SDNode *N = FindCallSeqStart(Node, NestLevel, MaxNest, TII);
587         assert(N && "Must find call sequence start");
588 
589         SUnit *Def = &SUnits[N->getNodeId()];
590         CallSeqEndForStart[Def] = SU;
591 
592         ++NumLiveRegs;
593         LiveRegDefs[CallResource] = Def;
594         LiveRegGens[CallResource] = SU;
595         break;
596       }
597 }
598 
599 /// Check to see if any of the pending instructions are ready to issue.  If
600 /// so, add them to the available queue.
601 void ScheduleDAGRRList::ReleasePending() {
602   if (DisableSchedCycles) {
603     assert(PendingQueue.empty() && "pending instrs not allowed in this mode");
604     return;
605   }
606 
607   // If the available queue is empty, it is safe to reset MinAvailableCycle.
608   if (AvailableQueue->empty())
609     MinAvailableCycle = std::numeric_limits<unsigned>::max();
610 
611   // Check to see if any of the pending instructions are ready to issue.  If
612   // so, add them to the available queue.
613   for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
614     unsigned ReadyCycle = PendingQueue[i]->getHeight();
615     if (ReadyCycle < MinAvailableCycle)
616       MinAvailableCycle = ReadyCycle;
617 
618     if (PendingQueue[i]->isAvailable) {
619       if (!isReady(PendingQueue[i]))
620           continue;
621       AvailableQueue->push(PendingQueue[i]);
622     }
623     PendingQueue[i]->isPending = false;
624     PendingQueue[i] = PendingQueue.back();
625     PendingQueue.pop_back();
626     --i; --e;
627   }
628 }
629 
630 /// Move the scheduler state forward by the specified number of Cycles.
631 void ScheduleDAGRRList::AdvanceToCycle(unsigned NextCycle) {
632   if (NextCycle <= CurCycle)
633     return;
634 
635   IssueCount = 0;
636   AvailableQueue->setCurCycle(NextCycle);
637   if (!HazardRec->isEnabled()) {
638     // Bypass lots of virtual calls in case of long latency.
639     CurCycle = NextCycle;
640   }
641   else {
642     for (; CurCycle != NextCycle; ++CurCycle) {
643       HazardRec->RecedeCycle();
644     }
645   }
646   // FIXME: Instead of visiting the pending Q each time, set a dirty flag on the
647   // available Q to release pending nodes at least once before popping.
648   ReleasePending();
649 }
650 
651 /// Move the scheduler state forward until the specified node's dependents are
652 /// ready and can be scheduled with no resource conflicts.
653 void ScheduleDAGRRList::AdvancePastStalls(SUnit *SU) {
654   if (DisableSchedCycles)
655     return;
656 
657   // FIXME: Nodes such as CopyFromReg probably should not advance the current
658   // cycle. Otherwise, we can wrongly mask real stalls. If the non-machine node
659   // has predecessors the cycle will be advanced when they are scheduled.
660   // But given the crude nature of modeling latency though such nodes, we
661   // currently need to treat these nodes like real instructions.
662   // if (!SU->getNode() || !SU->getNode()->isMachineOpcode()) return;
663 
664   unsigned ReadyCycle = SU->getHeight();
665 
666   // Bump CurCycle to account for latency. We assume the latency of other
667   // available instructions may be hidden by the stall (not a full pipe stall).
668   // This updates the hazard recognizer's cycle before reserving resources for
669   // this instruction.
670   AdvanceToCycle(ReadyCycle);
671 
672   // Calls are scheduled in their preceding cycle, so don't conflict with
673   // hazards from instructions after the call. EmitNode will reset the
674   // scoreboard state before emitting the call.
675   if (SU->isCall)
676     return;
677 
678   // FIXME: For resource conflicts in very long non-pipelined stages, we
679   // should probably skip ahead here to avoid useless scoreboard checks.
680   int Stalls = 0;
681   while (true) {
682     ScheduleHazardRecognizer::HazardType HT =
683       HazardRec->getHazardType(SU, -Stalls);
684 
685     if (HT == ScheduleHazardRecognizer::NoHazard)
686       break;
687 
688     ++Stalls;
689   }
690   AdvanceToCycle(CurCycle + Stalls);
691 }
692 
693 /// Record this SUnit in the HazardRecognizer.
694 /// Does not update CurCycle.
695 void ScheduleDAGRRList::EmitNode(SUnit *SU) {
696   if (!HazardRec->isEnabled())
697     return;
698 
699   // Check for phys reg copy.
700   if (!SU->getNode())
701     return;
702 
703   switch (SU->getNode()->getOpcode()) {
704   default:
705     assert(SU->getNode()->isMachineOpcode() &&
706            "This target-independent node should not be scheduled.");
707     break;
708   case ISD::MERGE_VALUES:
709   case ISD::TokenFactor:
710   case ISD::LIFETIME_START:
711   case ISD::LIFETIME_END:
712   case ISD::CopyToReg:
713   case ISD::CopyFromReg:
714   case ISD::EH_LABEL:
715     // Noops don't affect the scoreboard state. Copies are likely to be
716     // removed.
717     return;
718   case ISD::INLINEASM:
719   case ISD::INLINEASM_BR:
720     // For inline asm, clear the pipeline state.
721     HazardRec->Reset();
722     return;
723   }
724   if (SU->isCall) {
725     // Calls are scheduled with their preceding instructions. For bottom-up
726     // scheduling, clear the pipeline state before emitting.
727     HazardRec->Reset();
728   }
729 
730   HazardRec->EmitInstruction(SU);
731 }
732 
733 static void resetVRegCycle(SUnit *SU);
734 
735 /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
736 /// count of its predecessors. If a predecessor pending count is zero, add it to
737 /// the Available queue.
738 void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU) {
739   LLVM_DEBUG(dbgs() << "\n*** Scheduling [" << CurCycle << "]: ");
740   LLVM_DEBUG(dumpNode(*SU));
741 
742 #ifndef NDEBUG
743   if (CurCycle < SU->getHeight())
744     LLVM_DEBUG(dbgs() << "   Height [" << SU->getHeight()
745                       << "] pipeline stall!\n");
746 #endif
747 
748   // FIXME: Do not modify node height. It may interfere with
749   // backtracking. Instead add a "ready cycle" to SUnit. Before scheduling the
750   // node its ready cycle can aid heuristics, and after scheduling it can
751   // indicate the scheduled cycle.
752   SU->setHeightToAtLeast(CurCycle);
753 
754   // Reserve resources for the scheduled instruction.
755   EmitNode(SU);
756 
757   Sequence.push_back(SU);
758 
759   AvailableQueue->scheduledNode(SU);
760 
761   // If HazardRec is disabled, and each inst counts as one cycle, then
762   // advance CurCycle before ReleasePredecessors to avoid useless pushes to
763   // PendingQueue for schedulers that implement HasReadyFilter.
764   if (!HazardRec->isEnabled() && AvgIPC < 2)
765     AdvanceToCycle(CurCycle + 1);
766 
767   // Update liveness of predecessors before successors to avoid treating a
768   // two-address node as a live range def.
769   ReleasePredecessors(SU);
770 
771   // Release all the implicit physical register defs that are live.
772   for (SDep &Succ : SU->Succs) {
773     // LiveRegDegs[Succ.getReg()] != SU when SU is a two-address node.
774     if (Succ.isAssignedRegDep() && LiveRegDefs[Succ.getReg()] == SU) {
775       assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
776       --NumLiveRegs;
777       LiveRegDefs[Succ.getReg()] = nullptr;
778       LiveRegGens[Succ.getReg()] = nullptr;
779       releaseInterferences(Succ.getReg());
780     }
781   }
782   // Release the special call resource dependence, if this is the beginning
783   // of a call.
784   unsigned CallResource = TRI->getNumRegs();
785   if (LiveRegDefs[CallResource] == SU)
786     for (const SDNode *SUNode = SU->getNode(); SUNode;
787          SUNode = SUNode->getGluedNode()) {
788       if (SUNode->isMachineOpcode() &&
789           SUNode->getMachineOpcode() == TII->getCallFrameSetupOpcode()) {
790         assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
791         --NumLiveRegs;
792         LiveRegDefs[CallResource] = nullptr;
793         LiveRegGens[CallResource] = nullptr;
794         releaseInterferences(CallResource);
795       }
796     }
797 
798   resetVRegCycle(SU);
799 
800   SU->isScheduled = true;
801 
802   // Conditions under which the scheduler should eagerly advance the cycle:
803   // (1) No available instructions
804   // (2) All pipelines full, so available instructions must have hazards.
805   //
806   // If HazardRec is disabled, the cycle was pre-advanced before calling
807   // ReleasePredecessors. In that case, IssueCount should remain 0.
808   //
809   // Check AvailableQueue after ReleasePredecessors in case of zero latency.
810   if (HazardRec->isEnabled() || AvgIPC > 1) {
811     if (SU->getNode() && SU->getNode()->isMachineOpcode())
812       ++IssueCount;
813     if ((HazardRec->isEnabled() && HazardRec->atIssueLimit())
814         || (!HazardRec->isEnabled() && IssueCount == AvgIPC))
815       AdvanceToCycle(CurCycle + 1);
816   }
817 }
818 
819 /// CapturePred - This does the opposite of ReleasePred. Since SU is being
820 /// unscheduled, increase the succ left count of its predecessors. Remove
821 /// them from AvailableQueue if necessary.
822 void ScheduleDAGRRList::CapturePred(SDep *PredEdge) {
823   SUnit *PredSU = PredEdge->getSUnit();
824   if (PredSU->isAvailable) {
825     PredSU->isAvailable = false;
826     if (!PredSU->isPending)
827       AvailableQueue->remove(PredSU);
828   }
829 
830   assert(PredSU->NumSuccsLeft < std::numeric_limits<unsigned>::max() &&
831          "NumSuccsLeft will overflow!");
832   ++PredSU->NumSuccsLeft;
833 }
834 
835 /// UnscheduleNodeBottomUp - Remove the node from the schedule, update its and
836 /// its predecessor states to reflect the change.
837 void ScheduleDAGRRList::UnscheduleNodeBottomUp(SUnit *SU) {
838   LLVM_DEBUG(dbgs() << "*** Unscheduling [" << SU->getHeight() << "]: ");
839   LLVM_DEBUG(dumpNode(*SU));
840 
841   for (SDep &Pred : SU->Preds) {
842     CapturePred(&Pred);
843     if (Pred.isAssignedRegDep() && SU == LiveRegGens[Pred.getReg()]){
844       assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
845       assert(LiveRegDefs[Pred.getReg()] == Pred.getSUnit() &&
846              "Physical register dependency violated?");
847       --NumLiveRegs;
848       LiveRegDefs[Pred.getReg()] = nullptr;
849       LiveRegGens[Pred.getReg()] = nullptr;
850       releaseInterferences(Pred.getReg());
851     }
852   }
853 
854   // Reclaim the special call resource dependence, if this is the beginning
855   // of a call.
856   unsigned CallResource = TRI->getNumRegs();
857   for (const SDNode *SUNode = SU->getNode(); SUNode;
858        SUNode = SUNode->getGluedNode()) {
859     if (SUNode->isMachineOpcode() &&
860         SUNode->getMachineOpcode() == TII->getCallFrameSetupOpcode()) {
861       SUnit *SeqEnd = CallSeqEndForStart[SU];
862       assert(SeqEnd && "Call sequence start/end must be known");
863       assert(!LiveRegDefs[CallResource]);
864       assert(!LiveRegGens[CallResource]);
865       ++NumLiveRegs;
866       LiveRegDefs[CallResource] = SU;
867       LiveRegGens[CallResource] = SeqEnd;
868     }
869   }
870 
871   // Release the special call resource dependence, if this is the end
872   // of a call.
873   if (LiveRegGens[CallResource] == SU)
874     for (const SDNode *SUNode = SU->getNode(); SUNode;
875          SUNode = SUNode->getGluedNode()) {
876       if (SUNode->isMachineOpcode() &&
877           SUNode->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) {
878         assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
879         assert(LiveRegDefs[CallResource]);
880         assert(LiveRegGens[CallResource]);
881         --NumLiveRegs;
882         LiveRegDefs[CallResource] = nullptr;
883         LiveRegGens[CallResource] = nullptr;
884         releaseInterferences(CallResource);
885       }
886     }
887 
888   for (auto &Succ : SU->Succs) {
889     if (Succ.isAssignedRegDep()) {
890       auto Reg = Succ.getReg();
891       if (!LiveRegDefs[Reg])
892         ++NumLiveRegs;
893       // This becomes the nearest def. Note that an earlier def may still be
894       // pending if this is a two-address node.
895       LiveRegDefs[Reg] = SU;
896 
897       // Update LiveRegGen only if was empty before this unscheduling.
898       // This is to avoid incorrect updating LiveRegGen set in previous run.
899       if (!LiveRegGens[Reg]) {
900         // Find the successor with the lowest height.
901         LiveRegGens[Reg] = Succ.getSUnit();
902         for (auto &Succ2 : SU->Succs) {
903           if (Succ2.isAssignedRegDep() && Succ2.getReg() == Reg &&
904               Succ2.getSUnit()->getHeight() < LiveRegGens[Reg]->getHeight())
905             LiveRegGens[Reg] = Succ2.getSUnit();
906         }
907       }
908     }
909   }
910   if (SU->getHeight() < MinAvailableCycle)
911     MinAvailableCycle = SU->getHeight();
912 
913   SU->setHeightDirty();
914   SU->isScheduled = false;
915   SU->isAvailable = true;
916   if (!DisableSchedCycles && AvailableQueue->hasReadyFilter()) {
917     // Don't make available until backtracking is complete.
918     SU->isPending = true;
919     PendingQueue.push_back(SU);
920   }
921   else {
922     AvailableQueue->push(SU);
923   }
924   AvailableQueue->unscheduledNode(SU);
925 }
926 
927 /// After backtracking, the hazard checker needs to be restored to a state
928 /// corresponding the current cycle.
929 void ScheduleDAGRRList::RestoreHazardCheckerBottomUp() {
930   HazardRec->Reset();
931 
932   unsigned LookAhead = std::min((unsigned)Sequence.size(),
933                                 HazardRec->getMaxLookAhead());
934   if (LookAhead == 0)
935     return;
936 
937   std::vector<SUnit *>::const_iterator I = (Sequence.end() - LookAhead);
938   unsigned HazardCycle = (*I)->getHeight();
939   for (auto E = Sequence.end(); I != E; ++I) {
940     SUnit *SU = *I;
941     for (; SU->getHeight() > HazardCycle; ++HazardCycle) {
942       HazardRec->RecedeCycle();
943     }
944     EmitNode(SU);
945   }
946 }
947 
948 /// BacktrackBottomUp - Backtrack scheduling to a previous cycle specified in
949 /// BTCycle in order to schedule a specific node.
950 void ScheduleDAGRRList::BacktrackBottomUp(SUnit *SU, SUnit *BtSU) {
951   SUnit *OldSU = Sequence.back();
952   while (true) {
953     Sequence.pop_back();
954     // FIXME: use ready cycle instead of height
955     CurCycle = OldSU->getHeight();
956     UnscheduleNodeBottomUp(OldSU);
957     AvailableQueue->setCurCycle(CurCycle);
958     if (OldSU == BtSU)
959       break;
960     OldSU = Sequence.back();
961   }
962 
963   assert(!SU->isSucc(OldSU) && "Something is wrong!");
964 
965   RestoreHazardCheckerBottomUp();
966 
967   ReleasePending();
968 
969   ++NumBacktracks;
970 }
971 
972 static bool isOperandOf(const SUnit *SU, SDNode *N) {
973   for (const SDNode *SUNode = SU->getNode(); SUNode;
974        SUNode = SUNode->getGluedNode()) {
975     if (SUNode->isOperandOf(N))
976       return true;
977   }
978   return false;
979 }
980 
981 /// TryUnfold - Attempt to unfold
982 SUnit *ScheduleDAGRRList::TryUnfoldSU(SUnit *SU) {
983   SDNode *N = SU->getNode();
984   // Use while over if to ease fall through.
985   SmallVector<SDNode *, 2> NewNodes;
986   if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
987     return nullptr;
988 
989   assert(NewNodes.size() == 2 && "Expected a load folding node!");
990 
991   N = NewNodes[1];
992   SDNode *LoadNode = NewNodes[0];
993   unsigned NumVals = N->getNumValues();
994   unsigned OldNumVals = SU->getNode()->getNumValues();
995 
996   // LoadNode may already exist. This can happen when there is another
997   // load from the same location and producing the same type of value
998   // but it has different alignment or volatileness.
999   bool isNewLoad = true;
1000   SUnit *LoadSU;
1001   if (LoadNode->getNodeId() != -1) {
1002     LoadSU = &SUnits[LoadNode->getNodeId()];
1003     // If LoadSU has already been scheduled, we should clone it but
1004     // this would negate the benefit to unfolding so just return SU.
1005     if (LoadSU->isScheduled)
1006       return SU;
1007     isNewLoad = false;
1008   } else {
1009     LoadSU = CreateNewSUnit(LoadNode);
1010     LoadNode->setNodeId(LoadSU->NodeNum);
1011 
1012     InitNumRegDefsLeft(LoadSU);
1013     computeLatency(LoadSU);
1014   }
1015 
1016   bool isNewN = true;
1017   SUnit *NewSU;
1018   // This can only happen when isNewLoad is false.
1019   if (N->getNodeId() != -1) {
1020     NewSU = &SUnits[N->getNodeId()];
1021     // If NewSU has already been scheduled, we need to clone it, but this
1022     // negates the benefit to unfolding so just return SU.
1023     if (NewSU->isScheduled) {
1024       return SU;
1025     }
1026     isNewN = false;
1027   } else {
1028     NewSU = CreateNewSUnit(N);
1029     N->setNodeId(NewSU->NodeNum);
1030 
1031     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1032     for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
1033       if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
1034         NewSU->isTwoAddress = true;
1035         break;
1036       }
1037     }
1038     if (MCID.isCommutable())
1039       NewSU->isCommutable = true;
1040 
1041     InitNumRegDefsLeft(NewSU);
1042     computeLatency(NewSU);
1043   }
1044 
1045   LLVM_DEBUG(dbgs() << "Unfolding SU #" << SU->NodeNum << "\n");
1046 
1047   // Now that we are committed to unfolding replace DAG Uses.
1048   for (unsigned i = 0; i != NumVals; ++i)
1049     DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i));
1050   DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals - 1),
1051                                  SDValue(LoadNode, 1));
1052 
1053   // Record all the edges to and from the old SU, by category.
1054   SmallVector<SDep, 4> ChainPreds;
1055   SmallVector<SDep, 4> ChainSuccs;
1056   SmallVector<SDep, 4> LoadPreds;
1057   SmallVector<SDep, 4> NodePreds;
1058   SmallVector<SDep, 4> NodeSuccs;
1059   for (SDep &Pred : SU->Preds) {
1060     if (Pred.isCtrl())
1061       ChainPreds.push_back(Pred);
1062     else if (isOperandOf(Pred.getSUnit(), LoadNode))
1063       LoadPreds.push_back(Pred);
1064     else
1065       NodePreds.push_back(Pred);
1066   }
1067   for (SDep &Succ : SU->Succs) {
1068     if (Succ.isCtrl())
1069       ChainSuccs.push_back(Succ);
1070     else
1071       NodeSuccs.push_back(Succ);
1072   }
1073 
1074   // Now assign edges to the newly-created nodes.
1075   for (const SDep &Pred : ChainPreds) {
1076     RemovePred(SU, Pred);
1077     if (isNewLoad)
1078       AddPredQueued(LoadSU, Pred);
1079   }
1080   for (const SDep &Pred : LoadPreds) {
1081     RemovePred(SU, Pred);
1082     if (isNewLoad)
1083       AddPredQueued(LoadSU, Pred);
1084   }
1085   for (const SDep &Pred : NodePreds) {
1086     RemovePred(SU, Pred);
1087     AddPredQueued(NewSU, Pred);
1088   }
1089   for (SDep &D : NodeSuccs) {
1090     SUnit *SuccDep = D.getSUnit();
1091     D.setSUnit(SU);
1092     RemovePred(SuccDep, D);
1093     D.setSUnit(NewSU);
1094     AddPredQueued(SuccDep, D);
1095     // Balance register pressure.
1096     if (AvailableQueue->tracksRegPressure() && SuccDep->isScheduled &&
1097         !D.isCtrl() && NewSU->NumRegDefsLeft > 0)
1098       --NewSU->NumRegDefsLeft;
1099   }
1100   for (SDep &D : ChainSuccs) {
1101     SUnit *SuccDep = D.getSUnit();
1102     D.setSUnit(SU);
1103     RemovePred(SuccDep, D);
1104     if (isNewLoad) {
1105       D.setSUnit(LoadSU);
1106       AddPredQueued(SuccDep, D);
1107     }
1108   }
1109 
1110   // Add a data dependency to reflect that NewSU reads the value defined
1111   // by LoadSU.
1112   SDep D(LoadSU, SDep::Data, 0);
1113   D.setLatency(LoadSU->Latency);
1114   AddPredQueued(NewSU, D);
1115 
1116   if (isNewLoad)
1117     AvailableQueue->addNode(LoadSU);
1118   if (isNewN)
1119     AvailableQueue->addNode(NewSU);
1120 
1121   ++NumUnfolds;
1122 
1123   if (NewSU->NumSuccsLeft == 0)
1124     NewSU->isAvailable = true;
1125 
1126   return NewSU;
1127 }
1128 
1129 /// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
1130 /// successors to the newly created node.
1131 SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) {
1132   SDNode *N = SU->getNode();
1133   if (!N)
1134     return nullptr;
1135 
1136   LLVM_DEBUG(dbgs() << "Considering duplicating the SU\n");
1137   LLVM_DEBUG(dumpNode(*SU));
1138 
1139   if (N->getGluedNode() &&
1140       !TII->canCopyGluedNodeDuringSchedule(N)) {
1141     LLVM_DEBUG(
1142         dbgs()
1143         << "Giving up because it has incoming glue and the target does not "
1144            "want to copy it\n");
1145     return nullptr;
1146   }
1147 
1148   SUnit *NewSU;
1149   bool TryUnfold = false;
1150   for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
1151     MVT VT = N->getSimpleValueType(i);
1152     if (VT == MVT::Glue) {
1153       LLVM_DEBUG(dbgs() << "Giving up because it has outgoing glue\n");
1154       return nullptr;
1155     } else if (VT == MVT::Other)
1156       TryUnfold = true;
1157   }
1158   for (const SDValue &Op : N->op_values()) {
1159     MVT VT = Op.getNode()->getSimpleValueType(Op.getResNo());
1160     if (VT == MVT::Glue && !TII->canCopyGluedNodeDuringSchedule(N)) {
1161       LLVM_DEBUG(
1162           dbgs() << "Giving up because it one of the operands is glue and "
1163                     "the target does not want to copy it\n");
1164       return nullptr;
1165     }
1166   }
1167 
1168   // If possible unfold instruction.
1169   if (TryUnfold) {
1170     SUnit *UnfoldSU = TryUnfoldSU(SU);
1171     if (!UnfoldSU)
1172       return nullptr;
1173     SU = UnfoldSU;
1174     N = SU->getNode();
1175     // If this can be scheduled don't bother duplicating and just return
1176     if (SU->NumSuccsLeft == 0)
1177       return SU;
1178   }
1179 
1180   LLVM_DEBUG(dbgs() << "    Duplicating SU #" << SU->NodeNum << "\n");
1181   NewSU = CreateClone(SU);
1182 
1183   // New SUnit has the exact same predecessors.
1184   for (SDep &Pred : SU->Preds)
1185     if (!Pred.isArtificial())
1186       AddPredQueued(NewSU, Pred);
1187 
1188   // Make sure the clone comes after the original. (InstrEmitter assumes
1189   // this ordering.)
1190   AddPredQueued(NewSU, SDep(SU, SDep::Artificial));
1191 
1192   // Only copy scheduled successors. Cut them from old node's successor
1193   // list and move them over.
1194   SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
1195   for (SDep &Succ : SU->Succs) {
1196     if (Succ.isArtificial())
1197       continue;
1198     SUnit *SuccSU = Succ.getSUnit();
1199     if (SuccSU->isScheduled) {
1200       SDep D = Succ;
1201       D.setSUnit(NewSU);
1202       AddPredQueued(SuccSU, D);
1203       D.setSUnit(SU);
1204       DelDeps.emplace_back(SuccSU, D);
1205     }
1206   }
1207   for (const auto &[DelSU, DelD] : DelDeps)
1208     RemovePred(DelSU, DelD);
1209 
1210   AvailableQueue->updateNode(SU);
1211   AvailableQueue->addNode(NewSU);
1212 
1213   ++NumDups;
1214   return NewSU;
1215 }
1216 
1217 /// InsertCopiesAndMoveSuccs - Insert register copies and move all
1218 /// scheduled successors of the given SUnit to the last copy.
1219 void ScheduleDAGRRList::InsertCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
1220                                               const TargetRegisterClass *DestRC,
1221                                               const TargetRegisterClass *SrcRC,
1222                                               SmallVectorImpl<SUnit*> &Copies) {
1223   SUnit *CopyFromSU = CreateNewSUnit(nullptr);
1224   CopyFromSU->CopySrcRC = SrcRC;
1225   CopyFromSU->CopyDstRC = DestRC;
1226 
1227   SUnit *CopyToSU = CreateNewSUnit(nullptr);
1228   CopyToSU->CopySrcRC = DestRC;
1229   CopyToSU->CopyDstRC = SrcRC;
1230 
1231   // Only copy scheduled successors. Cut them from old node's successor
1232   // list and move them over.
1233   SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
1234   for (SDep &Succ : SU->Succs) {
1235     if (Succ.isArtificial())
1236       continue;
1237     SUnit *SuccSU = Succ.getSUnit();
1238     if (SuccSU->isScheduled) {
1239       SDep D = Succ;
1240       D.setSUnit(CopyToSU);
1241       AddPredQueued(SuccSU, D);
1242       DelDeps.emplace_back(SuccSU, Succ);
1243     }
1244     else {
1245       // Avoid scheduling the def-side copy before other successors. Otherwise,
1246       // we could introduce another physreg interference on the copy and
1247       // continue inserting copies indefinitely.
1248       AddPredQueued(SuccSU, SDep(CopyFromSU, SDep::Artificial));
1249     }
1250   }
1251   for (const auto &[DelSU, DelD] : DelDeps)
1252     RemovePred(DelSU, DelD);
1253 
1254   SDep FromDep(SU, SDep::Data, Reg);
1255   FromDep.setLatency(SU->Latency);
1256   AddPredQueued(CopyFromSU, FromDep);
1257   SDep ToDep(CopyFromSU, SDep::Data, 0);
1258   ToDep.setLatency(CopyFromSU->Latency);
1259   AddPredQueued(CopyToSU, ToDep);
1260 
1261   AvailableQueue->updateNode(SU);
1262   AvailableQueue->addNode(CopyFromSU);
1263   AvailableQueue->addNode(CopyToSU);
1264   Copies.push_back(CopyFromSU);
1265   Copies.push_back(CopyToSU);
1266 
1267   ++NumPRCopies;
1268 }
1269 
1270 /// getPhysicalRegisterVT - Returns the ValueType of the physical register
1271 /// definition of the specified node.
1272 /// FIXME: Move to SelectionDAG?
1273 static MVT getPhysicalRegisterVT(SDNode *N, unsigned Reg,
1274                                  const TargetInstrInfo *TII) {
1275   unsigned NumRes;
1276   if (N->getOpcode() == ISD::CopyFromReg) {
1277     // CopyFromReg has: "chain, Val, glue" so operand 1 gives the type.
1278     NumRes = 1;
1279   } else {
1280     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1281     assert(!MCID.implicit_defs().empty() &&
1282            "Physical reg def must be in implicit def list!");
1283     NumRes = MCID.getNumDefs();
1284     for (MCPhysReg ImpDef : MCID.implicit_defs()) {
1285       if (Reg == ImpDef)
1286         break;
1287       ++NumRes;
1288     }
1289   }
1290   return N->getSimpleValueType(NumRes);
1291 }
1292 
1293 /// CheckForLiveRegDef - Return true and update live register vector if the
1294 /// specified register def of the specified SUnit clobbers any "live" registers.
1295 static void CheckForLiveRegDef(SUnit *SU, unsigned Reg, SUnit **LiveRegDefs,
1296                                SmallSet<unsigned, 4> &RegAdded,
1297                                SmallVectorImpl<unsigned> &LRegs,
1298                                const TargetRegisterInfo *TRI,
1299                                const SDNode *Node = nullptr) {
1300   for (MCRegAliasIterator AliasI(Reg, TRI, true); AliasI.isValid(); ++AliasI) {
1301 
1302     // Check if Ref is live.
1303     if (!LiveRegDefs[*AliasI]) continue;
1304 
1305     // Allow multiple uses of the same def.
1306     if (LiveRegDefs[*AliasI] == SU) continue;
1307 
1308     // Allow multiple uses of same def
1309     if (Node && LiveRegDefs[*AliasI]->getNode() == Node)
1310       continue;
1311 
1312     // Add Reg to the set of interfering live regs.
1313     if (RegAdded.insert(*AliasI).second) {
1314       LRegs.push_back(*AliasI);
1315     }
1316   }
1317 }
1318 
1319 /// CheckForLiveRegDefMasked - Check for any live physregs that are clobbered
1320 /// by RegMask, and add them to LRegs.
1321 static void CheckForLiveRegDefMasked(SUnit *SU, const uint32_t *RegMask,
1322                                      ArrayRef<SUnit*> LiveRegDefs,
1323                                      SmallSet<unsigned, 4> &RegAdded,
1324                                      SmallVectorImpl<unsigned> &LRegs) {
1325   // Look at all live registers. Skip Reg0 and the special CallResource.
1326   for (unsigned i = 1, e = LiveRegDefs.size()-1; i != e; ++i) {
1327     if (!LiveRegDefs[i]) continue;
1328     if (LiveRegDefs[i] == SU) continue;
1329     if (!MachineOperand::clobbersPhysReg(RegMask, i)) continue;
1330     if (RegAdded.insert(i).second)
1331       LRegs.push_back(i);
1332   }
1333 }
1334 
1335 /// getNodeRegMask - Returns the register mask attached to an SDNode, if any.
1336 static const uint32_t *getNodeRegMask(const SDNode *N) {
1337   for (const SDValue &Op : N->op_values())
1338     if (const auto *RegOp = dyn_cast<RegisterMaskSDNode>(Op.getNode()))
1339       return RegOp->getRegMask();
1340   return nullptr;
1341 }
1342 
1343 /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
1344 /// scheduling of the given node to satisfy live physical register dependencies.
1345 /// If the specific node is the last one that's available to schedule, do
1346 /// whatever is necessary (i.e. backtracking or cloning) to make it possible.
1347 bool ScheduleDAGRRList::
1348 DelayForLiveRegsBottomUp(SUnit *SU, SmallVectorImpl<unsigned> &LRegs) {
1349   if (NumLiveRegs == 0)
1350     return false;
1351 
1352   SmallSet<unsigned, 4> RegAdded;
1353   // If this node would clobber any "live" register, then it's not ready.
1354   //
1355   // If SU is the currently live definition of the same register that it uses,
1356   // then we are free to schedule it.
1357   for (SDep &Pred : SU->Preds) {
1358     if (Pred.isAssignedRegDep() && LiveRegDefs[Pred.getReg()] != SU)
1359       CheckForLiveRegDef(Pred.getSUnit(), Pred.getReg(), LiveRegDefs.get(),
1360                          RegAdded, LRegs, TRI);
1361   }
1362 
1363   for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) {
1364     if (Node->getOpcode() == ISD::INLINEASM ||
1365         Node->getOpcode() == ISD::INLINEASM_BR) {
1366       // Inline asm can clobber physical defs.
1367       unsigned NumOps = Node->getNumOperands();
1368       if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
1369         --NumOps;  // Ignore the glue operand.
1370 
1371       for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
1372         unsigned Flags = Node->getConstantOperandVal(i);
1373         const InlineAsm::Flag F(Flags);
1374         unsigned NumVals = F.getNumOperandRegisters();
1375 
1376         ++i; // Skip the ID value.
1377         if (F.isRegDefKind() || F.isRegDefEarlyClobberKind() ||
1378             F.isClobberKind()) {
1379           // Check for def of register or earlyclobber register.
1380           for (; NumVals; --NumVals, ++i) {
1381             Register Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
1382             if (Reg.isPhysical())
1383               CheckForLiveRegDef(SU, Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI);
1384           }
1385         } else
1386           i += NumVals;
1387       }
1388       continue;
1389     }
1390 
1391     if (Node->getOpcode() == ISD::CopyToReg) {
1392       Register Reg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
1393       if (Reg.isPhysical()) {
1394         SDNode *SrcNode = Node->getOperand(2).getNode();
1395         CheckForLiveRegDef(SU, Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI,
1396                            SrcNode);
1397       }
1398     }
1399 
1400     if (!Node->isMachineOpcode())
1401       continue;
1402     // If we're in the middle of scheduling a call, don't begin scheduling
1403     // another call. Also, don't allow any physical registers to be live across
1404     // the call.
1405     if (Node->getMachineOpcode() == TII->getCallFrameDestroyOpcode()) {
1406       // Check the special calling-sequence resource.
1407       unsigned CallResource = TRI->getNumRegs();
1408       if (LiveRegDefs[CallResource]) {
1409         SDNode *Gen = LiveRegGens[CallResource]->getNode();
1410         while (SDNode *Glued = Gen->getGluedNode())
1411           Gen = Glued;
1412         if (!IsChainDependent(Gen, Node, 0, TII) &&
1413             RegAdded.insert(CallResource).second)
1414           LRegs.push_back(CallResource);
1415       }
1416     }
1417     if (const uint32_t *RegMask = getNodeRegMask(Node))
1418       CheckForLiveRegDefMasked(SU, RegMask,
1419                                ArrayRef(LiveRegDefs.get(), TRI->getNumRegs()),
1420                                RegAdded, LRegs);
1421 
1422     const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode());
1423     if (MCID.hasOptionalDef()) {
1424       // Most ARM instructions have an OptionalDef for CPSR, to model the S-bit.
1425       // This operand can be either a def of CPSR, if the S bit is set; or a use
1426       // of %noreg.  When the OptionalDef is set to a valid register, we need to
1427       // handle it in the same way as an ImplicitDef.
1428       for (unsigned i = 0; i < MCID.getNumDefs(); ++i)
1429         if (MCID.operands()[i].isOptionalDef()) {
1430           const SDValue &OptionalDef = Node->getOperand(i - Node->getNumValues());
1431           Register Reg = cast<RegisterSDNode>(OptionalDef)->getReg();
1432           CheckForLiveRegDef(SU, Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI);
1433         }
1434     }
1435     for (MCPhysReg Reg : MCID.implicit_defs())
1436       CheckForLiveRegDef(SU, Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI);
1437   }
1438 
1439   return !LRegs.empty();
1440 }
1441 
1442 void ScheduleDAGRRList::releaseInterferences(unsigned Reg) {
1443   // Add the nodes that aren't ready back onto the available list.
1444   for (unsigned i = Interferences.size(); i > 0; --i) {
1445     SUnit *SU = Interferences[i-1];
1446     LRegsMapT::iterator LRegsPos = LRegsMap.find(SU);
1447     if (Reg) {
1448       SmallVectorImpl<unsigned> &LRegs = LRegsPos->second;
1449       if (!is_contained(LRegs, Reg))
1450         continue;
1451     }
1452     SU->isPending = false;
1453     // The interfering node may no longer be available due to backtracking.
1454     // Furthermore, it may have been made available again, in which case it is
1455     // now already in the AvailableQueue.
1456     if (SU->isAvailable && !SU->NodeQueueId) {
1457       LLVM_DEBUG(dbgs() << "    Repushing SU #" << SU->NodeNum << '\n');
1458       AvailableQueue->push(SU);
1459     }
1460     if (i < Interferences.size())
1461       Interferences[i-1] = Interferences.back();
1462     Interferences.pop_back();
1463     LRegsMap.erase(LRegsPos);
1464   }
1465 }
1466 
1467 /// Return a node that can be scheduled in this cycle. Requirements:
1468 /// (1) Ready: latency has been satisfied
1469 /// (2) No Hazards: resources are available
1470 /// (3) No Interferences: may unschedule to break register interferences.
1471 SUnit *ScheduleDAGRRList::PickNodeToScheduleBottomUp() {
1472   SUnit *CurSU = AvailableQueue->empty() ? nullptr : AvailableQueue->pop();
1473   auto FindAvailableNode = [&]() {
1474     while (CurSU) {
1475       SmallVector<unsigned, 4> LRegs;
1476       if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
1477         break;
1478       LLVM_DEBUG(dbgs() << "    Interfering reg ";
1479                  if (LRegs[0] == TRI->getNumRegs()) dbgs() << "CallResource";
1480                  else dbgs() << printReg(LRegs[0], TRI);
1481                  dbgs() << " SU #" << CurSU->NodeNum << '\n');
1482       auto [LRegsIter, LRegsInserted] = LRegsMap.try_emplace(CurSU, LRegs);
1483       if (LRegsInserted) {
1484         CurSU->isPending = true;  // This SU is not in AvailableQueue right now.
1485         Interferences.push_back(CurSU);
1486       }
1487       else {
1488         assert(CurSU->isPending && "Interferences are pending");
1489         // Update the interference with current live regs.
1490         LRegsIter->second = LRegs;
1491       }
1492       CurSU = AvailableQueue->pop();
1493     }
1494   };
1495   FindAvailableNode();
1496   if (CurSU)
1497     return CurSU;
1498 
1499   // We query the topological order in the loop body, so make sure outstanding
1500   // updates are applied before entering it (we only enter the loop if there
1501   // are some interferences). If we make changes to the ordering, we exit
1502   // the loop.
1503 
1504   // All candidates are delayed due to live physical reg dependencies.
1505   // Try backtracking, code duplication, or inserting cross class copies
1506   // to resolve it.
1507   for (SUnit *TrySU : Interferences) {
1508     SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
1509 
1510     // Try unscheduling up to the point where it's safe to schedule
1511     // this node.
1512     SUnit *BtSU = nullptr;
1513     unsigned LiveCycle = std::numeric_limits<unsigned>::max();
1514     for (unsigned Reg : LRegs) {
1515       if (LiveRegGens[Reg]->getHeight() < LiveCycle) {
1516         BtSU = LiveRegGens[Reg];
1517         LiveCycle = BtSU->getHeight();
1518       }
1519     }
1520     if (!WillCreateCycle(TrySU, BtSU))  {
1521       // BacktrackBottomUp mutates Interferences!
1522       BacktrackBottomUp(TrySU, BtSU);
1523 
1524       // Force the current node to be scheduled before the node that
1525       // requires the physical reg dep.
1526       if (BtSU->isAvailable) {
1527         BtSU->isAvailable = false;
1528         if (!BtSU->isPending)
1529           AvailableQueue->remove(BtSU);
1530       }
1531       LLVM_DEBUG(dbgs() << "ARTIFICIAL edge from SU(" << BtSU->NodeNum
1532                         << ") to SU(" << TrySU->NodeNum << ")\n");
1533       AddPredQueued(TrySU, SDep(BtSU, SDep::Artificial));
1534 
1535       // If one or more successors has been unscheduled, then the current
1536       // node is no longer available.
1537       if (!TrySU->isAvailable || !TrySU->NodeQueueId) {
1538         LLVM_DEBUG(dbgs() << "TrySU not available; choosing node from queue\n");
1539         CurSU = AvailableQueue->pop();
1540       } else {
1541         LLVM_DEBUG(dbgs() << "TrySU available\n");
1542         // Available and in AvailableQueue
1543         AvailableQueue->remove(TrySU);
1544         CurSU = TrySU;
1545       }
1546       FindAvailableNode();
1547       // Interferences has been mutated. We must break.
1548       break;
1549     }
1550   }
1551 
1552   if (!CurSU) {
1553     // Can't backtrack. If it's too expensive to copy the value, then try
1554     // duplicate the nodes that produces these "too expensive to copy"
1555     // values to break the dependency. In case even that doesn't work,
1556     // insert cross class copies.
1557     // If it's not too expensive, i.e. cost != -1, issue copies.
1558     SUnit *TrySU = Interferences[0];
1559     SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
1560     assert(LRegs.size() == 1 && "Can't handle this yet!");
1561     unsigned Reg = LRegs[0];
1562     SUnit *LRDef = LiveRegDefs[Reg];
1563     MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
1564     const TargetRegisterClass *RC =
1565       TRI->getMinimalPhysRegClass(Reg, VT);
1566     const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
1567 
1568     // If cross copy register class is the same as RC, then it must be possible
1569     // copy the value directly. Do not try duplicate the def.
1570     // If cross copy register class is not the same as RC, then it's possible to
1571     // copy the value but it require cross register class copies and it is
1572     // expensive.
1573     // If cross copy register class is null, then it's not possible to copy
1574     // the value at all.
1575     SUnit *NewDef = nullptr;
1576     if (DestRC != RC) {
1577       NewDef = CopyAndMoveSuccessors(LRDef);
1578       if (!DestRC && !NewDef)
1579         report_fatal_error("Can't handle live physical register dependency!");
1580     }
1581     if (!NewDef) {
1582       // Issue copies, these can be expensive cross register class copies.
1583       SmallVector<SUnit*, 2> Copies;
1584       InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
1585       LLVM_DEBUG(dbgs() << "    Adding an edge from SU #" << TrySU->NodeNum
1586                         << " to SU #" << Copies.front()->NodeNum << "\n");
1587       AddPredQueued(TrySU, SDep(Copies.front(), SDep::Artificial));
1588       NewDef = Copies.back();
1589     }
1590 
1591     LLVM_DEBUG(dbgs() << "    Adding an edge from SU #" << NewDef->NodeNum
1592                       << " to SU #" << TrySU->NodeNum << "\n");
1593     LiveRegDefs[Reg] = NewDef;
1594     AddPredQueued(NewDef, SDep(TrySU, SDep::Artificial));
1595     TrySU->isAvailable = false;
1596     CurSU = NewDef;
1597   }
1598   assert(CurSU && "Unable to resolve live physical register dependencies!");
1599   return CurSU;
1600 }
1601 
1602 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
1603 /// schedulers.
1604 void ScheduleDAGRRList::ListScheduleBottomUp() {
1605   // Release any predecessors of the special Exit node.
1606   ReleasePredecessors(&ExitSU);
1607 
1608   // Add root to Available queue.
1609   if (!SUnits.empty()) {
1610     SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()];
1611     assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!");
1612     RootSU->isAvailable = true;
1613     AvailableQueue->push(RootSU);
1614   }
1615 
1616   // While Available queue is not empty, grab the node with the highest
1617   // priority. If it is not ready put it back.  Schedule the node.
1618   Sequence.reserve(SUnits.size());
1619   while (!AvailableQueue->empty() || !Interferences.empty()) {
1620     LLVM_DEBUG(dbgs() << "\nExamining Available:\n";
1621                AvailableQueue->dump(this));
1622 
1623     // Pick the best node to schedule taking all constraints into
1624     // consideration.
1625     SUnit *SU = PickNodeToScheduleBottomUp();
1626 
1627     AdvancePastStalls(SU);
1628 
1629     ScheduleNodeBottomUp(SU);
1630 
1631     while (AvailableQueue->empty() && !PendingQueue.empty()) {
1632       // Advance the cycle to free resources. Skip ahead to the next ready SU.
1633       assert(MinAvailableCycle < std::numeric_limits<unsigned>::max() &&
1634              "MinAvailableCycle uninitialized");
1635       AdvanceToCycle(std::max(CurCycle + 1, MinAvailableCycle));
1636     }
1637   }
1638 
1639   // Reverse the order if it is bottom up.
1640   std::reverse(Sequence.begin(), Sequence.end());
1641 
1642 #ifndef NDEBUG
1643   VerifyScheduledSequence(/*isBottomUp=*/true);
1644 #endif
1645 }
1646 
1647 namespace {
1648 
1649 class RegReductionPQBase;
1650 
1651 struct queue_sort {
1652   bool isReady(SUnit* SU, unsigned CurCycle) const { return true; }
1653 };
1654 
1655 #ifndef NDEBUG
1656 template<class SF>
1657 struct reverse_sort : public queue_sort {
1658   SF &SortFunc;
1659 
1660   reverse_sort(SF &sf) : SortFunc(sf) {}
1661 
1662   bool operator()(SUnit* left, SUnit* right) const {
1663     // reverse left/right rather than simply !SortFunc(left, right)
1664     // to expose different paths in the comparison logic.
1665     return SortFunc(right, left);
1666   }
1667 };
1668 #endif // NDEBUG
1669 
1670 /// bu_ls_rr_sort - Priority function for bottom up register pressure
1671 // reduction scheduler.
1672 struct bu_ls_rr_sort : public queue_sort {
1673   enum {
1674     IsBottomUp = true,
1675     HasReadyFilter = false
1676   };
1677 
1678   RegReductionPQBase *SPQ;
1679 
1680   bu_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
1681 
1682   bool operator()(SUnit* left, SUnit* right) const;
1683 };
1684 
1685 // src_ls_rr_sort - Priority function for source order scheduler.
1686 struct src_ls_rr_sort : public queue_sort {
1687   enum {
1688     IsBottomUp = true,
1689     HasReadyFilter = false
1690   };
1691 
1692   RegReductionPQBase *SPQ;
1693 
1694   src_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
1695 
1696   bool operator()(SUnit* left, SUnit* right) const;
1697 };
1698 
1699 // hybrid_ls_rr_sort - Priority function for hybrid scheduler.
1700 struct hybrid_ls_rr_sort : public queue_sort {
1701   enum {
1702     IsBottomUp = true,
1703     HasReadyFilter = false
1704   };
1705 
1706   RegReductionPQBase *SPQ;
1707 
1708   hybrid_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
1709 
1710   bool isReady(SUnit *SU, unsigned CurCycle) const;
1711 
1712   bool operator()(SUnit* left, SUnit* right) const;
1713 };
1714 
1715 // ilp_ls_rr_sort - Priority function for ILP (instruction level parallelism)
1716 // scheduler.
1717 struct ilp_ls_rr_sort : public queue_sort {
1718   enum {
1719     IsBottomUp = true,
1720     HasReadyFilter = false
1721   };
1722 
1723   RegReductionPQBase *SPQ;
1724 
1725   ilp_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
1726 
1727   bool isReady(SUnit *SU, unsigned CurCycle) const;
1728 
1729   bool operator()(SUnit* left, SUnit* right) const;
1730 };
1731 
1732 class RegReductionPQBase : public SchedulingPriorityQueue {
1733 protected:
1734   std::vector<SUnit *> Queue;
1735   unsigned CurQueueId = 0;
1736   bool TracksRegPressure;
1737   bool SrcOrder;
1738 
1739   // SUnits - The SUnits for the current graph.
1740   std::vector<SUnit> *SUnits = nullptr;
1741 
1742   MachineFunction &MF;
1743   const TargetInstrInfo *TII = nullptr;
1744   const TargetRegisterInfo *TRI = nullptr;
1745   const TargetLowering *TLI = nullptr;
1746   ScheduleDAGRRList *scheduleDAG = nullptr;
1747 
1748   // SethiUllmanNumbers - The SethiUllman number for each node.
1749   std::vector<unsigned> SethiUllmanNumbers;
1750 
1751   /// RegPressure - Tracking current reg pressure per register class.
1752   std::vector<unsigned> RegPressure;
1753 
1754   /// RegLimit - Tracking the number of allocatable registers per register
1755   /// class.
1756   std::vector<unsigned> RegLimit;
1757 
1758 public:
1759   RegReductionPQBase(MachineFunction &mf,
1760                      bool hasReadyFilter,
1761                      bool tracksrp,
1762                      bool srcorder,
1763                      const TargetInstrInfo *tii,
1764                      const TargetRegisterInfo *tri,
1765                      const TargetLowering *tli)
1766     : SchedulingPriorityQueue(hasReadyFilter), TracksRegPressure(tracksrp),
1767       SrcOrder(srcorder), MF(mf), TII(tii), TRI(tri), TLI(tli) {
1768     if (TracksRegPressure) {
1769       unsigned NumRC = TRI->getNumRegClasses();
1770       RegLimit.resize(NumRC);
1771       RegPressure.resize(NumRC);
1772       std::fill(RegLimit.begin(), RegLimit.end(), 0);
1773       std::fill(RegPressure.begin(), RegPressure.end(), 0);
1774       for (const TargetRegisterClass *RC : TRI->regclasses())
1775         RegLimit[RC->getID()] = tri->getRegPressureLimit(RC, MF);
1776     }
1777   }
1778 
1779   void setScheduleDAG(ScheduleDAGRRList *scheduleDag) {
1780     scheduleDAG = scheduleDag;
1781   }
1782 
1783   ScheduleHazardRecognizer* getHazardRec() {
1784     return scheduleDAG->getHazardRec();
1785   }
1786 
1787   void initNodes(std::vector<SUnit> &sunits) override;
1788 
1789   void addNode(const SUnit *SU) override;
1790 
1791   void updateNode(const SUnit *SU) override;
1792 
1793   void releaseState() override {
1794     SUnits = nullptr;
1795     SethiUllmanNumbers.clear();
1796     std::fill(RegPressure.begin(), RegPressure.end(), 0);
1797   }
1798 
1799   unsigned getNodePriority(const SUnit *SU) const;
1800 
1801   unsigned getNodeOrdering(const SUnit *SU) const {
1802     if (!SU->getNode()) return 0;
1803 
1804     return SU->getNode()->getIROrder();
1805   }
1806 
1807   bool empty() const override { return Queue.empty(); }
1808 
1809   void push(SUnit *U) override {
1810     assert(!U->NodeQueueId && "Node in the queue already");
1811     U->NodeQueueId = ++CurQueueId;
1812     Queue.push_back(U);
1813   }
1814 
1815   void remove(SUnit *SU) override {
1816     assert(!Queue.empty() && "Queue is empty!");
1817     assert(SU->NodeQueueId != 0 && "Not in queue!");
1818     std::vector<SUnit *>::iterator I = llvm::find(Queue, SU);
1819     if (I != std::prev(Queue.end()))
1820       std::swap(*I, Queue.back());
1821     Queue.pop_back();
1822     SU->NodeQueueId = 0;
1823   }
1824 
1825   bool tracksRegPressure() const override { return TracksRegPressure; }
1826 
1827   void dumpRegPressure() const;
1828 
1829   bool HighRegPressure(const SUnit *SU) const;
1830 
1831   bool MayReduceRegPressure(SUnit *SU) const;
1832 
1833   int RegPressureDiff(SUnit *SU, unsigned &LiveUses) const;
1834 
1835   void scheduledNode(SUnit *SU) override;
1836 
1837   void unscheduledNode(SUnit *SU) override;
1838 
1839 protected:
1840   bool canClobber(const SUnit *SU, const SUnit *Op);
1841   void AddPseudoTwoAddrDeps();
1842   void PrescheduleNodesWithMultipleUses();
1843   void CalculateSethiUllmanNumbers();
1844 };
1845 
1846 template<class SF>
1847 static SUnit *popFromQueueImpl(std::vector<SUnit *> &Q, SF &Picker) {
1848   unsigned BestIdx = 0;
1849   // Only compute the cost for the first 1000 items in the queue, to avoid
1850   // excessive compile-times for very large queues.
1851   for (unsigned I = 1, E = std::min(Q.size(), (decltype(Q.size()))1000); I != E;
1852        I++)
1853     if (Picker(Q[BestIdx], Q[I]))
1854       BestIdx = I;
1855   SUnit *V = Q[BestIdx];
1856   if (BestIdx + 1 != Q.size())
1857     std::swap(Q[BestIdx], Q.back());
1858   Q.pop_back();
1859   return V;
1860 }
1861 
1862 template<class SF>
1863 SUnit *popFromQueue(std::vector<SUnit *> &Q, SF &Picker, ScheduleDAG *DAG) {
1864 #ifndef NDEBUG
1865   if (DAG->StressSched) {
1866     reverse_sort<SF> RPicker(Picker);
1867     return popFromQueueImpl(Q, RPicker);
1868   }
1869 #endif
1870   (void)DAG;
1871   return popFromQueueImpl(Q, Picker);
1872 }
1873 
1874 //===----------------------------------------------------------------------===//
1875 //                RegReductionPriorityQueue Definition
1876 //===----------------------------------------------------------------------===//
1877 //
1878 // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers
1879 // to reduce register pressure.
1880 //
1881 template<class SF>
1882 class RegReductionPriorityQueue : public RegReductionPQBase {
1883   SF Picker;
1884 
1885 public:
1886   RegReductionPriorityQueue(MachineFunction &mf,
1887                             bool tracksrp,
1888                             bool srcorder,
1889                             const TargetInstrInfo *tii,
1890                             const TargetRegisterInfo *tri,
1891                             const TargetLowering *tli)
1892     : RegReductionPQBase(mf, SF::HasReadyFilter, tracksrp, srcorder,
1893                          tii, tri, tli),
1894       Picker(this) {}
1895 
1896   bool isBottomUp() const override { return SF::IsBottomUp; }
1897 
1898   bool isReady(SUnit *U) const override {
1899     return Picker.HasReadyFilter && Picker.isReady(U, getCurCycle());
1900   }
1901 
1902   SUnit *pop() override {
1903     if (Queue.empty()) return nullptr;
1904 
1905     SUnit *V = popFromQueue(Queue, Picker, scheduleDAG);
1906     V->NodeQueueId = 0;
1907     return V;
1908   }
1909 
1910 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1911   LLVM_DUMP_METHOD void dump(ScheduleDAG *DAG) const override {
1912     // Emulate pop() without clobbering NodeQueueIds.
1913     std::vector<SUnit *> DumpQueue = Queue;
1914     SF DumpPicker = Picker;
1915     while (!DumpQueue.empty()) {
1916       SUnit *SU = popFromQueue(DumpQueue, DumpPicker, scheduleDAG);
1917       dbgs() << "Height " << SU->getHeight() << ": ";
1918       DAG->dumpNode(*SU);
1919     }
1920   }
1921 #endif
1922 };
1923 
1924 using BURegReductionPriorityQueue = RegReductionPriorityQueue<bu_ls_rr_sort>;
1925 using SrcRegReductionPriorityQueue = RegReductionPriorityQueue<src_ls_rr_sort>;
1926 using HybridBURRPriorityQueue = RegReductionPriorityQueue<hybrid_ls_rr_sort>;
1927 using ILPBURRPriorityQueue = RegReductionPriorityQueue<ilp_ls_rr_sort>;
1928 
1929 } // end anonymous namespace
1930 
1931 //===----------------------------------------------------------------------===//
1932 //           Static Node Priority for Register Pressure Reduction
1933 //===----------------------------------------------------------------------===//
1934 
1935 // Check for special nodes that bypass scheduling heuristics.
1936 // Currently this pushes TokenFactor nodes down, but may be used for other
1937 // pseudo-ops as well.
1938 //
1939 // Return -1 to schedule right above left, 1 for left above right.
1940 // Return 0 if no bias exists.
1941 static int checkSpecialNodes(const SUnit *left, const SUnit *right) {
1942   bool LSchedLow = left->isScheduleLow;
1943   bool RSchedLow = right->isScheduleLow;
1944   if (LSchedLow != RSchedLow)
1945     return LSchedLow < RSchedLow ? 1 : -1;
1946   return 0;
1947 }
1948 
1949 /// CalcNodeSethiUllmanNumber - Compute Sethi Ullman number.
1950 /// Smaller number is the higher priority.
1951 static unsigned
1952 CalcNodeSethiUllmanNumber(const SUnit *SU, std::vector<unsigned> &SUNumbers) {
1953   if (SUNumbers[SU->NodeNum] != 0)
1954     return SUNumbers[SU->NodeNum];
1955 
1956   // Use WorkList to avoid stack overflow on excessively large IRs.
1957   struct WorkState {
1958     WorkState(const SUnit *SU) : SU(SU) {}
1959     const SUnit *SU;
1960     unsigned PredsProcessed = 0;
1961   };
1962 
1963   SmallVector<WorkState, 16> WorkList;
1964   WorkList.push_back(SU);
1965   while (!WorkList.empty()) {
1966     auto &Temp = WorkList.back();
1967     auto *TempSU = Temp.SU;
1968     bool AllPredsKnown = true;
1969     // Try to find a non-evaluated pred and push it into the processing stack.
1970     for (unsigned P = Temp.PredsProcessed; P < TempSU->Preds.size(); ++P) {
1971       auto &Pred = TempSU->Preds[P];
1972       if (Pred.isCtrl()) continue;  // ignore chain preds
1973       SUnit *PredSU = Pred.getSUnit();
1974       if (SUNumbers[PredSU->NodeNum] == 0) {
1975 #ifndef NDEBUG
1976         // In debug mode, check that we don't have such element in the stack.
1977         for (auto It : WorkList)
1978           assert(It.SU != PredSU && "Trying to push an element twice?");
1979 #endif
1980         // Next time start processing this one starting from the next pred.
1981         Temp.PredsProcessed = P + 1;
1982         WorkList.push_back(PredSU);
1983         AllPredsKnown = false;
1984         break;
1985       }
1986     }
1987 
1988     if (!AllPredsKnown)
1989       continue;
1990 
1991     // Once all preds are known, we can calculate the answer for this one.
1992     unsigned SethiUllmanNumber = 0;
1993     unsigned Extra = 0;
1994     for (const SDep &Pred : TempSU->Preds) {
1995       if (Pred.isCtrl()) continue;  // ignore chain preds
1996       SUnit *PredSU = Pred.getSUnit();
1997       unsigned PredSethiUllman = SUNumbers[PredSU->NodeNum];
1998       assert(PredSethiUllman > 0 && "We should have evaluated this pred!");
1999       if (PredSethiUllman > SethiUllmanNumber) {
2000         SethiUllmanNumber = PredSethiUllman;
2001         Extra = 0;
2002       } else if (PredSethiUllman == SethiUllmanNumber)
2003         ++Extra;
2004     }
2005 
2006     SethiUllmanNumber += Extra;
2007     if (SethiUllmanNumber == 0)
2008       SethiUllmanNumber = 1;
2009     SUNumbers[TempSU->NodeNum] = SethiUllmanNumber;
2010     WorkList.pop_back();
2011   }
2012 
2013   assert(SUNumbers[SU->NodeNum] > 0 && "SethiUllman should never be zero!");
2014   return SUNumbers[SU->NodeNum];
2015 }
2016 
2017 /// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all
2018 /// scheduling units.
2019 void RegReductionPQBase::CalculateSethiUllmanNumbers() {
2020   SethiUllmanNumbers.assign(SUnits->size(), 0);
2021 
2022   for (const SUnit &SU : *SUnits)
2023     CalcNodeSethiUllmanNumber(&SU, SethiUllmanNumbers);
2024 }
2025 
2026 void RegReductionPQBase::addNode(const SUnit *SU) {
2027   unsigned SUSize = SethiUllmanNumbers.size();
2028   if (SUnits->size() > SUSize)
2029     SethiUllmanNumbers.resize(SUSize*2, 0);
2030   CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
2031 }
2032 
2033 void RegReductionPQBase::updateNode(const SUnit *SU) {
2034   SethiUllmanNumbers[SU->NodeNum] = 0;
2035   CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
2036 }
2037 
2038 // Lower priority means schedule further down. For bottom-up scheduling, lower
2039 // priority SUs are scheduled before higher priority SUs.
2040 unsigned RegReductionPQBase::getNodePriority(const SUnit *SU) const {
2041   assert(SU->NodeNum < SethiUllmanNumbers.size());
2042   unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0;
2043   if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
2044     // CopyToReg should be close to its uses to facilitate coalescing and
2045     // avoid spilling.
2046     return 0;
2047   if (Opc == TargetOpcode::EXTRACT_SUBREG ||
2048       Opc == TargetOpcode::SUBREG_TO_REG ||
2049       Opc == TargetOpcode::INSERT_SUBREG)
2050     // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be
2051     // close to their uses to facilitate coalescing.
2052     return 0;
2053   if (SU->NumSuccs == 0 && SU->NumPreds != 0)
2054     // If SU does not have a register use, i.e. it doesn't produce a value
2055     // that would be consumed (e.g. store), then it terminates a chain of
2056     // computation.  Give it a large SethiUllman number so it will be
2057     // scheduled right before its predecessors that it doesn't lengthen
2058     // their live ranges.
2059     return 0xffff;
2060   if (SU->NumPreds == 0 && SU->NumSuccs != 0)
2061     // If SU does not have a register def, schedule it close to its uses
2062     // because it does not lengthen any live ranges.
2063     return 0;
2064 #if 1
2065   return SethiUllmanNumbers[SU->NodeNum];
2066 #else
2067   unsigned Priority = SethiUllmanNumbers[SU->NodeNum];
2068   if (SU->isCallOp) {
2069     // FIXME: This assumes all of the defs are used as call operands.
2070     int NP = (int)Priority - SU->getNode()->getNumValues();
2071     return (NP > 0) ? NP : 0;
2072   }
2073   return Priority;
2074 #endif
2075 }
2076 
2077 //===----------------------------------------------------------------------===//
2078 //                     Register Pressure Tracking
2079 //===----------------------------------------------------------------------===//
2080 
2081 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2082 LLVM_DUMP_METHOD void RegReductionPQBase::dumpRegPressure() const {
2083   for (const TargetRegisterClass *RC : TRI->regclasses()) {
2084     unsigned Id = RC->getID();
2085     unsigned RP = RegPressure[Id];
2086     if (!RP) continue;
2087     LLVM_DEBUG(dbgs() << TRI->getRegClassName(RC) << ": " << RP << " / "
2088                       << RegLimit[Id] << '\n');
2089   }
2090 }
2091 #endif
2092 
2093 bool RegReductionPQBase::HighRegPressure(const SUnit *SU) const {
2094   if (!TLI)
2095     return false;
2096 
2097   for (const SDep &Pred : SU->Preds) {
2098     if (Pred.isCtrl())
2099       continue;
2100     SUnit *PredSU = Pred.getSUnit();
2101     // NumRegDefsLeft is zero when enough uses of this node have been scheduled
2102     // to cover the number of registers defined (they are all live).
2103     if (PredSU->NumRegDefsLeft == 0) {
2104       continue;
2105     }
2106     for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
2107          RegDefPos.IsValid(); RegDefPos.Advance()) {
2108       unsigned RCId, Cost;
2109       GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
2110 
2111       if ((RegPressure[RCId] + Cost) >= RegLimit[RCId])
2112         return true;
2113     }
2114   }
2115   return false;
2116 }
2117 
2118 bool RegReductionPQBase::MayReduceRegPressure(SUnit *SU) const {
2119   const SDNode *N = SU->getNode();
2120 
2121   if (!N->isMachineOpcode() || !SU->NumSuccs)
2122     return false;
2123 
2124   unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2125   for (unsigned i = 0; i != NumDefs; ++i) {
2126     MVT VT = N->getSimpleValueType(i);
2127     if (!N->hasAnyUseOfValue(i))
2128       continue;
2129     unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2130     if (RegPressure[RCId] >= RegLimit[RCId])
2131       return true;
2132   }
2133   return false;
2134 }
2135 
2136 // Compute the register pressure contribution by this instruction by count up
2137 // for uses that are not live and down for defs. Only count register classes
2138 // that are already under high pressure. As a side effect, compute the number of
2139 // uses of registers that are already live.
2140 //
2141 // FIXME: This encompasses the logic in HighRegPressure and MayReduceRegPressure
2142 // so could probably be factored.
2143 int RegReductionPQBase::RegPressureDiff(SUnit *SU, unsigned &LiveUses) const {
2144   LiveUses = 0;
2145   int PDiff = 0;
2146   for (const SDep &Pred : SU->Preds) {
2147     if (Pred.isCtrl())
2148       continue;
2149     SUnit *PredSU = Pred.getSUnit();
2150     // NumRegDefsLeft is zero when enough uses of this node have been scheduled
2151     // to cover the number of registers defined (they are all live).
2152     if (PredSU->NumRegDefsLeft == 0) {
2153       if (PredSU->getNode()->isMachineOpcode())
2154         ++LiveUses;
2155       continue;
2156     }
2157     for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
2158          RegDefPos.IsValid(); RegDefPos.Advance()) {
2159       MVT VT = RegDefPos.GetValue();
2160       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2161       if (RegPressure[RCId] >= RegLimit[RCId])
2162         ++PDiff;
2163     }
2164   }
2165   const SDNode *N = SU->getNode();
2166 
2167   if (!N || !N->isMachineOpcode() || !SU->NumSuccs)
2168     return PDiff;
2169 
2170   unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2171   for (unsigned i = 0; i != NumDefs; ++i) {
2172     MVT VT = N->getSimpleValueType(i);
2173     if (!N->hasAnyUseOfValue(i))
2174       continue;
2175     unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2176     if (RegPressure[RCId] >= RegLimit[RCId])
2177       --PDiff;
2178   }
2179   return PDiff;
2180 }
2181 
2182 void RegReductionPQBase::scheduledNode(SUnit *SU) {
2183   if (!TracksRegPressure)
2184     return;
2185 
2186   if (!SU->getNode())
2187     return;
2188 
2189   for (const SDep &Pred : SU->Preds) {
2190     if (Pred.isCtrl())
2191       continue;
2192     SUnit *PredSU = Pred.getSUnit();
2193     // NumRegDefsLeft is zero when enough uses of this node have been scheduled
2194     // to cover the number of registers defined (they are all live).
2195     if (PredSU->NumRegDefsLeft == 0) {
2196       continue;
2197     }
2198     // FIXME: The ScheduleDAG currently loses information about which of a
2199     // node's values is consumed by each dependence. Consequently, if the node
2200     // defines multiple register classes, we don't know which to pressurize
2201     // here. Instead the following loop consumes the register defs in an
2202     // arbitrary order. At least it handles the common case of clustered loads
2203     // to the same class. For precise liveness, each SDep needs to indicate the
2204     // result number. But that tightly couples the ScheduleDAG with the
2205     // SelectionDAG making updates tricky. A simpler hack would be to attach a
2206     // value type or register class to SDep.
2207     //
2208     // The most important aspect of register tracking is balancing the increase
2209     // here with the reduction further below. Note that this SU may use multiple
2210     // defs in PredSU. The can't be determined here, but we've already
2211     // compensated by reducing NumRegDefsLeft in PredSU during
2212     // ScheduleDAGSDNodes::AddSchedEdges.
2213     --PredSU->NumRegDefsLeft;
2214     unsigned SkipRegDefs = PredSU->NumRegDefsLeft;
2215     for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
2216          RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
2217       if (SkipRegDefs)
2218         continue;
2219 
2220       unsigned RCId, Cost;
2221       GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
2222       RegPressure[RCId] += Cost;
2223       break;
2224     }
2225   }
2226 
2227   // We should have this assert, but there may be dead SDNodes that never
2228   // materialize as SUnits, so they don't appear to generate liveness.
2229   //assert(SU->NumRegDefsLeft == 0 && "not all regdefs have scheduled uses");
2230   int SkipRegDefs = (int)SU->NumRegDefsLeft;
2231   for (ScheduleDAGSDNodes::RegDefIter RegDefPos(SU, scheduleDAG);
2232        RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
2233     if (SkipRegDefs > 0)
2234       continue;
2235     unsigned RCId, Cost;
2236     GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
2237     if (RegPressure[RCId] < Cost) {
2238       // Register pressure tracking is imprecise. This can happen. But we try
2239       // hard not to let it happen because it likely results in poor scheduling.
2240       LLVM_DEBUG(dbgs() << "  SU(" << SU->NodeNum
2241                         << ") has too many regdefs\n");
2242       RegPressure[RCId] = 0;
2243     }
2244     else {
2245       RegPressure[RCId] -= Cost;
2246     }
2247   }
2248   LLVM_DEBUG(dumpRegPressure());
2249 }
2250 
2251 void RegReductionPQBase::unscheduledNode(SUnit *SU) {
2252   if (!TracksRegPressure)
2253     return;
2254 
2255   const SDNode *N = SU->getNode();
2256   if (!N) return;
2257 
2258   if (!N->isMachineOpcode()) {
2259     if (N->getOpcode() != ISD::CopyToReg)
2260       return;
2261   } else {
2262     unsigned Opc = N->getMachineOpcode();
2263     if (Opc == TargetOpcode::EXTRACT_SUBREG ||
2264         Opc == TargetOpcode::INSERT_SUBREG ||
2265         Opc == TargetOpcode::SUBREG_TO_REG ||
2266         Opc == TargetOpcode::REG_SEQUENCE ||
2267         Opc == TargetOpcode::IMPLICIT_DEF)
2268       return;
2269   }
2270 
2271   for (const SDep &Pred : SU->Preds) {
2272     if (Pred.isCtrl())
2273       continue;
2274     SUnit *PredSU = Pred.getSUnit();
2275     // NumSuccsLeft counts all deps. Don't compare it with NumSuccs which only
2276     // counts data deps.
2277     if (PredSU->NumSuccsLeft != PredSU->Succs.size())
2278       continue;
2279     const SDNode *PN = PredSU->getNode();
2280     if (!PN->isMachineOpcode()) {
2281       if (PN->getOpcode() == ISD::CopyFromReg) {
2282         MVT VT = PN->getSimpleValueType(0);
2283         unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2284         RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2285       }
2286       continue;
2287     }
2288     unsigned POpc = PN->getMachineOpcode();
2289     if (POpc == TargetOpcode::IMPLICIT_DEF)
2290       continue;
2291     if (POpc == TargetOpcode::EXTRACT_SUBREG ||
2292         POpc == TargetOpcode::INSERT_SUBREG ||
2293         POpc == TargetOpcode::SUBREG_TO_REG) {
2294       MVT VT = PN->getSimpleValueType(0);
2295       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2296       RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2297       continue;
2298     }
2299     if (POpc == TargetOpcode::REG_SEQUENCE) {
2300       unsigned DstRCIdx = PN->getConstantOperandVal(0);
2301       const TargetRegisterClass *RC = TRI->getRegClass(DstRCIdx);
2302       unsigned RCId = RC->getID();
2303       // REG_SEQUENCE is untyped, so getRepRegClassCostFor could not be used
2304       // here. Instead use the same constant as in GetCostForDef.
2305       RegPressure[RCId] += RegSequenceCost;
2306       continue;
2307     }
2308     unsigned NumDefs = TII->get(PN->getMachineOpcode()).getNumDefs();
2309     for (unsigned i = 0; i != NumDefs; ++i) {
2310       MVT VT = PN->getSimpleValueType(i);
2311       if (!PN->hasAnyUseOfValue(i))
2312         continue;
2313       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2314       if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT))
2315         // Register pressure tracking is imprecise. This can happen.
2316         RegPressure[RCId] = 0;
2317       else
2318         RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
2319     }
2320   }
2321 
2322   // Check for isMachineOpcode() as PrescheduleNodesWithMultipleUses()
2323   // may transfer data dependencies to CopyToReg.
2324   if (SU->NumSuccs && N->isMachineOpcode()) {
2325     unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2326     for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
2327       MVT VT = N->getSimpleValueType(i);
2328       if (VT == MVT::Glue || VT == MVT::Other)
2329         continue;
2330       if (!N->hasAnyUseOfValue(i))
2331         continue;
2332       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2333       RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2334     }
2335   }
2336 
2337   LLVM_DEBUG(dumpRegPressure());
2338 }
2339 
2340 //===----------------------------------------------------------------------===//
2341 //           Dynamic Node Priority for Register Pressure Reduction
2342 //===----------------------------------------------------------------------===//
2343 
2344 /// closestSucc - Returns the scheduled cycle of the successor which is
2345 /// closest to the current cycle.
2346 static unsigned closestSucc(const SUnit *SU) {
2347   unsigned MaxHeight = 0;
2348   for (const SDep &Succ : SU->Succs) {
2349     if (Succ.isCtrl()) continue;  // ignore chain succs
2350     unsigned Height = Succ.getSUnit()->getHeight();
2351     // If there are bunch of CopyToRegs stacked up, they should be considered
2352     // to be at the same position.
2353     if (Succ.getSUnit()->getNode() &&
2354         Succ.getSUnit()->getNode()->getOpcode() == ISD::CopyToReg)
2355       Height = closestSucc(Succ.getSUnit())+1;
2356     if (Height > MaxHeight)
2357       MaxHeight = Height;
2358   }
2359   return MaxHeight;
2360 }
2361 
2362 /// calcMaxScratches - Returns an cost estimate of the worse case requirement
2363 /// for scratch registers, i.e. number of data dependencies.
2364 static unsigned calcMaxScratches(const SUnit *SU) {
2365   unsigned Scratches = 0;
2366   for (const SDep &Pred : SU->Preds) {
2367     if (Pred.isCtrl()) continue;  // ignore chain preds
2368     Scratches++;
2369   }
2370   return Scratches;
2371 }
2372 
2373 /// hasOnlyLiveInOpers - Return true if SU has only value predecessors that are
2374 /// CopyFromReg from a virtual register.
2375 static bool hasOnlyLiveInOpers(const SUnit *SU) {
2376   bool RetVal = false;
2377   for (const SDep &Pred : SU->Preds) {
2378     if (Pred.isCtrl()) continue;
2379     const SUnit *PredSU = Pred.getSUnit();
2380     if (PredSU->getNode() &&
2381         PredSU->getNode()->getOpcode() == ISD::CopyFromReg) {
2382       Register Reg =
2383           cast<RegisterSDNode>(PredSU->getNode()->getOperand(1))->getReg();
2384       if (Reg.isVirtual()) {
2385         RetVal = true;
2386         continue;
2387       }
2388     }
2389     return false;
2390   }
2391   return RetVal;
2392 }
2393 
2394 /// hasOnlyLiveOutUses - Return true if SU has only value successors that are
2395 /// CopyToReg to a virtual register. This SU def is probably a liveout and
2396 /// it has no other use. It should be scheduled closer to the terminator.
2397 static bool hasOnlyLiveOutUses(const SUnit *SU) {
2398   bool RetVal = false;
2399   for (const SDep &Succ : SU->Succs) {
2400     if (Succ.isCtrl()) continue;
2401     const SUnit *SuccSU = Succ.getSUnit();
2402     if (SuccSU->getNode() && SuccSU->getNode()->getOpcode() == ISD::CopyToReg) {
2403       Register Reg =
2404           cast<RegisterSDNode>(SuccSU->getNode()->getOperand(1))->getReg();
2405       if (Reg.isVirtual()) {
2406         RetVal = true;
2407         continue;
2408       }
2409     }
2410     return false;
2411   }
2412   return RetVal;
2413 }
2414 
2415 // Set isVRegCycle for a node with only live in opers and live out uses. Also
2416 // set isVRegCycle for its CopyFromReg operands.
2417 //
2418 // This is only relevant for single-block loops, in which case the VRegCycle
2419 // node is likely an induction variable in which the operand and target virtual
2420 // registers should be coalesced (e.g. pre/post increment values). Setting the
2421 // isVRegCycle flag helps the scheduler prioritize other uses of the same
2422 // CopyFromReg so that this node becomes the virtual register "kill". This
2423 // avoids interference between the values live in and out of the block and
2424 // eliminates a copy inside the loop.
2425 static void initVRegCycle(SUnit *SU) {
2426   if (DisableSchedVRegCycle)
2427     return;
2428 
2429   if (!hasOnlyLiveInOpers(SU) || !hasOnlyLiveOutUses(SU))
2430     return;
2431 
2432   LLVM_DEBUG(dbgs() << "VRegCycle: SU(" << SU->NodeNum << ")\n");
2433 
2434   SU->isVRegCycle = true;
2435 
2436   for (const SDep &Pred : SU->Preds) {
2437     if (Pred.isCtrl()) continue;
2438     Pred.getSUnit()->isVRegCycle = true;
2439   }
2440 }
2441 
2442 // After scheduling the definition of a VRegCycle, clear the isVRegCycle flag of
2443 // CopyFromReg operands. We should no longer penalize other uses of this VReg.
2444 static void resetVRegCycle(SUnit *SU) {
2445   if (!SU->isVRegCycle)
2446     return;
2447 
2448   for (const SDep &Pred : SU->Preds) {
2449     if (Pred.isCtrl()) continue;  // ignore chain preds
2450     SUnit *PredSU = Pred.getSUnit();
2451     if (PredSU->isVRegCycle) {
2452       assert(PredSU->getNode()->getOpcode() == ISD::CopyFromReg &&
2453              "VRegCycle def must be CopyFromReg");
2454       Pred.getSUnit()->isVRegCycle = false;
2455     }
2456   }
2457 }
2458 
2459 // Return true if this SUnit uses a CopyFromReg node marked as a VRegCycle. This
2460 // means a node that defines the VRegCycle has not been scheduled yet.
2461 static bool hasVRegCycleUse(const SUnit *SU) {
2462   // If this SU also defines the VReg, don't hoist it as a "use".
2463   if (SU->isVRegCycle)
2464     return false;
2465 
2466   for (const SDep &Pred : SU->Preds) {
2467     if (Pred.isCtrl()) continue;  // ignore chain preds
2468     if (Pred.getSUnit()->isVRegCycle &&
2469         Pred.getSUnit()->getNode()->getOpcode() == ISD::CopyFromReg) {
2470       LLVM_DEBUG(dbgs() << "  VReg cycle use: SU (" << SU->NodeNum << ")\n");
2471       return true;
2472     }
2473   }
2474   return false;
2475 }
2476 
2477 // Check for either a dependence (latency) or resource (hazard) stall.
2478 //
2479 // Note: The ScheduleHazardRecognizer interface requires a non-const SU.
2480 static bool BUHasStall(SUnit *SU, int Height, RegReductionPQBase *SPQ) {
2481   if ((int)SPQ->getCurCycle() < Height) return true;
2482   if (SPQ->getHazardRec()->getHazardType(SU, 0)
2483       != ScheduleHazardRecognizer::NoHazard)
2484     return true;
2485   return false;
2486 }
2487 
2488 // Return -1 if left has higher priority, 1 if right has higher priority.
2489 // Return 0 if latency-based priority is equivalent.
2490 static int BUCompareLatency(SUnit *left, SUnit *right, bool checkPref,
2491                             RegReductionPQBase *SPQ) {
2492   // Scheduling an instruction that uses a VReg whose postincrement has not yet
2493   // been scheduled will induce a copy. Model this as an extra cycle of latency.
2494   int LPenalty = hasVRegCycleUse(left) ? 1 : 0;
2495   int RPenalty = hasVRegCycleUse(right) ? 1 : 0;
2496   int LHeight = (int)left->getHeight() + LPenalty;
2497   int RHeight = (int)right->getHeight() + RPenalty;
2498 
2499   bool LStall = (!checkPref || left->SchedulingPref == Sched::ILP) &&
2500     BUHasStall(left, LHeight, SPQ);
2501   bool RStall = (!checkPref || right->SchedulingPref == Sched::ILP) &&
2502     BUHasStall(right, RHeight, SPQ);
2503 
2504   // If scheduling one of the node will cause a pipeline stall, delay it.
2505   // If scheduling either one of the node will cause a pipeline stall, sort
2506   // them according to their height.
2507   if (LStall) {
2508     if (!RStall)
2509       return 1;
2510     if (LHeight != RHeight)
2511       return LHeight > RHeight ? 1 : -1;
2512   } else if (RStall)
2513     return -1;
2514 
2515   // If either node is scheduling for latency, sort them by height/depth
2516   // and latency.
2517   if (!checkPref || (left->SchedulingPref == Sched::ILP ||
2518                      right->SchedulingPref == Sched::ILP)) {
2519     // If neither instruction stalls (!LStall && !RStall) and HazardRecognizer
2520     // is enabled, grouping instructions by cycle, then its height is already
2521     // covered so only its depth matters. We also reach this point if both stall
2522     // but have the same height.
2523     if (!SPQ->getHazardRec()->isEnabled()) {
2524       if (LHeight != RHeight)
2525         return LHeight > RHeight ? 1 : -1;
2526     }
2527     int LDepth = left->getDepth() - LPenalty;
2528     int RDepth = right->getDepth() - RPenalty;
2529     if (LDepth != RDepth) {
2530       LLVM_DEBUG(dbgs() << "  Comparing latency of SU (" << left->NodeNum
2531                         << ") depth " << LDepth << " vs SU (" << right->NodeNum
2532                         << ") depth " << RDepth << "\n");
2533       return LDepth < RDepth ? 1 : -1;
2534     }
2535     if (left->Latency != right->Latency)
2536       return left->Latency > right->Latency ? 1 : -1;
2537   }
2538   return 0;
2539 }
2540 
2541 static bool BURRSort(SUnit *left, SUnit *right, RegReductionPQBase *SPQ) {
2542   // Schedule physical register definitions close to their use. This is
2543   // motivated by microarchitectures that can fuse cmp+jump macro-ops. But as
2544   // long as shortening physreg live ranges is generally good, we can defer
2545   // creating a subtarget hook.
2546   if (!DisableSchedPhysRegJoin) {
2547     bool LHasPhysReg = left->hasPhysRegDefs;
2548     bool RHasPhysReg = right->hasPhysRegDefs;
2549     if (LHasPhysReg != RHasPhysReg) {
2550       #ifndef NDEBUG
2551       static const char *const PhysRegMsg[] = { " has no physreg",
2552                                                 " defines a physreg" };
2553       #endif
2554       LLVM_DEBUG(dbgs() << "  SU (" << left->NodeNum << ") "
2555                         << PhysRegMsg[LHasPhysReg] << " SU(" << right->NodeNum
2556                         << ") " << PhysRegMsg[RHasPhysReg] << "\n");
2557       return LHasPhysReg < RHasPhysReg;
2558     }
2559   }
2560 
2561   // Prioritize by Sethi-Ulmann number and push CopyToReg nodes down.
2562   unsigned LPriority = SPQ->getNodePriority(left);
2563   unsigned RPriority = SPQ->getNodePriority(right);
2564 
2565   // Be really careful about hoisting call operands above previous calls.
2566   // Only allows it if it would reduce register pressure.
2567   if (left->isCall && right->isCallOp) {
2568     unsigned RNumVals = right->getNode()->getNumValues();
2569     RPriority = (RPriority > RNumVals) ? (RPriority - RNumVals) : 0;
2570   }
2571   if (right->isCall && left->isCallOp) {
2572     unsigned LNumVals = left->getNode()->getNumValues();
2573     LPriority = (LPriority > LNumVals) ? (LPriority - LNumVals) : 0;
2574   }
2575 
2576   if (LPriority != RPriority)
2577     return LPriority > RPriority;
2578 
2579   // One or both of the nodes are calls and their sethi-ullman numbers are the
2580   // same, then keep source order.
2581   if (left->isCall || right->isCall) {
2582     unsigned LOrder = SPQ->getNodeOrdering(left);
2583     unsigned ROrder = SPQ->getNodeOrdering(right);
2584 
2585     // Prefer an ordering where the lower the non-zero order number, the higher
2586     // the preference.
2587     if ((LOrder || ROrder) && LOrder != ROrder)
2588       return LOrder != 0 && (LOrder < ROrder || ROrder == 0);
2589   }
2590 
2591   // Try schedule def + use closer when Sethi-Ullman numbers are the same.
2592   // e.g.
2593   // t1 = op t2, c1
2594   // t3 = op t4, c2
2595   //
2596   // and the following instructions are both ready.
2597   // t2 = op c3
2598   // t4 = op c4
2599   //
2600   // Then schedule t2 = op first.
2601   // i.e.
2602   // t4 = op c4
2603   // t2 = op c3
2604   // t1 = op t2, c1
2605   // t3 = op t4, c2
2606   //
2607   // This creates more short live intervals.
2608   unsigned LDist = closestSucc(left);
2609   unsigned RDist = closestSucc(right);
2610   if (LDist != RDist)
2611     return LDist < RDist;
2612 
2613   // How many registers becomes live when the node is scheduled.
2614   unsigned LScratch = calcMaxScratches(left);
2615   unsigned RScratch = calcMaxScratches(right);
2616   if (LScratch != RScratch)
2617     return LScratch > RScratch;
2618 
2619   // Comparing latency against a call makes little sense unless the node
2620   // is register pressure-neutral.
2621   if ((left->isCall && RPriority > 0) || (right->isCall && LPriority > 0))
2622     return (left->NodeQueueId > right->NodeQueueId);
2623 
2624   // Do not compare latencies when one or both of the nodes are calls.
2625   if (!DisableSchedCycles &&
2626       !(left->isCall || right->isCall)) {
2627     int result = BUCompareLatency(left, right, false /*checkPref*/, SPQ);
2628     if (result != 0)
2629       return result > 0;
2630   }
2631   else {
2632     if (left->getHeight() != right->getHeight())
2633       return left->getHeight() > right->getHeight();
2634 
2635     if (left->getDepth() != right->getDepth())
2636       return left->getDepth() < right->getDepth();
2637   }
2638 
2639   assert(left->NodeQueueId && right->NodeQueueId &&
2640          "NodeQueueId cannot be zero");
2641   return (left->NodeQueueId > right->NodeQueueId);
2642 }
2643 
2644 // Bottom up
2645 bool bu_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2646   if (int res = checkSpecialNodes(left, right))
2647     return res > 0;
2648 
2649   return BURRSort(left, right, SPQ);
2650 }
2651 
2652 // Source order, otherwise bottom up.
2653 bool src_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2654   if (int res = checkSpecialNodes(left, right))
2655     return res > 0;
2656 
2657   unsigned LOrder = SPQ->getNodeOrdering(left);
2658   unsigned ROrder = SPQ->getNodeOrdering(right);
2659 
2660   // Prefer an ordering where the lower the non-zero order number, the higher
2661   // the preference.
2662   if ((LOrder || ROrder) && LOrder != ROrder)
2663     return LOrder != 0 && (LOrder < ROrder || ROrder == 0);
2664 
2665   return BURRSort(left, right, SPQ);
2666 }
2667 
2668 // If the time between now and when the instruction will be ready can cover
2669 // the spill code, then avoid adding it to the ready queue. This gives long
2670 // stalls highest priority and allows hoisting across calls. It should also
2671 // speed up processing the available queue.
2672 bool hybrid_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
2673   static const unsigned ReadyDelay = 3;
2674 
2675   if (SPQ->MayReduceRegPressure(SU)) return true;
2676 
2677   if (SU->getHeight() > (CurCycle + ReadyDelay)) return false;
2678 
2679   if (SPQ->getHazardRec()->getHazardType(SU, -ReadyDelay)
2680       != ScheduleHazardRecognizer::NoHazard)
2681     return false;
2682 
2683   return true;
2684 }
2685 
2686 // Return true if right should be scheduled with higher priority than left.
2687 bool hybrid_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2688   if (int res = checkSpecialNodes(left, right))
2689     return res > 0;
2690 
2691   if (left->isCall || right->isCall)
2692     // No way to compute latency of calls.
2693     return BURRSort(left, right, SPQ);
2694 
2695   bool LHigh = SPQ->HighRegPressure(left);
2696   bool RHigh = SPQ->HighRegPressure(right);
2697   // Avoid causing spills. If register pressure is high, schedule for
2698   // register pressure reduction.
2699   if (LHigh && !RHigh) {
2700     LLVM_DEBUG(dbgs() << "  pressure SU(" << left->NodeNum << ") > SU("
2701                       << right->NodeNum << ")\n");
2702     return true;
2703   }
2704   else if (!LHigh && RHigh) {
2705     LLVM_DEBUG(dbgs() << "  pressure SU(" << right->NodeNum << ") > SU("
2706                       << left->NodeNum << ")\n");
2707     return false;
2708   }
2709   if (!LHigh && !RHigh) {
2710     int result = BUCompareLatency(left, right, true /*checkPref*/, SPQ);
2711     if (result != 0)
2712       return result > 0;
2713   }
2714   return BURRSort(left, right, SPQ);
2715 }
2716 
2717 // Schedule as many instructions in each cycle as possible. So don't make an
2718 // instruction available unless it is ready in the current cycle.
2719 bool ilp_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
2720   if (SU->getHeight() > CurCycle) return false;
2721 
2722   if (SPQ->getHazardRec()->getHazardType(SU, 0)
2723       != ScheduleHazardRecognizer::NoHazard)
2724     return false;
2725 
2726   return true;
2727 }
2728 
2729 static bool canEnableCoalescing(SUnit *SU) {
2730   unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0;
2731   if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
2732     // CopyToReg should be close to its uses to facilitate coalescing and
2733     // avoid spilling.
2734     return true;
2735 
2736   if (Opc == TargetOpcode::EXTRACT_SUBREG ||
2737       Opc == TargetOpcode::SUBREG_TO_REG ||
2738       Opc == TargetOpcode::INSERT_SUBREG)
2739     // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be
2740     // close to their uses to facilitate coalescing.
2741     return true;
2742 
2743   if (SU->NumPreds == 0 && SU->NumSuccs != 0)
2744     // If SU does not have a register def, schedule it close to its uses
2745     // because it does not lengthen any live ranges.
2746     return true;
2747 
2748   return false;
2749 }
2750 
2751 // list-ilp is currently an experimental scheduler that allows various
2752 // heuristics to be enabled prior to the normal register reduction logic.
2753 bool ilp_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2754   if (int res = checkSpecialNodes(left, right))
2755     return res > 0;
2756 
2757   if (left->isCall || right->isCall)
2758     // No way to compute latency of calls.
2759     return BURRSort(left, right, SPQ);
2760 
2761   unsigned LLiveUses = 0, RLiveUses = 0;
2762   int LPDiff = 0, RPDiff = 0;
2763   if (!DisableSchedRegPressure || !DisableSchedLiveUses) {
2764     LPDiff = SPQ->RegPressureDiff(left, LLiveUses);
2765     RPDiff = SPQ->RegPressureDiff(right, RLiveUses);
2766   }
2767   if (!DisableSchedRegPressure && LPDiff != RPDiff) {
2768     LLVM_DEBUG(dbgs() << "RegPressureDiff SU(" << left->NodeNum
2769                       << "): " << LPDiff << " != SU(" << right->NodeNum
2770                       << "): " << RPDiff << "\n");
2771     return LPDiff > RPDiff;
2772   }
2773 
2774   if (!DisableSchedRegPressure && (LPDiff > 0 || RPDiff > 0)) {
2775     bool LReduce = canEnableCoalescing(left);
2776     bool RReduce = canEnableCoalescing(right);
2777     if (LReduce && !RReduce) return false;
2778     if (RReduce && !LReduce) return true;
2779   }
2780 
2781   if (!DisableSchedLiveUses && (LLiveUses != RLiveUses)) {
2782     LLVM_DEBUG(dbgs() << "Live uses SU(" << left->NodeNum << "): " << LLiveUses
2783                       << " != SU(" << right->NodeNum << "): " << RLiveUses
2784                       << "\n");
2785     return LLiveUses < RLiveUses;
2786   }
2787 
2788   if (!DisableSchedStalls) {
2789     bool LStall = BUHasStall(left, left->getHeight(), SPQ);
2790     bool RStall = BUHasStall(right, right->getHeight(), SPQ);
2791     if (LStall != RStall)
2792       return left->getHeight() > right->getHeight();
2793   }
2794 
2795   if (!DisableSchedCriticalPath) {
2796     int spread = (int)left->getDepth() - (int)right->getDepth();
2797     if (std::abs(spread) > MaxReorderWindow) {
2798       LLVM_DEBUG(dbgs() << "Depth of SU(" << left->NodeNum << "): "
2799                         << left->getDepth() << " != SU(" << right->NodeNum
2800                         << "): " << right->getDepth() << "\n");
2801       return left->getDepth() < right->getDepth();
2802     }
2803   }
2804 
2805   if (!DisableSchedHeight && left->getHeight() != right->getHeight()) {
2806     int spread = (int)left->getHeight() - (int)right->getHeight();
2807     if (std::abs(spread) > MaxReorderWindow)
2808       return left->getHeight() > right->getHeight();
2809   }
2810 
2811   return BURRSort(left, right, SPQ);
2812 }
2813 
2814 void RegReductionPQBase::initNodes(std::vector<SUnit> &sunits) {
2815   SUnits = &sunits;
2816   // Add pseudo dependency edges for two-address nodes.
2817   if (!Disable2AddrHack)
2818     AddPseudoTwoAddrDeps();
2819   // Reroute edges to nodes with multiple uses.
2820   if (!TracksRegPressure && !SrcOrder)
2821     PrescheduleNodesWithMultipleUses();
2822   // Calculate node priorities.
2823   CalculateSethiUllmanNumbers();
2824 
2825   // For single block loops, mark nodes that look like canonical IV increments.
2826   if (scheduleDAG->BB->isSuccessor(scheduleDAG->BB))
2827     for (SUnit &SU : sunits)
2828       initVRegCycle(&SU);
2829 }
2830 
2831 //===----------------------------------------------------------------------===//
2832 //                    Preschedule for Register Pressure
2833 //===----------------------------------------------------------------------===//
2834 
2835 bool RegReductionPQBase::canClobber(const SUnit *SU, const SUnit *Op) {
2836   if (SU->isTwoAddress) {
2837     unsigned Opc = SU->getNode()->getMachineOpcode();
2838     const MCInstrDesc &MCID = TII->get(Opc);
2839     unsigned NumRes = MCID.getNumDefs();
2840     unsigned NumOps = MCID.getNumOperands() - NumRes;
2841     for (unsigned i = 0; i != NumOps; ++i) {
2842       if (MCID.getOperandConstraint(i+NumRes, MCOI::TIED_TO) != -1) {
2843         SDNode *DU = SU->getNode()->getOperand(i).getNode();
2844         if (DU->getNodeId() != -1 &&
2845             Op->OrigNode == &(*SUnits)[DU->getNodeId()])
2846           return true;
2847       }
2848     }
2849   }
2850   return false;
2851 }
2852 
2853 /// canClobberReachingPhysRegUse - True if SU would clobber one of it's
2854 /// successor's explicit physregs whose definition can reach DepSU.
2855 /// i.e. DepSU should not be scheduled above SU.
2856 static bool canClobberReachingPhysRegUse(const SUnit *DepSU, const SUnit *SU,
2857                                          ScheduleDAGRRList *scheduleDAG,
2858                                          const TargetInstrInfo *TII,
2859                                          const TargetRegisterInfo *TRI) {
2860   ArrayRef<MCPhysReg> ImpDefs =
2861       TII->get(SU->getNode()->getMachineOpcode()).implicit_defs();
2862   const uint32_t *RegMask = getNodeRegMask(SU->getNode());
2863   if (ImpDefs.empty() && !RegMask)
2864     return false;
2865 
2866   for (const SDep &Succ : SU->Succs) {
2867     SUnit *SuccSU = Succ.getSUnit();
2868     for (const SDep &SuccPred : SuccSU->Preds) {
2869       if (!SuccPred.isAssignedRegDep())
2870         continue;
2871 
2872       if (RegMask &&
2873           MachineOperand::clobbersPhysReg(RegMask, SuccPred.getReg()) &&
2874           scheduleDAG->IsReachable(DepSU, SuccPred.getSUnit()))
2875         return true;
2876 
2877       for (MCPhysReg ImpDef : ImpDefs) {
2878         // Return true if SU clobbers this physical register use and the
2879         // definition of the register reaches from DepSU. IsReachable queries
2880         // a topological forward sort of the DAG (following the successors).
2881         if (TRI->regsOverlap(ImpDef, SuccPred.getReg()) &&
2882             scheduleDAG->IsReachable(DepSU, SuccPred.getSUnit()))
2883           return true;
2884       }
2885     }
2886   }
2887   return false;
2888 }
2889 
2890 /// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's
2891 /// physical register defs.
2892 static bool canClobberPhysRegDefs(const SUnit *SuccSU, const SUnit *SU,
2893                                   const TargetInstrInfo *TII,
2894                                   const TargetRegisterInfo *TRI) {
2895   SDNode *N = SuccSU->getNode();
2896   unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2897   ArrayRef<MCPhysReg> ImpDefs = TII->get(N->getMachineOpcode()).implicit_defs();
2898   assert(!ImpDefs.empty() && "Caller should check hasPhysRegDefs");
2899   for (const SDNode *SUNode = SU->getNode(); SUNode;
2900        SUNode = SUNode->getGluedNode()) {
2901     if (!SUNode->isMachineOpcode())
2902       continue;
2903     ArrayRef<MCPhysReg> SUImpDefs =
2904         TII->get(SUNode->getMachineOpcode()).implicit_defs();
2905     const uint32_t *SURegMask = getNodeRegMask(SUNode);
2906     if (SUImpDefs.empty() && !SURegMask)
2907       continue;
2908     for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
2909       MVT VT = N->getSimpleValueType(i);
2910       if (VT == MVT::Glue || VT == MVT::Other)
2911         continue;
2912       if (!N->hasAnyUseOfValue(i))
2913         continue;
2914       MCPhysReg Reg = ImpDefs[i - NumDefs];
2915       if (SURegMask && MachineOperand::clobbersPhysReg(SURegMask, Reg))
2916         return true;
2917       for (MCPhysReg SUReg : SUImpDefs) {
2918         if (TRI->regsOverlap(Reg, SUReg))
2919           return true;
2920       }
2921     }
2922   }
2923   return false;
2924 }
2925 
2926 /// PrescheduleNodesWithMultipleUses - Nodes with multiple uses
2927 /// are not handled well by the general register pressure reduction
2928 /// heuristics. When presented with code like this:
2929 ///
2930 ///      N
2931 ///    / |
2932 ///   /  |
2933 ///  U  store
2934 ///  |
2935 /// ...
2936 ///
2937 /// the heuristics tend to push the store up, but since the
2938 /// operand of the store has another use (U), this would increase
2939 /// the length of that other use (the U->N edge).
2940 ///
2941 /// This function transforms code like the above to route U's
2942 /// dependence through the store when possible, like this:
2943 ///
2944 ///      N
2945 ///      ||
2946 ///      ||
2947 ///     store
2948 ///       |
2949 ///       U
2950 ///       |
2951 ///      ...
2952 ///
2953 /// This results in the store being scheduled immediately
2954 /// after N, which shortens the U->N live range, reducing
2955 /// register pressure.
2956 void RegReductionPQBase::PrescheduleNodesWithMultipleUses() {
2957   // Visit all the nodes in topological order, working top-down.
2958   for (SUnit &SU : *SUnits) {
2959     // For now, only look at nodes with no data successors, such as stores.
2960     // These are especially important, due to the heuristics in
2961     // getNodePriority for nodes with no data successors.
2962     if (SU.NumSuccs != 0)
2963       continue;
2964     // For now, only look at nodes with exactly one data predecessor.
2965     if (SU.NumPreds != 1)
2966       continue;
2967     // Avoid prescheduling copies to virtual registers, which don't behave
2968     // like other nodes from the perspective of scheduling heuristics.
2969     if (SDNode *N = SU.getNode())
2970       if (N->getOpcode() == ISD::CopyToReg &&
2971           cast<RegisterSDNode>(N->getOperand(1))->getReg().isVirtual())
2972         continue;
2973 
2974     SDNode *PredFrameSetup = nullptr;
2975     for (const SDep &Pred : SU.Preds)
2976       if (Pred.isCtrl() && Pred.getSUnit()) {
2977         // Find the predecessor which is not data dependence.
2978         SDNode *PredND = Pred.getSUnit()->getNode();
2979 
2980         // If PredND is FrameSetup, we should not pre-scheduled the node,
2981         // or else, when bottom up scheduling, ADJCALLSTACKDOWN and
2982         // ADJCALLSTACKUP may hold CallResource too long and make other
2983         // calls can't be scheduled. If there's no other available node
2984         // to schedule, the schedular will try to rename the register by
2985         // creating copy to avoid the conflict which will fail because
2986         // CallResource is not a real physical register.
2987         if (PredND && PredND->isMachineOpcode() &&
2988             (PredND->getMachineOpcode() == TII->getCallFrameSetupOpcode())) {
2989           PredFrameSetup = PredND;
2990           break;
2991         }
2992       }
2993     // Skip the node has FrameSetup parent.
2994     if (PredFrameSetup != nullptr)
2995       continue;
2996 
2997     // Locate the single data predecessor.
2998     SUnit *PredSU = nullptr;
2999     for (const SDep &Pred : SU.Preds)
3000       if (!Pred.isCtrl()) {
3001         PredSU = Pred.getSUnit();
3002         break;
3003       }
3004     assert(PredSU);
3005 
3006     // Don't rewrite edges that carry physregs, because that requires additional
3007     // support infrastructure.
3008     if (PredSU->hasPhysRegDefs)
3009       continue;
3010     // Short-circuit the case where SU is PredSU's only data successor.
3011     if (PredSU->NumSuccs == 1)
3012       continue;
3013     // Avoid prescheduling to copies from virtual registers, which don't behave
3014     // like other nodes from the perspective of scheduling heuristics.
3015     if (SDNode *N = SU.getNode())
3016       if (N->getOpcode() == ISD::CopyFromReg &&
3017           cast<RegisterSDNode>(N->getOperand(1))->getReg().isVirtual())
3018         continue;
3019 
3020     // Perform checks on the successors of PredSU.
3021     for (const SDep &PredSucc : PredSU->Succs) {
3022       SUnit *PredSuccSU = PredSucc.getSUnit();
3023       if (PredSuccSU == &SU) continue;
3024       // If PredSU has another successor with no data successors, for
3025       // now don't attempt to choose either over the other.
3026       if (PredSuccSU->NumSuccs == 0)
3027         goto outer_loop_continue;
3028       // Don't break physical register dependencies.
3029       if (SU.hasPhysRegClobbers && PredSuccSU->hasPhysRegDefs)
3030         if (canClobberPhysRegDefs(PredSuccSU, &SU, TII, TRI))
3031           goto outer_loop_continue;
3032       // Don't introduce graph cycles.
3033       if (scheduleDAG->IsReachable(&SU, PredSuccSU))
3034         goto outer_loop_continue;
3035     }
3036 
3037     // Ok, the transformation is safe and the heuristics suggest it is
3038     // profitable. Update the graph.
3039     LLVM_DEBUG(
3040         dbgs() << "    Prescheduling SU #" << SU.NodeNum << " next to PredSU #"
3041                << PredSU->NodeNum
3042                << " to guide scheduling in the presence of multiple uses\n");
3043     for (unsigned i = 0; i != PredSU->Succs.size(); ++i) {
3044       SDep Edge = PredSU->Succs[i];
3045       assert(!Edge.isAssignedRegDep());
3046       SUnit *SuccSU = Edge.getSUnit();
3047       if (SuccSU != &SU) {
3048         Edge.setSUnit(PredSU);
3049         scheduleDAG->RemovePred(SuccSU, Edge);
3050         scheduleDAG->AddPredQueued(&SU, Edge);
3051         Edge.setSUnit(&SU);
3052         scheduleDAG->AddPredQueued(SuccSU, Edge);
3053         --i;
3054       }
3055     }
3056   outer_loop_continue:;
3057   }
3058 }
3059 
3060 /// AddPseudoTwoAddrDeps - If two nodes share an operand and one of them uses
3061 /// it as a def&use operand. Add a pseudo control edge from it to the other
3062 /// node (if it won't create a cycle) so the two-address one will be scheduled
3063 /// first (lower in the schedule). If both nodes are two-address, favor the
3064 /// one that has a CopyToReg use (more likely to be a loop induction update).
3065 /// If both are two-address, but one is commutable while the other is not
3066 /// commutable, favor the one that's not commutable.
3067 void RegReductionPQBase::AddPseudoTwoAddrDeps() {
3068   for (SUnit &SU : *SUnits) {
3069     if (!SU.isTwoAddress)
3070       continue;
3071 
3072     SDNode *Node = SU.getNode();
3073     if (!Node || !Node->isMachineOpcode() || SU.getNode()->getGluedNode())
3074       continue;
3075 
3076     bool isLiveOut = hasOnlyLiveOutUses(&SU);
3077     unsigned Opc = Node->getMachineOpcode();
3078     const MCInstrDesc &MCID = TII->get(Opc);
3079     unsigned NumRes = MCID.getNumDefs();
3080     unsigned NumOps = MCID.getNumOperands() - NumRes;
3081     for (unsigned j = 0; j != NumOps; ++j) {
3082       if (MCID.getOperandConstraint(j+NumRes, MCOI::TIED_TO) == -1)
3083         continue;
3084       SDNode *DU = SU.getNode()->getOperand(j).getNode();
3085       if (DU->getNodeId() == -1)
3086         continue;
3087       const SUnit *DUSU = &(*SUnits)[DU->getNodeId()];
3088       if (!DUSU)
3089         continue;
3090       for (const SDep &Succ : DUSU->Succs) {
3091         if (Succ.isCtrl())
3092           continue;
3093         SUnit *SuccSU = Succ.getSUnit();
3094         if (SuccSU == &SU)
3095           continue;
3096         // Be conservative. Ignore if nodes aren't at roughly the same
3097         // depth and height.
3098         if (SuccSU->getHeight() < SU.getHeight() &&
3099             (SU.getHeight() - SuccSU->getHeight()) > 1)
3100           continue;
3101         // Skip past COPY_TO_REGCLASS nodes, so that the pseudo edge
3102         // constrains whatever is using the copy, instead of the copy
3103         // itself. In the case that the copy is coalesced, this
3104         // preserves the intent of the pseudo two-address heurietics.
3105         while (SuccSU->Succs.size() == 1 &&
3106                SuccSU->getNode()->isMachineOpcode() &&
3107                SuccSU->getNode()->getMachineOpcode() ==
3108                  TargetOpcode::COPY_TO_REGCLASS)
3109           SuccSU = SuccSU->Succs.front().getSUnit();
3110         // Don't constrain non-instruction nodes.
3111         if (!SuccSU->getNode() || !SuccSU->getNode()->isMachineOpcode())
3112           continue;
3113         // Don't constrain nodes with physical register defs if the
3114         // predecessor can clobber them.
3115         if (SuccSU->hasPhysRegDefs && SU.hasPhysRegClobbers) {
3116           if (canClobberPhysRegDefs(SuccSU, &SU, TII, TRI))
3117             continue;
3118         }
3119         // Don't constrain EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG;
3120         // these may be coalesced away. We want them close to their uses.
3121         unsigned SuccOpc = SuccSU->getNode()->getMachineOpcode();
3122         if (SuccOpc == TargetOpcode::EXTRACT_SUBREG ||
3123             SuccOpc == TargetOpcode::INSERT_SUBREG ||
3124             SuccOpc == TargetOpcode::SUBREG_TO_REG)
3125           continue;
3126         if (!canClobberReachingPhysRegUse(SuccSU, &SU, scheduleDAG, TII, TRI) &&
3127             (!canClobber(SuccSU, DUSU) ||
3128              (isLiveOut && !hasOnlyLiveOutUses(SuccSU)) ||
3129              (!SU.isCommutable && SuccSU->isCommutable)) &&
3130             !scheduleDAG->IsReachable(SuccSU, &SU)) {
3131           LLVM_DEBUG(dbgs()
3132                      << "    Adding a pseudo-two-addr edge from SU #"
3133                      << SU.NodeNum << " to SU #" << SuccSU->NodeNum << "\n");
3134           scheduleDAG->AddPredQueued(&SU, SDep(SuccSU, SDep::Artificial));
3135         }
3136       }
3137     }
3138   }
3139 }
3140 
3141 //===----------------------------------------------------------------------===//
3142 //                         Public Constructor Functions
3143 //===----------------------------------------------------------------------===//
3144 
3145 ScheduleDAGSDNodes *llvm::createBURRListDAGScheduler(SelectionDAGISel *IS,
3146                                                      CodeGenOptLevel OptLevel) {
3147   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3148   const TargetInstrInfo *TII = STI.getInstrInfo();
3149   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3150 
3151   BURegReductionPriorityQueue *PQ =
3152     new BURegReductionPriorityQueue(*IS->MF, false, false, TII, TRI, nullptr);
3153   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
3154   PQ->setScheduleDAG(SD);
3155   return SD;
3156 }
3157 
3158 ScheduleDAGSDNodes *
3159 llvm::createSourceListDAGScheduler(SelectionDAGISel *IS,
3160                                    CodeGenOptLevel OptLevel) {
3161   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3162   const TargetInstrInfo *TII = STI.getInstrInfo();
3163   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3164 
3165   SrcRegReductionPriorityQueue *PQ =
3166     new SrcRegReductionPriorityQueue(*IS->MF, false, true, TII, TRI, nullptr);
3167   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
3168   PQ->setScheduleDAG(SD);
3169   return SD;
3170 }
3171 
3172 ScheduleDAGSDNodes *
3173 llvm::createHybridListDAGScheduler(SelectionDAGISel *IS,
3174                                    CodeGenOptLevel OptLevel) {
3175   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3176   const TargetInstrInfo *TII = STI.getInstrInfo();
3177   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3178   const TargetLowering *TLI = IS->TLI;
3179 
3180   HybridBURRPriorityQueue *PQ =
3181     new HybridBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI);
3182 
3183   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
3184   PQ->setScheduleDAG(SD);
3185   return SD;
3186 }
3187 
3188 ScheduleDAGSDNodes *llvm::createILPListDAGScheduler(SelectionDAGISel *IS,
3189                                                     CodeGenOptLevel OptLevel) {
3190   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3191   const TargetInstrInfo *TII = STI.getInstrInfo();
3192   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3193   const TargetLowering *TLI = IS->TLI;
3194 
3195   ILPBURRPriorityQueue *PQ =
3196     new ILPBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI);
3197   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
3198   PQ->setScheduleDAG(SD);
3199   return SD;
3200 }
3201