• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===----- ScheduleDAGRRList.cpp - Reg pressure reduction list scheduler --===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This implements bottom-up and top-down register pressure reduction list
11 // schedulers, using standard algorithms.  The basic approach uses a priority
12 // queue of available nodes to schedule.  One at a time, nodes are taken from
13 // the priority queue (thus in priority order), checked for legality to
14 // schedule, and emitted if legal.
15 //
16 //===----------------------------------------------------------------------===//
17 
18 #include "llvm/CodeGen/SchedulerRegistry.h"
19 #include "ScheduleDAGSDNodes.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
25 #include "llvm/CodeGen/SelectionDAGISel.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/InlineAsm.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Target/TargetInstrInfo.h"
32 #include "llvm/Target/TargetLowering.h"
33 #include "llvm/Target/TargetRegisterInfo.h"
34 #include "llvm/Target/TargetSubtargetInfo.h"
35 #include <climits>
36 using namespace llvm;
37 
38 #define DEBUG_TYPE "pre-RA-sched"
39 
40 STATISTIC(NumBacktracks, "Number of times scheduler backtracked");
41 STATISTIC(NumUnfolds,    "Number of nodes unfolded");
42 STATISTIC(NumDups,       "Number of duplicated nodes");
43 STATISTIC(NumPRCopies,   "Number of physical register copies");
44 
45 static RegisterScheduler
46   burrListDAGScheduler("list-burr",
47                        "Bottom-up register reduction list scheduling",
48                        createBURRListDAGScheduler);
49 static RegisterScheduler
50   sourceListDAGScheduler("source",
51                          "Similar to list-burr but schedules in source "
52                          "order when possible",
53                          createSourceListDAGScheduler);
54 
55 static RegisterScheduler
56   hybridListDAGScheduler("list-hybrid",
57                          "Bottom-up register pressure aware list scheduling "
58                          "which tries to balance latency and register pressure",
59                          createHybridListDAGScheduler);
60 
61 static RegisterScheduler
62   ILPListDAGScheduler("list-ilp",
63                       "Bottom-up register pressure aware list scheduling "
64                       "which tries to balance ILP and register pressure",
65                       createILPListDAGScheduler);
66 
67 static cl::opt<bool> DisableSchedCycles(
68   "disable-sched-cycles", cl::Hidden, cl::init(false),
69   cl::desc("Disable cycle-level precision during preRA scheduling"));
70 
71 // Temporary sched=list-ilp flags until the heuristics are robust.
72 // Some options are also available under sched=list-hybrid.
73 static cl::opt<bool> DisableSchedRegPressure(
74   "disable-sched-reg-pressure", cl::Hidden, cl::init(false),
75   cl::desc("Disable regpressure priority in sched=list-ilp"));
76 static cl::opt<bool> DisableSchedLiveUses(
77   "disable-sched-live-uses", cl::Hidden, cl::init(true),
78   cl::desc("Disable live use priority in sched=list-ilp"));
79 static cl::opt<bool> DisableSchedVRegCycle(
80   "disable-sched-vrcycle", cl::Hidden, cl::init(false),
81   cl::desc("Disable virtual register cycle interference checks"));
82 static cl::opt<bool> DisableSchedPhysRegJoin(
83   "disable-sched-physreg-join", cl::Hidden, cl::init(false),
84   cl::desc("Disable physreg def-use affinity"));
85 static cl::opt<bool> DisableSchedStalls(
86   "disable-sched-stalls", cl::Hidden, cl::init(true),
87   cl::desc("Disable no-stall priority in sched=list-ilp"));
88 static cl::opt<bool> DisableSchedCriticalPath(
89   "disable-sched-critical-path", cl::Hidden, cl::init(false),
90   cl::desc("Disable critical path priority in sched=list-ilp"));
91 static cl::opt<bool> DisableSchedHeight(
92   "disable-sched-height", cl::Hidden, cl::init(false),
93   cl::desc("Disable scheduled-height priority in sched=list-ilp"));
94 static cl::opt<bool> Disable2AddrHack(
95   "disable-2addr-hack", cl::Hidden, cl::init(true),
96   cl::desc("Disable scheduler's two-address hack"));
97 
98 static cl::opt<int> MaxReorderWindow(
99   "max-sched-reorder", cl::Hidden, cl::init(6),
100   cl::desc("Number of instructions to allow ahead of the critical path "
101            "in sched=list-ilp"));
102 
103 static cl::opt<unsigned> AvgIPC(
104   "sched-avg-ipc", cl::Hidden, cl::init(1),
105   cl::desc("Average inst/cycle whan no target itinerary exists."));
106 
107 namespace {
108 //===----------------------------------------------------------------------===//
109 /// ScheduleDAGRRList - The actual register reduction list scheduler
110 /// implementation.  This supports both top-down and bottom-up scheduling.
111 ///
112 class ScheduleDAGRRList : public ScheduleDAGSDNodes {
113 private:
114   /// NeedLatency - True if the scheduler will make use of latency information.
115   ///
116   bool NeedLatency;
117 
118   /// AvailableQueue - The priority queue to use for the available SUnits.
119   SchedulingPriorityQueue *AvailableQueue;
120 
121   /// PendingQueue - This contains all of the instructions whose operands have
122   /// been issued, but their results are not ready yet (due to the latency of
123   /// the operation).  Once the operands becomes available, the instruction is
124   /// added to the AvailableQueue.
125   std::vector<SUnit*> PendingQueue;
126 
127   /// HazardRec - The hazard recognizer to use.
128   ScheduleHazardRecognizer *HazardRec;
129 
130   /// CurCycle - The current scheduler state corresponds to this cycle.
131   unsigned CurCycle;
132 
133   /// MinAvailableCycle - Cycle of the soonest available instruction.
134   unsigned MinAvailableCycle;
135 
136   /// IssueCount - Count instructions issued in this cycle
137   /// Currently valid only for bottom-up scheduling.
138   unsigned IssueCount;
139 
140   /// LiveRegDefs - A set of physical registers and their definition
141   /// that are "live". These nodes must be scheduled before any other nodes that
142   /// modifies the registers can be scheduled.
143   unsigned NumLiveRegs;
144   std::unique_ptr<SUnit*[]> LiveRegDefs;
145   std::unique_ptr<SUnit*[]> LiveRegGens;
146 
147   // Collect interferences between physical register use/defs.
148   // Each interference is an SUnit and set of physical registers.
149   SmallVector<SUnit*, 4> Interferences;
150   typedef DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMapT;
151   LRegsMapT LRegsMap;
152 
153   /// Topo - A topological ordering for SUnits which permits fast IsReachable
154   /// and similar queries.
155   ScheduleDAGTopologicalSort Topo;
156 
157   // Hack to keep track of the inverse of FindCallSeqStart without more crazy
158   // DAG crawling.
159   DenseMap<SUnit*, SUnit*> CallSeqEndForStart;
160 
161 public:
ScheduleDAGRRList(MachineFunction & mf,bool needlatency,SchedulingPriorityQueue * availqueue,CodeGenOpt::Level OptLevel)162   ScheduleDAGRRList(MachineFunction &mf, bool needlatency,
163                     SchedulingPriorityQueue *availqueue,
164                     CodeGenOpt::Level OptLevel)
165     : ScheduleDAGSDNodes(mf),
166       NeedLatency(needlatency), AvailableQueue(availqueue), CurCycle(0),
167       Topo(SUnits, nullptr) {
168 
169     const TargetSubtargetInfo &STI = mf.getSubtarget();
170     if (DisableSchedCycles || !NeedLatency)
171       HazardRec = new ScheduleHazardRecognizer();
172     else
173       HazardRec = STI.getInstrInfo()->CreateTargetHazardRecognizer(&STI, this);
174   }
175 
~ScheduleDAGRRList()176   ~ScheduleDAGRRList() override {
177     delete HazardRec;
178     delete AvailableQueue;
179   }
180 
181   void Schedule() override;
182 
getHazardRec()183   ScheduleHazardRecognizer *getHazardRec() { return HazardRec; }
184 
185   /// IsReachable - Checks if SU is reachable from TargetSU.
IsReachable(const SUnit * SU,const SUnit * TargetSU)186   bool IsReachable(const SUnit *SU, const SUnit *TargetSU) {
187     return Topo.IsReachable(SU, TargetSU);
188   }
189 
190   /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
191   /// create a cycle.
WillCreateCycle(SUnit * SU,SUnit * TargetSU)192   bool WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
193     return Topo.WillCreateCycle(SU, TargetSU);
194   }
195 
196   /// AddPred - adds a predecessor edge to SUnit SU.
197   /// This returns true if this is a new predecessor.
198   /// Updates the topological ordering if required.
AddPred(SUnit * SU,const SDep & D)199   void AddPred(SUnit *SU, const SDep &D) {
200     Topo.AddPred(SU, D.getSUnit());
201     SU->addPred(D);
202   }
203 
204   /// RemovePred - removes a predecessor edge from SUnit SU.
205   /// This returns true if an edge was removed.
206   /// Updates the topological ordering if required.
RemovePred(SUnit * SU,const SDep & D)207   void RemovePred(SUnit *SU, const SDep &D) {
208     Topo.RemovePred(SU, D.getSUnit());
209     SU->removePred(D);
210   }
211 
212 private:
isReady(SUnit * SU)213   bool isReady(SUnit *SU) {
214     return DisableSchedCycles || !AvailableQueue->hasReadyFilter() ||
215       AvailableQueue->isReady(SU);
216   }
217 
218   void ReleasePred(SUnit *SU, const SDep *PredEdge);
219   void ReleasePredecessors(SUnit *SU);
220   void ReleasePending();
221   void AdvanceToCycle(unsigned NextCycle);
222   void AdvancePastStalls(SUnit *SU);
223   void EmitNode(SUnit *SU);
224   void ScheduleNodeBottomUp(SUnit*);
225   void CapturePred(SDep *PredEdge);
226   void UnscheduleNodeBottomUp(SUnit*);
227   void RestoreHazardCheckerBottomUp();
228   void BacktrackBottomUp(SUnit*, SUnit*);
229   SUnit *CopyAndMoveSuccessors(SUnit*);
230   void InsertCopiesAndMoveSuccs(SUnit*, unsigned,
231                                 const TargetRegisterClass*,
232                                 const TargetRegisterClass*,
233                                 SmallVectorImpl<SUnit*>&);
234   bool DelayForLiveRegsBottomUp(SUnit*, SmallVectorImpl<unsigned>&);
235 
236   void releaseInterferences(unsigned Reg = 0);
237 
238   SUnit *PickNodeToScheduleBottomUp();
239   void ListScheduleBottomUp();
240 
241   /// CreateNewSUnit - Creates a new SUnit and returns a pointer to it.
242   /// Updates the topological ordering if required.
CreateNewSUnit(SDNode * N)243   SUnit *CreateNewSUnit(SDNode *N) {
244     unsigned NumSUnits = SUnits.size();
245     SUnit *NewNode = newSUnit(N);
246     // Update the topological ordering.
247     if (NewNode->NodeNum >= NumSUnits)
248       Topo.InitDAGTopologicalSorting();
249     return NewNode;
250   }
251 
252   /// CreateClone - Creates a new SUnit from an existing one.
253   /// Updates the topological ordering if required.
CreateClone(SUnit * N)254   SUnit *CreateClone(SUnit *N) {
255     unsigned NumSUnits = SUnits.size();
256     SUnit *NewNode = Clone(N);
257     // Update the topological ordering.
258     if (NewNode->NodeNum >= NumSUnits)
259       Topo.InitDAGTopologicalSorting();
260     return NewNode;
261   }
262 
263   /// forceUnitLatencies - Register-pressure-reducing scheduling doesn't
264   /// need actual latency information but the hybrid scheduler does.
forceUnitLatencies() const265   bool forceUnitLatencies() const override {
266     return !NeedLatency;
267   }
268 };
269 }  // end anonymous namespace
270 
271 /// GetCostForDef - Looks up the register class and cost for a given definition.
272 /// Typically this just means looking up the representative register class,
273 /// but for untyped values (MVT::Untyped) it means inspecting the node's
274 /// opcode to determine what register class is being generated.
GetCostForDef(const ScheduleDAGSDNodes::RegDefIter & RegDefPos,const TargetLowering * TLI,const TargetInstrInfo * TII,const TargetRegisterInfo * TRI,unsigned & RegClass,unsigned & Cost,const MachineFunction & MF)275 static void GetCostForDef(const ScheduleDAGSDNodes::RegDefIter &RegDefPos,
276                           const TargetLowering *TLI,
277                           const TargetInstrInfo *TII,
278                           const TargetRegisterInfo *TRI,
279                           unsigned &RegClass, unsigned &Cost,
280                           const MachineFunction &MF) {
281   MVT VT = RegDefPos.GetValue();
282 
283   // Special handling for untyped values.  These values can only come from
284   // the expansion of custom DAG-to-DAG patterns.
285   if (VT == MVT::Untyped) {
286     const SDNode *Node = RegDefPos.GetNode();
287 
288     // Special handling for CopyFromReg of untyped values.
289     if (!Node->isMachineOpcode() && Node->getOpcode() == ISD::CopyFromReg) {
290       unsigned Reg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
291       const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(Reg);
292       RegClass = RC->getID();
293       Cost = 1;
294       return;
295     }
296 
297     unsigned Opcode = Node->getMachineOpcode();
298     if (Opcode == TargetOpcode::REG_SEQUENCE) {
299       unsigned DstRCIdx = cast<ConstantSDNode>(Node->getOperand(0))->getZExtValue();
300       const TargetRegisterClass *RC = TRI->getRegClass(DstRCIdx);
301       RegClass = RC->getID();
302       Cost = 1;
303       return;
304     }
305 
306     unsigned Idx = RegDefPos.GetIdx();
307     const MCInstrDesc Desc = TII->get(Opcode);
308     const TargetRegisterClass *RC = TII->getRegClass(Desc, Idx, TRI, MF);
309     RegClass = RC->getID();
310     // FIXME: Cost arbitrarily set to 1 because there doesn't seem to be a
311     // better way to determine it.
312     Cost = 1;
313   } else {
314     RegClass = TLI->getRepRegClassFor(VT)->getID();
315     Cost = TLI->getRepRegClassCostFor(VT);
316   }
317 }
318 
319 /// Schedule - Schedule the DAG using list scheduling.
Schedule()320 void ScheduleDAGRRList::Schedule() {
321   DEBUG(dbgs()
322         << "********** List Scheduling BB#" << BB->getNumber()
323         << " '" << BB->getName() << "' **********\n");
324 
325   CurCycle = 0;
326   IssueCount = 0;
327   MinAvailableCycle = DisableSchedCycles ? 0 : UINT_MAX;
328   NumLiveRegs = 0;
329   // Allocate slots for each physical register, plus one for a special register
330   // to track the virtual resource of a calling sequence.
331   LiveRegDefs.reset(new SUnit*[TRI->getNumRegs() + 1]());
332   LiveRegGens.reset(new SUnit*[TRI->getNumRegs() + 1]());
333   CallSeqEndForStart.clear();
334   assert(Interferences.empty() && LRegsMap.empty() && "stale Interferences");
335 
336   // Build the scheduling graph.
337   BuildSchedGraph(nullptr);
338 
339   DEBUG(for (SUnit &SU : SUnits)
340           SU.dumpAll(this));
341   Topo.InitDAGTopologicalSorting();
342 
343   AvailableQueue->initNodes(SUnits);
344 
345   HazardRec->Reset();
346 
347   // Execute the actual scheduling loop.
348   ListScheduleBottomUp();
349 
350   AvailableQueue->releaseState();
351 
352   DEBUG({
353       dbgs() << "*** Final schedule ***\n";
354       dumpSchedule();
355       dbgs() << '\n';
356     });
357 }
358 
359 //===----------------------------------------------------------------------===//
360 //  Bottom-Up Scheduling
361 //===----------------------------------------------------------------------===//
362 
363 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
364 /// the AvailableQueue if the count reaches zero. Also update its cycle bound.
ReleasePred(SUnit * SU,const SDep * PredEdge)365 void ScheduleDAGRRList::ReleasePred(SUnit *SU, const SDep *PredEdge) {
366   SUnit *PredSU = PredEdge->getSUnit();
367 
368 #ifndef NDEBUG
369   if (PredSU->NumSuccsLeft == 0) {
370     dbgs() << "*** Scheduling failed! ***\n";
371     PredSU->dump(this);
372     dbgs() << " has been released too many times!\n";
373     llvm_unreachable(nullptr);
374   }
375 #endif
376   --PredSU->NumSuccsLeft;
377 
378   if (!forceUnitLatencies()) {
379     // Updating predecessor's height. This is now the cycle when the
380     // predecessor can be scheduled without causing a pipeline stall.
381     PredSU->setHeightToAtLeast(SU->getHeight() + PredEdge->getLatency());
382   }
383 
384   // If all the node's successors are scheduled, this node is ready
385   // to be scheduled. Ignore the special EntrySU node.
386   if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU) {
387     PredSU->isAvailable = true;
388 
389     unsigned Height = PredSU->getHeight();
390     if (Height < MinAvailableCycle)
391       MinAvailableCycle = Height;
392 
393     if (isReady(PredSU)) {
394       AvailableQueue->push(PredSU);
395     }
396     // CapturePred and others may have left the node in the pending queue, avoid
397     // adding it twice.
398     else if (!PredSU->isPending) {
399       PredSU->isPending = true;
400       PendingQueue.push_back(PredSU);
401     }
402   }
403 }
404 
405 /// IsChainDependent - Test if Outer is reachable from Inner through
406 /// chain dependencies.
IsChainDependent(SDNode * Outer,SDNode * Inner,unsigned NestLevel,const TargetInstrInfo * TII)407 static bool IsChainDependent(SDNode *Outer, SDNode *Inner,
408                              unsigned NestLevel,
409                              const TargetInstrInfo *TII) {
410   SDNode *N = Outer;
411   for (;;) {
412     if (N == Inner)
413       return true;
414     // For a TokenFactor, examine each operand. There may be multiple ways
415     // to get to the CALLSEQ_BEGIN, but we need to find the path with the
416     // most nesting in order to ensure that we find the corresponding match.
417     if (N->getOpcode() == ISD::TokenFactor) {
418       for (const SDValue &Op : N->op_values())
419         if (IsChainDependent(Op.getNode(), Inner, NestLevel, TII))
420           return true;
421       return false;
422     }
423     // Check for a lowered CALLSEQ_BEGIN or CALLSEQ_END.
424     if (N->isMachineOpcode()) {
425       if (N->getMachineOpcode() ==
426           (unsigned)TII->getCallFrameDestroyOpcode()) {
427         ++NestLevel;
428       } else if (N->getMachineOpcode() ==
429                  (unsigned)TII->getCallFrameSetupOpcode()) {
430         if (NestLevel == 0)
431           return false;
432         --NestLevel;
433       }
434     }
435     // Otherwise, find the chain and continue climbing.
436     for (const SDValue &Op : N->op_values())
437       if (Op.getValueType() == MVT::Other) {
438         N = Op.getNode();
439         goto found_chain_operand;
440       }
441     return false;
442   found_chain_operand:;
443     if (N->getOpcode() == ISD::EntryToken)
444       return false;
445   }
446 }
447 
448 /// FindCallSeqStart - Starting from the (lowered) CALLSEQ_END node, locate
449 /// the corresponding (lowered) CALLSEQ_BEGIN node.
450 ///
451 /// NestLevel and MaxNested are used in recursion to indcate the current level
452 /// of nesting of CALLSEQ_BEGIN and CALLSEQ_END pairs, as well as the maximum
453 /// level seen so far.
454 ///
455 /// TODO: It would be better to give CALLSEQ_END an explicit operand to point
456 /// to the corresponding CALLSEQ_BEGIN to avoid needing to search for it.
457 static SDNode *
FindCallSeqStart(SDNode * N,unsigned & NestLevel,unsigned & MaxNest,const TargetInstrInfo * TII)458 FindCallSeqStart(SDNode *N, unsigned &NestLevel, unsigned &MaxNest,
459                  const TargetInstrInfo *TII) {
460   for (;;) {
461     // For a TokenFactor, examine each operand. There may be multiple ways
462     // to get to the CALLSEQ_BEGIN, but we need to find the path with the
463     // most nesting in order to ensure that we find the corresponding match.
464     if (N->getOpcode() == ISD::TokenFactor) {
465       SDNode *Best = nullptr;
466       unsigned BestMaxNest = MaxNest;
467       for (const SDValue &Op : N->op_values()) {
468         unsigned MyNestLevel = NestLevel;
469         unsigned MyMaxNest = MaxNest;
470         if (SDNode *New = FindCallSeqStart(Op.getNode(),
471                                            MyNestLevel, MyMaxNest, TII))
472           if (!Best || (MyMaxNest > BestMaxNest)) {
473             Best = New;
474             BestMaxNest = MyMaxNest;
475           }
476       }
477       assert(Best);
478       MaxNest = BestMaxNest;
479       return Best;
480     }
481     // Check for a lowered CALLSEQ_BEGIN or CALLSEQ_END.
482     if (N->isMachineOpcode()) {
483       if (N->getMachineOpcode() ==
484           (unsigned)TII->getCallFrameDestroyOpcode()) {
485         ++NestLevel;
486         MaxNest = std::max(MaxNest, NestLevel);
487       } else if (N->getMachineOpcode() ==
488                  (unsigned)TII->getCallFrameSetupOpcode()) {
489         assert(NestLevel != 0);
490         --NestLevel;
491         if (NestLevel == 0)
492           return N;
493       }
494     }
495     // Otherwise, find the chain and continue climbing.
496     for (const SDValue &Op : N->op_values())
497       if (Op.getValueType() == MVT::Other) {
498         N = Op.getNode();
499         goto found_chain_operand;
500       }
501     return nullptr;
502   found_chain_operand:;
503     if (N->getOpcode() == ISD::EntryToken)
504       return nullptr;
505   }
506 }
507 
508 /// Call ReleasePred for each predecessor, then update register live def/gen.
509 /// Always update LiveRegDefs for a register dependence even if the current SU
510 /// also defines the register. This effectively create one large live range
511 /// across a sequence of two-address node. This is important because the
512 /// entire chain must be scheduled together. Example:
513 ///
514 /// flags = (3) add
515 /// flags = (2) addc flags
516 /// flags = (1) addc flags
517 ///
518 /// results in
519 ///
520 /// LiveRegDefs[flags] = 3
521 /// LiveRegGens[flags] = 1
522 ///
523 /// If (2) addc is unscheduled, then (1) addc must also be unscheduled to avoid
524 /// interference on flags.
ReleasePredecessors(SUnit * SU)525 void ScheduleDAGRRList::ReleasePredecessors(SUnit *SU) {
526   // Bottom up: release predecessors
527   for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
528        I != E; ++I) {
529     ReleasePred(SU, &*I);
530     if (I->isAssignedRegDep()) {
531       // This is a physical register dependency and it's impossible or
532       // expensive to copy the register. Make sure nothing that can
533       // clobber the register is scheduled between the predecessor and
534       // this node.
535       SUnit *RegDef = LiveRegDefs[I->getReg()]; (void)RegDef;
536       assert((!RegDef || RegDef == SU || RegDef == I->getSUnit()) &&
537              "interference on register dependence");
538       LiveRegDefs[I->getReg()] = I->getSUnit();
539       if (!LiveRegGens[I->getReg()]) {
540         ++NumLiveRegs;
541         LiveRegGens[I->getReg()] = SU;
542       }
543     }
544   }
545 
546   // If we're scheduling a lowered CALLSEQ_END, find the corresponding
547   // CALLSEQ_BEGIN. Inject an artificial physical register dependence between
548   // these nodes, to prevent other calls from being interscheduled with them.
549   unsigned CallResource = TRI->getNumRegs();
550   if (!LiveRegDefs[CallResource])
551     for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode())
552       if (Node->isMachineOpcode() &&
553           Node->getMachineOpcode() == (unsigned)TII->getCallFrameDestroyOpcode()) {
554         unsigned NestLevel = 0;
555         unsigned MaxNest = 0;
556         SDNode *N = FindCallSeqStart(Node, NestLevel, MaxNest, TII);
557 
558         SUnit *Def = &SUnits[N->getNodeId()];
559         CallSeqEndForStart[Def] = SU;
560 
561         ++NumLiveRegs;
562         LiveRegDefs[CallResource] = Def;
563         LiveRegGens[CallResource] = SU;
564         break;
565       }
566 }
567 
568 /// Check to see if any of the pending instructions are ready to issue.  If
569 /// so, add them to the available queue.
ReleasePending()570 void ScheduleDAGRRList::ReleasePending() {
571   if (DisableSchedCycles) {
572     assert(PendingQueue.empty() && "pending instrs not allowed in this mode");
573     return;
574   }
575 
576   // If the available queue is empty, it is safe to reset MinAvailableCycle.
577   if (AvailableQueue->empty())
578     MinAvailableCycle = UINT_MAX;
579 
580   // Check to see if any of the pending instructions are ready to issue.  If
581   // so, add them to the available queue.
582   for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
583     unsigned ReadyCycle = PendingQueue[i]->getHeight();
584     if (ReadyCycle < MinAvailableCycle)
585       MinAvailableCycle = ReadyCycle;
586 
587     if (PendingQueue[i]->isAvailable) {
588       if (!isReady(PendingQueue[i]))
589           continue;
590       AvailableQueue->push(PendingQueue[i]);
591     }
592     PendingQueue[i]->isPending = false;
593     PendingQueue[i] = PendingQueue.back();
594     PendingQueue.pop_back();
595     --i; --e;
596   }
597 }
598 
599 /// Move the scheduler state forward by the specified number of Cycles.
AdvanceToCycle(unsigned NextCycle)600 void ScheduleDAGRRList::AdvanceToCycle(unsigned NextCycle) {
601   if (NextCycle <= CurCycle)
602     return;
603 
604   IssueCount = 0;
605   AvailableQueue->setCurCycle(NextCycle);
606   if (!HazardRec->isEnabled()) {
607     // Bypass lots of virtual calls in case of long latency.
608     CurCycle = NextCycle;
609   }
610   else {
611     for (; CurCycle != NextCycle; ++CurCycle) {
612       HazardRec->RecedeCycle();
613     }
614   }
615   // FIXME: Instead of visiting the pending Q each time, set a dirty flag on the
616   // available Q to release pending nodes at least once before popping.
617   ReleasePending();
618 }
619 
620 /// Move the scheduler state forward until the specified node's dependents are
621 /// ready and can be scheduled with no resource conflicts.
AdvancePastStalls(SUnit * SU)622 void ScheduleDAGRRList::AdvancePastStalls(SUnit *SU) {
623   if (DisableSchedCycles)
624     return;
625 
626   // FIXME: Nodes such as CopyFromReg probably should not advance the current
627   // cycle. Otherwise, we can wrongly mask real stalls. If the non-machine node
628   // has predecessors the cycle will be advanced when they are scheduled.
629   // But given the crude nature of modeling latency though such nodes, we
630   // currently need to treat these nodes like real instructions.
631   // if (!SU->getNode() || !SU->getNode()->isMachineOpcode()) return;
632 
633   unsigned ReadyCycle = SU->getHeight();
634 
635   // Bump CurCycle to account for latency. We assume the latency of other
636   // available instructions may be hidden by the stall (not a full pipe stall).
637   // This updates the hazard recognizer's cycle before reserving resources for
638   // this instruction.
639   AdvanceToCycle(ReadyCycle);
640 
641   // Calls are scheduled in their preceding cycle, so don't conflict with
642   // hazards from instructions after the call. EmitNode will reset the
643   // scoreboard state before emitting the call.
644   if (SU->isCall)
645     return;
646 
647   // FIXME: For resource conflicts in very long non-pipelined stages, we
648   // should probably skip ahead here to avoid useless scoreboard checks.
649   int Stalls = 0;
650   while (true) {
651     ScheduleHazardRecognizer::HazardType HT =
652       HazardRec->getHazardType(SU, -Stalls);
653 
654     if (HT == ScheduleHazardRecognizer::NoHazard)
655       break;
656 
657     ++Stalls;
658   }
659   AdvanceToCycle(CurCycle + Stalls);
660 }
661 
662 /// Record this SUnit in the HazardRecognizer.
663 /// Does not update CurCycle.
EmitNode(SUnit * SU)664 void ScheduleDAGRRList::EmitNode(SUnit *SU) {
665   if (!HazardRec->isEnabled())
666     return;
667 
668   // Check for phys reg copy.
669   if (!SU->getNode())
670     return;
671 
672   switch (SU->getNode()->getOpcode()) {
673   default:
674     assert(SU->getNode()->isMachineOpcode() &&
675            "This target-independent node should not be scheduled.");
676     break;
677   case ISD::MERGE_VALUES:
678   case ISD::TokenFactor:
679   case ISD::LIFETIME_START:
680   case ISD::LIFETIME_END:
681   case ISD::CopyToReg:
682   case ISD::CopyFromReg:
683   case ISD::EH_LABEL:
684     // Noops don't affect the scoreboard state. Copies are likely to be
685     // removed.
686     return;
687   case ISD::INLINEASM:
688     // For inline asm, clear the pipeline state.
689     HazardRec->Reset();
690     return;
691   }
692   if (SU->isCall) {
693     // Calls are scheduled with their preceding instructions. For bottom-up
694     // scheduling, clear the pipeline state before emitting.
695     HazardRec->Reset();
696   }
697 
698   HazardRec->EmitInstruction(SU);
699 }
700 
701 static void resetVRegCycle(SUnit *SU);
702 
703 /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
704 /// count of its predecessors. If a predecessor pending count is zero, add it to
705 /// the Available queue.
ScheduleNodeBottomUp(SUnit * SU)706 void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU) {
707   DEBUG(dbgs() << "\n*** Scheduling [" << CurCycle << "]: ");
708   DEBUG(SU->dump(this));
709 
710 #ifndef NDEBUG
711   if (CurCycle < SU->getHeight())
712     DEBUG(dbgs() << "   Height [" << SU->getHeight()
713           << "] pipeline stall!\n");
714 #endif
715 
716   // FIXME: Do not modify node height. It may interfere with
717   // backtracking. Instead add a "ready cycle" to SUnit. Before scheduling the
718   // node its ready cycle can aid heuristics, and after scheduling it can
719   // indicate the scheduled cycle.
720   SU->setHeightToAtLeast(CurCycle);
721 
722   // Reserve resources for the scheduled instruction.
723   EmitNode(SU);
724 
725   Sequence.push_back(SU);
726 
727   AvailableQueue->scheduledNode(SU);
728 
729   // If HazardRec is disabled, and each inst counts as one cycle, then
730   // advance CurCycle before ReleasePredecessors to avoid useless pushes to
731   // PendingQueue for schedulers that implement HasReadyFilter.
732   if (!HazardRec->isEnabled() && AvgIPC < 2)
733     AdvanceToCycle(CurCycle + 1);
734 
735   // Update liveness of predecessors before successors to avoid treating a
736   // two-address node as a live range def.
737   ReleasePredecessors(SU);
738 
739   // Release all the implicit physical register defs that are live.
740   for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
741        I != E; ++I) {
742     // LiveRegDegs[I->getReg()] != SU when SU is a two-address node.
743     if (I->isAssignedRegDep() && LiveRegDefs[I->getReg()] == SU) {
744       assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
745       --NumLiveRegs;
746       LiveRegDefs[I->getReg()] = nullptr;
747       LiveRegGens[I->getReg()] = nullptr;
748       releaseInterferences(I->getReg());
749     }
750   }
751   // Release the special call resource dependence, if this is the beginning
752   // of a call.
753   unsigned CallResource = TRI->getNumRegs();
754   if (LiveRegDefs[CallResource] == SU)
755     for (const SDNode *SUNode = SU->getNode(); SUNode;
756          SUNode = SUNode->getGluedNode()) {
757       if (SUNode->isMachineOpcode() &&
758           SUNode->getMachineOpcode() == (unsigned)TII->getCallFrameSetupOpcode()) {
759         assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
760         --NumLiveRegs;
761         LiveRegDefs[CallResource] = nullptr;
762         LiveRegGens[CallResource] = nullptr;
763         releaseInterferences(CallResource);
764       }
765     }
766 
767   resetVRegCycle(SU);
768 
769   SU->isScheduled = true;
770 
771   // Conditions under which the scheduler should eagerly advance the cycle:
772   // (1) No available instructions
773   // (2) All pipelines full, so available instructions must have hazards.
774   //
775   // If HazardRec is disabled, the cycle was pre-advanced before calling
776   // ReleasePredecessors. In that case, IssueCount should remain 0.
777   //
778   // Check AvailableQueue after ReleasePredecessors in case of zero latency.
779   if (HazardRec->isEnabled() || AvgIPC > 1) {
780     if (SU->getNode() && SU->getNode()->isMachineOpcode())
781       ++IssueCount;
782     if ((HazardRec->isEnabled() && HazardRec->atIssueLimit())
783         || (!HazardRec->isEnabled() && IssueCount == AvgIPC))
784       AdvanceToCycle(CurCycle + 1);
785   }
786 }
787 
788 /// CapturePred - This does the opposite of ReleasePred. Since SU is being
789 /// unscheduled, incrcease the succ left count of its predecessors. Remove
790 /// them from AvailableQueue if necessary.
CapturePred(SDep * PredEdge)791 void ScheduleDAGRRList::CapturePred(SDep *PredEdge) {
792   SUnit *PredSU = PredEdge->getSUnit();
793   if (PredSU->isAvailable) {
794     PredSU->isAvailable = false;
795     if (!PredSU->isPending)
796       AvailableQueue->remove(PredSU);
797   }
798 
799   assert(PredSU->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!");
800   ++PredSU->NumSuccsLeft;
801 }
802 
803 /// UnscheduleNodeBottomUp - Remove the node from the schedule, update its and
804 /// its predecessor states to reflect the change.
UnscheduleNodeBottomUp(SUnit * SU)805 void ScheduleDAGRRList::UnscheduleNodeBottomUp(SUnit *SU) {
806   DEBUG(dbgs() << "*** Unscheduling [" << SU->getHeight() << "]: ");
807   DEBUG(SU->dump(this));
808 
809   for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
810        I != E; ++I) {
811     CapturePred(&*I);
812     if (I->isAssignedRegDep() && SU == LiveRegGens[I->getReg()]){
813       assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
814       assert(LiveRegDefs[I->getReg()] == I->getSUnit() &&
815              "Physical register dependency violated?");
816       --NumLiveRegs;
817       LiveRegDefs[I->getReg()] = nullptr;
818       LiveRegGens[I->getReg()] = nullptr;
819       releaseInterferences(I->getReg());
820     }
821   }
822 
823   // Reclaim the special call resource dependence, if this is the beginning
824   // of a call.
825   unsigned CallResource = TRI->getNumRegs();
826   for (const SDNode *SUNode = SU->getNode(); SUNode;
827        SUNode = SUNode->getGluedNode()) {
828     if (SUNode->isMachineOpcode() &&
829         SUNode->getMachineOpcode() == (unsigned)TII->getCallFrameSetupOpcode()) {
830       ++NumLiveRegs;
831       LiveRegDefs[CallResource] = SU;
832       LiveRegGens[CallResource] = CallSeqEndForStart[SU];
833     }
834   }
835 
836   // Release the special call resource dependence, if this is the end
837   // of a call.
838   if (LiveRegGens[CallResource] == SU)
839     for (const SDNode *SUNode = SU->getNode(); SUNode;
840          SUNode = SUNode->getGluedNode()) {
841       if (SUNode->isMachineOpcode() &&
842           SUNode->getMachineOpcode() == (unsigned)TII->getCallFrameDestroyOpcode()) {
843         assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
844         --NumLiveRegs;
845         LiveRegDefs[CallResource] = nullptr;
846         LiveRegGens[CallResource] = nullptr;
847         releaseInterferences(CallResource);
848       }
849     }
850 
851   for (auto &Succ : SU->Succs) {
852     if (Succ.isAssignedRegDep()) {
853       auto Reg = Succ.getReg();
854       if (!LiveRegDefs[Reg])
855         ++NumLiveRegs;
856       // This becomes the nearest def. Note that an earlier def may still be
857       // pending if this is a two-address node.
858       LiveRegDefs[Reg] = SU;
859 
860       // Update LiveRegGen only if was empty before this unscheduling.
861       // This is to avoid incorrect updating LiveRegGen set in previous run.
862       if (!LiveRegGens[Reg]) {
863         // Find the successor with the lowest height.
864         LiveRegGens[Reg] = Succ.getSUnit();
865         for (auto &Succ2 : SU->Succs) {
866           if (Succ2.isAssignedRegDep() && Succ2.getReg() == Reg &&
867               Succ2.getSUnit()->getHeight() < LiveRegGens[Reg]->getHeight())
868             LiveRegGens[Reg] = Succ2.getSUnit();
869         }
870       }
871     }
872   }
873   if (SU->getHeight() < MinAvailableCycle)
874     MinAvailableCycle = SU->getHeight();
875 
876   SU->setHeightDirty();
877   SU->isScheduled = false;
878   SU->isAvailable = true;
879   if (!DisableSchedCycles && AvailableQueue->hasReadyFilter()) {
880     // Don't make available until backtracking is complete.
881     SU->isPending = true;
882     PendingQueue.push_back(SU);
883   }
884   else {
885     AvailableQueue->push(SU);
886   }
887   AvailableQueue->unscheduledNode(SU);
888 }
889 
890 /// After backtracking, the hazard checker needs to be restored to a state
891 /// corresponding the current cycle.
RestoreHazardCheckerBottomUp()892 void ScheduleDAGRRList::RestoreHazardCheckerBottomUp() {
893   HazardRec->Reset();
894 
895   unsigned LookAhead = std::min((unsigned)Sequence.size(),
896                                 HazardRec->getMaxLookAhead());
897   if (LookAhead == 0)
898     return;
899 
900   std::vector<SUnit*>::const_iterator I = (Sequence.end() - LookAhead);
901   unsigned HazardCycle = (*I)->getHeight();
902   for (std::vector<SUnit*>::const_iterator E = Sequence.end(); I != E; ++I) {
903     SUnit *SU = *I;
904     for (; SU->getHeight() > HazardCycle; ++HazardCycle) {
905       HazardRec->RecedeCycle();
906     }
907     EmitNode(SU);
908   }
909 }
910 
911 /// BacktrackBottomUp - Backtrack scheduling to a previous cycle specified in
912 /// BTCycle in order to schedule a specific node.
BacktrackBottomUp(SUnit * SU,SUnit * BtSU)913 void ScheduleDAGRRList::BacktrackBottomUp(SUnit *SU, SUnit *BtSU) {
914   SUnit *OldSU = Sequence.back();
915   while (true) {
916     Sequence.pop_back();
917     // FIXME: use ready cycle instead of height
918     CurCycle = OldSU->getHeight();
919     UnscheduleNodeBottomUp(OldSU);
920     AvailableQueue->setCurCycle(CurCycle);
921     if (OldSU == BtSU)
922       break;
923     OldSU = Sequence.back();
924   }
925 
926   assert(!SU->isSucc(OldSU) && "Something is wrong!");
927 
928   RestoreHazardCheckerBottomUp();
929 
930   ReleasePending();
931 
932   ++NumBacktracks;
933 }
934 
isOperandOf(const SUnit * SU,SDNode * N)935 static bool isOperandOf(const SUnit *SU, SDNode *N) {
936   for (const SDNode *SUNode = SU->getNode(); SUNode;
937        SUNode = SUNode->getGluedNode()) {
938     if (SUNode->isOperandOf(N))
939       return true;
940   }
941   return false;
942 }
943 
944 /// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
945 /// successors to the newly created node.
CopyAndMoveSuccessors(SUnit * SU)946 SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) {
947   SDNode *N = SU->getNode();
948   if (!N)
949     return nullptr;
950 
951   if (SU->getNode()->getGluedNode())
952     return nullptr;
953 
954   SUnit *NewSU;
955   bool TryUnfold = false;
956   for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
957     MVT VT = N->getSimpleValueType(i);
958     if (VT == MVT::Glue)
959       return nullptr;
960     else if (VT == MVT::Other)
961       TryUnfold = true;
962   }
963   for (const SDValue &Op : N->op_values()) {
964     MVT VT = Op.getNode()->getSimpleValueType(Op.getResNo());
965     if (VT == MVT::Glue)
966       return nullptr;
967   }
968 
969   if (TryUnfold) {
970     SmallVector<SDNode*, 2> NewNodes;
971     if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
972       return nullptr;
973 
974     // unfolding an x86 DEC64m operation results in store, dec, load which
975     // can't be handled here so quit
976     if (NewNodes.size() == 3)
977       return nullptr;
978 
979     DEBUG(dbgs() << "Unfolding SU #" << SU->NodeNum << "\n");
980     assert(NewNodes.size() == 2 && "Expected a load folding node!");
981 
982     N = NewNodes[1];
983     SDNode *LoadNode = NewNodes[0];
984     unsigned NumVals = N->getNumValues();
985     unsigned OldNumVals = SU->getNode()->getNumValues();
986     for (unsigned i = 0; i != NumVals; ++i)
987       DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i));
988     DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals-1),
989                                    SDValue(LoadNode, 1));
990 
991     // LoadNode may already exist. This can happen when there is another
992     // load from the same location and producing the same type of value
993     // but it has different alignment or volatileness.
994     bool isNewLoad = true;
995     SUnit *LoadSU;
996     if (LoadNode->getNodeId() != -1) {
997       LoadSU = &SUnits[LoadNode->getNodeId()];
998       isNewLoad = false;
999     } else {
1000       LoadSU = CreateNewSUnit(LoadNode);
1001       LoadNode->setNodeId(LoadSU->NodeNum);
1002 
1003       InitNumRegDefsLeft(LoadSU);
1004       computeLatency(LoadSU);
1005     }
1006 
1007     SUnit *NewSU = CreateNewSUnit(N);
1008     assert(N->getNodeId() == -1 && "Node already inserted!");
1009     N->setNodeId(NewSU->NodeNum);
1010 
1011     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1012     for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
1013       if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
1014         NewSU->isTwoAddress = true;
1015         break;
1016       }
1017     }
1018     if (MCID.isCommutable())
1019       NewSU->isCommutable = true;
1020 
1021     InitNumRegDefsLeft(NewSU);
1022     computeLatency(NewSU);
1023 
1024     // Record all the edges to and from the old SU, by category.
1025     SmallVector<SDep, 4> ChainPreds;
1026     SmallVector<SDep, 4> ChainSuccs;
1027     SmallVector<SDep, 4> LoadPreds;
1028     SmallVector<SDep, 4> NodePreds;
1029     SmallVector<SDep, 4> NodeSuccs;
1030     for (SDep &Pred : SU->Preds) {
1031       if (Pred.isCtrl())
1032         ChainPreds.push_back(Pred);
1033       else if (isOperandOf(Pred.getSUnit(), LoadNode))
1034         LoadPreds.push_back(Pred);
1035       else
1036         NodePreds.push_back(Pred);
1037     }
1038     for (SDep &Succ : SU->Succs) {
1039       if (Succ.isCtrl())
1040         ChainSuccs.push_back(Succ);
1041       else
1042         NodeSuccs.push_back(Succ);
1043     }
1044 
1045     // Now assign edges to the newly-created nodes.
1046     for (const SDep &Pred : ChainPreds) {
1047       RemovePred(SU, Pred);
1048       if (isNewLoad)
1049         AddPred(LoadSU, Pred);
1050     }
1051     for (const SDep &Pred : LoadPreds) {
1052       RemovePred(SU, Pred);
1053       if (isNewLoad)
1054         AddPred(LoadSU, Pred);
1055     }
1056     for (const SDep &Pred : NodePreds) {
1057       RemovePred(SU, Pred);
1058       AddPred(NewSU, Pred);
1059     }
1060     for (SDep D : NodeSuccs) {
1061       SUnit *SuccDep = D.getSUnit();
1062       D.setSUnit(SU);
1063       RemovePred(SuccDep, D);
1064       D.setSUnit(NewSU);
1065       AddPred(SuccDep, D);
1066       // Balance register pressure.
1067       if (AvailableQueue->tracksRegPressure() && SuccDep->isScheduled
1068           && !D.isCtrl() && NewSU->NumRegDefsLeft > 0)
1069         --NewSU->NumRegDefsLeft;
1070     }
1071     for (SDep D : ChainSuccs) {
1072       SUnit *SuccDep = D.getSUnit();
1073       D.setSUnit(SU);
1074       RemovePred(SuccDep, D);
1075       if (isNewLoad) {
1076         D.setSUnit(LoadSU);
1077         AddPred(SuccDep, D);
1078       }
1079     }
1080 
1081     // Add a data dependency to reflect that NewSU reads the value defined
1082     // by LoadSU.
1083     SDep D(LoadSU, SDep::Data, 0);
1084     D.setLatency(LoadSU->Latency);
1085     AddPred(NewSU, D);
1086 
1087     if (isNewLoad)
1088       AvailableQueue->addNode(LoadSU);
1089     AvailableQueue->addNode(NewSU);
1090 
1091     ++NumUnfolds;
1092 
1093     if (NewSU->NumSuccsLeft == 0) {
1094       NewSU->isAvailable = true;
1095       return NewSU;
1096     }
1097     SU = NewSU;
1098   }
1099 
1100   DEBUG(dbgs() << "    Duplicating SU #" << SU->NodeNum << "\n");
1101   NewSU = CreateClone(SU);
1102 
1103   // New SUnit has the exact same predecessors.
1104   for (SDep &Pred : SU->Preds)
1105     if (!Pred.isArtificial())
1106       AddPred(NewSU, Pred);
1107 
1108   // Only copy scheduled successors. Cut them from old node's successor
1109   // list and move them over.
1110   SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
1111   for (SDep &Succ : SU->Succs) {
1112     if (Succ.isArtificial())
1113       continue;
1114     SUnit *SuccSU = Succ.getSUnit();
1115     if (SuccSU->isScheduled) {
1116       SDep D = Succ;
1117       D.setSUnit(NewSU);
1118       AddPred(SuccSU, D);
1119       D.setSUnit(SU);
1120       DelDeps.push_back(std::make_pair(SuccSU, D));
1121     }
1122   }
1123   for (auto &DelDep : DelDeps)
1124     RemovePred(DelDep.first, DelDep.second);
1125 
1126   AvailableQueue->updateNode(SU);
1127   AvailableQueue->addNode(NewSU);
1128 
1129   ++NumDups;
1130   return NewSU;
1131 }
1132 
1133 /// InsertCopiesAndMoveSuccs - Insert register copies and move all
1134 /// scheduled successors of the given SUnit to the last copy.
InsertCopiesAndMoveSuccs(SUnit * SU,unsigned Reg,const TargetRegisterClass * DestRC,const TargetRegisterClass * SrcRC,SmallVectorImpl<SUnit * > & Copies)1135 void ScheduleDAGRRList::InsertCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
1136                                               const TargetRegisterClass *DestRC,
1137                                               const TargetRegisterClass *SrcRC,
1138                                               SmallVectorImpl<SUnit*> &Copies) {
1139   SUnit *CopyFromSU = CreateNewSUnit(nullptr);
1140   CopyFromSU->CopySrcRC = SrcRC;
1141   CopyFromSU->CopyDstRC = DestRC;
1142 
1143   SUnit *CopyToSU = CreateNewSUnit(nullptr);
1144   CopyToSU->CopySrcRC = DestRC;
1145   CopyToSU->CopyDstRC = SrcRC;
1146 
1147   // Only copy scheduled successors. Cut them from old node's successor
1148   // list and move them over.
1149   SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
1150   for (SDep &Succ : SU->Succs) {
1151     if (Succ.isArtificial())
1152       continue;
1153     SUnit *SuccSU = Succ.getSUnit();
1154     if (SuccSU->isScheduled) {
1155       SDep D = Succ;
1156       D.setSUnit(CopyToSU);
1157       AddPred(SuccSU, D);
1158       DelDeps.push_back(std::make_pair(SuccSU, Succ));
1159     }
1160     else {
1161       // Avoid scheduling the def-side copy before other successors. Otherwise
1162       // we could introduce another physreg interference on the copy and
1163       // continue inserting copies indefinitely.
1164       AddPred(SuccSU, SDep(CopyFromSU, SDep::Artificial));
1165     }
1166   }
1167   for (auto &DelDep : DelDeps)
1168     RemovePred(DelDep.first, DelDep.second);
1169 
1170   SDep FromDep(SU, SDep::Data, Reg);
1171   FromDep.setLatency(SU->Latency);
1172   AddPred(CopyFromSU, FromDep);
1173   SDep ToDep(CopyFromSU, SDep::Data, 0);
1174   ToDep.setLatency(CopyFromSU->Latency);
1175   AddPred(CopyToSU, ToDep);
1176 
1177   AvailableQueue->updateNode(SU);
1178   AvailableQueue->addNode(CopyFromSU);
1179   AvailableQueue->addNode(CopyToSU);
1180   Copies.push_back(CopyFromSU);
1181   Copies.push_back(CopyToSU);
1182 
1183   ++NumPRCopies;
1184 }
1185 
1186 /// getPhysicalRegisterVT - Returns the ValueType of the physical register
1187 /// definition of the specified node.
1188 /// FIXME: Move to SelectionDAG?
getPhysicalRegisterVT(SDNode * N,unsigned Reg,const TargetInstrInfo * TII)1189 static MVT getPhysicalRegisterVT(SDNode *N, unsigned Reg,
1190                                  const TargetInstrInfo *TII) {
1191   unsigned NumRes;
1192   if (N->getOpcode() == ISD::CopyFromReg) {
1193     // CopyFromReg has: "chain, Val, glue" so operand 1 gives the type.
1194     NumRes = 1;
1195   } else {
1196     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1197     assert(MCID.ImplicitDefs && "Physical reg def must be in implicit def list!");
1198     NumRes = MCID.getNumDefs();
1199     for (const MCPhysReg *ImpDef = MCID.getImplicitDefs(); *ImpDef; ++ImpDef) {
1200       if (Reg == *ImpDef)
1201         break;
1202       ++NumRes;
1203     }
1204   }
1205   return N->getSimpleValueType(NumRes);
1206 }
1207 
1208 /// CheckForLiveRegDef - Return true and update live register vector if the
1209 /// specified register def of the specified SUnit clobbers any "live" registers.
CheckForLiveRegDef(SUnit * SU,unsigned Reg,SUnit ** LiveRegDefs,SmallSet<unsigned,4> & RegAdded,SmallVectorImpl<unsigned> & LRegs,const TargetRegisterInfo * TRI)1210 static void CheckForLiveRegDef(SUnit *SU, unsigned Reg,
1211                                SUnit **LiveRegDefs,
1212                                SmallSet<unsigned, 4> &RegAdded,
1213                                SmallVectorImpl<unsigned> &LRegs,
1214                                const TargetRegisterInfo *TRI) {
1215   for (MCRegAliasIterator AliasI(Reg, TRI, true); AliasI.isValid(); ++AliasI) {
1216 
1217     // Check if Ref is live.
1218     if (!LiveRegDefs[*AliasI]) continue;
1219 
1220     // Allow multiple uses of the same def.
1221     if (LiveRegDefs[*AliasI] == SU) continue;
1222 
1223     // Add Reg to the set of interfering live regs.
1224     if (RegAdded.insert(*AliasI).second) {
1225       LRegs.push_back(*AliasI);
1226     }
1227   }
1228 }
1229 
1230 /// CheckForLiveRegDefMasked - Check for any live physregs that are clobbered
1231 /// by RegMask, and add them to LRegs.
CheckForLiveRegDefMasked(SUnit * SU,const uint32_t * RegMask,ArrayRef<SUnit * > LiveRegDefs,SmallSet<unsigned,4> & RegAdded,SmallVectorImpl<unsigned> & LRegs)1232 static void CheckForLiveRegDefMasked(SUnit *SU, const uint32_t *RegMask,
1233                                      ArrayRef<SUnit*> LiveRegDefs,
1234                                      SmallSet<unsigned, 4> &RegAdded,
1235                                      SmallVectorImpl<unsigned> &LRegs) {
1236   // Look at all live registers. Skip Reg0 and the special CallResource.
1237   for (unsigned i = 1, e = LiveRegDefs.size()-1; i != e; ++i) {
1238     if (!LiveRegDefs[i]) continue;
1239     if (LiveRegDefs[i] == SU) continue;
1240     if (!MachineOperand::clobbersPhysReg(RegMask, i)) continue;
1241     if (RegAdded.insert(i).second)
1242       LRegs.push_back(i);
1243   }
1244 }
1245 
1246 /// getNodeRegMask - Returns the register mask attached to an SDNode, if any.
getNodeRegMask(const SDNode * N)1247 static const uint32_t *getNodeRegMask(const SDNode *N) {
1248   for (const SDValue &Op : N->op_values())
1249     if (const auto *RegOp = dyn_cast<RegisterMaskSDNode>(Op.getNode()))
1250       return RegOp->getRegMask();
1251   return nullptr;
1252 }
1253 
1254 /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
1255 /// scheduling of the given node to satisfy live physical register dependencies.
1256 /// If the specific node is the last one that's available to schedule, do
1257 /// whatever is necessary (i.e. backtracking or cloning) to make it possible.
1258 bool ScheduleDAGRRList::
DelayForLiveRegsBottomUp(SUnit * SU,SmallVectorImpl<unsigned> & LRegs)1259 DelayForLiveRegsBottomUp(SUnit *SU, SmallVectorImpl<unsigned> &LRegs) {
1260   if (NumLiveRegs == 0)
1261     return false;
1262 
1263   SmallSet<unsigned, 4> RegAdded;
1264   // If this node would clobber any "live" register, then it's not ready.
1265   //
1266   // If SU is the currently live definition of the same register that it uses,
1267   // then we are free to schedule it.
1268   for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1269        I != E; ++I) {
1270     if (I->isAssignedRegDep() && LiveRegDefs[I->getReg()] != SU)
1271       CheckForLiveRegDef(I->getSUnit(), I->getReg(), LiveRegDefs.get(),
1272                          RegAdded, LRegs, TRI);
1273   }
1274 
1275   for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) {
1276     if (Node->getOpcode() == ISD::INLINEASM) {
1277       // Inline asm can clobber physical defs.
1278       unsigned NumOps = Node->getNumOperands();
1279       if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
1280         --NumOps;  // Ignore the glue operand.
1281 
1282       for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
1283         unsigned Flags =
1284           cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
1285         unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
1286 
1287         ++i; // Skip the ID value.
1288         if (InlineAsm::isRegDefKind(Flags) ||
1289             InlineAsm::isRegDefEarlyClobberKind(Flags) ||
1290             InlineAsm::isClobberKind(Flags)) {
1291           // Check for def of register or earlyclobber register.
1292           for (; NumVals; --NumVals, ++i) {
1293             unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
1294             if (TargetRegisterInfo::isPhysicalRegister(Reg))
1295               CheckForLiveRegDef(SU, Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI);
1296           }
1297         } else
1298           i += NumVals;
1299       }
1300       continue;
1301     }
1302 
1303     if (!Node->isMachineOpcode())
1304       continue;
1305     // If we're in the middle of scheduling a call, don't begin scheduling
1306     // another call. Also, don't allow any physical registers to be live across
1307     // the call.
1308     if (Node->getMachineOpcode() == (unsigned)TII->getCallFrameDestroyOpcode()) {
1309       // Check the special calling-sequence resource.
1310       unsigned CallResource = TRI->getNumRegs();
1311       if (LiveRegDefs[CallResource]) {
1312         SDNode *Gen = LiveRegGens[CallResource]->getNode();
1313         while (SDNode *Glued = Gen->getGluedNode())
1314           Gen = Glued;
1315         if (!IsChainDependent(Gen, Node, 0, TII) &&
1316             RegAdded.insert(CallResource).second)
1317           LRegs.push_back(CallResource);
1318       }
1319     }
1320     if (const uint32_t *RegMask = getNodeRegMask(Node))
1321       CheckForLiveRegDefMasked(SU, RegMask,
1322                                makeArrayRef(LiveRegDefs.get(), TRI->getNumRegs()),
1323                                RegAdded, LRegs);
1324 
1325     const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode());
1326     if (!MCID.ImplicitDefs)
1327       continue;
1328     for (const MCPhysReg *Reg = MCID.getImplicitDefs(); *Reg; ++Reg)
1329       CheckForLiveRegDef(SU, *Reg, LiveRegDefs.get(), RegAdded, LRegs, TRI);
1330   }
1331 
1332   return !LRegs.empty();
1333 }
1334 
releaseInterferences(unsigned Reg)1335 void ScheduleDAGRRList::releaseInterferences(unsigned Reg) {
1336   // Add the nodes that aren't ready back onto the available list.
1337   for (unsigned i = Interferences.size(); i > 0; --i) {
1338     SUnit *SU = Interferences[i-1];
1339     LRegsMapT::iterator LRegsPos = LRegsMap.find(SU);
1340     if (Reg) {
1341       SmallVectorImpl<unsigned> &LRegs = LRegsPos->second;
1342       if (std::find(LRegs.begin(), LRegs.end(), Reg) == LRegs.end())
1343         continue;
1344     }
1345     SU->isPending = false;
1346     // The interfering node may no longer be available due to backtracking.
1347     // Furthermore, it may have been made available again, in which case it is
1348     // now already in the AvailableQueue.
1349     if (SU->isAvailable && !SU->NodeQueueId) {
1350       DEBUG(dbgs() << "    Repushing SU #" << SU->NodeNum << '\n');
1351       AvailableQueue->push(SU);
1352     }
1353     if (i < Interferences.size())
1354       Interferences[i-1] = Interferences.back();
1355     Interferences.pop_back();
1356     LRegsMap.erase(LRegsPos);
1357   }
1358 }
1359 
1360 /// Return a node that can be scheduled in this cycle. Requirements:
1361 /// (1) Ready: latency has been satisfied
1362 /// (2) No Hazards: resources are available
1363 /// (3) No Interferences: may unschedule to break register interferences.
PickNodeToScheduleBottomUp()1364 SUnit *ScheduleDAGRRList::PickNodeToScheduleBottomUp() {
1365   SUnit *CurSU = AvailableQueue->empty() ? nullptr : AvailableQueue->pop();
1366   while (CurSU) {
1367     SmallVector<unsigned, 4> LRegs;
1368     if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
1369       break;
1370     DEBUG(dbgs() << "    Interfering reg " <<
1371           (LRegs[0] == TRI->getNumRegs() ? "CallResource"
1372            : TRI->getName(LRegs[0]))
1373            << " SU #" << CurSU->NodeNum << '\n');
1374     std::pair<LRegsMapT::iterator, bool> LRegsPair =
1375       LRegsMap.insert(std::make_pair(CurSU, LRegs));
1376     if (LRegsPair.second) {
1377       CurSU->isPending = true;  // This SU is not in AvailableQueue right now.
1378       Interferences.push_back(CurSU);
1379     }
1380     else {
1381       assert(CurSU->isPending && "Interferences are pending");
1382       // Update the interference with current live regs.
1383       LRegsPair.first->second = LRegs;
1384     }
1385     CurSU = AvailableQueue->pop();
1386   }
1387   if (CurSU)
1388     return CurSU;
1389 
1390   // All candidates are delayed due to live physical reg dependencies.
1391   // Try backtracking, code duplication, or inserting cross class copies
1392   // to resolve it.
1393   for (SUnit *TrySU : Interferences) {
1394     SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
1395 
1396     // Try unscheduling up to the point where it's safe to schedule
1397     // this node.
1398     SUnit *BtSU = nullptr;
1399     unsigned LiveCycle = UINT_MAX;
1400     for (unsigned Reg : LRegs) {
1401       if (LiveRegGens[Reg]->getHeight() < LiveCycle) {
1402         BtSU = LiveRegGens[Reg];
1403         LiveCycle = BtSU->getHeight();
1404       }
1405     }
1406     if (!WillCreateCycle(TrySU, BtSU))  {
1407       // BacktrackBottomUp mutates Interferences!
1408       BacktrackBottomUp(TrySU, BtSU);
1409 
1410       // Force the current node to be scheduled before the node that
1411       // requires the physical reg dep.
1412       if (BtSU->isAvailable) {
1413         BtSU->isAvailable = false;
1414         if (!BtSU->isPending)
1415           AvailableQueue->remove(BtSU);
1416       }
1417       DEBUG(dbgs() << "ARTIFICIAL edge from SU(" << BtSU->NodeNum << ") to SU("
1418             << TrySU->NodeNum << ")\n");
1419       AddPred(TrySU, SDep(BtSU, SDep::Artificial));
1420 
1421       // If one or more successors has been unscheduled, then the current
1422       // node is no longer available.
1423       if (!TrySU->isAvailable || !TrySU->NodeQueueId)
1424         CurSU = AvailableQueue->pop();
1425       else {
1426         // Available and in AvailableQueue
1427         AvailableQueue->remove(TrySU);
1428         CurSU = TrySU;
1429       }
1430       // Interferences has been mutated. We must break.
1431       break;
1432     }
1433   }
1434 
1435   if (!CurSU) {
1436     // Can't backtrack. If it's too expensive to copy the value, then try
1437     // duplicate the nodes that produces these "too expensive to copy"
1438     // values to break the dependency. In case even that doesn't work,
1439     // insert cross class copies.
1440     // If it's not too expensive, i.e. cost != -1, issue copies.
1441     SUnit *TrySU = Interferences[0];
1442     SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
1443     assert(LRegs.size() == 1 && "Can't handle this yet!");
1444     unsigned Reg = LRegs[0];
1445     SUnit *LRDef = LiveRegDefs[Reg];
1446     MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
1447     const TargetRegisterClass *RC =
1448       TRI->getMinimalPhysRegClass(Reg, VT);
1449     const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
1450 
1451     // If cross copy register class is the same as RC, then it must be possible
1452     // copy the value directly. Do not try duplicate the def.
1453     // If cross copy register class is not the same as RC, then it's possible to
1454     // copy the value but it require cross register class copies and it is
1455     // expensive.
1456     // If cross copy register class is null, then it's not possible to copy
1457     // the value at all.
1458     SUnit *NewDef = nullptr;
1459     if (DestRC != RC) {
1460       NewDef = CopyAndMoveSuccessors(LRDef);
1461       if (!DestRC && !NewDef)
1462         report_fatal_error("Can't handle live physical register dependency!");
1463     }
1464     if (!NewDef) {
1465       // Issue copies, these can be expensive cross register class copies.
1466       SmallVector<SUnit*, 2> Copies;
1467       InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
1468       DEBUG(dbgs() << "    Adding an edge from SU #" << TrySU->NodeNum
1469             << " to SU #" << Copies.front()->NodeNum << "\n");
1470       AddPred(TrySU, SDep(Copies.front(), SDep::Artificial));
1471       NewDef = Copies.back();
1472     }
1473 
1474     DEBUG(dbgs() << "    Adding an edge from SU #" << NewDef->NodeNum
1475           << " to SU #" << TrySU->NodeNum << "\n");
1476     LiveRegDefs[Reg] = NewDef;
1477     AddPred(NewDef, SDep(TrySU, SDep::Artificial));
1478     TrySU->isAvailable = false;
1479     CurSU = NewDef;
1480   }
1481   assert(CurSU && "Unable to resolve live physical register dependencies!");
1482   return CurSU;
1483 }
1484 
1485 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
1486 /// schedulers.
ListScheduleBottomUp()1487 void ScheduleDAGRRList::ListScheduleBottomUp() {
1488   // Release any predecessors of the special Exit node.
1489   ReleasePredecessors(&ExitSU);
1490 
1491   // Add root to Available queue.
1492   if (!SUnits.empty()) {
1493     SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()];
1494     assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!");
1495     RootSU->isAvailable = true;
1496     AvailableQueue->push(RootSU);
1497   }
1498 
1499   // While Available queue is not empty, grab the node with the highest
1500   // priority. If it is not ready put it back.  Schedule the node.
1501   Sequence.reserve(SUnits.size());
1502   while (!AvailableQueue->empty() || !Interferences.empty()) {
1503     DEBUG(dbgs() << "\nExamining Available:\n";
1504           AvailableQueue->dump(this));
1505 
1506     // Pick the best node to schedule taking all constraints into
1507     // consideration.
1508     SUnit *SU = PickNodeToScheduleBottomUp();
1509 
1510     AdvancePastStalls(SU);
1511 
1512     ScheduleNodeBottomUp(SU);
1513 
1514     while (AvailableQueue->empty() && !PendingQueue.empty()) {
1515       // Advance the cycle to free resources. Skip ahead to the next ready SU.
1516       assert(MinAvailableCycle < UINT_MAX && "MinAvailableCycle uninitialized");
1517       AdvanceToCycle(std::max(CurCycle + 1, MinAvailableCycle));
1518     }
1519   }
1520 
1521   // Reverse the order if it is bottom up.
1522   std::reverse(Sequence.begin(), Sequence.end());
1523 
1524 #ifndef NDEBUG
1525   VerifyScheduledSequence(/*isBottomUp=*/true);
1526 #endif
1527 }
1528 
1529 //===----------------------------------------------------------------------===//
1530 //                RegReductionPriorityQueue Definition
1531 //===----------------------------------------------------------------------===//
1532 //
1533 // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers
1534 // to reduce register pressure.
1535 //
1536 namespace {
1537 class RegReductionPQBase;
1538 
1539 struct queue_sort : public std::binary_function<SUnit*, SUnit*, bool> {
isReady__anon7133554f0211::queue_sort1540   bool isReady(SUnit* SU, unsigned CurCycle) const { return true; }
1541 };
1542 
1543 #ifndef NDEBUG
1544 template<class SF>
1545 struct reverse_sort : public queue_sort {
1546   SF &SortFunc;
reverse_sort__anon7133554f0211::reverse_sort1547   reverse_sort(SF &sf) : SortFunc(sf) {}
1548 
operator ()__anon7133554f0211::reverse_sort1549   bool operator()(SUnit* left, SUnit* right) const {
1550     // reverse left/right rather than simply !SortFunc(left, right)
1551     // to expose different paths in the comparison logic.
1552     return SortFunc(right, left);
1553   }
1554 };
1555 #endif // NDEBUG
1556 
1557 /// bu_ls_rr_sort - Priority function for bottom up register pressure
1558 // reduction scheduler.
1559 struct bu_ls_rr_sort : public queue_sort {
1560   enum {
1561     IsBottomUp = true,
1562     HasReadyFilter = false
1563   };
1564 
1565   RegReductionPQBase *SPQ;
bu_ls_rr_sort__anon7133554f0211::bu_ls_rr_sort1566   bu_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
1567 
1568   bool operator()(SUnit* left, SUnit* right) const;
1569 };
1570 
1571 // src_ls_rr_sort - Priority function for source order scheduler.
1572 struct src_ls_rr_sort : public queue_sort {
1573   enum {
1574     IsBottomUp = true,
1575     HasReadyFilter = false
1576   };
1577 
1578   RegReductionPQBase *SPQ;
src_ls_rr_sort__anon7133554f0211::src_ls_rr_sort1579   src_ls_rr_sort(RegReductionPQBase *spq)
1580     : SPQ(spq) {}
1581 
1582   bool operator()(SUnit* left, SUnit* right) const;
1583 };
1584 
1585 // hybrid_ls_rr_sort - Priority function for hybrid scheduler.
1586 struct hybrid_ls_rr_sort : public queue_sort {
1587   enum {
1588     IsBottomUp = true,
1589     HasReadyFilter = false
1590   };
1591 
1592   RegReductionPQBase *SPQ;
hybrid_ls_rr_sort__anon7133554f0211::hybrid_ls_rr_sort1593   hybrid_ls_rr_sort(RegReductionPQBase *spq)
1594     : SPQ(spq) {}
1595 
1596   bool isReady(SUnit *SU, unsigned CurCycle) const;
1597 
1598   bool operator()(SUnit* left, SUnit* right) const;
1599 };
1600 
1601 // ilp_ls_rr_sort - Priority function for ILP (instruction level parallelism)
1602 // scheduler.
1603 struct ilp_ls_rr_sort : public queue_sort {
1604   enum {
1605     IsBottomUp = true,
1606     HasReadyFilter = false
1607   };
1608 
1609   RegReductionPQBase *SPQ;
ilp_ls_rr_sort__anon7133554f0211::ilp_ls_rr_sort1610   ilp_ls_rr_sort(RegReductionPQBase *spq)
1611     : SPQ(spq) {}
1612 
1613   bool isReady(SUnit *SU, unsigned CurCycle) const;
1614 
1615   bool operator()(SUnit* left, SUnit* right) const;
1616 };
1617 
1618 class RegReductionPQBase : public SchedulingPriorityQueue {
1619 protected:
1620   std::vector<SUnit*> Queue;
1621   unsigned CurQueueId;
1622   bool TracksRegPressure;
1623   bool SrcOrder;
1624 
1625   // SUnits - The SUnits for the current graph.
1626   std::vector<SUnit> *SUnits;
1627 
1628   MachineFunction &MF;
1629   const TargetInstrInfo *TII;
1630   const TargetRegisterInfo *TRI;
1631   const TargetLowering *TLI;
1632   ScheduleDAGRRList *scheduleDAG;
1633 
1634   // SethiUllmanNumbers - The SethiUllman number for each node.
1635   std::vector<unsigned> SethiUllmanNumbers;
1636 
1637   /// RegPressure - Tracking current reg pressure per register class.
1638   ///
1639   std::vector<unsigned> RegPressure;
1640 
1641   /// RegLimit - Tracking the number of allocatable registers per register
1642   /// class.
1643   std::vector<unsigned> RegLimit;
1644 
1645 public:
RegReductionPQBase(MachineFunction & mf,bool hasReadyFilter,bool tracksrp,bool srcorder,const TargetInstrInfo * tii,const TargetRegisterInfo * tri,const TargetLowering * tli)1646   RegReductionPQBase(MachineFunction &mf,
1647                      bool hasReadyFilter,
1648                      bool tracksrp,
1649                      bool srcorder,
1650                      const TargetInstrInfo *tii,
1651                      const TargetRegisterInfo *tri,
1652                      const TargetLowering *tli)
1653     : SchedulingPriorityQueue(hasReadyFilter),
1654       CurQueueId(0), TracksRegPressure(tracksrp), SrcOrder(srcorder),
1655       MF(mf), TII(tii), TRI(tri), TLI(tli), scheduleDAG(nullptr) {
1656     if (TracksRegPressure) {
1657       unsigned NumRC = TRI->getNumRegClasses();
1658       RegLimit.resize(NumRC);
1659       RegPressure.resize(NumRC);
1660       std::fill(RegLimit.begin(), RegLimit.end(), 0);
1661       std::fill(RegPressure.begin(), RegPressure.end(), 0);
1662       for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
1663              E = TRI->regclass_end(); I != E; ++I)
1664         RegLimit[(*I)->getID()] = tri->getRegPressureLimit(*I, MF);
1665     }
1666   }
1667 
setScheduleDAG(ScheduleDAGRRList * scheduleDag)1668   void setScheduleDAG(ScheduleDAGRRList *scheduleDag) {
1669     scheduleDAG = scheduleDag;
1670   }
1671 
getHazardRec()1672   ScheduleHazardRecognizer* getHazardRec() {
1673     return scheduleDAG->getHazardRec();
1674   }
1675 
1676   void initNodes(std::vector<SUnit> &sunits) override;
1677 
1678   void addNode(const SUnit *SU) override;
1679 
1680   void updateNode(const SUnit *SU) override;
1681 
releaseState()1682   void releaseState() override {
1683     SUnits = nullptr;
1684     SethiUllmanNumbers.clear();
1685     std::fill(RegPressure.begin(), RegPressure.end(), 0);
1686   }
1687 
1688   unsigned getNodePriority(const SUnit *SU) const;
1689 
getNodeOrdering(const SUnit * SU) const1690   unsigned getNodeOrdering(const SUnit *SU) const {
1691     if (!SU->getNode()) return 0;
1692 
1693     return SU->getNode()->getIROrder();
1694   }
1695 
empty() const1696   bool empty() const override { return Queue.empty(); }
1697 
push(SUnit * U)1698   void push(SUnit *U) override {
1699     assert(!U->NodeQueueId && "Node in the queue already");
1700     U->NodeQueueId = ++CurQueueId;
1701     Queue.push_back(U);
1702   }
1703 
remove(SUnit * SU)1704   void remove(SUnit *SU) override {
1705     assert(!Queue.empty() && "Queue is empty!");
1706     assert(SU->NodeQueueId != 0 && "Not in queue!");
1707     std::vector<SUnit *>::iterator I = std::find(Queue.begin(), Queue.end(),
1708                                                  SU);
1709     if (I != std::prev(Queue.end()))
1710       std::swap(*I, Queue.back());
1711     Queue.pop_back();
1712     SU->NodeQueueId = 0;
1713   }
1714 
tracksRegPressure() const1715   bool tracksRegPressure() const override { return TracksRegPressure; }
1716 
1717   void dumpRegPressure() const;
1718 
1719   bool HighRegPressure(const SUnit *SU) const;
1720 
1721   bool MayReduceRegPressure(SUnit *SU) const;
1722 
1723   int RegPressureDiff(SUnit *SU, unsigned &LiveUses) const;
1724 
1725   void scheduledNode(SUnit *SU) override;
1726 
1727   void unscheduledNode(SUnit *SU) override;
1728 
1729 protected:
1730   bool canClobber(const SUnit *SU, const SUnit *Op);
1731   void AddPseudoTwoAddrDeps();
1732   void PrescheduleNodesWithMultipleUses();
1733   void CalculateSethiUllmanNumbers();
1734 };
1735 
1736 template<class SF>
popFromQueueImpl(std::vector<SUnit * > & Q,SF & Picker)1737 static SUnit *popFromQueueImpl(std::vector<SUnit*> &Q, SF &Picker) {
1738   std::vector<SUnit *>::iterator Best = Q.begin();
1739   for (std::vector<SUnit *>::iterator I = std::next(Q.begin()),
1740          E = Q.end(); I != E; ++I)
1741     if (Picker(*Best, *I))
1742       Best = I;
1743   SUnit *V = *Best;
1744   if (Best != std::prev(Q.end()))
1745     std::swap(*Best, Q.back());
1746   Q.pop_back();
1747   return V;
1748 }
1749 
1750 template<class SF>
popFromQueue(std::vector<SUnit * > & Q,SF & Picker,ScheduleDAG * DAG)1751 SUnit *popFromQueue(std::vector<SUnit*> &Q, SF &Picker, ScheduleDAG *DAG) {
1752 #ifndef NDEBUG
1753   if (DAG->StressSched) {
1754     reverse_sort<SF> RPicker(Picker);
1755     return popFromQueueImpl(Q, RPicker);
1756   }
1757 #endif
1758   (void)DAG;
1759   return popFromQueueImpl(Q, Picker);
1760 }
1761 
1762 template<class SF>
1763 class RegReductionPriorityQueue : public RegReductionPQBase {
1764   SF Picker;
1765 
1766 public:
RegReductionPriorityQueue(MachineFunction & mf,bool tracksrp,bool srcorder,const TargetInstrInfo * tii,const TargetRegisterInfo * tri,const TargetLowering * tli)1767   RegReductionPriorityQueue(MachineFunction &mf,
1768                             bool tracksrp,
1769                             bool srcorder,
1770                             const TargetInstrInfo *tii,
1771                             const TargetRegisterInfo *tri,
1772                             const TargetLowering *tli)
1773     : RegReductionPQBase(mf, SF::HasReadyFilter, tracksrp, srcorder,
1774                          tii, tri, tli),
1775       Picker(this) {}
1776 
isBottomUp() const1777   bool isBottomUp() const override { return SF::IsBottomUp; }
1778 
isReady(SUnit * U) const1779   bool isReady(SUnit *U) const override {
1780     return Picker.HasReadyFilter && Picker.isReady(U, getCurCycle());
1781   }
1782 
pop()1783   SUnit *pop() override {
1784     if (Queue.empty()) return nullptr;
1785 
1786     SUnit *V = popFromQueue(Queue, Picker, scheduleDAG);
1787     V->NodeQueueId = 0;
1788     return V;
1789   }
1790 
1791 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dump(ScheduleDAG * DAG) const1792   void dump(ScheduleDAG *DAG) const override {
1793     // Emulate pop() without clobbering NodeQueueIds.
1794     std::vector<SUnit*> DumpQueue = Queue;
1795     SF DumpPicker = Picker;
1796     while (!DumpQueue.empty()) {
1797       SUnit *SU = popFromQueue(DumpQueue, DumpPicker, scheduleDAG);
1798       dbgs() << "Height " << SU->getHeight() << ": ";
1799       SU->dump(DAG);
1800     }
1801   }
1802 #endif
1803 };
1804 
1805 typedef RegReductionPriorityQueue<bu_ls_rr_sort>
1806 BURegReductionPriorityQueue;
1807 
1808 typedef RegReductionPriorityQueue<src_ls_rr_sort>
1809 SrcRegReductionPriorityQueue;
1810 
1811 typedef RegReductionPriorityQueue<hybrid_ls_rr_sort>
1812 HybridBURRPriorityQueue;
1813 
1814 typedef RegReductionPriorityQueue<ilp_ls_rr_sort>
1815 ILPBURRPriorityQueue;
1816 } // end anonymous namespace
1817 
1818 //===----------------------------------------------------------------------===//
1819 //           Static Node Priority for Register Pressure Reduction
1820 //===----------------------------------------------------------------------===//
1821 
1822 // Check for special nodes that bypass scheduling heuristics.
1823 // Currently this pushes TokenFactor nodes down, but may be used for other
1824 // pseudo-ops as well.
1825 //
1826 // Return -1 to schedule right above left, 1 for left above right.
1827 // Return 0 if no bias exists.
checkSpecialNodes(const SUnit * left,const SUnit * right)1828 static int checkSpecialNodes(const SUnit *left, const SUnit *right) {
1829   bool LSchedLow = left->isScheduleLow;
1830   bool RSchedLow = right->isScheduleLow;
1831   if (LSchedLow != RSchedLow)
1832     return LSchedLow < RSchedLow ? 1 : -1;
1833   return 0;
1834 }
1835 
1836 /// CalcNodeSethiUllmanNumber - Compute Sethi Ullman number.
1837 /// Smaller number is the higher priority.
1838 static unsigned
CalcNodeSethiUllmanNumber(const SUnit * SU,std::vector<unsigned> & SUNumbers)1839 CalcNodeSethiUllmanNumber(const SUnit *SU, std::vector<unsigned> &SUNumbers) {
1840   unsigned &SethiUllmanNumber = SUNumbers[SU->NodeNum];
1841   if (SethiUllmanNumber != 0)
1842     return SethiUllmanNumber;
1843 
1844   unsigned Extra = 0;
1845   for (const SDep &Pred : SU->Preds) {
1846     if (Pred.isCtrl()) continue;  // ignore chain preds
1847     SUnit *PredSU = Pred.getSUnit();
1848     unsigned PredSethiUllman = CalcNodeSethiUllmanNumber(PredSU, SUNumbers);
1849     if (PredSethiUllman > SethiUllmanNumber) {
1850       SethiUllmanNumber = PredSethiUllman;
1851       Extra = 0;
1852     } else if (PredSethiUllman == SethiUllmanNumber)
1853       ++Extra;
1854   }
1855 
1856   SethiUllmanNumber += Extra;
1857 
1858   if (SethiUllmanNumber == 0)
1859     SethiUllmanNumber = 1;
1860 
1861   return SethiUllmanNumber;
1862 }
1863 
1864 /// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all
1865 /// scheduling units.
CalculateSethiUllmanNumbers()1866 void RegReductionPQBase::CalculateSethiUllmanNumbers() {
1867   SethiUllmanNumbers.assign(SUnits->size(), 0);
1868 
1869   for (const SUnit &SU : *SUnits)
1870     CalcNodeSethiUllmanNumber(&SU, SethiUllmanNumbers);
1871 }
1872 
addNode(const SUnit * SU)1873 void RegReductionPQBase::addNode(const SUnit *SU) {
1874   unsigned SUSize = SethiUllmanNumbers.size();
1875   if (SUnits->size() > SUSize)
1876     SethiUllmanNumbers.resize(SUSize*2, 0);
1877   CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
1878 }
1879 
updateNode(const SUnit * SU)1880 void RegReductionPQBase::updateNode(const SUnit *SU) {
1881   SethiUllmanNumbers[SU->NodeNum] = 0;
1882   CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
1883 }
1884 
1885 // Lower priority means schedule further down. For bottom-up scheduling, lower
1886 // priority SUs are scheduled before higher priority SUs.
getNodePriority(const SUnit * SU) const1887 unsigned RegReductionPQBase::getNodePriority(const SUnit *SU) const {
1888   assert(SU->NodeNum < SethiUllmanNumbers.size());
1889   unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0;
1890   if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
1891     // CopyToReg should be close to its uses to facilitate coalescing and
1892     // avoid spilling.
1893     return 0;
1894   if (Opc == TargetOpcode::EXTRACT_SUBREG ||
1895       Opc == TargetOpcode::SUBREG_TO_REG ||
1896       Opc == TargetOpcode::INSERT_SUBREG)
1897     // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be
1898     // close to their uses to facilitate coalescing.
1899     return 0;
1900   if (SU->NumSuccs == 0 && SU->NumPreds != 0)
1901     // If SU does not have a register use, i.e. it doesn't produce a value
1902     // that would be consumed (e.g. store), then it terminates a chain of
1903     // computation.  Give it a large SethiUllman number so it will be
1904     // scheduled right before its predecessors that it doesn't lengthen
1905     // their live ranges.
1906     return 0xffff;
1907   if (SU->NumPreds == 0 && SU->NumSuccs != 0)
1908     // If SU does not have a register def, schedule it close to its uses
1909     // because it does not lengthen any live ranges.
1910     return 0;
1911 #if 1
1912   return SethiUllmanNumbers[SU->NodeNum];
1913 #else
1914   unsigned Priority = SethiUllmanNumbers[SU->NodeNum];
1915   if (SU->isCallOp) {
1916     // FIXME: This assumes all of the defs are used as call operands.
1917     int NP = (int)Priority - SU->getNode()->getNumValues();
1918     return (NP > 0) ? NP : 0;
1919   }
1920   return Priority;
1921 #endif
1922 }
1923 
1924 //===----------------------------------------------------------------------===//
1925 //                     Register Pressure Tracking
1926 //===----------------------------------------------------------------------===//
1927 
dumpRegPressure() const1928 void RegReductionPQBase::dumpRegPressure() const {
1929 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1930   for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
1931          E = TRI->regclass_end(); I != E; ++I) {
1932     const TargetRegisterClass *RC = *I;
1933     unsigned Id = RC->getID();
1934     unsigned RP = RegPressure[Id];
1935     if (!RP) continue;
1936     DEBUG(dbgs() << TRI->getRegClassName(RC) << ": " << RP << " / "
1937           << RegLimit[Id] << '\n');
1938   }
1939 #endif
1940 }
1941 
HighRegPressure(const SUnit * SU) const1942 bool RegReductionPQBase::HighRegPressure(const SUnit *SU) const {
1943   if (!TLI)
1944     return false;
1945 
1946   for (const SDep &Pred : SU->Preds) {
1947     if (Pred.isCtrl())
1948       continue;
1949     SUnit *PredSU = Pred.getSUnit();
1950     // NumRegDefsLeft is zero when enough uses of this node have been scheduled
1951     // to cover the number of registers defined (they are all live).
1952     if (PredSU->NumRegDefsLeft == 0) {
1953       continue;
1954     }
1955     for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
1956          RegDefPos.IsValid(); RegDefPos.Advance()) {
1957       unsigned RCId, Cost;
1958       GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
1959 
1960       if ((RegPressure[RCId] + Cost) >= RegLimit[RCId])
1961         return true;
1962     }
1963   }
1964   return false;
1965 }
1966 
MayReduceRegPressure(SUnit * SU) const1967 bool RegReductionPQBase::MayReduceRegPressure(SUnit *SU) const {
1968   const SDNode *N = SU->getNode();
1969 
1970   if (!N->isMachineOpcode() || !SU->NumSuccs)
1971     return false;
1972 
1973   unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
1974   for (unsigned i = 0; i != NumDefs; ++i) {
1975     MVT VT = N->getSimpleValueType(i);
1976     if (!N->hasAnyUseOfValue(i))
1977       continue;
1978     unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1979     if (RegPressure[RCId] >= RegLimit[RCId])
1980       return true;
1981   }
1982   return false;
1983 }
1984 
1985 // Compute the register pressure contribution by this instruction by count up
1986 // for uses that are not live and down for defs. Only count register classes
1987 // that are already under high pressure. As a side effect, compute the number of
1988 // uses of registers that are already live.
1989 //
1990 // FIXME: This encompasses the logic in HighRegPressure and MayReduceRegPressure
1991 // so could probably be factored.
RegPressureDiff(SUnit * SU,unsigned & LiveUses) const1992 int RegReductionPQBase::RegPressureDiff(SUnit *SU, unsigned &LiveUses) const {
1993   LiveUses = 0;
1994   int PDiff = 0;
1995   for (const SDep &Pred : SU->Preds) {
1996     if (Pred.isCtrl())
1997       continue;
1998     SUnit *PredSU = Pred.getSUnit();
1999     // NumRegDefsLeft is zero when enough uses of this node have been scheduled
2000     // to cover the number of registers defined (they are all live).
2001     if (PredSU->NumRegDefsLeft == 0) {
2002       if (PredSU->getNode()->isMachineOpcode())
2003         ++LiveUses;
2004       continue;
2005     }
2006     for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
2007          RegDefPos.IsValid(); RegDefPos.Advance()) {
2008       MVT VT = RegDefPos.GetValue();
2009       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2010       if (RegPressure[RCId] >= RegLimit[RCId])
2011         ++PDiff;
2012     }
2013   }
2014   const SDNode *N = SU->getNode();
2015 
2016   if (!N || !N->isMachineOpcode() || !SU->NumSuccs)
2017     return PDiff;
2018 
2019   unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2020   for (unsigned i = 0; i != NumDefs; ++i) {
2021     MVT VT = N->getSimpleValueType(i);
2022     if (!N->hasAnyUseOfValue(i))
2023       continue;
2024     unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2025     if (RegPressure[RCId] >= RegLimit[RCId])
2026       --PDiff;
2027   }
2028   return PDiff;
2029 }
2030 
scheduledNode(SUnit * SU)2031 void RegReductionPQBase::scheduledNode(SUnit *SU) {
2032   if (!TracksRegPressure)
2033     return;
2034 
2035   if (!SU->getNode())
2036     return;
2037 
2038   for (const SDep &Pred : SU->Preds) {
2039     if (Pred.isCtrl())
2040       continue;
2041     SUnit *PredSU = Pred.getSUnit();
2042     // NumRegDefsLeft is zero when enough uses of this node have been scheduled
2043     // to cover the number of registers defined (they are all live).
2044     if (PredSU->NumRegDefsLeft == 0) {
2045       continue;
2046     }
2047     // FIXME: The ScheduleDAG currently loses information about which of a
2048     // node's values is consumed by each dependence. Consequently, if the node
2049     // defines multiple register classes, we don't know which to pressurize
2050     // here. Instead the following loop consumes the register defs in an
2051     // arbitrary order. At least it handles the common case of clustered loads
2052     // to the same class. For precise liveness, each SDep needs to indicate the
2053     // result number. But that tightly couples the ScheduleDAG with the
2054     // SelectionDAG making updates tricky. A simpler hack would be to attach a
2055     // value type or register class to SDep.
2056     //
2057     // The most important aspect of register tracking is balancing the increase
2058     // here with the reduction further below. Note that this SU may use multiple
2059     // defs in PredSU. The can't be determined here, but we've already
2060     // compensated by reducing NumRegDefsLeft in PredSU during
2061     // ScheduleDAGSDNodes::AddSchedEdges.
2062     --PredSU->NumRegDefsLeft;
2063     unsigned SkipRegDefs = PredSU->NumRegDefsLeft;
2064     for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
2065          RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
2066       if (SkipRegDefs)
2067         continue;
2068 
2069       unsigned RCId, Cost;
2070       GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
2071       RegPressure[RCId] += Cost;
2072       break;
2073     }
2074   }
2075 
2076   // We should have this assert, but there may be dead SDNodes that never
2077   // materialize as SUnits, so they don't appear to generate liveness.
2078   //assert(SU->NumRegDefsLeft == 0 && "not all regdefs have scheduled uses");
2079   int SkipRegDefs = (int)SU->NumRegDefsLeft;
2080   for (ScheduleDAGSDNodes::RegDefIter RegDefPos(SU, scheduleDAG);
2081        RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
2082     if (SkipRegDefs > 0)
2083       continue;
2084     unsigned RCId, Cost;
2085     GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
2086     if (RegPressure[RCId] < Cost) {
2087       // Register pressure tracking is imprecise. This can happen. But we try
2088       // hard not to let it happen because it likely results in poor scheduling.
2089       DEBUG(dbgs() << "  SU(" << SU->NodeNum << ") has too many regdefs\n");
2090       RegPressure[RCId] = 0;
2091     }
2092     else {
2093       RegPressure[RCId] -= Cost;
2094     }
2095   }
2096   dumpRegPressure();
2097 }
2098 
unscheduledNode(SUnit * SU)2099 void RegReductionPQBase::unscheduledNode(SUnit *SU) {
2100   if (!TracksRegPressure)
2101     return;
2102 
2103   const SDNode *N = SU->getNode();
2104   if (!N) return;
2105 
2106   if (!N->isMachineOpcode()) {
2107     if (N->getOpcode() != ISD::CopyToReg)
2108       return;
2109   } else {
2110     unsigned Opc = N->getMachineOpcode();
2111     if (Opc == TargetOpcode::EXTRACT_SUBREG ||
2112         Opc == TargetOpcode::INSERT_SUBREG ||
2113         Opc == TargetOpcode::SUBREG_TO_REG ||
2114         Opc == TargetOpcode::REG_SEQUENCE ||
2115         Opc == TargetOpcode::IMPLICIT_DEF)
2116       return;
2117   }
2118 
2119   for (const SDep &Pred : SU->Preds) {
2120     if (Pred.isCtrl())
2121       continue;
2122     SUnit *PredSU = Pred.getSUnit();
2123     // NumSuccsLeft counts all deps. Don't compare it with NumSuccs which only
2124     // counts data deps.
2125     if (PredSU->NumSuccsLeft != PredSU->Succs.size())
2126       continue;
2127     const SDNode *PN = PredSU->getNode();
2128     if (!PN->isMachineOpcode()) {
2129       if (PN->getOpcode() == ISD::CopyFromReg) {
2130         MVT VT = PN->getSimpleValueType(0);
2131         unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2132         RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2133       }
2134       continue;
2135     }
2136     unsigned POpc = PN->getMachineOpcode();
2137     if (POpc == TargetOpcode::IMPLICIT_DEF)
2138       continue;
2139     if (POpc == TargetOpcode::EXTRACT_SUBREG ||
2140         POpc == TargetOpcode::INSERT_SUBREG ||
2141         POpc == TargetOpcode::SUBREG_TO_REG) {
2142       MVT VT = PN->getSimpleValueType(0);
2143       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2144       RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2145       continue;
2146     }
2147     unsigned NumDefs = TII->get(PN->getMachineOpcode()).getNumDefs();
2148     for (unsigned i = 0; i != NumDefs; ++i) {
2149       MVT VT = PN->getSimpleValueType(i);
2150       if (!PN->hasAnyUseOfValue(i))
2151         continue;
2152       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2153       if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT))
2154         // Register pressure tracking is imprecise. This can happen.
2155         RegPressure[RCId] = 0;
2156       else
2157         RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
2158     }
2159   }
2160 
2161   // Check for isMachineOpcode() as PrescheduleNodesWithMultipleUses()
2162   // may transfer data dependencies to CopyToReg.
2163   if (SU->NumSuccs && N->isMachineOpcode()) {
2164     unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2165     for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
2166       MVT VT = N->getSimpleValueType(i);
2167       if (VT == MVT::Glue || VT == MVT::Other)
2168         continue;
2169       if (!N->hasAnyUseOfValue(i))
2170         continue;
2171       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2172       RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2173     }
2174   }
2175 
2176   dumpRegPressure();
2177 }
2178 
2179 //===----------------------------------------------------------------------===//
2180 //           Dynamic Node Priority for Register Pressure Reduction
2181 //===----------------------------------------------------------------------===//
2182 
2183 /// closestSucc - Returns the scheduled cycle of the successor which is
2184 /// closest to the current cycle.
closestSucc(const SUnit * SU)2185 static unsigned closestSucc(const SUnit *SU) {
2186   unsigned MaxHeight = 0;
2187   for (const SDep &Succ : SU->Succs) {
2188     if (Succ.isCtrl()) continue;  // ignore chain succs
2189     unsigned Height = Succ.getSUnit()->getHeight();
2190     // If there are bunch of CopyToRegs stacked up, they should be considered
2191     // to be at the same position.
2192     if (Succ.getSUnit()->getNode() &&
2193         Succ.getSUnit()->getNode()->getOpcode() == ISD::CopyToReg)
2194       Height = closestSucc(Succ.getSUnit())+1;
2195     if (Height > MaxHeight)
2196       MaxHeight = Height;
2197   }
2198   return MaxHeight;
2199 }
2200 
2201 /// calcMaxScratches - Returns an cost estimate of the worse case requirement
2202 /// for scratch registers, i.e. number of data dependencies.
calcMaxScratches(const SUnit * SU)2203 static unsigned calcMaxScratches(const SUnit *SU) {
2204   unsigned Scratches = 0;
2205   for (const SDep &Pred : SU->Preds) {
2206     if (Pred.isCtrl()) continue;  // ignore chain preds
2207     Scratches++;
2208   }
2209   return Scratches;
2210 }
2211 
2212 /// hasOnlyLiveInOpers - Return true if SU has only value predecessors that are
2213 /// CopyFromReg from a virtual register.
hasOnlyLiveInOpers(const SUnit * SU)2214 static bool hasOnlyLiveInOpers(const SUnit *SU) {
2215   bool RetVal = false;
2216   for (const SDep &Pred : SU->Preds) {
2217     if (Pred.isCtrl()) continue;
2218     const SUnit *PredSU = Pred.getSUnit();
2219     if (PredSU->getNode() &&
2220         PredSU->getNode()->getOpcode() == ISD::CopyFromReg) {
2221       unsigned Reg =
2222         cast<RegisterSDNode>(PredSU->getNode()->getOperand(1))->getReg();
2223       if (TargetRegisterInfo::isVirtualRegister(Reg)) {
2224         RetVal = true;
2225         continue;
2226       }
2227     }
2228     return false;
2229   }
2230   return RetVal;
2231 }
2232 
2233 /// hasOnlyLiveOutUses - Return true if SU has only value successors that are
2234 /// CopyToReg to a virtual register. This SU def is probably a liveout and
2235 /// it has no other use. It should be scheduled closer to the terminator.
hasOnlyLiveOutUses(const SUnit * SU)2236 static bool hasOnlyLiveOutUses(const SUnit *SU) {
2237   bool RetVal = false;
2238   for (const SDep &Succ : SU->Succs) {
2239     if (Succ.isCtrl()) continue;
2240     const SUnit *SuccSU = Succ.getSUnit();
2241     if (SuccSU->getNode() && SuccSU->getNode()->getOpcode() == ISD::CopyToReg) {
2242       unsigned Reg =
2243         cast<RegisterSDNode>(SuccSU->getNode()->getOperand(1))->getReg();
2244       if (TargetRegisterInfo::isVirtualRegister(Reg)) {
2245         RetVal = true;
2246         continue;
2247       }
2248     }
2249     return false;
2250   }
2251   return RetVal;
2252 }
2253 
2254 // Set isVRegCycle for a node with only live in opers and live out uses. Also
2255 // set isVRegCycle for its CopyFromReg operands.
2256 //
2257 // This is only relevant for single-block loops, in which case the VRegCycle
2258 // node is likely an induction variable in which the operand and target virtual
2259 // registers should be coalesced (e.g. pre/post increment values). Setting the
2260 // isVRegCycle flag helps the scheduler prioritize other uses of the same
2261 // CopyFromReg so that this node becomes the virtual register "kill". This
2262 // avoids interference between the values live in and out of the block and
2263 // eliminates a copy inside the loop.
initVRegCycle(SUnit * SU)2264 static void initVRegCycle(SUnit *SU) {
2265   if (DisableSchedVRegCycle)
2266     return;
2267 
2268   if (!hasOnlyLiveInOpers(SU) || !hasOnlyLiveOutUses(SU))
2269     return;
2270 
2271   DEBUG(dbgs() << "VRegCycle: SU(" << SU->NodeNum << ")\n");
2272 
2273   SU->isVRegCycle = true;
2274 
2275   for (const SDep &Pred : SU->Preds) {
2276     if (Pred.isCtrl()) continue;
2277     Pred.getSUnit()->isVRegCycle = true;
2278   }
2279 }
2280 
2281 // After scheduling the definition of a VRegCycle, clear the isVRegCycle flag of
2282 // CopyFromReg operands. We should no longer penalize other uses of this VReg.
resetVRegCycle(SUnit * SU)2283 static void resetVRegCycle(SUnit *SU) {
2284   if (!SU->isVRegCycle)
2285     return;
2286 
2287   for (const SDep &Pred : SU->Preds) {
2288     if (Pred.isCtrl()) continue;  // ignore chain preds
2289     SUnit *PredSU = Pred.getSUnit();
2290     if (PredSU->isVRegCycle) {
2291       assert(PredSU->getNode()->getOpcode() == ISD::CopyFromReg &&
2292              "VRegCycle def must be CopyFromReg");
2293       Pred.getSUnit()->isVRegCycle = false;
2294     }
2295   }
2296 }
2297 
2298 // Return true if this SUnit uses a CopyFromReg node marked as a VRegCycle. This
2299 // means a node that defines the VRegCycle has not been scheduled yet.
hasVRegCycleUse(const SUnit * SU)2300 static bool hasVRegCycleUse(const SUnit *SU) {
2301   // If this SU also defines the VReg, don't hoist it as a "use".
2302   if (SU->isVRegCycle)
2303     return false;
2304 
2305   for (const SDep &Pred : SU->Preds) {
2306     if (Pred.isCtrl()) continue;  // ignore chain preds
2307     if (Pred.getSUnit()->isVRegCycle &&
2308         Pred.getSUnit()->getNode()->getOpcode() == ISD::CopyFromReg) {
2309       DEBUG(dbgs() << "  VReg cycle use: SU (" << SU->NodeNum << ")\n");
2310       return true;
2311     }
2312   }
2313   return false;
2314 }
2315 
2316 // Check for either a dependence (latency) or resource (hazard) stall.
2317 //
2318 // Note: The ScheduleHazardRecognizer interface requires a non-const SU.
BUHasStall(SUnit * SU,int Height,RegReductionPQBase * SPQ)2319 static bool BUHasStall(SUnit *SU, int Height, RegReductionPQBase *SPQ) {
2320   if ((int)SPQ->getCurCycle() < Height) return true;
2321   if (SPQ->getHazardRec()->getHazardType(SU, 0)
2322       != ScheduleHazardRecognizer::NoHazard)
2323     return true;
2324   return false;
2325 }
2326 
2327 // Return -1 if left has higher priority, 1 if right has higher priority.
2328 // Return 0 if latency-based priority is equivalent.
BUCompareLatency(SUnit * left,SUnit * right,bool checkPref,RegReductionPQBase * SPQ)2329 static int BUCompareLatency(SUnit *left, SUnit *right, bool checkPref,
2330                             RegReductionPQBase *SPQ) {
2331   // Scheduling an instruction that uses a VReg whose postincrement has not yet
2332   // been scheduled will induce a copy. Model this as an extra cycle of latency.
2333   int LPenalty = hasVRegCycleUse(left) ? 1 : 0;
2334   int RPenalty = hasVRegCycleUse(right) ? 1 : 0;
2335   int LHeight = (int)left->getHeight() + LPenalty;
2336   int RHeight = (int)right->getHeight() + RPenalty;
2337 
2338   bool LStall = (!checkPref || left->SchedulingPref == Sched::ILP) &&
2339     BUHasStall(left, LHeight, SPQ);
2340   bool RStall = (!checkPref || right->SchedulingPref == Sched::ILP) &&
2341     BUHasStall(right, RHeight, SPQ);
2342 
2343   // If scheduling one of the node will cause a pipeline stall, delay it.
2344   // If scheduling either one of the node will cause a pipeline stall, sort
2345   // them according to their height.
2346   if (LStall) {
2347     if (!RStall)
2348       return 1;
2349     if (LHeight != RHeight)
2350       return LHeight > RHeight ? 1 : -1;
2351   } else if (RStall)
2352     return -1;
2353 
2354   // If either node is scheduling for latency, sort them by height/depth
2355   // and latency.
2356   if (!checkPref || (left->SchedulingPref == Sched::ILP ||
2357                      right->SchedulingPref == Sched::ILP)) {
2358     // If neither instruction stalls (!LStall && !RStall) and HazardRecognizer
2359     // is enabled, grouping instructions by cycle, then its height is already
2360     // covered so only its depth matters. We also reach this point if both stall
2361     // but have the same height.
2362     if (!SPQ->getHazardRec()->isEnabled()) {
2363       if (LHeight != RHeight)
2364         return LHeight > RHeight ? 1 : -1;
2365     }
2366     int LDepth = left->getDepth() - LPenalty;
2367     int RDepth = right->getDepth() - RPenalty;
2368     if (LDepth != RDepth) {
2369       DEBUG(dbgs() << "  Comparing latency of SU (" << left->NodeNum
2370             << ") depth " << LDepth << " vs SU (" << right->NodeNum
2371             << ") depth " << RDepth << "\n");
2372       return LDepth < RDepth ? 1 : -1;
2373     }
2374     if (left->Latency != right->Latency)
2375       return left->Latency > right->Latency ? 1 : -1;
2376   }
2377   return 0;
2378 }
2379 
BURRSort(SUnit * left,SUnit * right,RegReductionPQBase * SPQ)2380 static bool BURRSort(SUnit *left, SUnit *right, RegReductionPQBase *SPQ) {
2381   // Schedule physical register definitions close to their use. This is
2382   // motivated by microarchitectures that can fuse cmp+jump macro-ops. But as
2383   // long as shortening physreg live ranges is generally good, we can defer
2384   // creating a subtarget hook.
2385   if (!DisableSchedPhysRegJoin) {
2386     bool LHasPhysReg = left->hasPhysRegDefs;
2387     bool RHasPhysReg = right->hasPhysRegDefs;
2388     if (LHasPhysReg != RHasPhysReg) {
2389       #ifndef NDEBUG
2390       static const char *const PhysRegMsg[] = { " has no physreg",
2391                                                 " defines a physreg" };
2392       #endif
2393       DEBUG(dbgs() << "  SU (" << left->NodeNum << ") "
2394             << PhysRegMsg[LHasPhysReg] << " SU(" << right->NodeNum << ") "
2395             << PhysRegMsg[RHasPhysReg] << "\n");
2396       return LHasPhysReg < RHasPhysReg;
2397     }
2398   }
2399 
2400   // Prioritize by Sethi-Ulmann number and push CopyToReg nodes down.
2401   unsigned LPriority = SPQ->getNodePriority(left);
2402   unsigned RPriority = SPQ->getNodePriority(right);
2403 
2404   // Be really careful about hoisting call operands above previous calls.
2405   // Only allows it if it would reduce register pressure.
2406   if (left->isCall && right->isCallOp) {
2407     unsigned RNumVals = right->getNode()->getNumValues();
2408     RPriority = (RPriority > RNumVals) ? (RPriority - RNumVals) : 0;
2409   }
2410   if (right->isCall && left->isCallOp) {
2411     unsigned LNumVals = left->getNode()->getNumValues();
2412     LPriority = (LPriority > LNumVals) ? (LPriority - LNumVals) : 0;
2413   }
2414 
2415   if (LPriority != RPriority)
2416     return LPriority > RPriority;
2417 
2418   // One or both of the nodes are calls and their sethi-ullman numbers are the
2419   // same, then keep source order.
2420   if (left->isCall || right->isCall) {
2421     unsigned LOrder = SPQ->getNodeOrdering(left);
2422     unsigned ROrder = SPQ->getNodeOrdering(right);
2423 
2424     // Prefer an ordering where the lower the non-zero order number, the higher
2425     // the preference.
2426     if ((LOrder || ROrder) && LOrder != ROrder)
2427       return LOrder != 0 && (LOrder < ROrder || ROrder == 0);
2428   }
2429 
2430   // Try schedule def + use closer when Sethi-Ullman numbers are the same.
2431   // e.g.
2432   // t1 = op t2, c1
2433   // t3 = op t4, c2
2434   //
2435   // and the following instructions are both ready.
2436   // t2 = op c3
2437   // t4 = op c4
2438   //
2439   // Then schedule t2 = op first.
2440   // i.e.
2441   // t4 = op c4
2442   // t2 = op c3
2443   // t1 = op t2, c1
2444   // t3 = op t4, c2
2445   //
2446   // This creates more short live intervals.
2447   unsigned LDist = closestSucc(left);
2448   unsigned RDist = closestSucc(right);
2449   if (LDist != RDist)
2450     return LDist < RDist;
2451 
2452   // How many registers becomes live when the node is scheduled.
2453   unsigned LScratch = calcMaxScratches(left);
2454   unsigned RScratch = calcMaxScratches(right);
2455   if (LScratch != RScratch)
2456     return LScratch > RScratch;
2457 
2458   // Comparing latency against a call makes little sense unless the node
2459   // is register pressure-neutral.
2460   if ((left->isCall && RPriority > 0) || (right->isCall && LPriority > 0))
2461     return (left->NodeQueueId > right->NodeQueueId);
2462 
2463   // Do not compare latencies when one or both of the nodes are calls.
2464   if (!DisableSchedCycles &&
2465       !(left->isCall || right->isCall)) {
2466     int result = BUCompareLatency(left, right, false /*checkPref*/, SPQ);
2467     if (result != 0)
2468       return result > 0;
2469   }
2470   else {
2471     if (left->getHeight() != right->getHeight())
2472       return left->getHeight() > right->getHeight();
2473 
2474     if (left->getDepth() != right->getDepth())
2475       return left->getDepth() < right->getDepth();
2476   }
2477 
2478   assert(left->NodeQueueId && right->NodeQueueId &&
2479          "NodeQueueId cannot be zero");
2480   return (left->NodeQueueId > right->NodeQueueId);
2481 }
2482 
2483 // Bottom up
operator ()(SUnit * left,SUnit * right) const2484 bool bu_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2485   if (int res = checkSpecialNodes(left, right))
2486     return res > 0;
2487 
2488   return BURRSort(left, right, SPQ);
2489 }
2490 
2491 // Source order, otherwise bottom up.
operator ()(SUnit * left,SUnit * right) const2492 bool src_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2493   if (int res = checkSpecialNodes(left, right))
2494     return res > 0;
2495 
2496   unsigned LOrder = SPQ->getNodeOrdering(left);
2497   unsigned ROrder = SPQ->getNodeOrdering(right);
2498 
2499   // Prefer an ordering where the lower the non-zero order number, the higher
2500   // the preference.
2501   if ((LOrder || ROrder) && LOrder != ROrder)
2502     return LOrder != 0 && (LOrder < ROrder || ROrder == 0);
2503 
2504   return BURRSort(left, right, SPQ);
2505 }
2506 
2507 // If the time between now and when the instruction will be ready can cover
2508 // the spill code, then avoid adding it to the ready queue. This gives long
2509 // stalls highest priority and allows hoisting across calls. It should also
2510 // speed up processing the available queue.
isReady(SUnit * SU,unsigned CurCycle) const2511 bool hybrid_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
2512   static const unsigned ReadyDelay = 3;
2513 
2514   if (SPQ->MayReduceRegPressure(SU)) return true;
2515 
2516   if (SU->getHeight() > (CurCycle + ReadyDelay)) return false;
2517 
2518   if (SPQ->getHazardRec()->getHazardType(SU, -ReadyDelay)
2519       != ScheduleHazardRecognizer::NoHazard)
2520     return false;
2521 
2522   return true;
2523 }
2524 
2525 // Return true if right should be scheduled with higher priority than left.
operator ()(SUnit * left,SUnit * right) const2526 bool hybrid_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2527   if (int res = checkSpecialNodes(left, right))
2528     return res > 0;
2529 
2530   if (left->isCall || right->isCall)
2531     // No way to compute latency of calls.
2532     return BURRSort(left, right, SPQ);
2533 
2534   bool LHigh = SPQ->HighRegPressure(left);
2535   bool RHigh = SPQ->HighRegPressure(right);
2536   // Avoid causing spills. If register pressure is high, schedule for
2537   // register pressure reduction.
2538   if (LHigh && !RHigh) {
2539     DEBUG(dbgs() << "  pressure SU(" << left->NodeNum << ") > SU("
2540           << right->NodeNum << ")\n");
2541     return true;
2542   }
2543   else if (!LHigh && RHigh) {
2544     DEBUG(dbgs() << "  pressure SU(" << right->NodeNum << ") > SU("
2545           << left->NodeNum << ")\n");
2546     return false;
2547   }
2548   if (!LHigh && !RHigh) {
2549     int result = BUCompareLatency(left, right, true /*checkPref*/, SPQ);
2550     if (result != 0)
2551       return result > 0;
2552   }
2553   return BURRSort(left, right, SPQ);
2554 }
2555 
2556 // Schedule as many instructions in each cycle as possible. So don't make an
2557 // instruction available unless it is ready in the current cycle.
isReady(SUnit * SU,unsigned CurCycle) const2558 bool ilp_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
2559   if (SU->getHeight() > CurCycle) return false;
2560 
2561   if (SPQ->getHazardRec()->getHazardType(SU, 0)
2562       != ScheduleHazardRecognizer::NoHazard)
2563     return false;
2564 
2565   return true;
2566 }
2567 
canEnableCoalescing(SUnit * SU)2568 static bool canEnableCoalescing(SUnit *SU) {
2569   unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0;
2570   if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
2571     // CopyToReg should be close to its uses to facilitate coalescing and
2572     // avoid spilling.
2573     return true;
2574 
2575   if (Opc == TargetOpcode::EXTRACT_SUBREG ||
2576       Opc == TargetOpcode::SUBREG_TO_REG ||
2577       Opc == TargetOpcode::INSERT_SUBREG)
2578     // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be
2579     // close to their uses to facilitate coalescing.
2580     return true;
2581 
2582   if (SU->NumPreds == 0 && SU->NumSuccs != 0)
2583     // If SU does not have a register def, schedule it close to its uses
2584     // because it does not lengthen any live ranges.
2585     return true;
2586 
2587   return false;
2588 }
2589 
2590 // list-ilp is currently an experimental scheduler that allows various
2591 // heuristics to be enabled prior to the normal register reduction logic.
operator ()(SUnit * left,SUnit * right) const2592 bool ilp_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2593   if (int res = checkSpecialNodes(left, right))
2594     return res > 0;
2595 
2596   if (left->isCall || right->isCall)
2597     // No way to compute latency of calls.
2598     return BURRSort(left, right, SPQ);
2599 
2600   unsigned LLiveUses = 0, RLiveUses = 0;
2601   int LPDiff = 0, RPDiff = 0;
2602   if (!DisableSchedRegPressure || !DisableSchedLiveUses) {
2603     LPDiff = SPQ->RegPressureDiff(left, LLiveUses);
2604     RPDiff = SPQ->RegPressureDiff(right, RLiveUses);
2605   }
2606   if (!DisableSchedRegPressure && LPDiff != RPDiff) {
2607     DEBUG(dbgs() << "RegPressureDiff SU(" << left->NodeNum << "): " << LPDiff
2608           << " != SU(" << right->NodeNum << "): " << RPDiff << "\n");
2609     return LPDiff > RPDiff;
2610   }
2611 
2612   if (!DisableSchedRegPressure && (LPDiff > 0 || RPDiff > 0)) {
2613     bool LReduce = canEnableCoalescing(left);
2614     bool RReduce = canEnableCoalescing(right);
2615     if (LReduce && !RReduce) return false;
2616     if (RReduce && !LReduce) return true;
2617   }
2618 
2619   if (!DisableSchedLiveUses && (LLiveUses != RLiveUses)) {
2620     DEBUG(dbgs() << "Live uses SU(" << left->NodeNum << "): " << LLiveUses
2621           << " != SU(" << right->NodeNum << "): " << RLiveUses << "\n");
2622     return LLiveUses < RLiveUses;
2623   }
2624 
2625   if (!DisableSchedStalls) {
2626     bool LStall = BUHasStall(left, left->getHeight(), SPQ);
2627     bool RStall = BUHasStall(right, right->getHeight(), SPQ);
2628     if (LStall != RStall)
2629       return left->getHeight() > right->getHeight();
2630   }
2631 
2632   if (!DisableSchedCriticalPath) {
2633     int spread = (int)left->getDepth() - (int)right->getDepth();
2634     if (std::abs(spread) > MaxReorderWindow) {
2635       DEBUG(dbgs() << "Depth of SU(" << left->NodeNum << "): "
2636             << left->getDepth() << " != SU(" << right->NodeNum << "): "
2637             << right->getDepth() << "\n");
2638       return left->getDepth() < right->getDepth();
2639     }
2640   }
2641 
2642   if (!DisableSchedHeight && left->getHeight() != right->getHeight()) {
2643     int spread = (int)left->getHeight() - (int)right->getHeight();
2644     if (std::abs(spread) > MaxReorderWindow)
2645       return left->getHeight() > right->getHeight();
2646   }
2647 
2648   return BURRSort(left, right, SPQ);
2649 }
2650 
initNodes(std::vector<SUnit> & sunits)2651 void RegReductionPQBase::initNodes(std::vector<SUnit> &sunits) {
2652   SUnits = &sunits;
2653   // Add pseudo dependency edges for two-address nodes.
2654   if (!Disable2AddrHack)
2655     AddPseudoTwoAddrDeps();
2656   // Reroute edges to nodes with multiple uses.
2657   if (!TracksRegPressure && !SrcOrder)
2658     PrescheduleNodesWithMultipleUses();
2659   // Calculate node priorities.
2660   CalculateSethiUllmanNumbers();
2661 
2662   // For single block loops, mark nodes that look like canonical IV increments.
2663   if (scheduleDAG->BB->isSuccessor(scheduleDAG->BB))
2664     for (SUnit &SU : sunits)
2665       initVRegCycle(&SU);
2666 }
2667 
2668 //===----------------------------------------------------------------------===//
2669 //                    Preschedule for Register Pressure
2670 //===----------------------------------------------------------------------===//
2671 
canClobber(const SUnit * SU,const SUnit * Op)2672 bool RegReductionPQBase::canClobber(const SUnit *SU, const SUnit *Op) {
2673   if (SU->isTwoAddress) {
2674     unsigned Opc = SU->getNode()->getMachineOpcode();
2675     const MCInstrDesc &MCID = TII->get(Opc);
2676     unsigned NumRes = MCID.getNumDefs();
2677     unsigned NumOps = MCID.getNumOperands() - NumRes;
2678     for (unsigned i = 0; i != NumOps; ++i) {
2679       if (MCID.getOperandConstraint(i+NumRes, MCOI::TIED_TO) != -1) {
2680         SDNode *DU = SU->getNode()->getOperand(i).getNode();
2681         if (DU->getNodeId() != -1 &&
2682             Op->OrigNode == &(*SUnits)[DU->getNodeId()])
2683           return true;
2684       }
2685     }
2686   }
2687   return false;
2688 }
2689 
2690 /// canClobberReachingPhysRegUse - True if SU would clobber one of it's
2691 /// successor's explicit physregs whose definition can reach DepSU.
2692 /// i.e. DepSU should not be scheduled above SU.
canClobberReachingPhysRegUse(const SUnit * DepSU,const SUnit * SU,ScheduleDAGRRList * scheduleDAG,const TargetInstrInfo * TII,const TargetRegisterInfo * TRI)2693 static bool canClobberReachingPhysRegUse(const SUnit *DepSU, const SUnit *SU,
2694                                          ScheduleDAGRRList *scheduleDAG,
2695                                          const TargetInstrInfo *TII,
2696                                          const TargetRegisterInfo *TRI) {
2697   const MCPhysReg *ImpDefs
2698     = TII->get(SU->getNode()->getMachineOpcode()).getImplicitDefs();
2699   const uint32_t *RegMask = getNodeRegMask(SU->getNode());
2700   if(!ImpDefs && !RegMask)
2701     return false;
2702 
2703   for (const SDep &Succ : SU->Succs) {
2704     SUnit *SuccSU = Succ.getSUnit();
2705     for (const SDep &SuccPred : SuccSU->Preds) {
2706       if (!SuccPred.isAssignedRegDep())
2707         continue;
2708 
2709       if (RegMask &&
2710           MachineOperand::clobbersPhysReg(RegMask, SuccPred.getReg()) &&
2711           scheduleDAG->IsReachable(DepSU, SuccPred.getSUnit()))
2712         return true;
2713 
2714       if (ImpDefs)
2715         for (const MCPhysReg *ImpDef = ImpDefs; *ImpDef; ++ImpDef)
2716           // Return true if SU clobbers this physical register use and the
2717           // definition of the register reaches from DepSU. IsReachable queries
2718           // a topological forward sort of the DAG (following the successors).
2719           if (TRI->regsOverlap(*ImpDef, SuccPred.getReg()) &&
2720               scheduleDAG->IsReachable(DepSU, SuccPred.getSUnit()))
2721             return true;
2722     }
2723   }
2724   return false;
2725 }
2726 
2727 /// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's
2728 /// physical register defs.
canClobberPhysRegDefs(const SUnit * SuccSU,const SUnit * SU,const TargetInstrInfo * TII,const TargetRegisterInfo * TRI)2729 static bool canClobberPhysRegDefs(const SUnit *SuccSU, const SUnit *SU,
2730                                   const TargetInstrInfo *TII,
2731                                   const TargetRegisterInfo *TRI) {
2732   SDNode *N = SuccSU->getNode();
2733   unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2734   const MCPhysReg *ImpDefs = TII->get(N->getMachineOpcode()).getImplicitDefs();
2735   assert(ImpDefs && "Caller should check hasPhysRegDefs");
2736   for (const SDNode *SUNode = SU->getNode(); SUNode;
2737        SUNode = SUNode->getGluedNode()) {
2738     if (!SUNode->isMachineOpcode())
2739       continue;
2740     const MCPhysReg *SUImpDefs =
2741       TII->get(SUNode->getMachineOpcode()).getImplicitDefs();
2742     const uint32_t *SURegMask = getNodeRegMask(SUNode);
2743     if (!SUImpDefs && !SURegMask)
2744       continue;
2745     for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
2746       MVT VT = N->getSimpleValueType(i);
2747       if (VT == MVT::Glue || VT == MVT::Other)
2748         continue;
2749       if (!N->hasAnyUseOfValue(i))
2750         continue;
2751       unsigned Reg = ImpDefs[i - NumDefs];
2752       if (SURegMask && MachineOperand::clobbersPhysReg(SURegMask, Reg))
2753         return true;
2754       if (!SUImpDefs)
2755         continue;
2756       for (;*SUImpDefs; ++SUImpDefs) {
2757         unsigned SUReg = *SUImpDefs;
2758         if (TRI->regsOverlap(Reg, SUReg))
2759           return true;
2760       }
2761     }
2762   }
2763   return false;
2764 }
2765 
2766 /// PrescheduleNodesWithMultipleUses - Nodes with multiple uses
2767 /// are not handled well by the general register pressure reduction
2768 /// heuristics. When presented with code like this:
2769 ///
2770 ///      N
2771 ///    / |
2772 ///   /  |
2773 ///  U  store
2774 ///  |
2775 /// ...
2776 ///
2777 /// the heuristics tend to push the store up, but since the
2778 /// operand of the store has another use (U), this would increase
2779 /// the length of that other use (the U->N edge).
2780 ///
2781 /// This function transforms code like the above to route U's
2782 /// dependence through the store when possible, like this:
2783 ///
2784 ///      N
2785 ///      ||
2786 ///      ||
2787 ///     store
2788 ///       |
2789 ///       U
2790 ///       |
2791 ///      ...
2792 ///
2793 /// This results in the store being scheduled immediately
2794 /// after N, which shortens the U->N live range, reducing
2795 /// register pressure.
2796 ///
PrescheduleNodesWithMultipleUses()2797 void RegReductionPQBase::PrescheduleNodesWithMultipleUses() {
2798   // Visit all the nodes in topological order, working top-down.
2799   for (SUnit &SU : *SUnits) {
2800     // For now, only look at nodes with no data successors, such as stores.
2801     // These are especially important, due to the heuristics in
2802     // getNodePriority for nodes with no data successors.
2803     if (SU.NumSuccs != 0)
2804       continue;
2805     // For now, only look at nodes with exactly one data predecessor.
2806     if (SU.NumPreds != 1)
2807       continue;
2808     // Avoid prescheduling copies to virtual registers, which don't behave
2809     // like other nodes from the perspective of scheduling heuristics.
2810     if (SDNode *N = SU.getNode())
2811       if (N->getOpcode() == ISD::CopyToReg &&
2812           TargetRegisterInfo::isVirtualRegister
2813             (cast<RegisterSDNode>(N->getOperand(1))->getReg()))
2814         continue;
2815 
2816     // Locate the single data predecessor.
2817     SUnit *PredSU = nullptr;
2818     for (const SDep &Pred : SU.Preds)
2819       if (!Pred.isCtrl()) {
2820         PredSU = Pred.getSUnit();
2821         break;
2822       }
2823     assert(PredSU);
2824 
2825     // Don't rewrite edges that carry physregs, because that requires additional
2826     // support infrastructure.
2827     if (PredSU->hasPhysRegDefs)
2828       continue;
2829     // Short-circuit the case where SU is PredSU's only data successor.
2830     if (PredSU->NumSuccs == 1)
2831       continue;
2832     // Avoid prescheduling to copies from virtual registers, which don't behave
2833     // like other nodes from the perspective of scheduling heuristics.
2834     if (SDNode *N = SU.getNode())
2835       if (N->getOpcode() == ISD::CopyFromReg &&
2836           TargetRegisterInfo::isVirtualRegister
2837             (cast<RegisterSDNode>(N->getOperand(1))->getReg()))
2838         continue;
2839 
2840     // Perform checks on the successors of PredSU.
2841     for (const SDep &PredSucc : PredSU->Succs) {
2842       SUnit *PredSuccSU = PredSucc.getSUnit();
2843       if (PredSuccSU == &SU) continue;
2844       // If PredSU has another successor with no data successors, for
2845       // now don't attempt to choose either over the other.
2846       if (PredSuccSU->NumSuccs == 0)
2847         goto outer_loop_continue;
2848       // Don't break physical register dependencies.
2849       if (SU.hasPhysRegClobbers && PredSuccSU->hasPhysRegDefs)
2850         if (canClobberPhysRegDefs(PredSuccSU, &SU, TII, TRI))
2851           goto outer_loop_continue;
2852       // Don't introduce graph cycles.
2853       if (scheduleDAG->IsReachable(&SU, PredSuccSU))
2854         goto outer_loop_continue;
2855     }
2856 
2857     // Ok, the transformation is safe and the heuristics suggest it is
2858     // profitable. Update the graph.
2859     DEBUG(dbgs() << "    Prescheduling SU #" << SU.NodeNum
2860                  << " next to PredSU #" << PredSU->NodeNum
2861                  << " to guide scheduling in the presence of multiple uses\n");
2862     for (unsigned i = 0; i != PredSU->Succs.size(); ++i) {
2863       SDep Edge = PredSU->Succs[i];
2864       assert(!Edge.isAssignedRegDep());
2865       SUnit *SuccSU = Edge.getSUnit();
2866       if (SuccSU != &SU) {
2867         Edge.setSUnit(PredSU);
2868         scheduleDAG->RemovePred(SuccSU, Edge);
2869         scheduleDAG->AddPred(&SU, Edge);
2870         Edge.setSUnit(&SU);
2871         scheduleDAG->AddPred(SuccSU, Edge);
2872         --i;
2873       }
2874     }
2875   outer_loop_continue:;
2876   }
2877 }
2878 
2879 /// AddPseudoTwoAddrDeps - If two nodes share an operand and one of them uses
2880 /// it as a def&use operand. Add a pseudo control edge from it to the other
2881 /// node (if it won't create a cycle) so the two-address one will be scheduled
2882 /// first (lower in the schedule). If both nodes are two-address, favor the
2883 /// one that has a CopyToReg use (more likely to be a loop induction update).
2884 /// If both are two-address, but one is commutable while the other is not
2885 /// commutable, favor the one that's not commutable.
AddPseudoTwoAddrDeps()2886 void RegReductionPQBase::AddPseudoTwoAddrDeps() {
2887   for (SUnit &SU : *SUnits) {
2888     if (!SU.isTwoAddress)
2889       continue;
2890 
2891     SDNode *Node = SU.getNode();
2892     if (!Node || !Node->isMachineOpcode() || SU.getNode()->getGluedNode())
2893       continue;
2894 
2895     bool isLiveOut = hasOnlyLiveOutUses(&SU);
2896     unsigned Opc = Node->getMachineOpcode();
2897     const MCInstrDesc &MCID = TII->get(Opc);
2898     unsigned NumRes = MCID.getNumDefs();
2899     unsigned NumOps = MCID.getNumOperands() - NumRes;
2900     for (unsigned j = 0; j != NumOps; ++j) {
2901       if (MCID.getOperandConstraint(j+NumRes, MCOI::TIED_TO) == -1)
2902         continue;
2903       SDNode *DU = SU.getNode()->getOperand(j).getNode();
2904       if (DU->getNodeId() == -1)
2905         continue;
2906       const SUnit *DUSU = &(*SUnits)[DU->getNodeId()];
2907       if (!DUSU)
2908         continue;
2909       for (const SDep &Succ : DUSU->Succs) {
2910         if (Succ.isCtrl())
2911           continue;
2912         SUnit *SuccSU = Succ.getSUnit();
2913         if (SuccSU == &SU)
2914           continue;
2915         // Be conservative. Ignore if nodes aren't at roughly the same
2916         // depth and height.
2917         if (SuccSU->getHeight() < SU.getHeight() &&
2918             (SU.getHeight() - SuccSU->getHeight()) > 1)
2919           continue;
2920         // Skip past COPY_TO_REGCLASS nodes, so that the pseudo edge
2921         // constrains whatever is using the copy, instead of the copy
2922         // itself. In the case that the copy is coalesced, this
2923         // preserves the intent of the pseudo two-address heurietics.
2924         while (SuccSU->Succs.size() == 1 &&
2925                SuccSU->getNode()->isMachineOpcode() &&
2926                SuccSU->getNode()->getMachineOpcode() ==
2927                  TargetOpcode::COPY_TO_REGCLASS)
2928           SuccSU = SuccSU->Succs.front().getSUnit();
2929         // Don't constrain non-instruction nodes.
2930         if (!SuccSU->getNode() || !SuccSU->getNode()->isMachineOpcode())
2931           continue;
2932         // Don't constrain nodes with physical register defs if the
2933         // predecessor can clobber them.
2934         if (SuccSU->hasPhysRegDefs && SU.hasPhysRegClobbers) {
2935           if (canClobberPhysRegDefs(SuccSU, &SU, TII, TRI))
2936             continue;
2937         }
2938         // Don't constrain EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG;
2939         // these may be coalesced away. We want them close to their uses.
2940         unsigned SuccOpc = SuccSU->getNode()->getMachineOpcode();
2941         if (SuccOpc == TargetOpcode::EXTRACT_SUBREG ||
2942             SuccOpc == TargetOpcode::INSERT_SUBREG ||
2943             SuccOpc == TargetOpcode::SUBREG_TO_REG)
2944           continue;
2945         if (!canClobberReachingPhysRegUse(SuccSU, &SU, scheduleDAG, TII, TRI) &&
2946             (!canClobber(SuccSU, DUSU) ||
2947              (isLiveOut && !hasOnlyLiveOutUses(SuccSU)) ||
2948              (!SU.isCommutable && SuccSU->isCommutable)) &&
2949             !scheduleDAG->IsReachable(SuccSU, &SU)) {
2950           DEBUG(dbgs() << "    Adding a pseudo-two-addr edge from SU #"
2951                        << SU.NodeNum << " to SU #" << SuccSU->NodeNum << "\n");
2952           scheduleDAG->AddPred(&SU, SDep(SuccSU, SDep::Artificial));
2953         }
2954       }
2955     }
2956   }
2957 }
2958 
2959 //===----------------------------------------------------------------------===//
2960 //                         Public Constructor Functions
2961 //===----------------------------------------------------------------------===//
2962 
2963 llvm::ScheduleDAGSDNodes *
createBURRListDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level OptLevel)2964 llvm::createBURRListDAGScheduler(SelectionDAGISel *IS,
2965                                  CodeGenOpt::Level OptLevel) {
2966   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
2967   const TargetInstrInfo *TII = STI.getInstrInfo();
2968   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
2969 
2970   BURegReductionPriorityQueue *PQ =
2971     new BURegReductionPriorityQueue(*IS->MF, false, false, TII, TRI, nullptr);
2972   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
2973   PQ->setScheduleDAG(SD);
2974   return SD;
2975 }
2976 
2977 llvm::ScheduleDAGSDNodes *
createSourceListDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level OptLevel)2978 llvm::createSourceListDAGScheduler(SelectionDAGISel *IS,
2979                                    CodeGenOpt::Level OptLevel) {
2980   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
2981   const TargetInstrInfo *TII = STI.getInstrInfo();
2982   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
2983 
2984   SrcRegReductionPriorityQueue *PQ =
2985     new SrcRegReductionPriorityQueue(*IS->MF, false, true, TII, TRI, nullptr);
2986   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
2987   PQ->setScheduleDAG(SD);
2988   return SD;
2989 }
2990 
2991 llvm::ScheduleDAGSDNodes *
createHybridListDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level OptLevel)2992 llvm::createHybridListDAGScheduler(SelectionDAGISel *IS,
2993                                    CodeGenOpt::Level OptLevel) {
2994   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
2995   const TargetInstrInfo *TII = STI.getInstrInfo();
2996   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
2997   const TargetLowering *TLI = IS->TLI;
2998 
2999   HybridBURRPriorityQueue *PQ =
3000     new HybridBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI);
3001 
3002   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
3003   PQ->setScheduleDAG(SD);
3004   return SD;
3005 }
3006 
3007 llvm::ScheduleDAGSDNodes *
createILPListDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level OptLevel)3008 llvm::createILPListDAGScheduler(SelectionDAGISel *IS,
3009                                 CodeGenOpt::Level OptLevel) {
3010   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3011   const TargetInstrInfo *TII = STI.getInstrInfo();
3012   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3013   const TargetLowering *TLI = IS->TLI;
3014 
3015   ILPBURRPriorityQueue *PQ =
3016     new ILPBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI);
3017   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
3018   PQ->setScheduleDAG(SD);
3019   return SD;
3020 }
3021