• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===----- ScheduleDAGFast.cpp - Fast poor 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 a fast scheduler.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "InstrEmitter.h"
15 #include "ScheduleDAGSDNodes.h"
16 #include "SDNodeDbgValue.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/CodeGen/SchedulerRegistry.h"
21 #include "llvm/CodeGen/SelectionDAGISel.h"
22 #include "llvm/CodeGen/TargetInstrInfo.h"
23 #include "llvm/CodeGen/TargetRegisterInfo.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/InlineAsm.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/ErrorHandling.h"
28 #include "llvm/Support/raw_ostream.h"
29 using namespace llvm;
30 
31 #define DEBUG_TYPE "pre-RA-sched"
32 
33 STATISTIC(NumUnfolds,    "Number of nodes unfolded");
34 STATISTIC(NumDups,       "Number of duplicated nodes");
35 STATISTIC(NumPRCopies,   "Number of physical copies");
36 
37 static RegisterScheduler
38   fastDAGScheduler("fast", "Fast suboptimal list scheduling",
39                    createFastDAGScheduler);
40 static RegisterScheduler
41   linearizeDAGScheduler("linearize", "Linearize DAG, no scheduling",
42                         createDAGLinearizer);
43 
44 
45 namespace {
46   /// FastPriorityQueue - A degenerate priority queue that considers
47   /// all nodes to have the same priority.
48   ///
49   struct FastPriorityQueue {
50     SmallVector<SUnit *, 16> Queue;
51 
empty__anon53d202c10111::FastPriorityQueue52     bool empty() const { return Queue.empty(); }
53 
push__anon53d202c10111::FastPriorityQueue54     void push(SUnit *U) {
55       Queue.push_back(U);
56     }
57 
pop__anon53d202c10111::FastPriorityQueue58     SUnit *pop() {
59       if (empty()) return nullptr;
60       SUnit *V = Queue.back();
61       Queue.pop_back();
62       return V;
63     }
64   };
65 
66 //===----------------------------------------------------------------------===//
67 /// ScheduleDAGFast - The actual "fast" list scheduler implementation.
68 ///
69 class ScheduleDAGFast : public ScheduleDAGSDNodes {
70 private:
71   /// AvailableQueue - The priority queue to use for the available SUnits.
72   FastPriorityQueue AvailableQueue;
73 
74   /// LiveRegDefs - A set of physical registers and their definition
75   /// that are "live". These nodes must be scheduled before any other nodes that
76   /// modifies the registers can be scheduled.
77   unsigned NumLiveRegs;
78   std::vector<SUnit*> LiveRegDefs;
79   std::vector<unsigned> LiveRegCycles;
80 
81 public:
ScheduleDAGFast(MachineFunction & mf)82   ScheduleDAGFast(MachineFunction &mf)
83     : ScheduleDAGSDNodes(mf) {}
84 
85   void Schedule() override;
86 
87   /// AddPred - adds a predecessor edge to SUnit SU.
88   /// This returns true if this is a new predecessor.
AddPred(SUnit * SU,const SDep & D)89   void AddPred(SUnit *SU, const SDep &D) {
90     SU->addPred(D);
91   }
92 
93   /// RemovePred - removes a predecessor edge from SUnit SU.
94   /// This returns true if an edge was removed.
RemovePred(SUnit * SU,const SDep & D)95   void RemovePred(SUnit *SU, const SDep &D) {
96     SU->removePred(D);
97   }
98 
99 private:
100   void ReleasePred(SUnit *SU, SDep *PredEdge);
101   void ReleasePredecessors(SUnit *SU, unsigned CurCycle);
102   void ScheduleNodeBottomUp(SUnit*, unsigned);
103   SUnit *CopyAndMoveSuccessors(SUnit*);
104   void InsertCopiesAndMoveSuccs(SUnit*, unsigned,
105                                 const TargetRegisterClass*,
106                                 const TargetRegisterClass*,
107                                 SmallVectorImpl<SUnit*>&);
108   bool DelayForLiveRegsBottomUp(SUnit*, SmallVectorImpl<unsigned>&);
109   void ListScheduleBottomUp();
110 
111   /// forceUnitLatencies - The fast scheduler doesn't care about real latencies.
forceUnitLatencies() const112   bool forceUnitLatencies() const override { return true; }
113 };
114 }  // end anonymous namespace
115 
116 
117 /// Schedule - Schedule the DAG using list scheduling.
Schedule()118 void ScheduleDAGFast::Schedule() {
119   LLVM_DEBUG(dbgs() << "********** List Scheduling **********\n");
120 
121   NumLiveRegs = 0;
122   LiveRegDefs.resize(TRI->getNumRegs(), nullptr);
123   LiveRegCycles.resize(TRI->getNumRegs(), 0);
124 
125   // Build the scheduling graph.
126   BuildSchedGraph(nullptr);
127 
128   LLVM_DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su) SUnits[su]
129                  .dumpAll(this));
130 
131   // Execute the actual scheduling loop.
132   ListScheduleBottomUp();
133 }
134 
135 //===----------------------------------------------------------------------===//
136 //  Bottom-Up Scheduling
137 //===----------------------------------------------------------------------===//
138 
139 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
140 /// the AvailableQueue if the count reaches zero. Also update its cycle bound.
ReleasePred(SUnit * SU,SDep * PredEdge)141 void ScheduleDAGFast::ReleasePred(SUnit *SU, SDep *PredEdge) {
142   SUnit *PredSU = PredEdge->getSUnit();
143 
144 #ifndef NDEBUG
145   if (PredSU->NumSuccsLeft == 0) {
146     dbgs() << "*** Scheduling failed! ***\n";
147     PredSU->dump(this);
148     dbgs() << " has been released too many times!\n";
149     llvm_unreachable(nullptr);
150   }
151 #endif
152   --PredSU->NumSuccsLeft;
153 
154   // If all the node's successors are scheduled, this node is ready
155   // to be scheduled. Ignore the special EntrySU node.
156   if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU) {
157     PredSU->isAvailable = true;
158     AvailableQueue.push(PredSU);
159   }
160 }
161 
ReleasePredecessors(SUnit * SU,unsigned CurCycle)162 void ScheduleDAGFast::ReleasePredecessors(SUnit *SU, unsigned CurCycle) {
163   // Bottom up: release predecessors
164   for (SDep &Pred : SU->Preds) {
165     ReleasePred(SU, &Pred);
166     if (Pred.isAssignedRegDep()) {
167       // This is a physical register dependency and it's impossible or
168       // expensive to copy the register. Make sure nothing that can
169       // clobber the register is scheduled between the predecessor and
170       // this node.
171       if (!LiveRegDefs[Pred.getReg()]) {
172         ++NumLiveRegs;
173         LiveRegDefs[Pred.getReg()] = Pred.getSUnit();
174         LiveRegCycles[Pred.getReg()] = CurCycle;
175       }
176     }
177   }
178 }
179 
180 /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
181 /// count of its predecessors. If a predecessor pending count is zero, add it to
182 /// the Available queue.
ScheduleNodeBottomUp(SUnit * SU,unsigned CurCycle)183 void ScheduleDAGFast::ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle) {
184   LLVM_DEBUG(dbgs() << "*** Scheduling [" << CurCycle << "]: ");
185   LLVM_DEBUG(SU->dump(this));
186 
187   assert(CurCycle >= SU->getHeight() && "Node scheduled below its height!");
188   SU->setHeightToAtLeast(CurCycle);
189   Sequence.push_back(SU);
190 
191   ReleasePredecessors(SU, CurCycle);
192 
193   // Release all the implicit physical register defs that are live.
194   for (SDep &Succ : SU->Succs) {
195     if (Succ.isAssignedRegDep()) {
196       if (LiveRegCycles[Succ.getReg()] == Succ.getSUnit()->getHeight()) {
197         assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
198         assert(LiveRegDefs[Succ.getReg()] == SU &&
199                "Physical register dependency violated?");
200         --NumLiveRegs;
201         LiveRegDefs[Succ.getReg()] = nullptr;
202         LiveRegCycles[Succ.getReg()] = 0;
203       }
204     }
205   }
206 
207   SU->isScheduled = true;
208 }
209 
210 /// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
211 /// successors to the newly created node.
CopyAndMoveSuccessors(SUnit * SU)212 SUnit *ScheduleDAGFast::CopyAndMoveSuccessors(SUnit *SU) {
213   if (SU->getNode()->getGluedNode())
214     return nullptr;
215 
216   SDNode *N = SU->getNode();
217   if (!N)
218     return nullptr;
219 
220   SUnit *NewSU;
221   bool TryUnfold = false;
222   for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
223     MVT VT = N->getSimpleValueType(i);
224     if (VT == MVT::Glue)
225       return nullptr;
226     else if (VT == MVT::Other)
227       TryUnfold = true;
228   }
229   for (const SDValue &Op : N->op_values()) {
230     MVT VT = Op.getNode()->getSimpleValueType(Op.getResNo());
231     if (VT == MVT::Glue)
232       return nullptr;
233   }
234 
235   if (TryUnfold) {
236     SmallVector<SDNode*, 2> NewNodes;
237     if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
238       return nullptr;
239 
240     LLVM_DEBUG(dbgs() << "Unfolding SU # " << SU->NodeNum << "\n");
241     assert(NewNodes.size() == 2 && "Expected a load folding node!");
242 
243     N = NewNodes[1];
244     SDNode *LoadNode = NewNodes[0];
245     unsigned NumVals = N->getNumValues();
246     unsigned OldNumVals = SU->getNode()->getNumValues();
247     for (unsigned i = 0; i != NumVals; ++i)
248       DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i));
249     DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals-1),
250                                    SDValue(LoadNode, 1));
251 
252     SUnit *NewSU = newSUnit(N);
253     assert(N->getNodeId() == -1 && "Node already inserted!");
254     N->setNodeId(NewSU->NodeNum);
255 
256     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
257     for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
258       if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
259         NewSU->isTwoAddress = true;
260         break;
261       }
262     }
263     if (MCID.isCommutable())
264       NewSU->isCommutable = true;
265 
266     // LoadNode may already exist. This can happen when there is another
267     // load from the same location and producing the same type of value
268     // but it has different alignment or volatileness.
269     bool isNewLoad = true;
270     SUnit *LoadSU;
271     if (LoadNode->getNodeId() != -1) {
272       LoadSU = &SUnits[LoadNode->getNodeId()];
273       isNewLoad = false;
274     } else {
275       LoadSU = newSUnit(LoadNode);
276       LoadNode->setNodeId(LoadSU->NodeNum);
277     }
278 
279     SDep ChainPred;
280     SmallVector<SDep, 4> ChainSuccs;
281     SmallVector<SDep, 4> LoadPreds;
282     SmallVector<SDep, 4> NodePreds;
283     SmallVector<SDep, 4> NodeSuccs;
284     for (SDep &Pred : SU->Preds) {
285       if (Pred.isCtrl())
286         ChainPred = Pred;
287       else if (Pred.getSUnit()->getNode() &&
288                Pred.getSUnit()->getNode()->isOperandOf(LoadNode))
289         LoadPreds.push_back(Pred);
290       else
291         NodePreds.push_back(Pred);
292     }
293     for (SDep &Succ : SU->Succs) {
294       if (Succ.isCtrl())
295         ChainSuccs.push_back(Succ);
296       else
297         NodeSuccs.push_back(Succ);
298     }
299 
300     if (ChainPred.getSUnit()) {
301       RemovePred(SU, ChainPred);
302       if (isNewLoad)
303         AddPred(LoadSU, ChainPred);
304     }
305     for (unsigned i = 0, e = LoadPreds.size(); i != e; ++i) {
306       const SDep &Pred = LoadPreds[i];
307       RemovePred(SU, Pred);
308       if (isNewLoad) {
309         AddPred(LoadSU, Pred);
310       }
311     }
312     for (unsigned i = 0, e = NodePreds.size(); i != e; ++i) {
313       const SDep &Pred = NodePreds[i];
314       RemovePred(SU, Pred);
315       AddPred(NewSU, Pred);
316     }
317     for (unsigned i = 0, e = NodeSuccs.size(); i != e; ++i) {
318       SDep D = NodeSuccs[i];
319       SUnit *SuccDep = D.getSUnit();
320       D.setSUnit(SU);
321       RemovePred(SuccDep, D);
322       D.setSUnit(NewSU);
323       AddPred(SuccDep, D);
324     }
325     for (unsigned i = 0, e = ChainSuccs.size(); i != e; ++i) {
326       SDep D = ChainSuccs[i];
327       SUnit *SuccDep = D.getSUnit();
328       D.setSUnit(SU);
329       RemovePred(SuccDep, D);
330       if (isNewLoad) {
331         D.setSUnit(LoadSU);
332         AddPred(SuccDep, D);
333       }
334     }
335     if (isNewLoad) {
336       SDep D(LoadSU, SDep::Barrier);
337       D.setLatency(LoadSU->Latency);
338       AddPred(NewSU, D);
339     }
340 
341     ++NumUnfolds;
342 
343     if (NewSU->NumSuccsLeft == 0) {
344       NewSU->isAvailable = true;
345       return NewSU;
346     }
347     SU = NewSU;
348   }
349 
350   LLVM_DEBUG(dbgs() << "Duplicating SU # " << SU->NodeNum << "\n");
351   NewSU = Clone(SU);
352 
353   // New SUnit has the exact same predecessors.
354   for (SDep &Pred : SU->Preds)
355     if (!Pred.isArtificial())
356       AddPred(NewSU, Pred);
357 
358   // Only copy scheduled successors. Cut them from old node's successor
359   // list and move them over.
360   SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
361   for (SDep &Succ : SU->Succs) {
362     if (Succ.isArtificial())
363       continue;
364     SUnit *SuccSU = Succ.getSUnit();
365     if (SuccSU->isScheduled) {
366       SDep D = Succ;
367       D.setSUnit(NewSU);
368       AddPred(SuccSU, D);
369       D.setSUnit(SU);
370       DelDeps.push_back(std::make_pair(SuccSU, D));
371     }
372   }
373   for (unsigned i = 0, e = DelDeps.size(); i != e; ++i)
374     RemovePred(DelDeps[i].first, DelDeps[i].second);
375 
376   ++NumDups;
377   return NewSU;
378 }
379 
380 /// InsertCopiesAndMoveSuccs - Insert register copies and move all
381 /// scheduled successors of the given SUnit to the last copy.
InsertCopiesAndMoveSuccs(SUnit * SU,unsigned Reg,const TargetRegisterClass * DestRC,const TargetRegisterClass * SrcRC,SmallVectorImpl<SUnit * > & Copies)382 void ScheduleDAGFast::InsertCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
383                                               const TargetRegisterClass *DestRC,
384                                               const TargetRegisterClass *SrcRC,
385                                               SmallVectorImpl<SUnit*> &Copies) {
386   SUnit *CopyFromSU = newSUnit(static_cast<SDNode *>(nullptr));
387   CopyFromSU->CopySrcRC = SrcRC;
388   CopyFromSU->CopyDstRC = DestRC;
389 
390   SUnit *CopyToSU = newSUnit(static_cast<SDNode *>(nullptr));
391   CopyToSU->CopySrcRC = DestRC;
392   CopyToSU->CopyDstRC = SrcRC;
393 
394   // Only copy scheduled successors. Cut them from old node's successor
395   // list and move them over.
396   SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
397   for (SDep &Succ : SU->Succs) {
398     if (Succ.isArtificial())
399       continue;
400     SUnit *SuccSU = Succ.getSUnit();
401     if (SuccSU->isScheduled) {
402       SDep D = Succ;
403       D.setSUnit(CopyToSU);
404       AddPred(SuccSU, D);
405       DelDeps.push_back(std::make_pair(SuccSU, Succ));
406     }
407   }
408   for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) {
409     RemovePred(DelDeps[i].first, DelDeps[i].second);
410   }
411   SDep FromDep(SU, SDep::Data, Reg);
412   FromDep.setLatency(SU->Latency);
413   AddPred(CopyFromSU, FromDep);
414   SDep ToDep(CopyFromSU, SDep::Data, 0);
415   ToDep.setLatency(CopyFromSU->Latency);
416   AddPred(CopyToSU, ToDep);
417 
418   Copies.push_back(CopyFromSU);
419   Copies.push_back(CopyToSU);
420 
421   ++NumPRCopies;
422 }
423 
424 /// getPhysicalRegisterVT - Returns the ValueType of the physical register
425 /// definition of the specified node.
426 /// FIXME: Move to SelectionDAG?
getPhysicalRegisterVT(SDNode * N,unsigned Reg,const TargetInstrInfo * TII)427 static MVT getPhysicalRegisterVT(SDNode *N, unsigned Reg,
428                                  const TargetInstrInfo *TII) {
429   unsigned NumRes;
430   if (N->getOpcode() == ISD::CopyFromReg) {
431     // CopyFromReg has: "chain, Val, glue" so operand 1 gives the type.
432     NumRes = 1;
433   } else {
434     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
435     assert(MCID.ImplicitDefs && "Physical reg def must be in implicit def list!");
436     NumRes = MCID.getNumDefs();
437     for (const MCPhysReg *ImpDef = MCID.getImplicitDefs(); *ImpDef; ++ImpDef) {
438       if (Reg == *ImpDef)
439         break;
440       ++NumRes;
441     }
442   }
443   return N->getSimpleValueType(NumRes);
444 }
445 
446 /// CheckForLiveRegDef - Return true and update live register vector if the
447 /// 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)448 static bool CheckForLiveRegDef(SUnit *SU, unsigned Reg,
449                                std::vector<SUnit*> &LiveRegDefs,
450                                SmallSet<unsigned, 4> &RegAdded,
451                                SmallVectorImpl<unsigned> &LRegs,
452                                const TargetRegisterInfo *TRI) {
453   bool Added = false;
454   for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) {
455     if (LiveRegDefs[*AI] && LiveRegDefs[*AI] != SU) {
456       if (RegAdded.insert(*AI).second) {
457         LRegs.push_back(*AI);
458         Added = true;
459       }
460     }
461   }
462   return Added;
463 }
464 
465 /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
466 /// scheduling of the given node to satisfy live physical register dependencies.
467 /// If the specific node is the last one that's available to schedule, do
468 /// whatever is necessary (i.e. backtracking or cloning) to make it possible.
DelayForLiveRegsBottomUp(SUnit * SU,SmallVectorImpl<unsigned> & LRegs)469 bool ScheduleDAGFast::DelayForLiveRegsBottomUp(SUnit *SU,
470                                               SmallVectorImpl<unsigned> &LRegs){
471   if (NumLiveRegs == 0)
472     return false;
473 
474   SmallSet<unsigned, 4> RegAdded;
475   // If this node would clobber any "live" register, then it's not ready.
476   for (SDep &Pred : SU->Preds) {
477     if (Pred.isAssignedRegDep()) {
478       CheckForLiveRegDef(Pred.getSUnit(), Pred.getReg(), LiveRegDefs,
479                          RegAdded, LRegs, TRI);
480     }
481   }
482 
483   for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) {
484     if (Node->getOpcode() == ISD::INLINEASM) {
485       // Inline asm can clobber physical defs.
486       unsigned NumOps = Node->getNumOperands();
487       if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
488         --NumOps;  // Ignore the glue operand.
489 
490       for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
491         unsigned Flags =
492           cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
493         unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
494 
495         ++i; // Skip the ID value.
496         if (InlineAsm::isRegDefKind(Flags) ||
497             InlineAsm::isRegDefEarlyClobberKind(Flags) ||
498             InlineAsm::isClobberKind(Flags)) {
499           // Check for def of register or earlyclobber register.
500           for (; NumVals; --NumVals, ++i) {
501             unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
502             if (TargetRegisterInfo::isPhysicalRegister(Reg))
503               CheckForLiveRegDef(SU, Reg, LiveRegDefs, RegAdded, LRegs, TRI);
504           }
505         } else
506           i += NumVals;
507       }
508       continue;
509     }
510     if (!Node->isMachineOpcode())
511       continue;
512     const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode());
513     if (!MCID.ImplicitDefs)
514       continue;
515     for (const MCPhysReg *Reg = MCID.getImplicitDefs(); *Reg; ++Reg) {
516       CheckForLiveRegDef(SU, *Reg, LiveRegDefs, RegAdded, LRegs, TRI);
517     }
518   }
519   return !LRegs.empty();
520 }
521 
522 
523 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
524 /// schedulers.
ListScheduleBottomUp()525 void ScheduleDAGFast::ListScheduleBottomUp() {
526   unsigned CurCycle = 0;
527 
528   // Release any predecessors of the special Exit node.
529   ReleasePredecessors(&ExitSU, CurCycle);
530 
531   // Add root to Available queue.
532   if (!SUnits.empty()) {
533     SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()];
534     assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!");
535     RootSU->isAvailable = true;
536     AvailableQueue.push(RootSU);
537   }
538 
539   // While Available queue is not empty, grab the node with the highest
540   // priority. If it is not ready put it back.  Schedule the node.
541   SmallVector<SUnit*, 4> NotReady;
542   DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMap;
543   Sequence.reserve(SUnits.size());
544   while (!AvailableQueue.empty()) {
545     bool Delayed = false;
546     LRegsMap.clear();
547     SUnit *CurSU = AvailableQueue.pop();
548     while (CurSU) {
549       SmallVector<unsigned, 4> LRegs;
550       if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
551         break;
552       Delayed = true;
553       LRegsMap.insert(std::make_pair(CurSU, LRegs));
554 
555       CurSU->isPending = true;  // This SU is not in AvailableQueue right now.
556       NotReady.push_back(CurSU);
557       CurSU = AvailableQueue.pop();
558     }
559 
560     // All candidates are delayed due to live physical reg dependencies.
561     // Try code duplication or inserting cross class copies
562     // to resolve it.
563     if (Delayed && !CurSU) {
564       if (!CurSU) {
565         // Try duplicating the nodes that produces these
566         // "expensive to copy" values to break the dependency. In case even
567         // that doesn't work, insert cross class copies.
568         SUnit *TrySU = NotReady[0];
569         SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
570         assert(LRegs.size() == 1 && "Can't handle this yet!");
571         unsigned Reg = LRegs[0];
572         SUnit *LRDef = LiveRegDefs[Reg];
573         MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
574         const TargetRegisterClass *RC =
575           TRI->getMinimalPhysRegClass(Reg, VT);
576         const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
577 
578         // If cross copy register class is the same as RC, then it must be
579         // possible copy the value directly. Do not try duplicate the def.
580         // If cross copy register class is not the same as RC, then it's
581         // possible to copy the value but it require cross register class copies
582         // and it is expensive.
583         // If cross copy register class is null, then it's not possible to copy
584         // the value at all.
585         SUnit *NewDef = nullptr;
586         if (DestRC != RC) {
587           NewDef = CopyAndMoveSuccessors(LRDef);
588           if (!DestRC && !NewDef)
589             report_fatal_error("Can't handle live physical "
590                                "register dependency!");
591         }
592         if (!NewDef) {
593           // Issue copies, these can be expensive cross register class copies.
594           SmallVector<SUnit*, 2> Copies;
595           InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
596           LLVM_DEBUG(dbgs() << "Adding an edge from SU # " << TrySU->NodeNum
597                             << " to SU #" << Copies.front()->NodeNum << "\n");
598           AddPred(TrySU, SDep(Copies.front(), SDep::Artificial));
599           NewDef = Copies.back();
600         }
601 
602         LLVM_DEBUG(dbgs() << "Adding an edge from SU # " << NewDef->NodeNum
603                           << " to SU #" << TrySU->NodeNum << "\n");
604         LiveRegDefs[Reg] = NewDef;
605         AddPred(NewDef, SDep(TrySU, SDep::Artificial));
606         TrySU->isAvailable = false;
607         CurSU = NewDef;
608       }
609 
610       if (!CurSU) {
611         llvm_unreachable("Unable to resolve live physical register dependencies!");
612       }
613     }
614 
615     // Add the nodes that aren't ready back onto the available list.
616     for (unsigned i = 0, e = NotReady.size(); i != e; ++i) {
617       NotReady[i]->isPending = false;
618       // May no longer be available due to backtracking.
619       if (NotReady[i]->isAvailable)
620         AvailableQueue.push(NotReady[i]);
621     }
622     NotReady.clear();
623 
624     if (CurSU)
625       ScheduleNodeBottomUp(CurSU, CurCycle);
626     ++CurCycle;
627   }
628 
629   // Reverse the order since it is bottom up.
630   std::reverse(Sequence.begin(), Sequence.end());
631 
632 #ifndef NDEBUG
633   VerifyScheduledSequence(/*isBottomUp=*/true);
634 #endif
635 }
636 
637 
638 namespace {
639 //===----------------------------------------------------------------------===//
640 // ScheduleDAGLinearize - No scheduling scheduler, it simply linearize the
641 // DAG in topological order.
642 // IMPORTANT: this may not work for targets with phyreg dependency.
643 //
644 class ScheduleDAGLinearize : public ScheduleDAGSDNodes {
645 public:
ScheduleDAGLinearize(MachineFunction & mf)646   ScheduleDAGLinearize(MachineFunction &mf) : ScheduleDAGSDNodes(mf) {}
647 
648   void Schedule() override;
649 
650   MachineBasicBlock *
651     EmitSchedule(MachineBasicBlock::iterator &InsertPos) override;
652 
653 private:
654   std::vector<SDNode*> Sequence;
655   DenseMap<SDNode*, SDNode*> GluedMap;  // Cache glue to its user
656 
657   void ScheduleNode(SDNode *N);
658 };
659 } // end anonymous namespace
660 
ScheduleNode(SDNode * N)661 void ScheduleDAGLinearize::ScheduleNode(SDNode *N) {
662   if (N->getNodeId() != 0)
663     llvm_unreachable(nullptr);
664 
665   if (!N->isMachineOpcode() &&
666       (N->getOpcode() == ISD::EntryToken || isPassiveNode(N)))
667     // These nodes do not need to be translated into MIs.
668     return;
669 
670   LLVM_DEBUG(dbgs() << "\n*** Scheduling: ");
671   LLVM_DEBUG(N->dump(DAG));
672   Sequence.push_back(N);
673 
674   unsigned NumOps = N->getNumOperands();
675   if (unsigned NumLeft = NumOps) {
676     SDNode *GluedOpN = nullptr;
677     do {
678       const SDValue &Op = N->getOperand(NumLeft-1);
679       SDNode *OpN = Op.getNode();
680 
681       if (NumLeft == NumOps && Op.getValueType() == MVT::Glue) {
682         // Schedule glue operand right above N.
683         GluedOpN = OpN;
684         assert(OpN->getNodeId() != 0 && "Glue operand not ready?");
685         OpN->setNodeId(0);
686         ScheduleNode(OpN);
687         continue;
688       }
689 
690       if (OpN == GluedOpN)
691         // Glue operand is already scheduled.
692         continue;
693 
694       DenseMap<SDNode*, SDNode*>::iterator DI = GluedMap.find(OpN);
695       if (DI != GluedMap.end() && DI->second != N)
696         // Users of glues are counted against the glued users.
697         OpN = DI->second;
698 
699       unsigned Degree = OpN->getNodeId();
700       assert(Degree > 0 && "Predecessor over-released!");
701       OpN->setNodeId(--Degree);
702       if (Degree == 0)
703         ScheduleNode(OpN);
704     } while (--NumLeft);
705   }
706 }
707 
708 /// findGluedUser - Find the representative use of a glue value by walking
709 /// the use chain.
findGluedUser(SDNode * N)710 static SDNode *findGluedUser(SDNode *N) {
711   while (SDNode *Glued = N->getGluedUser())
712     N = Glued;
713   return N;
714 }
715 
Schedule()716 void ScheduleDAGLinearize::Schedule() {
717   LLVM_DEBUG(dbgs() << "********** DAG Linearization **********\n");
718 
719   SmallVector<SDNode*, 8> Glues;
720   unsigned DAGSize = 0;
721   for (SDNode &Node : DAG->allnodes()) {
722     SDNode *N = &Node;
723 
724     // Use node id to record degree.
725     unsigned Degree = N->use_size();
726     N->setNodeId(Degree);
727     unsigned NumVals = N->getNumValues();
728     if (NumVals && N->getValueType(NumVals-1) == MVT::Glue &&
729         N->hasAnyUseOfValue(NumVals-1)) {
730       SDNode *User = findGluedUser(N);
731       if (User) {
732         Glues.push_back(N);
733         GluedMap.insert(std::make_pair(N, User));
734       }
735     }
736 
737     if (N->isMachineOpcode() ||
738         (N->getOpcode() != ISD::EntryToken && !isPassiveNode(N)))
739       ++DAGSize;
740   }
741 
742   for (unsigned i = 0, e = Glues.size(); i != e; ++i) {
743     SDNode *Glue = Glues[i];
744     SDNode *GUser = GluedMap[Glue];
745     unsigned Degree = Glue->getNodeId();
746     unsigned UDegree = GUser->getNodeId();
747 
748     // Glue user must be scheduled together with the glue operand. So other
749     // users of the glue operand must be treated as its users.
750     SDNode *ImmGUser = Glue->getGluedUser();
751     for (const SDNode *U : Glue->uses())
752       if (U == ImmGUser)
753         --Degree;
754     GUser->setNodeId(UDegree + Degree);
755     Glue->setNodeId(1);
756   }
757 
758   Sequence.reserve(DAGSize);
759   ScheduleNode(DAG->getRoot().getNode());
760 }
761 
762 MachineBasicBlock*
EmitSchedule(MachineBasicBlock::iterator & InsertPos)763 ScheduleDAGLinearize::EmitSchedule(MachineBasicBlock::iterator &InsertPos) {
764   InstrEmitter Emitter(BB, InsertPos);
765   DenseMap<SDValue, unsigned> VRBaseMap;
766 
767   LLVM_DEBUG({ dbgs() << "\n*** Final schedule ***\n"; });
768 
769   unsigned NumNodes = Sequence.size();
770   MachineBasicBlock *BB = Emitter.getBlock();
771   for (unsigned i = 0; i != NumNodes; ++i) {
772     SDNode *N = Sequence[NumNodes-i-1];
773     LLVM_DEBUG(N->dump(DAG));
774     Emitter.EmitNode(N, false, false, VRBaseMap);
775 
776     // Emit any debug values associated with the node.
777     if (N->getHasDebugValue()) {
778       MachineBasicBlock::iterator InsertPos = Emitter.getInsertPos();
779       for (auto DV : DAG->GetDbgValues(N)) {
780         if (DV->isInvalidated())
781           continue;
782         if (auto *DbgMI = Emitter.EmitDbgValue(DV, VRBaseMap))
783           BB->insert(InsertPos, DbgMI);
784         DV->setIsInvalidated();
785       }
786     }
787   }
788 
789   LLVM_DEBUG(dbgs() << '\n');
790 
791   InsertPos = Emitter.getInsertPos();
792   return Emitter.getBlock();
793 }
794 
795 //===----------------------------------------------------------------------===//
796 //                         Public Constructor Functions
797 //===----------------------------------------------------------------------===//
798 
799 llvm::ScheduleDAGSDNodes *
createFastDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level)800 llvm::createFastDAGScheduler(SelectionDAGISel *IS, CodeGenOpt::Level) {
801   return new ScheduleDAGFast(*IS->MF);
802 }
803 
804 llvm::ScheduleDAGSDNodes *
createDAGLinearizer(SelectionDAGISel * IS,CodeGenOpt::Level)805 llvm::createDAGLinearizer(SelectionDAGISel *IS, CodeGenOpt::Level) {
806   return new ScheduleDAGLinearize(*IS->MF);
807 }
808