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::vector<SUnit*> LiveRegDefs;
145   std::vector<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.resize(TRI->getNumRegs() + 1, nullptr);
332   LiveRegGens.resize(TRI->getNumRegs() + 1, nullptr);
333   CallSeqEndForStart.clear();
334   assert(Interferences.empty() && LRegsMap.empty() && "stale Interferences");
335 
336   // Build the scheduling graph.
337   BuildSchedGraph(nullptr);
338 
339   DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
340           SUnits[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 (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
419         if (IsChainDependent(N->getOperand(i).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 (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
437       if (N->getOperand(i).getValueType() == MVT::Other) {
438         N = N->getOperand(i).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 (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
468         unsigned MyNestLevel = NestLevel;
469         unsigned MyMaxNest = MaxNest;
470         if (SDNode *New = FindCallSeqStart(N->getOperand(i).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 (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
497       if (N->getOperand(i).getValueType() == MVT::Other) {
498         N = N->getOperand(i).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 (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
852        I != E; ++I) {
853     if (I->isAssignedRegDep()) {
854       if (!LiveRegDefs[I->getReg()])
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[I->getReg()] = SU;
859       if (LiveRegGens[I->getReg()] == nullptr ||
860           I->getSUnit()->getHeight() < LiveRegGens[I->getReg()]->getHeight())
861         LiveRegGens[I->getReg()] = I->getSUnit();
862     }
863   }
864   if (SU->getHeight() < MinAvailableCycle)
865     MinAvailableCycle = SU->getHeight();
866 
867   SU->setHeightDirty();
868   SU->isScheduled = false;
869   SU->isAvailable = true;
870   if (!DisableSchedCycles && AvailableQueue->hasReadyFilter()) {
871     // Don't make available until backtracking is complete.
872     SU->isPending = true;
873     PendingQueue.push_back(SU);
874   }
875   else {
876     AvailableQueue->push(SU);
877   }
878   AvailableQueue->unscheduledNode(SU);
879 }
880 
881 /// After backtracking, the hazard checker needs to be restored to a state
882 /// corresponding the current cycle.
RestoreHazardCheckerBottomUp()883 void ScheduleDAGRRList::RestoreHazardCheckerBottomUp() {
884   HazardRec->Reset();
885 
886   unsigned LookAhead = std::min((unsigned)Sequence.size(),
887                                 HazardRec->getMaxLookAhead());
888   if (LookAhead == 0)
889     return;
890 
891   std::vector<SUnit*>::const_iterator I = (Sequence.end() - LookAhead);
892   unsigned HazardCycle = (*I)->getHeight();
893   for (std::vector<SUnit*>::const_iterator E = Sequence.end(); I != E; ++I) {
894     SUnit *SU = *I;
895     for (; SU->getHeight() > HazardCycle; ++HazardCycle) {
896       HazardRec->RecedeCycle();
897     }
898     EmitNode(SU);
899   }
900 }
901 
902 /// BacktrackBottomUp - Backtrack scheduling to a previous cycle specified in
903 /// BTCycle in order to schedule a specific node.
BacktrackBottomUp(SUnit * SU,SUnit * BtSU)904 void ScheduleDAGRRList::BacktrackBottomUp(SUnit *SU, SUnit *BtSU) {
905   SUnit *OldSU = Sequence.back();
906   while (true) {
907     Sequence.pop_back();
908     // FIXME: use ready cycle instead of height
909     CurCycle = OldSU->getHeight();
910     UnscheduleNodeBottomUp(OldSU);
911     AvailableQueue->setCurCycle(CurCycle);
912     if (OldSU == BtSU)
913       break;
914     OldSU = Sequence.back();
915   }
916 
917   assert(!SU->isSucc(OldSU) && "Something is wrong!");
918 
919   RestoreHazardCheckerBottomUp();
920 
921   ReleasePending();
922 
923   ++NumBacktracks;
924 }
925 
isOperandOf(const SUnit * SU,SDNode * N)926 static bool isOperandOf(const SUnit *SU, SDNode *N) {
927   for (const SDNode *SUNode = SU->getNode(); SUNode;
928        SUNode = SUNode->getGluedNode()) {
929     if (SUNode->isOperandOf(N))
930       return true;
931   }
932   return false;
933 }
934 
935 /// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
936 /// successors to the newly created node.
CopyAndMoveSuccessors(SUnit * SU)937 SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) {
938   SDNode *N = SU->getNode();
939   if (!N)
940     return nullptr;
941 
942   if (SU->getNode()->getGluedNode())
943     return nullptr;
944 
945   SUnit *NewSU;
946   bool TryUnfold = false;
947   for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
948     MVT VT = N->getSimpleValueType(i);
949     if (VT == MVT::Glue)
950       return nullptr;
951     else if (VT == MVT::Other)
952       TryUnfold = true;
953   }
954   for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
955     const SDValue &Op = N->getOperand(i);
956     MVT VT = Op.getNode()->getSimpleValueType(Op.getResNo());
957     if (VT == MVT::Glue)
958       return nullptr;
959   }
960 
961   if (TryUnfold) {
962     SmallVector<SDNode*, 2> NewNodes;
963     if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
964       return nullptr;
965 
966     // unfolding an x86 DEC64m operation results in store, dec, load which
967     // can't be handled here so quit
968     if (NewNodes.size() == 3)
969       return nullptr;
970 
971     DEBUG(dbgs() << "Unfolding SU #" << SU->NodeNum << "\n");
972     assert(NewNodes.size() == 2 && "Expected a load folding node!");
973 
974     N = NewNodes[1];
975     SDNode *LoadNode = NewNodes[0];
976     unsigned NumVals = N->getNumValues();
977     unsigned OldNumVals = SU->getNode()->getNumValues();
978     for (unsigned i = 0; i != NumVals; ++i)
979       DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i));
980     DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals-1),
981                                    SDValue(LoadNode, 1));
982 
983     // LoadNode may already exist. This can happen when there is another
984     // load from the same location and producing the same type of value
985     // but it has different alignment or volatileness.
986     bool isNewLoad = true;
987     SUnit *LoadSU;
988     if (LoadNode->getNodeId() != -1) {
989       LoadSU = &SUnits[LoadNode->getNodeId()];
990       isNewLoad = false;
991     } else {
992       LoadSU = CreateNewSUnit(LoadNode);
993       LoadNode->setNodeId(LoadSU->NodeNum);
994 
995       InitNumRegDefsLeft(LoadSU);
996       computeLatency(LoadSU);
997     }
998 
999     SUnit *NewSU = CreateNewSUnit(N);
1000     assert(N->getNodeId() == -1 && "Node already inserted!");
1001     N->setNodeId(NewSU->NodeNum);
1002 
1003     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1004     for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
1005       if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
1006         NewSU->isTwoAddress = true;
1007         break;
1008       }
1009     }
1010     if (MCID.isCommutable())
1011       NewSU->isCommutable = true;
1012 
1013     InitNumRegDefsLeft(NewSU);
1014     computeLatency(NewSU);
1015 
1016     // Record all the edges to and from the old SU, by category.
1017     SmallVector<SDep, 4> ChainPreds;
1018     SmallVector<SDep, 4> ChainSuccs;
1019     SmallVector<SDep, 4> LoadPreds;
1020     SmallVector<SDep, 4> NodePreds;
1021     SmallVector<SDep, 4> NodeSuccs;
1022     for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1023          I != E; ++I) {
1024       if (I->isCtrl())
1025         ChainPreds.push_back(*I);
1026       else if (isOperandOf(I->getSUnit(), LoadNode))
1027         LoadPreds.push_back(*I);
1028       else
1029         NodePreds.push_back(*I);
1030     }
1031     for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1032          I != E; ++I) {
1033       if (I->isCtrl())
1034         ChainSuccs.push_back(*I);
1035       else
1036         NodeSuccs.push_back(*I);
1037     }
1038 
1039     // Now assign edges to the newly-created nodes.
1040     for (unsigned i = 0, e = ChainPreds.size(); i != e; ++i) {
1041       const SDep &Pred = ChainPreds[i];
1042       RemovePred(SU, Pred);
1043       if (isNewLoad)
1044         AddPred(LoadSU, Pred);
1045     }
1046     for (unsigned i = 0, e = LoadPreds.size(); i != e; ++i) {
1047       const SDep &Pred = LoadPreds[i];
1048       RemovePred(SU, Pred);
1049       if (isNewLoad)
1050         AddPred(LoadSU, Pred);
1051     }
1052     for (unsigned i = 0, e = NodePreds.size(); i != e; ++i) {
1053       const SDep &Pred = NodePreds[i];
1054       RemovePred(SU, Pred);
1055       AddPred(NewSU, Pred);
1056     }
1057     for (unsigned i = 0, e = NodeSuccs.size(); i != e; ++i) {
1058       SDep D = NodeSuccs[i];
1059       SUnit *SuccDep = D.getSUnit();
1060       D.setSUnit(SU);
1061       RemovePred(SuccDep, D);
1062       D.setSUnit(NewSU);
1063       AddPred(SuccDep, D);
1064       // Balance register pressure.
1065       if (AvailableQueue->tracksRegPressure() && SuccDep->isScheduled
1066           && !D.isCtrl() && NewSU->NumRegDefsLeft > 0)
1067         --NewSU->NumRegDefsLeft;
1068     }
1069     for (unsigned i = 0, e = ChainSuccs.size(); i != e; ++i) {
1070       SDep D = ChainSuccs[i];
1071       SUnit *SuccDep = D.getSUnit();
1072       D.setSUnit(SU);
1073       RemovePred(SuccDep, D);
1074       if (isNewLoad) {
1075         D.setSUnit(LoadSU);
1076         AddPred(SuccDep, D);
1077       }
1078     }
1079 
1080     // Add a data dependency to reflect that NewSU reads the value defined
1081     // by LoadSU.
1082     SDep D(LoadSU, SDep::Data, 0);
1083     D.setLatency(LoadSU->Latency);
1084     AddPred(NewSU, D);
1085 
1086     if (isNewLoad)
1087       AvailableQueue->addNode(LoadSU);
1088     AvailableQueue->addNode(NewSU);
1089 
1090     ++NumUnfolds;
1091 
1092     if (NewSU->NumSuccsLeft == 0) {
1093       NewSU->isAvailable = true;
1094       return NewSU;
1095     }
1096     SU = NewSU;
1097   }
1098 
1099   DEBUG(dbgs() << "    Duplicating SU #" << SU->NodeNum << "\n");
1100   NewSU = CreateClone(SU);
1101 
1102   // New SUnit has the exact same predecessors.
1103   for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1104        I != E; ++I)
1105     if (!I->isArtificial())
1106       AddPred(NewSU, *I);
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 (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1112        I != E; ++I) {
1113     if (I->isArtificial())
1114       continue;
1115     SUnit *SuccSU = I->getSUnit();
1116     if (SuccSU->isScheduled) {
1117       SDep D = *I;
1118       D.setSUnit(NewSU);
1119       AddPred(SuccSU, D);
1120       D.setSUnit(SU);
1121       DelDeps.push_back(std::make_pair(SuccSU, D));
1122     }
1123   }
1124   for (unsigned i = 0, e = DelDeps.size(); i != e; ++i)
1125     RemovePred(DelDeps[i].first, DelDeps[i].second);
1126 
1127   AvailableQueue->updateNode(SU);
1128   AvailableQueue->addNode(NewSU);
1129 
1130   ++NumDups;
1131   return NewSU;
1132 }
1133 
1134 /// InsertCopiesAndMoveSuccs - Insert register copies and move all
1135 /// scheduled successors of the given SUnit to the last copy.
InsertCopiesAndMoveSuccs(SUnit * SU,unsigned Reg,const TargetRegisterClass * DestRC,const TargetRegisterClass * SrcRC,SmallVectorImpl<SUnit * > & Copies)1136 void ScheduleDAGRRList::InsertCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
1137                                               const TargetRegisterClass *DestRC,
1138                                               const TargetRegisterClass *SrcRC,
1139                                               SmallVectorImpl<SUnit*> &Copies) {
1140   SUnit *CopyFromSU = CreateNewSUnit(nullptr);
1141   CopyFromSU->CopySrcRC = SrcRC;
1142   CopyFromSU->CopyDstRC = DestRC;
1143 
1144   SUnit *CopyToSU = CreateNewSUnit(nullptr);
1145   CopyToSU->CopySrcRC = DestRC;
1146   CopyToSU->CopyDstRC = SrcRC;
1147 
1148   // Only copy scheduled successors. Cut them from old node's successor
1149   // list and move them over.
1150   SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
1151   for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1152        I != E; ++I) {
1153     if (I->isArtificial())
1154       continue;
1155     SUnit *SuccSU = I->getSUnit();
1156     if (SuccSU->isScheduled) {
1157       SDep D = *I;
1158       D.setSUnit(CopyToSU);
1159       AddPred(SuccSU, D);
1160       DelDeps.push_back(std::make_pair(SuccSU, *I));
1161     }
1162     else {
1163       // Avoid scheduling the def-side copy before other successors. Otherwise
1164       // we could introduce another physreg interference on the copy and
1165       // continue inserting copies indefinitely.
1166       AddPred(SuccSU, SDep(CopyFromSU, SDep::Artificial));
1167     }
1168   }
1169   for (unsigned i = 0, e = DelDeps.size(); i != e; ++i)
1170     RemovePred(DelDeps[i].first, DelDeps[i].second);
1171 
1172   SDep FromDep(SU, SDep::Data, Reg);
1173   FromDep.setLatency(SU->Latency);
1174   AddPred(CopyFromSU, FromDep);
1175   SDep ToDep(CopyFromSU, SDep::Data, 0);
1176   ToDep.setLatency(CopyFromSU->Latency);
1177   AddPred(CopyToSU, ToDep);
1178 
1179   AvailableQueue->updateNode(SU);
1180   AvailableQueue->addNode(CopyFromSU);
1181   AvailableQueue->addNode(CopyToSU);
1182   Copies.push_back(CopyFromSU);
1183   Copies.push_back(CopyToSU);
1184 
1185   ++NumPRCopies;
1186 }
1187 
1188 /// getPhysicalRegisterVT - Returns the ValueType of the physical register
1189 /// definition of the specified node.
1190 /// FIXME: Move to SelectionDAG?
getPhysicalRegisterVT(SDNode * N,unsigned Reg,const TargetInstrInfo * TII)1191 static MVT getPhysicalRegisterVT(SDNode *N, unsigned Reg,
1192                                  const TargetInstrInfo *TII) {
1193   unsigned NumRes;
1194   if (N->getOpcode() == ISD::CopyFromReg) {
1195     // CopyFromReg has: "chain, Val, glue" so operand 1 gives the type.
1196     NumRes = 1;
1197   } else {
1198     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1199     assert(MCID.ImplicitDefs && "Physical reg def must be in implicit def list!");
1200     NumRes = MCID.getNumDefs();
1201     for (const uint16_t *ImpDef = MCID.getImplicitDefs(); *ImpDef; ++ImpDef) {
1202       if (Reg == *ImpDef)
1203         break;
1204       ++NumRes;
1205     }
1206   }
1207   return N->getSimpleValueType(NumRes);
1208 }
1209 
1210 /// CheckForLiveRegDef - Return true and update live register vector if the
1211 /// specified register def of the specified SUnit clobbers any "live" registers.
CheckForLiveRegDef(SUnit * SU,unsigned Reg,std::vector<SUnit * > & LiveRegDefs,SmallSet<unsigned,4> & RegAdded,SmallVectorImpl<unsigned> & LRegs,const TargetRegisterInfo * TRI)1212 static void CheckForLiveRegDef(SUnit *SU, unsigned Reg,
1213                                std::vector<SUnit*> &LiveRegDefs,
1214                                SmallSet<unsigned, 4> &RegAdded,
1215                                SmallVectorImpl<unsigned> &LRegs,
1216                                const TargetRegisterInfo *TRI) {
1217   for (MCRegAliasIterator AliasI(Reg, TRI, true); AliasI.isValid(); ++AliasI) {
1218 
1219     // Check if Ref is live.
1220     if (!LiveRegDefs[*AliasI]) continue;
1221 
1222     // Allow multiple uses of the same def.
1223     if (LiveRegDefs[*AliasI] == SU) continue;
1224 
1225     // Add Reg to the set of interfering live regs.
1226     if (RegAdded.insert(*AliasI).second) {
1227       LRegs.push_back(*AliasI);
1228     }
1229   }
1230 }
1231 
1232 /// CheckForLiveRegDefMasked - Check for any live physregs that are clobbered
1233 /// by RegMask, and add them to LRegs.
CheckForLiveRegDefMasked(SUnit * SU,const uint32_t * RegMask,std::vector<SUnit * > & LiveRegDefs,SmallSet<unsigned,4> & RegAdded,SmallVectorImpl<unsigned> & LRegs)1234 static void CheckForLiveRegDefMasked(SUnit *SU, const uint32_t *RegMask,
1235                                      std::vector<SUnit*> &LiveRegDefs,
1236                                      SmallSet<unsigned, 4> &RegAdded,
1237                                      SmallVectorImpl<unsigned> &LRegs) {
1238   // Look at all live registers. Skip Reg0 and the special CallResource.
1239   for (unsigned i = 1, e = LiveRegDefs.size()-1; i != e; ++i) {
1240     if (!LiveRegDefs[i]) continue;
1241     if (LiveRegDefs[i] == SU) continue;
1242     if (!MachineOperand::clobbersPhysReg(RegMask, i)) continue;
1243     if (RegAdded.insert(i).second)
1244       LRegs.push_back(i);
1245   }
1246 }
1247 
1248 /// getNodeRegMask - Returns the register mask attached to an SDNode, if any.
getNodeRegMask(const SDNode * N)1249 static const uint32_t *getNodeRegMask(const SDNode *N) {
1250   for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1251     if (const RegisterMaskSDNode *Op =
1252         dyn_cast<RegisterMaskSDNode>(N->getOperand(i).getNode()))
1253       return Op->getRegMask();
1254   return nullptr;
1255 }
1256 
1257 /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
1258 /// scheduling of the given node to satisfy live physical register dependencies.
1259 /// If the specific node is the last one that's available to schedule, do
1260 /// whatever is necessary (i.e. backtracking or cloning) to make it possible.
1261 bool ScheduleDAGRRList::
DelayForLiveRegsBottomUp(SUnit * SU,SmallVectorImpl<unsigned> & LRegs)1262 DelayForLiveRegsBottomUp(SUnit *SU, SmallVectorImpl<unsigned> &LRegs) {
1263   if (NumLiveRegs == 0)
1264     return false;
1265 
1266   SmallSet<unsigned, 4> RegAdded;
1267   // If this node would clobber any "live" register, then it's not ready.
1268   //
1269   // If SU is the currently live definition of the same register that it uses,
1270   // then we are free to schedule it.
1271   for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1272        I != E; ++I) {
1273     if (I->isAssignedRegDep() && LiveRegDefs[I->getReg()] != SU)
1274       CheckForLiveRegDef(I->getSUnit(), I->getReg(), LiveRegDefs,
1275                          RegAdded, LRegs, TRI);
1276   }
1277 
1278   for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) {
1279     if (Node->getOpcode() == ISD::INLINEASM) {
1280       // Inline asm can clobber physical defs.
1281       unsigned NumOps = Node->getNumOperands();
1282       if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
1283         --NumOps;  // Ignore the glue operand.
1284 
1285       for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
1286         unsigned Flags =
1287           cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
1288         unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
1289 
1290         ++i; // Skip the ID value.
1291         if (InlineAsm::isRegDefKind(Flags) ||
1292             InlineAsm::isRegDefEarlyClobberKind(Flags) ||
1293             InlineAsm::isClobberKind(Flags)) {
1294           // Check for def of register or earlyclobber register.
1295           for (; NumVals; --NumVals, ++i) {
1296             unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
1297             if (TargetRegisterInfo::isPhysicalRegister(Reg))
1298               CheckForLiveRegDef(SU, Reg, LiveRegDefs, RegAdded, LRegs, TRI);
1299           }
1300         } else
1301           i += NumVals;
1302       }
1303       continue;
1304     }
1305 
1306     if (!Node->isMachineOpcode())
1307       continue;
1308     // If we're in the middle of scheduling a call, don't begin scheduling
1309     // another call. Also, don't allow any physical registers to be live across
1310     // the call.
1311     if (Node->getMachineOpcode() == (unsigned)TII->getCallFrameDestroyOpcode()) {
1312       // Check the special calling-sequence resource.
1313       unsigned CallResource = TRI->getNumRegs();
1314       if (LiveRegDefs[CallResource]) {
1315         SDNode *Gen = LiveRegGens[CallResource]->getNode();
1316         while (SDNode *Glued = Gen->getGluedNode())
1317           Gen = Glued;
1318         if (!IsChainDependent(Gen, Node, 0, TII) &&
1319             RegAdded.insert(CallResource).second)
1320           LRegs.push_back(CallResource);
1321       }
1322     }
1323     if (const uint32_t *RegMask = getNodeRegMask(Node))
1324       CheckForLiveRegDefMasked(SU, RegMask, LiveRegDefs, RegAdded, LRegs);
1325 
1326     const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode());
1327     if (!MCID.ImplicitDefs)
1328       continue;
1329     for (const uint16_t *Reg = MCID.getImplicitDefs(); *Reg; ++Reg)
1330       CheckForLiveRegDef(SU, *Reg, LiveRegDefs, RegAdded, LRegs, TRI);
1331   }
1332 
1333   return !LRegs.empty();
1334 }
1335 
releaseInterferences(unsigned Reg)1336 void ScheduleDAGRRList::releaseInterferences(unsigned Reg) {
1337   // Add the nodes that aren't ready back onto the available list.
1338   for (unsigned i = Interferences.size(); i > 0; --i) {
1339     SUnit *SU = Interferences[i-1];
1340     LRegsMapT::iterator LRegsPos = LRegsMap.find(SU);
1341     if (Reg) {
1342       SmallVectorImpl<unsigned> &LRegs = LRegsPos->second;
1343       if (std::find(LRegs.begin(), LRegs.end(), Reg) == LRegs.end())
1344         continue;
1345     }
1346     SU->isPending = false;
1347     // The interfering node may no longer be available due to backtracking.
1348     // Furthermore, it may have been made available again, in which case it is
1349     // now already in the AvailableQueue.
1350     if (SU->isAvailable && !SU->NodeQueueId) {
1351       DEBUG(dbgs() << "    Repushing SU #" << SU->NodeNum << '\n');
1352       AvailableQueue->push(SU);
1353     }
1354     if (i < Interferences.size())
1355       Interferences[i-1] = Interferences.back();
1356     Interferences.pop_back();
1357     LRegsMap.erase(LRegsPos);
1358   }
1359 }
1360 
1361 /// Return a node that can be scheduled in this cycle. Requirements:
1362 /// (1) Ready: latency has been satisfied
1363 /// (2) No Hazards: resources are available
1364 /// (3) No Interferences: may unschedule to break register interferences.
PickNodeToScheduleBottomUp()1365 SUnit *ScheduleDAGRRList::PickNodeToScheduleBottomUp() {
1366   SUnit *CurSU = AvailableQueue->empty() ? nullptr : AvailableQueue->pop();
1367   while (CurSU) {
1368     SmallVector<unsigned, 4> LRegs;
1369     if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
1370       break;
1371     DEBUG(dbgs() << "    Interfering reg " <<
1372           (LRegs[0] == TRI->getNumRegs() ? "CallResource"
1373            : TRI->getName(LRegs[0]))
1374            << " SU #" << CurSU->NodeNum << '\n');
1375     std::pair<LRegsMapT::iterator, bool> LRegsPair =
1376       LRegsMap.insert(std::make_pair(CurSU, LRegs));
1377     if (LRegsPair.second) {
1378       CurSU->isPending = true;  // This SU is not in AvailableQueue right now.
1379       Interferences.push_back(CurSU);
1380     }
1381     else {
1382       assert(CurSU->isPending && "Interferences are pending");
1383       // Update the interference with current live regs.
1384       LRegsPair.first->second = LRegs;
1385     }
1386     CurSU = AvailableQueue->pop();
1387   }
1388   if (CurSU)
1389     return CurSU;
1390 
1391   // All candidates are delayed due to live physical reg dependencies.
1392   // Try backtracking, code duplication, or inserting cross class copies
1393   // to resolve it.
1394   for (unsigned i = 0, e = Interferences.size(); i != e; ++i) {
1395     SUnit *TrySU = Interferences[i];
1396     SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
1397 
1398     // Try unscheduling up to the point where it's safe to schedule
1399     // this node.
1400     SUnit *BtSU = nullptr;
1401     unsigned LiveCycle = UINT_MAX;
1402     for (unsigned j = 0, ee = LRegs.size(); j != ee; ++j) {
1403       unsigned Reg = LRegs[j];
1404       if (LiveRegGens[Reg]->getHeight() < LiveCycle) {
1405         BtSU = LiveRegGens[Reg];
1406         LiveCycle = BtSU->getHeight();
1407       }
1408     }
1409     if (!WillCreateCycle(TrySU, BtSU))  {
1410       // BacktrackBottomUp mutates Interferences!
1411       BacktrackBottomUp(TrySU, BtSU);
1412 
1413       // Force the current node to be scheduled before the node that
1414       // requires the physical reg dep.
1415       if (BtSU->isAvailable) {
1416         BtSU->isAvailable = false;
1417         if (!BtSU->isPending)
1418           AvailableQueue->remove(BtSU);
1419       }
1420       DEBUG(dbgs() << "ARTIFICIAL edge from SU(" << BtSU->NodeNum << ") to SU("
1421             << TrySU->NodeNum << ")\n");
1422       AddPred(TrySU, SDep(BtSU, SDep::Artificial));
1423 
1424       // If one or more successors has been unscheduled, then the current
1425       // node is no longer available.
1426       if (!TrySU->isAvailable || !TrySU->NodeQueueId)
1427         CurSU = AvailableQueue->pop();
1428       else {
1429         // Available and in AvailableQueue
1430         AvailableQueue->remove(TrySU);
1431         CurSU = TrySU;
1432       }
1433       // Interferences has been mutated. We must break.
1434       break;
1435     }
1436   }
1437 
1438   if (!CurSU) {
1439     // Can't backtrack. If it's too expensive to copy the value, then try
1440     // duplicate the nodes that produces these "too expensive to copy"
1441     // values to break the dependency. In case even that doesn't work,
1442     // insert cross class copies.
1443     // If it's not too expensive, i.e. cost != -1, issue copies.
1444     SUnit *TrySU = Interferences[0];
1445     SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
1446     assert(LRegs.size() == 1 && "Can't handle this yet!");
1447     unsigned Reg = LRegs[0];
1448     SUnit *LRDef = LiveRegDefs[Reg];
1449     MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
1450     const TargetRegisterClass *RC =
1451       TRI->getMinimalPhysRegClass(Reg, VT);
1452     const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
1453 
1454     // If cross copy register class is the same as RC, then it must be possible
1455     // copy the value directly. Do not try duplicate the def.
1456     // If cross copy register class is not the same as RC, then it's possible to
1457     // copy the value but it require cross register class copies and it is
1458     // expensive.
1459     // If cross copy register class is null, then it's not possible to copy
1460     // the value at all.
1461     SUnit *NewDef = nullptr;
1462     if (DestRC != RC) {
1463       NewDef = CopyAndMoveSuccessors(LRDef);
1464       if (!DestRC && !NewDef)
1465         report_fatal_error("Can't handle live physical register dependency!");
1466     }
1467     if (!NewDef) {
1468       // Issue copies, these can be expensive cross register class copies.
1469       SmallVector<SUnit*, 2> Copies;
1470       InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
1471       DEBUG(dbgs() << "    Adding an edge from SU #" << TrySU->NodeNum
1472             << " to SU #" << Copies.front()->NodeNum << "\n");
1473       AddPred(TrySU, SDep(Copies.front(), SDep::Artificial));
1474       NewDef = Copies.back();
1475     }
1476 
1477     DEBUG(dbgs() << "    Adding an edge from SU #" << NewDef->NodeNum
1478           << " to SU #" << TrySU->NodeNum << "\n");
1479     LiveRegDefs[Reg] = NewDef;
1480     AddPred(NewDef, SDep(TrySU, SDep::Artificial));
1481     TrySU->isAvailable = false;
1482     CurSU = NewDef;
1483   }
1484   assert(CurSU && "Unable to resolve live physical register dependencies!");
1485   return CurSU;
1486 }
1487 
1488 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
1489 /// schedulers.
ListScheduleBottomUp()1490 void ScheduleDAGRRList::ListScheduleBottomUp() {
1491   // Release any predecessors of the special Exit node.
1492   ReleasePredecessors(&ExitSU);
1493 
1494   // Add root to Available queue.
1495   if (!SUnits.empty()) {
1496     SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()];
1497     assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!");
1498     RootSU->isAvailable = true;
1499     AvailableQueue->push(RootSU);
1500   }
1501 
1502   // While Available queue is not empty, grab the node with the highest
1503   // priority. If it is not ready put it back.  Schedule the node.
1504   Sequence.reserve(SUnits.size());
1505   while (!AvailableQueue->empty() || !Interferences.empty()) {
1506     DEBUG(dbgs() << "\nExamining Available:\n";
1507           AvailableQueue->dump(this));
1508 
1509     // Pick the best node to schedule taking all constraints into
1510     // consideration.
1511     SUnit *SU = PickNodeToScheduleBottomUp();
1512 
1513     AdvancePastStalls(SU);
1514 
1515     ScheduleNodeBottomUp(SU);
1516 
1517     while (AvailableQueue->empty() && !PendingQueue.empty()) {
1518       // Advance the cycle to free resources. Skip ahead to the next ready SU.
1519       assert(MinAvailableCycle < UINT_MAX && "MinAvailableCycle uninitialized");
1520       AdvanceToCycle(std::max(CurCycle + 1, MinAvailableCycle));
1521     }
1522   }
1523 
1524   // Reverse the order if it is bottom up.
1525   std::reverse(Sequence.begin(), Sequence.end());
1526 
1527 #ifndef NDEBUG
1528   VerifyScheduledSequence(/*isBottomUp=*/true);
1529 #endif
1530 }
1531 
1532 //===----------------------------------------------------------------------===//
1533 //                RegReductionPriorityQueue Definition
1534 //===----------------------------------------------------------------------===//
1535 //
1536 // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers
1537 // to reduce register pressure.
1538 //
1539 namespace {
1540 class RegReductionPQBase;
1541 
1542 struct queue_sort : public std::binary_function<SUnit*, SUnit*, bool> {
isReady__anon884f54d00211::queue_sort1543   bool isReady(SUnit* SU, unsigned CurCycle) const { return true; }
1544 };
1545 
1546 #ifndef NDEBUG
1547 template<class SF>
1548 struct reverse_sort : public queue_sort {
1549   SF &SortFunc;
reverse_sort__anon884f54d00211::reverse_sort1550   reverse_sort(SF &sf) : SortFunc(sf) {}
1551 
operator ()__anon884f54d00211::reverse_sort1552   bool operator()(SUnit* left, SUnit* right) const {
1553     // reverse left/right rather than simply !SortFunc(left, right)
1554     // to expose different paths in the comparison logic.
1555     return SortFunc(right, left);
1556   }
1557 };
1558 #endif // NDEBUG
1559 
1560 /// bu_ls_rr_sort - Priority function for bottom up register pressure
1561 // reduction scheduler.
1562 struct bu_ls_rr_sort : public queue_sort {
1563   enum {
1564     IsBottomUp = true,
1565     HasReadyFilter = false
1566   };
1567 
1568   RegReductionPQBase *SPQ;
bu_ls_rr_sort__anon884f54d00211::bu_ls_rr_sort1569   bu_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
1570 
1571   bool operator()(SUnit* left, SUnit* right) const;
1572 };
1573 
1574 // src_ls_rr_sort - Priority function for source order scheduler.
1575 struct src_ls_rr_sort : public queue_sort {
1576   enum {
1577     IsBottomUp = true,
1578     HasReadyFilter = false
1579   };
1580 
1581   RegReductionPQBase *SPQ;
src_ls_rr_sort__anon884f54d00211::src_ls_rr_sort1582   src_ls_rr_sort(RegReductionPQBase *spq)
1583     : SPQ(spq) {}
1584 
1585   bool operator()(SUnit* left, SUnit* right) const;
1586 };
1587 
1588 // hybrid_ls_rr_sort - Priority function for hybrid scheduler.
1589 struct hybrid_ls_rr_sort : public queue_sort {
1590   enum {
1591     IsBottomUp = true,
1592     HasReadyFilter = false
1593   };
1594 
1595   RegReductionPQBase *SPQ;
hybrid_ls_rr_sort__anon884f54d00211::hybrid_ls_rr_sort1596   hybrid_ls_rr_sort(RegReductionPQBase *spq)
1597     : SPQ(spq) {}
1598 
1599   bool isReady(SUnit *SU, unsigned CurCycle) const;
1600 
1601   bool operator()(SUnit* left, SUnit* right) const;
1602 };
1603 
1604 // ilp_ls_rr_sort - Priority function for ILP (instruction level parallelism)
1605 // scheduler.
1606 struct ilp_ls_rr_sort : public queue_sort {
1607   enum {
1608     IsBottomUp = true,
1609     HasReadyFilter = false
1610   };
1611 
1612   RegReductionPQBase *SPQ;
ilp_ls_rr_sort__anon884f54d00211::ilp_ls_rr_sort1613   ilp_ls_rr_sort(RegReductionPQBase *spq)
1614     : SPQ(spq) {}
1615 
1616   bool isReady(SUnit *SU, unsigned CurCycle) const;
1617 
1618   bool operator()(SUnit* left, SUnit* right) const;
1619 };
1620 
1621 class RegReductionPQBase : public SchedulingPriorityQueue {
1622 protected:
1623   std::vector<SUnit*> Queue;
1624   unsigned CurQueueId;
1625   bool TracksRegPressure;
1626   bool SrcOrder;
1627 
1628   // SUnits - The SUnits for the current graph.
1629   std::vector<SUnit> *SUnits;
1630 
1631   MachineFunction &MF;
1632   const TargetInstrInfo *TII;
1633   const TargetRegisterInfo *TRI;
1634   const TargetLowering *TLI;
1635   ScheduleDAGRRList *scheduleDAG;
1636 
1637   // SethiUllmanNumbers - The SethiUllman number for each node.
1638   std::vector<unsigned> SethiUllmanNumbers;
1639 
1640   /// RegPressure - Tracking current reg pressure per register class.
1641   ///
1642   std::vector<unsigned> RegPressure;
1643 
1644   /// RegLimit - Tracking the number of allocatable registers per register
1645   /// class.
1646   std::vector<unsigned> RegLimit;
1647 
1648 public:
RegReductionPQBase(MachineFunction & mf,bool hasReadyFilter,bool tracksrp,bool srcorder,const TargetInstrInfo * tii,const TargetRegisterInfo * tri,const TargetLowering * tli)1649   RegReductionPQBase(MachineFunction &mf,
1650                      bool hasReadyFilter,
1651                      bool tracksrp,
1652                      bool srcorder,
1653                      const TargetInstrInfo *tii,
1654                      const TargetRegisterInfo *tri,
1655                      const TargetLowering *tli)
1656     : SchedulingPriorityQueue(hasReadyFilter),
1657       CurQueueId(0), TracksRegPressure(tracksrp), SrcOrder(srcorder),
1658       MF(mf), TII(tii), TRI(tri), TLI(tli), scheduleDAG(nullptr) {
1659     if (TracksRegPressure) {
1660       unsigned NumRC = TRI->getNumRegClasses();
1661       RegLimit.resize(NumRC);
1662       RegPressure.resize(NumRC);
1663       std::fill(RegLimit.begin(), RegLimit.end(), 0);
1664       std::fill(RegPressure.begin(), RegPressure.end(), 0);
1665       for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
1666              E = TRI->regclass_end(); I != E; ++I)
1667         RegLimit[(*I)->getID()] = tri->getRegPressureLimit(*I, MF);
1668     }
1669   }
1670 
setScheduleDAG(ScheduleDAGRRList * scheduleDag)1671   void setScheduleDAG(ScheduleDAGRRList *scheduleDag) {
1672     scheduleDAG = scheduleDag;
1673   }
1674 
getHazardRec()1675   ScheduleHazardRecognizer* getHazardRec() {
1676     return scheduleDAG->getHazardRec();
1677   }
1678 
1679   void initNodes(std::vector<SUnit> &sunits) override;
1680 
1681   void addNode(const SUnit *SU) override;
1682 
1683   void updateNode(const SUnit *SU) override;
1684 
releaseState()1685   void releaseState() override {
1686     SUnits = nullptr;
1687     SethiUllmanNumbers.clear();
1688     std::fill(RegPressure.begin(), RegPressure.end(), 0);
1689   }
1690 
1691   unsigned getNodePriority(const SUnit *SU) const;
1692 
getNodeOrdering(const SUnit * SU) const1693   unsigned getNodeOrdering(const SUnit *SU) const {
1694     if (!SU->getNode()) return 0;
1695 
1696     return SU->getNode()->getIROrder();
1697   }
1698 
empty() const1699   bool empty() const override { return Queue.empty(); }
1700 
push(SUnit * U)1701   void push(SUnit *U) override {
1702     assert(!U->NodeQueueId && "Node in the queue already");
1703     U->NodeQueueId = ++CurQueueId;
1704     Queue.push_back(U);
1705   }
1706 
remove(SUnit * SU)1707   void remove(SUnit *SU) override {
1708     assert(!Queue.empty() && "Queue is empty!");
1709     assert(SU->NodeQueueId != 0 && "Not in queue!");
1710     std::vector<SUnit *>::iterator I = std::find(Queue.begin(), Queue.end(),
1711                                                  SU);
1712     if (I != std::prev(Queue.end()))
1713       std::swap(*I, Queue.back());
1714     Queue.pop_back();
1715     SU->NodeQueueId = 0;
1716   }
1717 
tracksRegPressure() const1718   bool tracksRegPressure() const override { return TracksRegPressure; }
1719 
1720   void dumpRegPressure() const;
1721 
1722   bool HighRegPressure(const SUnit *SU) const;
1723 
1724   bool MayReduceRegPressure(SUnit *SU) const;
1725 
1726   int RegPressureDiff(SUnit *SU, unsigned &LiveUses) const;
1727 
1728   void scheduledNode(SUnit *SU) override;
1729 
1730   void unscheduledNode(SUnit *SU) override;
1731 
1732 protected:
1733   bool canClobber(const SUnit *SU, const SUnit *Op);
1734   void AddPseudoTwoAddrDeps();
1735   void PrescheduleNodesWithMultipleUses();
1736   void CalculateSethiUllmanNumbers();
1737 };
1738 
1739 template<class SF>
popFromQueueImpl(std::vector<SUnit * > & Q,SF & Picker)1740 static SUnit *popFromQueueImpl(std::vector<SUnit*> &Q, SF &Picker) {
1741   std::vector<SUnit *>::iterator Best = Q.begin();
1742   for (std::vector<SUnit *>::iterator I = std::next(Q.begin()),
1743          E = Q.end(); I != E; ++I)
1744     if (Picker(*Best, *I))
1745       Best = I;
1746   SUnit *V = *Best;
1747   if (Best != std::prev(Q.end()))
1748     std::swap(*Best, Q.back());
1749   Q.pop_back();
1750   return V;
1751 }
1752 
1753 template<class SF>
popFromQueue(std::vector<SUnit * > & Q,SF & Picker,ScheduleDAG * DAG)1754 SUnit *popFromQueue(std::vector<SUnit*> &Q, SF &Picker, ScheduleDAG *DAG) {
1755 #ifndef NDEBUG
1756   if (DAG->StressSched) {
1757     reverse_sort<SF> RPicker(Picker);
1758     return popFromQueueImpl(Q, RPicker);
1759   }
1760 #endif
1761   (void)DAG;
1762   return popFromQueueImpl(Q, Picker);
1763 }
1764 
1765 template<class SF>
1766 class RegReductionPriorityQueue : public RegReductionPQBase {
1767   SF Picker;
1768 
1769 public:
RegReductionPriorityQueue(MachineFunction & mf,bool tracksrp,bool srcorder,const TargetInstrInfo * tii,const TargetRegisterInfo * tri,const TargetLowering * tli)1770   RegReductionPriorityQueue(MachineFunction &mf,
1771                             bool tracksrp,
1772                             bool srcorder,
1773                             const TargetInstrInfo *tii,
1774                             const TargetRegisterInfo *tri,
1775                             const TargetLowering *tli)
1776     : RegReductionPQBase(mf, SF::HasReadyFilter, tracksrp, srcorder,
1777                          tii, tri, tli),
1778       Picker(this) {}
1779 
isBottomUp() const1780   bool isBottomUp() const override { return SF::IsBottomUp; }
1781 
isReady(SUnit * U) const1782   bool isReady(SUnit *U) const override {
1783     return Picker.HasReadyFilter && Picker.isReady(U, getCurCycle());
1784   }
1785 
pop()1786   SUnit *pop() override {
1787     if (Queue.empty()) return nullptr;
1788 
1789     SUnit *V = popFromQueue(Queue, Picker, scheduleDAG);
1790     V->NodeQueueId = 0;
1791     return V;
1792   }
1793 
1794 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dump(ScheduleDAG * DAG) const1795   void dump(ScheduleDAG *DAG) const override {
1796     // Emulate pop() without clobbering NodeQueueIds.
1797     std::vector<SUnit*> DumpQueue = Queue;
1798     SF DumpPicker = Picker;
1799     while (!DumpQueue.empty()) {
1800       SUnit *SU = popFromQueue(DumpQueue, DumpPicker, scheduleDAG);
1801       dbgs() << "Height " << SU->getHeight() << ": ";
1802       SU->dump(DAG);
1803     }
1804   }
1805 #endif
1806 };
1807 
1808 typedef RegReductionPriorityQueue<bu_ls_rr_sort>
1809 BURegReductionPriorityQueue;
1810 
1811 typedef RegReductionPriorityQueue<src_ls_rr_sort>
1812 SrcRegReductionPriorityQueue;
1813 
1814 typedef RegReductionPriorityQueue<hybrid_ls_rr_sort>
1815 HybridBURRPriorityQueue;
1816 
1817 typedef RegReductionPriorityQueue<ilp_ls_rr_sort>
1818 ILPBURRPriorityQueue;
1819 } // end anonymous namespace
1820 
1821 //===----------------------------------------------------------------------===//
1822 //           Static Node Priority for Register Pressure Reduction
1823 //===----------------------------------------------------------------------===//
1824 
1825 // Check for special nodes that bypass scheduling heuristics.
1826 // Currently this pushes TokenFactor nodes down, but may be used for other
1827 // pseudo-ops as well.
1828 //
1829 // Return -1 to schedule right above left, 1 for left above right.
1830 // Return 0 if no bias exists.
checkSpecialNodes(const SUnit * left,const SUnit * right)1831 static int checkSpecialNodes(const SUnit *left, const SUnit *right) {
1832   bool LSchedLow = left->isScheduleLow;
1833   bool RSchedLow = right->isScheduleLow;
1834   if (LSchedLow != RSchedLow)
1835     return LSchedLow < RSchedLow ? 1 : -1;
1836   return 0;
1837 }
1838 
1839 /// CalcNodeSethiUllmanNumber - Compute Sethi Ullman number.
1840 /// Smaller number is the higher priority.
1841 static unsigned
CalcNodeSethiUllmanNumber(const SUnit * SU,std::vector<unsigned> & SUNumbers)1842 CalcNodeSethiUllmanNumber(const SUnit *SU, std::vector<unsigned> &SUNumbers) {
1843   unsigned &SethiUllmanNumber = SUNumbers[SU->NodeNum];
1844   if (SethiUllmanNumber != 0)
1845     return SethiUllmanNumber;
1846 
1847   unsigned Extra = 0;
1848   for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1849        I != E; ++I) {
1850     if (I->isCtrl()) continue;  // ignore chain preds
1851     SUnit *PredSU = I->getSUnit();
1852     unsigned PredSethiUllman = CalcNodeSethiUllmanNumber(PredSU, SUNumbers);
1853     if (PredSethiUllman > SethiUllmanNumber) {
1854       SethiUllmanNumber = PredSethiUllman;
1855       Extra = 0;
1856     } else if (PredSethiUllman == SethiUllmanNumber)
1857       ++Extra;
1858   }
1859 
1860   SethiUllmanNumber += Extra;
1861 
1862   if (SethiUllmanNumber == 0)
1863     SethiUllmanNumber = 1;
1864 
1865   return SethiUllmanNumber;
1866 }
1867 
1868 /// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all
1869 /// scheduling units.
CalculateSethiUllmanNumbers()1870 void RegReductionPQBase::CalculateSethiUllmanNumbers() {
1871   SethiUllmanNumbers.assign(SUnits->size(), 0);
1872 
1873   for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
1874     CalcNodeSethiUllmanNumber(&(*SUnits)[i], SethiUllmanNumbers);
1875 }
1876 
addNode(const SUnit * SU)1877 void RegReductionPQBase::addNode(const SUnit *SU) {
1878   unsigned SUSize = SethiUllmanNumbers.size();
1879   if (SUnits->size() > SUSize)
1880     SethiUllmanNumbers.resize(SUSize*2, 0);
1881   CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
1882 }
1883 
updateNode(const SUnit * SU)1884 void RegReductionPQBase::updateNode(const SUnit *SU) {
1885   SethiUllmanNumbers[SU->NodeNum] = 0;
1886   CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
1887 }
1888 
1889 // Lower priority means schedule further down. For bottom-up scheduling, lower
1890 // priority SUs are scheduled before higher priority SUs.
getNodePriority(const SUnit * SU) const1891 unsigned RegReductionPQBase::getNodePriority(const SUnit *SU) const {
1892   assert(SU->NodeNum < SethiUllmanNumbers.size());
1893   unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0;
1894   if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
1895     // CopyToReg should be close to its uses to facilitate coalescing and
1896     // avoid spilling.
1897     return 0;
1898   if (Opc == TargetOpcode::EXTRACT_SUBREG ||
1899       Opc == TargetOpcode::SUBREG_TO_REG ||
1900       Opc == TargetOpcode::INSERT_SUBREG)
1901     // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be
1902     // close to their uses to facilitate coalescing.
1903     return 0;
1904   if (SU->NumSuccs == 0 && SU->NumPreds != 0)
1905     // If SU does not have a register use, i.e. it doesn't produce a value
1906     // that would be consumed (e.g. store), then it terminates a chain of
1907     // computation.  Give it a large SethiUllman number so it will be
1908     // scheduled right before its predecessors that it doesn't lengthen
1909     // their live ranges.
1910     return 0xffff;
1911   if (SU->NumPreds == 0 && SU->NumSuccs != 0)
1912     // If SU does not have a register def, schedule it close to its uses
1913     // because it does not lengthen any live ranges.
1914     return 0;
1915 #if 1
1916   return SethiUllmanNumbers[SU->NodeNum];
1917 #else
1918   unsigned Priority = SethiUllmanNumbers[SU->NodeNum];
1919   if (SU->isCallOp) {
1920     // FIXME: This assumes all of the defs are used as call operands.
1921     int NP = (int)Priority - SU->getNode()->getNumValues();
1922     return (NP > 0) ? NP : 0;
1923   }
1924   return Priority;
1925 #endif
1926 }
1927 
1928 //===----------------------------------------------------------------------===//
1929 //                     Register Pressure Tracking
1930 //===----------------------------------------------------------------------===//
1931 
dumpRegPressure() const1932 void RegReductionPQBase::dumpRegPressure() const {
1933 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1934   for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
1935          E = TRI->regclass_end(); I != E; ++I) {
1936     const TargetRegisterClass *RC = *I;
1937     unsigned Id = RC->getID();
1938     unsigned RP = RegPressure[Id];
1939     if (!RP) continue;
1940     DEBUG(dbgs() << TRI->getRegClassName(RC) << ": " << RP << " / "
1941           << RegLimit[Id] << '\n');
1942   }
1943 #endif
1944 }
1945 
HighRegPressure(const SUnit * SU) const1946 bool RegReductionPQBase::HighRegPressure(const SUnit *SU) const {
1947   if (!TLI)
1948     return false;
1949 
1950   for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
1951        I != E; ++I) {
1952     if (I->isCtrl())
1953       continue;
1954     SUnit *PredSU = I->getSUnit();
1955     // NumRegDefsLeft is zero when enough uses of this node have been scheduled
1956     // to cover the number of registers defined (they are all live).
1957     if (PredSU->NumRegDefsLeft == 0) {
1958       continue;
1959     }
1960     for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
1961          RegDefPos.IsValid(); RegDefPos.Advance()) {
1962       unsigned RCId, Cost;
1963       GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
1964 
1965       if ((RegPressure[RCId] + Cost) >= RegLimit[RCId])
1966         return true;
1967     }
1968   }
1969   return false;
1970 }
1971 
MayReduceRegPressure(SUnit * SU) const1972 bool RegReductionPQBase::MayReduceRegPressure(SUnit *SU) const {
1973   const SDNode *N = SU->getNode();
1974 
1975   if (!N->isMachineOpcode() || !SU->NumSuccs)
1976     return false;
1977 
1978   unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
1979   for (unsigned i = 0; i != NumDefs; ++i) {
1980     MVT VT = N->getSimpleValueType(i);
1981     if (!N->hasAnyUseOfValue(i))
1982       continue;
1983     unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1984     if (RegPressure[RCId] >= RegLimit[RCId])
1985       return true;
1986   }
1987   return false;
1988 }
1989 
1990 // Compute the register pressure contribution by this instruction by count up
1991 // for uses that are not live and down for defs. Only count register classes
1992 // that are already under high pressure. As a side effect, compute the number of
1993 // uses of registers that are already live.
1994 //
1995 // FIXME: This encompasses the logic in HighRegPressure and MayReduceRegPressure
1996 // so could probably be factored.
RegPressureDiff(SUnit * SU,unsigned & LiveUses) const1997 int RegReductionPQBase::RegPressureDiff(SUnit *SU, unsigned &LiveUses) const {
1998   LiveUses = 0;
1999   int PDiff = 0;
2000   for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
2001        I != E; ++I) {
2002     if (I->isCtrl())
2003       continue;
2004     SUnit *PredSU = I->getSUnit();
2005     // NumRegDefsLeft is zero when enough uses of this node have been scheduled
2006     // to cover the number of registers defined (they are all live).
2007     if (PredSU->NumRegDefsLeft == 0) {
2008       if (PredSU->getNode()->isMachineOpcode())
2009         ++LiveUses;
2010       continue;
2011     }
2012     for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
2013          RegDefPos.IsValid(); RegDefPos.Advance()) {
2014       MVT VT = RegDefPos.GetValue();
2015       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2016       if (RegPressure[RCId] >= RegLimit[RCId])
2017         ++PDiff;
2018     }
2019   }
2020   const SDNode *N = SU->getNode();
2021 
2022   if (!N || !N->isMachineOpcode() || !SU->NumSuccs)
2023     return PDiff;
2024 
2025   unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2026   for (unsigned i = 0; i != NumDefs; ++i) {
2027     MVT VT = N->getSimpleValueType(i);
2028     if (!N->hasAnyUseOfValue(i))
2029       continue;
2030     unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2031     if (RegPressure[RCId] >= RegLimit[RCId])
2032       --PDiff;
2033   }
2034   return PDiff;
2035 }
2036 
scheduledNode(SUnit * SU)2037 void RegReductionPQBase::scheduledNode(SUnit *SU) {
2038   if (!TracksRegPressure)
2039     return;
2040 
2041   if (!SU->getNode())
2042     return;
2043 
2044   for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
2045        I != E; ++I) {
2046     if (I->isCtrl())
2047       continue;
2048     SUnit *PredSU = I->getSUnit();
2049     // NumRegDefsLeft is zero when enough uses of this node have been scheduled
2050     // to cover the number of registers defined (they are all live).
2051     if (PredSU->NumRegDefsLeft == 0) {
2052       continue;
2053     }
2054     // FIXME: The ScheduleDAG currently loses information about which of a
2055     // node's values is consumed by each dependence. Consequently, if the node
2056     // defines multiple register classes, we don't know which to pressurize
2057     // here. Instead the following loop consumes the register defs in an
2058     // arbitrary order. At least it handles the common case of clustered loads
2059     // to the same class. For precise liveness, each SDep needs to indicate the
2060     // result number. But that tightly couples the ScheduleDAG with the
2061     // SelectionDAG making updates tricky. A simpler hack would be to attach a
2062     // value type or register class to SDep.
2063     //
2064     // The most important aspect of register tracking is balancing the increase
2065     // here with the reduction further below. Note that this SU may use multiple
2066     // defs in PredSU. The can't be determined here, but we've already
2067     // compensated by reducing NumRegDefsLeft in PredSU during
2068     // ScheduleDAGSDNodes::AddSchedEdges.
2069     --PredSU->NumRegDefsLeft;
2070     unsigned SkipRegDefs = PredSU->NumRegDefsLeft;
2071     for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
2072          RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
2073       if (SkipRegDefs)
2074         continue;
2075 
2076       unsigned RCId, Cost;
2077       GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
2078       RegPressure[RCId] += Cost;
2079       break;
2080     }
2081   }
2082 
2083   // We should have this assert, but there may be dead SDNodes that never
2084   // materialize as SUnits, so they don't appear to generate liveness.
2085   //assert(SU->NumRegDefsLeft == 0 && "not all regdefs have scheduled uses");
2086   int SkipRegDefs = (int)SU->NumRegDefsLeft;
2087   for (ScheduleDAGSDNodes::RegDefIter RegDefPos(SU, scheduleDAG);
2088        RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
2089     if (SkipRegDefs > 0)
2090       continue;
2091     unsigned RCId, Cost;
2092     GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
2093     if (RegPressure[RCId] < Cost) {
2094       // Register pressure tracking is imprecise. This can happen. But we try
2095       // hard not to let it happen because it likely results in poor scheduling.
2096       DEBUG(dbgs() << "  SU(" << SU->NodeNum << ") has too many regdefs\n");
2097       RegPressure[RCId] = 0;
2098     }
2099     else {
2100       RegPressure[RCId] -= Cost;
2101     }
2102   }
2103   dumpRegPressure();
2104 }
2105 
unscheduledNode(SUnit * SU)2106 void RegReductionPQBase::unscheduledNode(SUnit *SU) {
2107   if (!TracksRegPressure)
2108     return;
2109 
2110   const SDNode *N = SU->getNode();
2111   if (!N) return;
2112 
2113   if (!N->isMachineOpcode()) {
2114     if (N->getOpcode() != ISD::CopyToReg)
2115       return;
2116   } else {
2117     unsigned Opc = N->getMachineOpcode();
2118     if (Opc == TargetOpcode::EXTRACT_SUBREG ||
2119         Opc == TargetOpcode::INSERT_SUBREG ||
2120         Opc == TargetOpcode::SUBREG_TO_REG ||
2121         Opc == TargetOpcode::REG_SEQUENCE ||
2122         Opc == TargetOpcode::IMPLICIT_DEF)
2123       return;
2124   }
2125 
2126   for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
2127        I != E; ++I) {
2128     if (I->isCtrl())
2129       continue;
2130     SUnit *PredSU = I->getSUnit();
2131     // NumSuccsLeft counts all deps. Don't compare it with NumSuccs which only
2132     // counts data deps.
2133     if (PredSU->NumSuccsLeft != PredSU->Succs.size())
2134       continue;
2135     const SDNode *PN = PredSU->getNode();
2136     if (!PN->isMachineOpcode()) {
2137       if (PN->getOpcode() == ISD::CopyFromReg) {
2138         MVT VT = PN->getSimpleValueType(0);
2139         unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2140         RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2141       }
2142       continue;
2143     }
2144     unsigned POpc = PN->getMachineOpcode();
2145     if (POpc == TargetOpcode::IMPLICIT_DEF)
2146       continue;
2147     if (POpc == TargetOpcode::EXTRACT_SUBREG ||
2148         POpc == TargetOpcode::INSERT_SUBREG ||
2149         POpc == TargetOpcode::SUBREG_TO_REG) {
2150       MVT VT = PN->getSimpleValueType(0);
2151       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2152       RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2153       continue;
2154     }
2155     unsigned NumDefs = TII->get(PN->getMachineOpcode()).getNumDefs();
2156     for (unsigned i = 0; i != NumDefs; ++i) {
2157       MVT VT = PN->getSimpleValueType(i);
2158       if (!PN->hasAnyUseOfValue(i))
2159         continue;
2160       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2161       if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT))
2162         // Register pressure tracking is imprecise. This can happen.
2163         RegPressure[RCId] = 0;
2164       else
2165         RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
2166     }
2167   }
2168 
2169   // Check for isMachineOpcode() as PrescheduleNodesWithMultipleUses()
2170   // may transfer data dependencies to CopyToReg.
2171   if (SU->NumSuccs && N->isMachineOpcode()) {
2172     unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2173     for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
2174       MVT VT = N->getSimpleValueType(i);
2175       if (VT == MVT::Glue || VT == MVT::Other)
2176         continue;
2177       if (!N->hasAnyUseOfValue(i))
2178         continue;
2179       unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2180       RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2181     }
2182   }
2183 
2184   dumpRegPressure();
2185 }
2186 
2187 //===----------------------------------------------------------------------===//
2188 //           Dynamic Node Priority for Register Pressure Reduction
2189 //===----------------------------------------------------------------------===//
2190 
2191 /// closestSucc - Returns the scheduled cycle of the successor which is
2192 /// closest to the current cycle.
closestSucc(const SUnit * SU)2193 static unsigned closestSucc(const SUnit *SU) {
2194   unsigned MaxHeight = 0;
2195   for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
2196        I != E; ++I) {
2197     if (I->isCtrl()) continue;  // ignore chain succs
2198     unsigned Height = I->getSUnit()->getHeight();
2199     // If there are bunch of CopyToRegs stacked up, they should be considered
2200     // to be at the same position.
2201     if (I->getSUnit()->getNode() &&
2202         I->getSUnit()->getNode()->getOpcode() == ISD::CopyToReg)
2203       Height = closestSucc(I->getSUnit())+1;
2204     if (Height > MaxHeight)
2205       MaxHeight = Height;
2206   }
2207   return MaxHeight;
2208 }
2209 
2210 /// calcMaxScratches - Returns an cost estimate of the worse case requirement
2211 /// for scratch registers, i.e. number of data dependencies.
calcMaxScratches(const SUnit * SU)2212 static unsigned calcMaxScratches(const SUnit *SU) {
2213   unsigned Scratches = 0;
2214   for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
2215        I != E; ++I) {
2216     if (I->isCtrl()) continue;  // ignore chain preds
2217     Scratches++;
2218   }
2219   return Scratches;
2220 }
2221 
2222 /// hasOnlyLiveInOpers - Return true if SU has only value predecessors that are
2223 /// CopyFromReg from a virtual register.
hasOnlyLiveInOpers(const SUnit * SU)2224 static bool hasOnlyLiveInOpers(const SUnit *SU) {
2225   bool RetVal = false;
2226   for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
2227        I != E; ++I) {
2228     if (I->isCtrl()) continue;
2229     const SUnit *PredSU = I->getSUnit();
2230     if (PredSU->getNode() &&
2231         PredSU->getNode()->getOpcode() == ISD::CopyFromReg) {
2232       unsigned Reg =
2233         cast<RegisterSDNode>(PredSU->getNode()->getOperand(1))->getReg();
2234       if (TargetRegisterInfo::isVirtualRegister(Reg)) {
2235         RetVal = true;
2236         continue;
2237       }
2238     }
2239     return false;
2240   }
2241   return RetVal;
2242 }
2243 
2244 /// hasOnlyLiveOutUses - Return true if SU has only value successors that are
2245 /// CopyToReg to a virtual register. This SU def is probably a liveout and
2246 /// it has no other use. It should be scheduled closer to the terminator.
hasOnlyLiveOutUses(const SUnit * SU)2247 static bool hasOnlyLiveOutUses(const SUnit *SU) {
2248   bool RetVal = false;
2249   for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
2250        I != E; ++I) {
2251     if (I->isCtrl()) continue;
2252     const SUnit *SuccSU = I->getSUnit();
2253     if (SuccSU->getNode() && SuccSU->getNode()->getOpcode() == ISD::CopyToReg) {
2254       unsigned Reg =
2255         cast<RegisterSDNode>(SuccSU->getNode()->getOperand(1))->getReg();
2256       if (TargetRegisterInfo::isVirtualRegister(Reg)) {
2257         RetVal = true;
2258         continue;
2259       }
2260     }
2261     return false;
2262   }
2263   return RetVal;
2264 }
2265 
2266 // Set isVRegCycle for a node with only live in opers and live out uses. Also
2267 // set isVRegCycle for its CopyFromReg operands.
2268 //
2269 // This is only relevant for single-block loops, in which case the VRegCycle
2270 // node is likely an induction variable in which the operand and target virtual
2271 // registers should be coalesced (e.g. pre/post increment values). Setting the
2272 // isVRegCycle flag helps the scheduler prioritize other uses of the same
2273 // CopyFromReg so that this node becomes the virtual register "kill". This
2274 // avoids interference between the values live in and out of the block and
2275 // eliminates a copy inside the loop.
initVRegCycle(SUnit * SU)2276 static void initVRegCycle(SUnit *SU) {
2277   if (DisableSchedVRegCycle)
2278     return;
2279 
2280   if (!hasOnlyLiveInOpers(SU) || !hasOnlyLiveOutUses(SU))
2281     return;
2282 
2283   DEBUG(dbgs() << "VRegCycle: SU(" << SU->NodeNum << ")\n");
2284 
2285   SU->isVRegCycle = true;
2286 
2287   for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
2288        I != E; ++I) {
2289     if (I->isCtrl()) continue;
2290     I->getSUnit()->isVRegCycle = true;
2291   }
2292 }
2293 
2294 // After scheduling the definition of a VRegCycle, clear the isVRegCycle flag of
2295 // CopyFromReg operands. We should no longer penalize other uses of this VReg.
resetVRegCycle(SUnit * SU)2296 static void resetVRegCycle(SUnit *SU) {
2297   if (!SU->isVRegCycle)
2298     return;
2299 
2300   for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
2301        I != E; ++I) {
2302     if (I->isCtrl()) continue;  // ignore chain preds
2303     SUnit *PredSU = I->getSUnit();
2304     if (PredSU->isVRegCycle) {
2305       assert(PredSU->getNode()->getOpcode() == ISD::CopyFromReg &&
2306              "VRegCycle def must be CopyFromReg");
2307       I->getSUnit()->isVRegCycle = 0;
2308     }
2309   }
2310 }
2311 
2312 // Return true if this SUnit uses a CopyFromReg node marked as a VRegCycle. This
2313 // means a node that defines the VRegCycle has not been scheduled yet.
hasVRegCycleUse(const SUnit * SU)2314 static bool hasVRegCycleUse(const SUnit *SU) {
2315   // If this SU also defines the VReg, don't hoist it as a "use".
2316   if (SU->isVRegCycle)
2317     return false;
2318 
2319   for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
2320        I != E; ++I) {
2321     if (I->isCtrl()) continue;  // ignore chain preds
2322     if (I->getSUnit()->isVRegCycle &&
2323         I->getSUnit()->getNode()->getOpcode() == ISD::CopyFromReg) {
2324       DEBUG(dbgs() << "  VReg cycle use: SU (" << SU->NodeNum << ")\n");
2325       return true;
2326     }
2327   }
2328   return false;
2329 }
2330 
2331 // Check for either a dependence (latency) or resource (hazard) stall.
2332 //
2333 // Note: The ScheduleHazardRecognizer interface requires a non-const SU.
BUHasStall(SUnit * SU,int Height,RegReductionPQBase * SPQ)2334 static bool BUHasStall(SUnit *SU, int Height, RegReductionPQBase *SPQ) {
2335   if ((int)SPQ->getCurCycle() < Height) return true;
2336   if (SPQ->getHazardRec()->getHazardType(SU, 0)
2337       != ScheduleHazardRecognizer::NoHazard)
2338     return true;
2339   return false;
2340 }
2341 
2342 // Return -1 if left has higher priority, 1 if right has higher priority.
2343 // Return 0 if latency-based priority is equivalent.
BUCompareLatency(SUnit * left,SUnit * right,bool checkPref,RegReductionPQBase * SPQ)2344 static int BUCompareLatency(SUnit *left, SUnit *right, bool checkPref,
2345                             RegReductionPQBase *SPQ) {
2346   // Scheduling an instruction that uses a VReg whose postincrement has not yet
2347   // been scheduled will induce a copy. Model this as an extra cycle of latency.
2348   int LPenalty = hasVRegCycleUse(left) ? 1 : 0;
2349   int RPenalty = hasVRegCycleUse(right) ? 1 : 0;
2350   int LHeight = (int)left->getHeight() + LPenalty;
2351   int RHeight = (int)right->getHeight() + RPenalty;
2352 
2353   bool LStall = (!checkPref || left->SchedulingPref == Sched::ILP) &&
2354     BUHasStall(left, LHeight, SPQ);
2355   bool RStall = (!checkPref || right->SchedulingPref == Sched::ILP) &&
2356     BUHasStall(right, RHeight, SPQ);
2357 
2358   // If scheduling one of the node will cause a pipeline stall, delay it.
2359   // If scheduling either one of the node will cause a pipeline stall, sort
2360   // them according to their height.
2361   if (LStall) {
2362     if (!RStall)
2363       return 1;
2364     if (LHeight != RHeight)
2365       return LHeight > RHeight ? 1 : -1;
2366   } else if (RStall)
2367     return -1;
2368 
2369   // If either node is scheduling for latency, sort them by height/depth
2370   // and latency.
2371   if (!checkPref || (left->SchedulingPref == Sched::ILP ||
2372                      right->SchedulingPref == Sched::ILP)) {
2373     // If neither instruction stalls (!LStall && !RStall) and HazardRecognizer
2374     // is enabled, grouping instructions by cycle, then its height is already
2375     // covered so only its depth matters. We also reach this point if both stall
2376     // but have the same height.
2377     if (!SPQ->getHazardRec()->isEnabled()) {
2378       if (LHeight != RHeight)
2379         return LHeight > RHeight ? 1 : -1;
2380     }
2381     int LDepth = left->getDepth() - LPenalty;
2382     int RDepth = right->getDepth() - RPenalty;
2383     if (LDepth != RDepth) {
2384       DEBUG(dbgs() << "  Comparing latency of SU (" << left->NodeNum
2385             << ") depth " << LDepth << " vs SU (" << right->NodeNum
2386             << ") depth " << RDepth << "\n");
2387       return LDepth < RDepth ? 1 : -1;
2388     }
2389     if (left->Latency != right->Latency)
2390       return left->Latency > right->Latency ? 1 : -1;
2391   }
2392   return 0;
2393 }
2394 
BURRSort(SUnit * left,SUnit * right,RegReductionPQBase * SPQ)2395 static bool BURRSort(SUnit *left, SUnit *right, RegReductionPQBase *SPQ) {
2396   // Schedule physical register definitions close to their use. This is
2397   // motivated by microarchitectures that can fuse cmp+jump macro-ops. But as
2398   // long as shortening physreg live ranges is generally good, we can defer
2399   // creating a subtarget hook.
2400   if (!DisableSchedPhysRegJoin) {
2401     bool LHasPhysReg = left->hasPhysRegDefs;
2402     bool RHasPhysReg = right->hasPhysRegDefs;
2403     if (LHasPhysReg != RHasPhysReg) {
2404       #ifndef NDEBUG
2405       static const char *const PhysRegMsg[] = { " has no physreg",
2406                                                 " defines a physreg" };
2407       #endif
2408       DEBUG(dbgs() << "  SU (" << left->NodeNum << ") "
2409             << PhysRegMsg[LHasPhysReg] << " SU(" << right->NodeNum << ") "
2410             << PhysRegMsg[RHasPhysReg] << "\n");
2411       return LHasPhysReg < RHasPhysReg;
2412     }
2413   }
2414 
2415   // Prioritize by Sethi-Ulmann number and push CopyToReg nodes down.
2416   unsigned LPriority = SPQ->getNodePriority(left);
2417   unsigned RPriority = SPQ->getNodePriority(right);
2418 
2419   // Be really careful about hoisting call operands above previous calls.
2420   // Only allows it if it would reduce register pressure.
2421   if (left->isCall && right->isCallOp) {
2422     unsigned RNumVals = right->getNode()->getNumValues();
2423     RPriority = (RPriority > RNumVals) ? (RPriority - RNumVals) : 0;
2424   }
2425   if (right->isCall && left->isCallOp) {
2426     unsigned LNumVals = left->getNode()->getNumValues();
2427     LPriority = (LPriority > LNumVals) ? (LPriority - LNumVals) : 0;
2428   }
2429 
2430   if (LPriority != RPriority)
2431     return LPriority > RPriority;
2432 
2433   // One or both of the nodes are calls and their sethi-ullman numbers are the
2434   // same, then keep source order.
2435   if (left->isCall || right->isCall) {
2436     unsigned LOrder = SPQ->getNodeOrdering(left);
2437     unsigned ROrder = SPQ->getNodeOrdering(right);
2438 
2439     // Prefer an ordering where the lower the non-zero order number, the higher
2440     // the preference.
2441     if ((LOrder || ROrder) && LOrder != ROrder)
2442       return LOrder != 0 && (LOrder < ROrder || ROrder == 0);
2443   }
2444 
2445   // Try schedule def + use closer when Sethi-Ullman numbers are the same.
2446   // e.g.
2447   // t1 = op t2, c1
2448   // t3 = op t4, c2
2449   //
2450   // and the following instructions are both ready.
2451   // t2 = op c3
2452   // t4 = op c4
2453   //
2454   // Then schedule t2 = op first.
2455   // i.e.
2456   // t4 = op c4
2457   // t2 = op c3
2458   // t1 = op t2, c1
2459   // t3 = op t4, c2
2460   //
2461   // This creates more short live intervals.
2462   unsigned LDist = closestSucc(left);
2463   unsigned RDist = closestSucc(right);
2464   if (LDist != RDist)
2465     return LDist < RDist;
2466 
2467   // How many registers becomes live when the node is scheduled.
2468   unsigned LScratch = calcMaxScratches(left);
2469   unsigned RScratch = calcMaxScratches(right);
2470   if (LScratch != RScratch)
2471     return LScratch > RScratch;
2472 
2473   // Comparing latency against a call makes little sense unless the node
2474   // is register pressure-neutral.
2475   if ((left->isCall && RPriority > 0) || (right->isCall && LPriority > 0))
2476     return (left->NodeQueueId > right->NodeQueueId);
2477 
2478   // Do not compare latencies when one or both of the nodes are calls.
2479   if (!DisableSchedCycles &&
2480       !(left->isCall || right->isCall)) {
2481     int result = BUCompareLatency(left, right, false /*checkPref*/, SPQ);
2482     if (result != 0)
2483       return result > 0;
2484   }
2485   else {
2486     if (left->getHeight() != right->getHeight())
2487       return left->getHeight() > right->getHeight();
2488 
2489     if (left->getDepth() != right->getDepth())
2490       return left->getDepth() < right->getDepth();
2491   }
2492 
2493   assert(left->NodeQueueId && right->NodeQueueId &&
2494          "NodeQueueId cannot be zero");
2495   return (left->NodeQueueId > right->NodeQueueId);
2496 }
2497 
2498 // Bottom up
operator ()(SUnit * left,SUnit * right) const2499 bool bu_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2500   if (int res = checkSpecialNodes(left, right))
2501     return res > 0;
2502 
2503   return BURRSort(left, right, SPQ);
2504 }
2505 
2506 // Source order, otherwise bottom up.
operator ()(SUnit * left,SUnit * right) const2507 bool src_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2508   if (int res = checkSpecialNodes(left, right))
2509     return res > 0;
2510 
2511   unsigned LOrder = SPQ->getNodeOrdering(left);
2512   unsigned ROrder = SPQ->getNodeOrdering(right);
2513 
2514   // Prefer an ordering where the lower the non-zero order number, the higher
2515   // the preference.
2516   if ((LOrder || ROrder) && LOrder != ROrder)
2517     return LOrder != 0 && (LOrder < ROrder || ROrder == 0);
2518 
2519   return BURRSort(left, right, SPQ);
2520 }
2521 
2522 // If the time between now and when the instruction will be ready can cover
2523 // the spill code, then avoid adding it to the ready queue. This gives long
2524 // stalls highest priority and allows hoisting across calls. It should also
2525 // speed up processing the available queue.
isReady(SUnit * SU,unsigned CurCycle) const2526 bool hybrid_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
2527   static const unsigned ReadyDelay = 3;
2528 
2529   if (SPQ->MayReduceRegPressure(SU)) return true;
2530 
2531   if (SU->getHeight() > (CurCycle + ReadyDelay)) return false;
2532 
2533   if (SPQ->getHazardRec()->getHazardType(SU, -ReadyDelay)
2534       != ScheduleHazardRecognizer::NoHazard)
2535     return false;
2536 
2537   return true;
2538 }
2539 
2540 // Return true if right should be scheduled with higher priority than left.
operator ()(SUnit * left,SUnit * right) const2541 bool hybrid_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2542   if (int res = checkSpecialNodes(left, right))
2543     return res > 0;
2544 
2545   if (left->isCall || right->isCall)
2546     // No way to compute latency of calls.
2547     return BURRSort(left, right, SPQ);
2548 
2549   bool LHigh = SPQ->HighRegPressure(left);
2550   bool RHigh = SPQ->HighRegPressure(right);
2551   // Avoid causing spills. If register pressure is high, schedule for
2552   // register pressure reduction.
2553   if (LHigh && !RHigh) {
2554     DEBUG(dbgs() << "  pressure SU(" << left->NodeNum << ") > SU("
2555           << right->NodeNum << ")\n");
2556     return true;
2557   }
2558   else if (!LHigh && RHigh) {
2559     DEBUG(dbgs() << "  pressure SU(" << right->NodeNum << ") > SU("
2560           << left->NodeNum << ")\n");
2561     return false;
2562   }
2563   if (!LHigh && !RHigh) {
2564     int result = BUCompareLatency(left, right, true /*checkPref*/, SPQ);
2565     if (result != 0)
2566       return result > 0;
2567   }
2568   return BURRSort(left, right, SPQ);
2569 }
2570 
2571 // Schedule as many instructions in each cycle as possible. So don't make an
2572 // instruction available unless it is ready in the current cycle.
isReady(SUnit * SU,unsigned CurCycle) const2573 bool ilp_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
2574   if (SU->getHeight() > CurCycle) return false;
2575 
2576   if (SPQ->getHazardRec()->getHazardType(SU, 0)
2577       != ScheduleHazardRecognizer::NoHazard)
2578     return false;
2579 
2580   return true;
2581 }
2582 
canEnableCoalescing(SUnit * SU)2583 static bool canEnableCoalescing(SUnit *SU) {
2584   unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0;
2585   if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
2586     // CopyToReg should be close to its uses to facilitate coalescing and
2587     // avoid spilling.
2588     return true;
2589 
2590   if (Opc == TargetOpcode::EXTRACT_SUBREG ||
2591       Opc == TargetOpcode::SUBREG_TO_REG ||
2592       Opc == TargetOpcode::INSERT_SUBREG)
2593     // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be
2594     // close to their uses to facilitate coalescing.
2595     return true;
2596 
2597   if (SU->NumPreds == 0 && SU->NumSuccs != 0)
2598     // If SU does not have a register def, schedule it close to its uses
2599     // because it does not lengthen any live ranges.
2600     return true;
2601 
2602   return false;
2603 }
2604 
2605 // list-ilp is currently an experimental scheduler that allows various
2606 // heuristics to be enabled prior to the normal register reduction logic.
operator ()(SUnit * left,SUnit * right) const2607 bool ilp_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2608   if (int res = checkSpecialNodes(left, right))
2609     return res > 0;
2610 
2611   if (left->isCall || right->isCall)
2612     // No way to compute latency of calls.
2613     return BURRSort(left, right, SPQ);
2614 
2615   unsigned LLiveUses = 0, RLiveUses = 0;
2616   int LPDiff = 0, RPDiff = 0;
2617   if (!DisableSchedRegPressure || !DisableSchedLiveUses) {
2618     LPDiff = SPQ->RegPressureDiff(left, LLiveUses);
2619     RPDiff = SPQ->RegPressureDiff(right, RLiveUses);
2620   }
2621   if (!DisableSchedRegPressure && LPDiff != RPDiff) {
2622     DEBUG(dbgs() << "RegPressureDiff SU(" << left->NodeNum << "): " << LPDiff
2623           << " != SU(" << right->NodeNum << "): " << RPDiff << "\n");
2624     return LPDiff > RPDiff;
2625   }
2626 
2627   if (!DisableSchedRegPressure && (LPDiff > 0 || RPDiff > 0)) {
2628     bool LReduce = canEnableCoalescing(left);
2629     bool RReduce = canEnableCoalescing(right);
2630     if (LReduce && !RReduce) return false;
2631     if (RReduce && !LReduce) return true;
2632   }
2633 
2634   if (!DisableSchedLiveUses && (LLiveUses != RLiveUses)) {
2635     DEBUG(dbgs() << "Live uses SU(" << left->NodeNum << "): " << LLiveUses
2636           << " != SU(" << right->NodeNum << "): " << RLiveUses << "\n");
2637     return LLiveUses < RLiveUses;
2638   }
2639 
2640   if (!DisableSchedStalls) {
2641     bool LStall = BUHasStall(left, left->getHeight(), SPQ);
2642     bool RStall = BUHasStall(right, right->getHeight(), SPQ);
2643     if (LStall != RStall)
2644       return left->getHeight() > right->getHeight();
2645   }
2646 
2647   if (!DisableSchedCriticalPath) {
2648     int spread = (int)left->getDepth() - (int)right->getDepth();
2649     if (std::abs(spread) > MaxReorderWindow) {
2650       DEBUG(dbgs() << "Depth of SU(" << left->NodeNum << "): "
2651             << left->getDepth() << " != SU(" << right->NodeNum << "): "
2652             << right->getDepth() << "\n");
2653       return left->getDepth() < right->getDepth();
2654     }
2655   }
2656 
2657   if (!DisableSchedHeight && left->getHeight() != right->getHeight()) {
2658     int spread = (int)left->getHeight() - (int)right->getHeight();
2659     if (std::abs(spread) > MaxReorderWindow)
2660       return left->getHeight() > right->getHeight();
2661   }
2662 
2663   return BURRSort(left, right, SPQ);
2664 }
2665 
initNodes(std::vector<SUnit> & sunits)2666 void RegReductionPQBase::initNodes(std::vector<SUnit> &sunits) {
2667   SUnits = &sunits;
2668   // Add pseudo dependency edges for two-address nodes.
2669   if (!Disable2AddrHack)
2670     AddPseudoTwoAddrDeps();
2671   // Reroute edges to nodes with multiple uses.
2672   if (!TracksRegPressure && !SrcOrder)
2673     PrescheduleNodesWithMultipleUses();
2674   // Calculate node priorities.
2675   CalculateSethiUllmanNumbers();
2676 
2677   // For single block loops, mark nodes that look like canonical IV increments.
2678   if (scheduleDAG->BB->isSuccessor(scheduleDAG->BB)) {
2679     for (unsigned i = 0, e = sunits.size(); i != e; ++i) {
2680       initVRegCycle(&sunits[i]);
2681     }
2682   }
2683 }
2684 
2685 //===----------------------------------------------------------------------===//
2686 //                    Preschedule for Register Pressure
2687 //===----------------------------------------------------------------------===//
2688 
canClobber(const SUnit * SU,const SUnit * Op)2689 bool RegReductionPQBase::canClobber(const SUnit *SU, const SUnit *Op) {
2690   if (SU->isTwoAddress) {
2691     unsigned Opc = SU->getNode()->getMachineOpcode();
2692     const MCInstrDesc &MCID = TII->get(Opc);
2693     unsigned NumRes = MCID.getNumDefs();
2694     unsigned NumOps = MCID.getNumOperands() - NumRes;
2695     for (unsigned i = 0; i != NumOps; ++i) {
2696       if (MCID.getOperandConstraint(i+NumRes, MCOI::TIED_TO) != -1) {
2697         SDNode *DU = SU->getNode()->getOperand(i).getNode();
2698         if (DU->getNodeId() != -1 &&
2699             Op->OrigNode == &(*SUnits)[DU->getNodeId()])
2700           return true;
2701       }
2702     }
2703   }
2704   return false;
2705 }
2706 
2707 /// canClobberReachingPhysRegUse - True if SU would clobber one of it's
2708 /// successor's explicit physregs whose definition can reach DepSU.
2709 /// i.e. DepSU should not be scheduled above SU.
canClobberReachingPhysRegUse(const SUnit * DepSU,const SUnit * SU,ScheduleDAGRRList * scheduleDAG,const TargetInstrInfo * TII,const TargetRegisterInfo * TRI)2710 static bool canClobberReachingPhysRegUse(const SUnit *DepSU, const SUnit *SU,
2711                                          ScheduleDAGRRList *scheduleDAG,
2712                                          const TargetInstrInfo *TII,
2713                                          const TargetRegisterInfo *TRI) {
2714   const uint16_t *ImpDefs
2715     = TII->get(SU->getNode()->getMachineOpcode()).getImplicitDefs();
2716   const uint32_t *RegMask = getNodeRegMask(SU->getNode());
2717   if(!ImpDefs && !RegMask)
2718     return false;
2719 
2720   for (SUnit::const_succ_iterator SI = SU->Succs.begin(), SE = SU->Succs.end();
2721        SI != SE; ++SI) {
2722     SUnit *SuccSU = SI->getSUnit();
2723     for (SUnit::const_pred_iterator PI = SuccSU->Preds.begin(),
2724            PE = SuccSU->Preds.end(); PI != PE; ++PI) {
2725       if (!PI->isAssignedRegDep())
2726         continue;
2727 
2728       if (RegMask && MachineOperand::clobbersPhysReg(RegMask, PI->getReg()) &&
2729           scheduleDAG->IsReachable(DepSU, PI->getSUnit()))
2730         return true;
2731 
2732       if (ImpDefs)
2733         for (const uint16_t *ImpDef = ImpDefs; *ImpDef; ++ImpDef)
2734           // Return true if SU clobbers this physical register use and the
2735           // definition of the register reaches from DepSU. IsReachable queries
2736           // a topological forward sort of the DAG (following the successors).
2737           if (TRI->regsOverlap(*ImpDef, PI->getReg()) &&
2738               scheduleDAG->IsReachable(DepSU, PI->getSUnit()))
2739             return true;
2740     }
2741   }
2742   return false;
2743 }
2744 
2745 /// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's
2746 /// physical register defs.
canClobberPhysRegDefs(const SUnit * SuccSU,const SUnit * SU,const TargetInstrInfo * TII,const TargetRegisterInfo * TRI)2747 static bool canClobberPhysRegDefs(const SUnit *SuccSU, const SUnit *SU,
2748                                   const TargetInstrInfo *TII,
2749                                   const TargetRegisterInfo *TRI) {
2750   SDNode *N = SuccSU->getNode();
2751   unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2752   const uint16_t *ImpDefs = TII->get(N->getMachineOpcode()).getImplicitDefs();
2753   assert(ImpDefs && "Caller should check hasPhysRegDefs");
2754   for (const SDNode *SUNode = SU->getNode(); SUNode;
2755        SUNode = SUNode->getGluedNode()) {
2756     if (!SUNode->isMachineOpcode())
2757       continue;
2758     const uint16_t *SUImpDefs =
2759       TII->get(SUNode->getMachineOpcode()).getImplicitDefs();
2760     const uint32_t *SURegMask = getNodeRegMask(SUNode);
2761     if (!SUImpDefs && !SURegMask)
2762       continue;
2763     for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
2764       MVT VT = N->getSimpleValueType(i);
2765       if (VT == MVT::Glue || VT == MVT::Other)
2766         continue;
2767       if (!N->hasAnyUseOfValue(i))
2768         continue;
2769       unsigned Reg = ImpDefs[i - NumDefs];
2770       if (SURegMask && MachineOperand::clobbersPhysReg(SURegMask, Reg))
2771         return true;
2772       if (!SUImpDefs)
2773         continue;
2774       for (;*SUImpDefs; ++SUImpDefs) {
2775         unsigned SUReg = *SUImpDefs;
2776         if (TRI->regsOverlap(Reg, SUReg))
2777           return true;
2778       }
2779     }
2780   }
2781   return false;
2782 }
2783 
2784 /// PrescheduleNodesWithMultipleUses - Nodes with multiple uses
2785 /// are not handled well by the general register pressure reduction
2786 /// heuristics. When presented with code like this:
2787 ///
2788 ///      N
2789 ///    / |
2790 ///   /  |
2791 ///  U  store
2792 ///  |
2793 /// ...
2794 ///
2795 /// the heuristics tend to push the store up, but since the
2796 /// operand of the store has another use (U), this would increase
2797 /// the length of that other use (the U->N edge).
2798 ///
2799 /// This function transforms code like the above to route U's
2800 /// dependence through the store when possible, like this:
2801 ///
2802 ///      N
2803 ///      ||
2804 ///      ||
2805 ///     store
2806 ///       |
2807 ///       U
2808 ///       |
2809 ///      ...
2810 ///
2811 /// This results in the store being scheduled immediately
2812 /// after N, which shortens the U->N live range, reducing
2813 /// register pressure.
2814 ///
PrescheduleNodesWithMultipleUses()2815 void RegReductionPQBase::PrescheduleNodesWithMultipleUses() {
2816   // Visit all the nodes in topological order, working top-down.
2817   for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
2818     SUnit *SU = &(*SUnits)[i];
2819     // For now, only look at nodes with no data successors, such as stores.
2820     // These are especially important, due to the heuristics in
2821     // getNodePriority for nodes with no data successors.
2822     if (SU->NumSuccs != 0)
2823       continue;
2824     // For now, only look at nodes with exactly one data predecessor.
2825     if (SU->NumPreds != 1)
2826       continue;
2827     // Avoid prescheduling copies to virtual registers, which don't behave
2828     // like other nodes from the perspective of scheduling heuristics.
2829     if (SDNode *N = SU->getNode())
2830       if (N->getOpcode() == ISD::CopyToReg &&
2831           TargetRegisterInfo::isVirtualRegister
2832             (cast<RegisterSDNode>(N->getOperand(1))->getReg()))
2833         continue;
2834 
2835     // Locate the single data predecessor.
2836     SUnit *PredSU = nullptr;
2837     for (SUnit::const_pred_iterator II = SU->Preds.begin(),
2838          EE = SU->Preds.end(); II != EE; ++II)
2839       if (!II->isCtrl()) {
2840         PredSU = II->getSUnit();
2841         break;
2842       }
2843     assert(PredSU);
2844 
2845     // Don't rewrite edges that carry physregs, because that requires additional
2846     // support infrastructure.
2847     if (PredSU->hasPhysRegDefs)
2848       continue;
2849     // Short-circuit the case where SU is PredSU's only data successor.
2850     if (PredSU->NumSuccs == 1)
2851       continue;
2852     // Avoid prescheduling to copies from virtual registers, which don't behave
2853     // like other nodes from the perspective of scheduling heuristics.
2854     if (SDNode *N = SU->getNode())
2855       if (N->getOpcode() == ISD::CopyFromReg &&
2856           TargetRegisterInfo::isVirtualRegister
2857             (cast<RegisterSDNode>(N->getOperand(1))->getReg()))
2858         continue;
2859 
2860     // Perform checks on the successors of PredSU.
2861     for (SUnit::const_succ_iterator II = PredSU->Succs.begin(),
2862          EE = PredSU->Succs.end(); II != EE; ++II) {
2863       SUnit *PredSuccSU = II->getSUnit();
2864       if (PredSuccSU == SU) continue;
2865       // If PredSU has another successor with no data successors, for
2866       // now don't attempt to choose either over the other.
2867       if (PredSuccSU->NumSuccs == 0)
2868         goto outer_loop_continue;
2869       // Don't break physical register dependencies.
2870       if (SU->hasPhysRegClobbers && PredSuccSU->hasPhysRegDefs)
2871         if (canClobberPhysRegDefs(PredSuccSU, SU, TII, TRI))
2872           goto outer_loop_continue;
2873       // Don't introduce graph cycles.
2874       if (scheduleDAG->IsReachable(SU, PredSuccSU))
2875         goto outer_loop_continue;
2876     }
2877 
2878     // Ok, the transformation is safe and the heuristics suggest it is
2879     // profitable. Update the graph.
2880     DEBUG(dbgs() << "    Prescheduling SU #" << SU->NodeNum
2881                  << " next to PredSU #" << PredSU->NodeNum
2882                  << " to guide scheduling in the presence of multiple uses\n");
2883     for (unsigned i = 0; i != PredSU->Succs.size(); ++i) {
2884       SDep Edge = PredSU->Succs[i];
2885       assert(!Edge.isAssignedRegDep());
2886       SUnit *SuccSU = Edge.getSUnit();
2887       if (SuccSU != SU) {
2888         Edge.setSUnit(PredSU);
2889         scheduleDAG->RemovePred(SuccSU, Edge);
2890         scheduleDAG->AddPred(SU, Edge);
2891         Edge.setSUnit(SU);
2892         scheduleDAG->AddPred(SuccSU, Edge);
2893         --i;
2894       }
2895     }
2896   outer_loop_continue:;
2897   }
2898 }
2899 
2900 /// AddPseudoTwoAddrDeps - If two nodes share an operand and one of them uses
2901 /// it as a def&use operand. Add a pseudo control edge from it to the other
2902 /// node (if it won't create a cycle) so the two-address one will be scheduled
2903 /// first (lower in the schedule). If both nodes are two-address, favor the
2904 /// one that has a CopyToReg use (more likely to be a loop induction update).
2905 /// If both are two-address, but one is commutable while the other is not
2906 /// commutable, favor the one that's not commutable.
AddPseudoTwoAddrDeps()2907 void RegReductionPQBase::AddPseudoTwoAddrDeps() {
2908   for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
2909     SUnit *SU = &(*SUnits)[i];
2910     if (!SU->isTwoAddress)
2911       continue;
2912 
2913     SDNode *Node = SU->getNode();
2914     if (!Node || !Node->isMachineOpcode() || SU->getNode()->getGluedNode())
2915       continue;
2916 
2917     bool isLiveOut = hasOnlyLiveOutUses(SU);
2918     unsigned Opc = Node->getMachineOpcode();
2919     const MCInstrDesc &MCID = TII->get(Opc);
2920     unsigned NumRes = MCID.getNumDefs();
2921     unsigned NumOps = MCID.getNumOperands() - NumRes;
2922     for (unsigned j = 0; j != NumOps; ++j) {
2923       if (MCID.getOperandConstraint(j+NumRes, MCOI::TIED_TO) == -1)
2924         continue;
2925       SDNode *DU = SU->getNode()->getOperand(j).getNode();
2926       if (DU->getNodeId() == -1)
2927         continue;
2928       const SUnit *DUSU = &(*SUnits)[DU->getNodeId()];
2929       if (!DUSU) continue;
2930       for (SUnit::const_succ_iterator I = DUSU->Succs.begin(),
2931            E = DUSU->Succs.end(); I != E; ++I) {
2932         if (I->isCtrl()) continue;
2933         SUnit *SuccSU = I->getSUnit();
2934         if (SuccSU == SU)
2935           continue;
2936         // Be conservative. Ignore if nodes aren't at roughly the same
2937         // depth and height.
2938         if (SuccSU->getHeight() < SU->getHeight() &&
2939             (SU->getHeight() - SuccSU->getHeight()) > 1)
2940           continue;
2941         // Skip past COPY_TO_REGCLASS nodes, so that the pseudo edge
2942         // constrains whatever is using the copy, instead of the copy
2943         // itself. In the case that the copy is coalesced, this
2944         // preserves the intent of the pseudo two-address heurietics.
2945         while (SuccSU->Succs.size() == 1 &&
2946                SuccSU->getNode()->isMachineOpcode() &&
2947                SuccSU->getNode()->getMachineOpcode() ==
2948                  TargetOpcode::COPY_TO_REGCLASS)
2949           SuccSU = SuccSU->Succs.front().getSUnit();
2950         // Don't constrain non-instruction nodes.
2951         if (!SuccSU->getNode() || !SuccSU->getNode()->isMachineOpcode())
2952           continue;
2953         // Don't constrain nodes with physical register defs if the
2954         // predecessor can clobber them.
2955         if (SuccSU->hasPhysRegDefs && SU->hasPhysRegClobbers) {
2956           if (canClobberPhysRegDefs(SuccSU, SU, TII, TRI))
2957             continue;
2958         }
2959         // Don't constrain EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG;
2960         // these may be coalesced away. We want them close to their uses.
2961         unsigned SuccOpc = SuccSU->getNode()->getMachineOpcode();
2962         if (SuccOpc == TargetOpcode::EXTRACT_SUBREG ||
2963             SuccOpc == TargetOpcode::INSERT_SUBREG ||
2964             SuccOpc == TargetOpcode::SUBREG_TO_REG)
2965           continue;
2966         if (!canClobberReachingPhysRegUse(SuccSU, SU, scheduleDAG, TII, TRI) &&
2967             (!canClobber(SuccSU, DUSU) ||
2968              (isLiveOut && !hasOnlyLiveOutUses(SuccSU)) ||
2969              (!SU->isCommutable && SuccSU->isCommutable)) &&
2970             !scheduleDAG->IsReachable(SuccSU, SU)) {
2971           DEBUG(dbgs() << "    Adding a pseudo-two-addr edge from SU #"
2972                        << SU->NodeNum << " to SU #" << SuccSU->NodeNum << "\n");
2973           scheduleDAG->AddPred(SU, SDep(SuccSU, SDep::Artificial));
2974         }
2975       }
2976     }
2977   }
2978 }
2979 
2980 //===----------------------------------------------------------------------===//
2981 //                         Public Constructor Functions
2982 //===----------------------------------------------------------------------===//
2983 
2984 llvm::ScheduleDAGSDNodes *
createBURRListDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level OptLevel)2985 llvm::createBURRListDAGScheduler(SelectionDAGISel *IS,
2986                                  CodeGenOpt::Level OptLevel) {
2987   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
2988   const TargetInstrInfo *TII = STI.getInstrInfo();
2989   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
2990 
2991   BURegReductionPriorityQueue *PQ =
2992     new BURegReductionPriorityQueue(*IS->MF, false, false, TII, TRI, nullptr);
2993   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
2994   PQ->setScheduleDAG(SD);
2995   return SD;
2996 }
2997 
2998 llvm::ScheduleDAGSDNodes *
createSourceListDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level OptLevel)2999 llvm::createSourceListDAGScheduler(SelectionDAGISel *IS,
3000                                    CodeGenOpt::Level OptLevel) {
3001   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3002   const TargetInstrInfo *TII = STI.getInstrInfo();
3003   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3004 
3005   SrcRegReductionPriorityQueue *PQ =
3006     new SrcRegReductionPriorityQueue(*IS->MF, false, true, TII, TRI, nullptr);
3007   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
3008   PQ->setScheduleDAG(SD);
3009   return SD;
3010 }
3011 
3012 llvm::ScheduleDAGSDNodes *
createHybridListDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level OptLevel)3013 llvm::createHybridListDAGScheduler(SelectionDAGISel *IS,
3014                                    CodeGenOpt::Level OptLevel) {
3015   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3016   const TargetInstrInfo *TII = STI.getInstrInfo();
3017   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3018   const TargetLowering *TLI = IS->TLI;
3019 
3020   HybridBURRPriorityQueue *PQ =
3021     new HybridBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI);
3022 
3023   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
3024   PQ->setScheduleDAG(SD);
3025   return SD;
3026 }
3027 
3028 llvm::ScheduleDAGSDNodes *
createILPListDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level OptLevel)3029 llvm::createILPListDAGScheduler(SelectionDAGISel *IS,
3030                                 CodeGenOpt::Level OptLevel) {
3031   const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3032   const TargetInstrInfo *TII = STI.getInstrInfo();
3033   const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3034   const TargetLowering *TLI = IS->TLI;
3035 
3036   ILPBURRPriorityQueue *PQ =
3037     new ILPBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI);
3038   ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
3039   PQ->setScheduleDAG(SD);
3040   return SD;
3041 }
3042