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 "llvm/CodeGen/SchedulerRegistry.h"
15 #include "InstrEmitter.h"
16 #include "ScheduleDAGSDNodes.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/CodeGen/SelectionDAGISel.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/InlineAsm.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/ErrorHandling.h"
25 #include "llvm/Support/raw_ostream.h"
26 #include "llvm/Target/TargetInstrInfo.h"
27 #include "llvm/Target/TargetRegisterInfo.h"
28 using namespace llvm;
29 
30 #define DEBUG_TYPE "pre-RA-sched"
31 
32 STATISTIC(NumUnfolds,    "Number of nodes unfolded");
33 STATISTIC(NumDups,       "Number of duplicated nodes");
34 STATISTIC(NumPRCopies,   "Number of physical copies");
35 
36 static RegisterScheduler
37   fastDAGScheduler("fast", "Fast suboptimal list scheduling",
38                    createFastDAGScheduler);
39 static RegisterScheduler
40   linearizeDAGScheduler("linearize", "Linearize DAG, no scheduling",
41                         createDAGLinearizer);
42 
43 
44 namespace {
45   /// FastPriorityQueue - A degenerate priority queue that considers
46   /// all nodes to have the same priority.
47   ///
48   struct FastPriorityQueue {
49     SmallVector<SUnit *, 16> Queue;
50 
empty__anon0612c6de0111::FastPriorityQueue51     bool empty() const { return Queue.empty(); }
52 
push__anon0612c6de0111::FastPriorityQueue53     void push(SUnit *U) {
54       Queue.push_back(U);
55     }
56 
pop__anon0612c6de0111::FastPriorityQueue57     SUnit *pop() {
58       if (empty()) return nullptr;
59       SUnit *V = Queue.back();
60       Queue.pop_back();
61       return V;
62     }
63   };
64 
65 //===----------------------------------------------------------------------===//
66 /// ScheduleDAGFast - The actual "fast" list scheduler implementation.
67 ///
68 class ScheduleDAGFast : public ScheduleDAGSDNodes {
69 private:
70   /// AvailableQueue - The priority queue to use for the available SUnits.
71   FastPriorityQueue AvailableQueue;
72 
73   /// LiveRegDefs - A set of physical registers and their definition
74   /// that are "live". These nodes must be scheduled before any other nodes that
75   /// modifies the registers can be scheduled.
76   unsigned NumLiveRegs;
77   std::vector<SUnit*> LiveRegDefs;
78   std::vector<unsigned> LiveRegCycles;
79 
80 public:
ScheduleDAGFast(MachineFunction & mf)81   ScheduleDAGFast(MachineFunction &mf)
82     : ScheduleDAGSDNodes(mf) {}
83 
84   void Schedule() override;
85 
86   /// AddPred - adds a predecessor edge to SUnit SU.
87   /// This returns true if this is a new predecessor.
AddPred(SUnit * SU,const SDep & D)88   void AddPred(SUnit *SU, const SDep &D) {
89     SU->addPred(D);
90   }
91 
92   /// RemovePred - removes a predecessor edge from SUnit SU.
93   /// This returns true if an edge was removed.
RemovePred(SUnit * SU,const SDep & D)94   void RemovePred(SUnit *SU, const SDep &D) {
95     SU->removePred(D);
96   }
97 
98 private:
99   void ReleasePred(SUnit *SU, SDep *PredEdge);
100   void ReleasePredecessors(SUnit *SU, unsigned CurCycle);
101   void ScheduleNodeBottomUp(SUnit*, unsigned);
102   SUnit *CopyAndMoveSuccessors(SUnit*);
103   void InsertCopiesAndMoveSuccs(SUnit*, unsigned,
104                                 const TargetRegisterClass*,
105                                 const TargetRegisterClass*,
106                                 SmallVectorImpl<SUnit*>&);
107   bool DelayForLiveRegsBottomUp(SUnit*, SmallVectorImpl<unsigned>&);
108   void ListScheduleBottomUp();
109 
110   /// forceUnitLatencies - The fast scheduler doesn't care about real latencies.
forceUnitLatencies() const111   bool forceUnitLatencies() const override { return true; }
112 };
113 }  // end anonymous namespace
114 
115 
116 /// Schedule - Schedule the DAG using list scheduling.
Schedule()117 void ScheduleDAGFast::Schedule() {
118   DEBUG(dbgs() << "********** List Scheduling **********\n");
119 
120   NumLiveRegs = 0;
121   LiveRegDefs.resize(TRI->getNumRegs(), nullptr);
122   LiveRegCycles.resize(TRI->getNumRegs(), 0);
123 
124   // Build the scheduling graph.
125   BuildSchedGraph(nullptr);
126 
127   DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
128           SUnits[su].dumpAll(this));
129 
130   // Execute the actual scheduling loop.
131   ListScheduleBottomUp();
132 }
133 
134 //===----------------------------------------------------------------------===//
135 //  Bottom-Up Scheduling
136 //===----------------------------------------------------------------------===//
137 
138 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
139 /// the AvailableQueue if the count reaches zero. Also update its cycle bound.
ReleasePred(SUnit * SU,SDep * PredEdge)140 void ScheduleDAGFast::ReleasePred(SUnit *SU, SDep *PredEdge) {
141   SUnit *PredSU = PredEdge->getSUnit();
142 
143 #ifndef NDEBUG
144   if (PredSU->NumSuccsLeft == 0) {
145     dbgs() << "*** Scheduling failed! ***\n";
146     PredSU->dump(this);
147     dbgs() << " has been released too many times!\n";
148     llvm_unreachable(nullptr);
149   }
150 #endif
151   --PredSU->NumSuccsLeft;
152 
153   // If all the node's successors are scheduled, this node is ready
154   // to be scheduled. Ignore the special EntrySU node.
155   if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU) {
156     PredSU->isAvailable = true;
157     AvailableQueue.push(PredSU);
158   }
159 }
160 
ReleasePredecessors(SUnit * SU,unsigned CurCycle)161 void ScheduleDAGFast::ReleasePredecessors(SUnit *SU, unsigned CurCycle) {
162   // Bottom up: release predecessors
163   for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
164        I != E; ++I) {
165     ReleasePred(SU, &*I);
166     if (I->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[I->getReg()]) {
172         ++NumLiveRegs;
173         LiveRegDefs[I->getReg()] = I->getSUnit();
174         LiveRegCycles[I->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   DEBUG(dbgs() << "*** Scheduling [" << CurCycle << "]: ");
185   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 (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
195        I != E; ++I) {
196     if (I->isAssignedRegDep()) {
197       if (LiveRegCycles[I->getReg()] == I->getSUnit()->getHeight()) {
198         assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
199         assert(LiveRegDefs[I->getReg()] == SU &&
200                "Physical register dependency violated?");
201         --NumLiveRegs;
202         LiveRegDefs[I->getReg()] = nullptr;
203         LiveRegCycles[I->getReg()] = 0;
204       }
205     }
206   }
207 
208   SU->isScheduled = true;
209 }
210 
211 /// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
212 /// successors to the newly created node.
CopyAndMoveSuccessors(SUnit * SU)213 SUnit *ScheduleDAGFast::CopyAndMoveSuccessors(SUnit *SU) {
214   if (SU->getNode()->getGluedNode())
215     return nullptr;
216 
217   SDNode *N = SU->getNode();
218   if (!N)
219     return nullptr;
220 
221   SUnit *NewSU;
222   bool TryUnfold = false;
223   for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
224     MVT VT = N->getSimpleValueType(i);
225     if (VT == MVT::Glue)
226       return nullptr;
227     else if (VT == MVT::Other)
228       TryUnfold = true;
229   }
230   for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
231     const SDValue &Op = N->getOperand(i);
232     MVT VT = Op.getNode()->getSimpleValueType(Op.getResNo());
233     if (VT == MVT::Glue)
234       return nullptr;
235   }
236 
237   if (TryUnfold) {
238     SmallVector<SDNode*, 2> NewNodes;
239     if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
240       return nullptr;
241 
242     DEBUG(dbgs() << "Unfolding SU # " << SU->NodeNum << "\n");
243     assert(NewNodes.size() == 2 && "Expected a load folding node!");
244 
245     N = NewNodes[1];
246     SDNode *LoadNode = NewNodes[0];
247     unsigned NumVals = N->getNumValues();
248     unsigned OldNumVals = SU->getNode()->getNumValues();
249     for (unsigned i = 0; i != NumVals; ++i)
250       DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i));
251     DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals-1),
252                                    SDValue(LoadNode, 1));
253 
254     SUnit *NewSU = newSUnit(N);
255     assert(N->getNodeId() == -1 && "Node already inserted!");
256     N->setNodeId(NewSU->NodeNum);
257 
258     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
259     for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
260       if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
261         NewSU->isTwoAddress = true;
262         break;
263       }
264     }
265     if (MCID.isCommutable())
266       NewSU->isCommutable = true;
267 
268     // LoadNode may already exist. This can happen when there is another
269     // load from the same location and producing the same type of value
270     // but it has different alignment or volatileness.
271     bool isNewLoad = true;
272     SUnit *LoadSU;
273     if (LoadNode->getNodeId() != -1) {
274       LoadSU = &SUnits[LoadNode->getNodeId()];
275       isNewLoad = false;
276     } else {
277       LoadSU = newSUnit(LoadNode);
278       LoadNode->setNodeId(LoadSU->NodeNum);
279     }
280 
281     SDep ChainPred;
282     SmallVector<SDep, 4> ChainSuccs;
283     SmallVector<SDep, 4> LoadPreds;
284     SmallVector<SDep, 4> NodePreds;
285     SmallVector<SDep, 4> NodeSuccs;
286     for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
287          I != E; ++I) {
288       if (I->isCtrl())
289         ChainPred = *I;
290       else if (I->getSUnit()->getNode() &&
291                I->getSUnit()->getNode()->isOperandOf(LoadNode))
292         LoadPreds.push_back(*I);
293       else
294         NodePreds.push_back(*I);
295     }
296     for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
297          I != E; ++I) {
298       if (I->isCtrl())
299         ChainSuccs.push_back(*I);
300       else
301         NodeSuccs.push_back(*I);
302     }
303 
304     if (ChainPred.getSUnit()) {
305       RemovePred(SU, ChainPred);
306       if (isNewLoad)
307         AddPred(LoadSU, ChainPred);
308     }
309     for (unsigned i = 0, e = LoadPreds.size(); i != e; ++i) {
310       const SDep &Pred = LoadPreds[i];
311       RemovePred(SU, Pred);
312       if (isNewLoad) {
313         AddPred(LoadSU, Pred);
314       }
315     }
316     for (unsigned i = 0, e = NodePreds.size(); i != e; ++i) {
317       const SDep &Pred = NodePreds[i];
318       RemovePred(SU, Pred);
319       AddPred(NewSU, Pred);
320     }
321     for (unsigned i = 0, e = NodeSuccs.size(); i != e; ++i) {
322       SDep D = NodeSuccs[i];
323       SUnit *SuccDep = D.getSUnit();
324       D.setSUnit(SU);
325       RemovePred(SuccDep, D);
326       D.setSUnit(NewSU);
327       AddPred(SuccDep, D);
328     }
329     for (unsigned i = 0, e = ChainSuccs.size(); i != e; ++i) {
330       SDep D = ChainSuccs[i];
331       SUnit *SuccDep = D.getSUnit();
332       D.setSUnit(SU);
333       RemovePred(SuccDep, D);
334       if (isNewLoad) {
335         D.setSUnit(LoadSU);
336         AddPred(SuccDep, D);
337       }
338     }
339     if (isNewLoad) {
340       SDep D(LoadSU, SDep::Barrier);
341       D.setLatency(LoadSU->Latency);
342       AddPred(NewSU, D);
343     }
344 
345     ++NumUnfolds;
346 
347     if (NewSU->NumSuccsLeft == 0) {
348       NewSU->isAvailable = true;
349       return NewSU;
350     }
351     SU = NewSU;
352   }
353 
354   DEBUG(dbgs() << "Duplicating SU # " << SU->NodeNum << "\n");
355   NewSU = Clone(SU);
356 
357   // New SUnit has the exact same predecessors.
358   for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
359        I != E; ++I)
360     if (!I->isArtificial())
361       AddPred(NewSU, *I);
362 
363   // Only copy scheduled successors. Cut them from old node's successor
364   // list and move them over.
365   SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
366   for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
367        I != E; ++I) {
368     if (I->isArtificial())
369       continue;
370     SUnit *SuccSU = I->getSUnit();
371     if (SuccSU->isScheduled) {
372       SDep D = *I;
373       D.setSUnit(NewSU);
374       AddPred(SuccSU, D);
375       D.setSUnit(SU);
376       DelDeps.push_back(std::make_pair(SuccSU, D));
377     }
378   }
379   for (unsigned i = 0, e = DelDeps.size(); i != e; ++i)
380     RemovePred(DelDeps[i].first, DelDeps[i].second);
381 
382   ++NumDups;
383   return NewSU;
384 }
385 
386 /// InsertCopiesAndMoveSuccs - Insert register copies and move all
387 /// scheduled successors of the given SUnit to the last copy.
InsertCopiesAndMoveSuccs(SUnit * SU,unsigned Reg,const TargetRegisterClass * DestRC,const TargetRegisterClass * SrcRC,SmallVectorImpl<SUnit * > & Copies)388 void ScheduleDAGFast::InsertCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
389                                               const TargetRegisterClass *DestRC,
390                                               const TargetRegisterClass *SrcRC,
391                                               SmallVectorImpl<SUnit*> &Copies) {
392   SUnit *CopyFromSU = newSUnit(static_cast<SDNode *>(nullptr));
393   CopyFromSU->CopySrcRC = SrcRC;
394   CopyFromSU->CopyDstRC = DestRC;
395 
396   SUnit *CopyToSU = newSUnit(static_cast<SDNode *>(nullptr));
397   CopyToSU->CopySrcRC = DestRC;
398   CopyToSU->CopyDstRC = SrcRC;
399 
400   // Only copy scheduled successors. Cut them from old node's successor
401   // list and move them over.
402   SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
403   for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
404        I != E; ++I) {
405     if (I->isArtificial())
406       continue;
407     SUnit *SuccSU = I->getSUnit();
408     if (SuccSU->isScheduled) {
409       SDep D = *I;
410       D.setSUnit(CopyToSU);
411       AddPred(SuccSU, D);
412       DelDeps.push_back(std::make_pair(SuccSU, *I));
413     }
414   }
415   for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) {
416     RemovePred(DelDeps[i].first, DelDeps[i].second);
417   }
418   SDep FromDep(SU, SDep::Data, Reg);
419   FromDep.setLatency(SU->Latency);
420   AddPred(CopyFromSU, FromDep);
421   SDep ToDep(CopyFromSU, SDep::Data, 0);
422   ToDep.setLatency(CopyFromSU->Latency);
423   AddPred(CopyToSU, ToDep);
424 
425   Copies.push_back(CopyFromSU);
426   Copies.push_back(CopyToSU);
427 
428   ++NumPRCopies;
429 }
430 
431 /// getPhysicalRegisterVT - Returns the ValueType of the physical register
432 /// definition of the specified node.
433 /// FIXME: Move to SelectionDAG?
getPhysicalRegisterVT(SDNode * N,unsigned Reg,const TargetInstrInfo * TII)434 static MVT getPhysicalRegisterVT(SDNode *N, unsigned Reg,
435                                  const TargetInstrInfo *TII) {
436   unsigned NumRes;
437   if (N->getOpcode() == ISD::CopyFromReg) {
438     // CopyFromReg has: "chain, Val, glue" so operand 1 gives the type.
439     NumRes = 1;
440   } else {
441     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
442     assert(MCID.ImplicitDefs && "Physical reg def must be in implicit def list!");
443     NumRes = MCID.getNumDefs();
444     for (const uint16_t *ImpDef = MCID.getImplicitDefs(); *ImpDef; ++ImpDef) {
445       if (Reg == *ImpDef)
446         break;
447       ++NumRes;
448     }
449   }
450   return N->getSimpleValueType(NumRes);
451 }
452 
453 /// CheckForLiveRegDef - Return true and update live register vector if the
454 /// 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)455 static bool CheckForLiveRegDef(SUnit *SU, unsigned Reg,
456                                std::vector<SUnit*> &LiveRegDefs,
457                                SmallSet<unsigned, 4> &RegAdded,
458                                SmallVectorImpl<unsigned> &LRegs,
459                                const TargetRegisterInfo *TRI) {
460   bool Added = false;
461   for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) {
462     if (LiveRegDefs[*AI] && LiveRegDefs[*AI] != SU) {
463       if (RegAdded.insert(*AI).second) {
464         LRegs.push_back(*AI);
465         Added = true;
466       }
467     }
468   }
469   return Added;
470 }
471 
472 /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
473 /// scheduling of the given node to satisfy live physical register dependencies.
474 /// If the specific node is the last one that's available to schedule, do
475 /// whatever is necessary (i.e. backtracking or cloning) to make it possible.
DelayForLiveRegsBottomUp(SUnit * SU,SmallVectorImpl<unsigned> & LRegs)476 bool ScheduleDAGFast::DelayForLiveRegsBottomUp(SUnit *SU,
477                                               SmallVectorImpl<unsigned> &LRegs){
478   if (NumLiveRegs == 0)
479     return false;
480 
481   SmallSet<unsigned, 4> RegAdded;
482   // If this node would clobber any "live" register, then it's not ready.
483   for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
484        I != E; ++I) {
485     if (I->isAssignedRegDep()) {
486       CheckForLiveRegDef(I->getSUnit(), I->getReg(), LiveRegDefs,
487                          RegAdded, LRegs, TRI);
488     }
489   }
490 
491   for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) {
492     if (Node->getOpcode() == ISD::INLINEASM) {
493       // Inline asm can clobber physical defs.
494       unsigned NumOps = Node->getNumOperands();
495       if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
496         --NumOps;  // Ignore the glue operand.
497 
498       for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
499         unsigned Flags =
500           cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
501         unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
502 
503         ++i; // Skip the ID value.
504         if (InlineAsm::isRegDefKind(Flags) ||
505             InlineAsm::isRegDefEarlyClobberKind(Flags) ||
506             InlineAsm::isClobberKind(Flags)) {
507           // Check for def of register or earlyclobber register.
508           for (; NumVals; --NumVals, ++i) {
509             unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
510             if (TargetRegisterInfo::isPhysicalRegister(Reg))
511               CheckForLiveRegDef(SU, Reg, LiveRegDefs, RegAdded, LRegs, TRI);
512           }
513         } else
514           i += NumVals;
515       }
516       continue;
517     }
518     if (!Node->isMachineOpcode())
519       continue;
520     const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode());
521     if (!MCID.ImplicitDefs)
522       continue;
523     for (const uint16_t *Reg = MCID.getImplicitDefs(); *Reg; ++Reg) {
524       CheckForLiveRegDef(SU, *Reg, LiveRegDefs, RegAdded, LRegs, TRI);
525     }
526   }
527   return !LRegs.empty();
528 }
529 
530 
531 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
532 /// schedulers.
ListScheduleBottomUp()533 void ScheduleDAGFast::ListScheduleBottomUp() {
534   unsigned CurCycle = 0;
535 
536   // Release any predecessors of the special Exit node.
537   ReleasePredecessors(&ExitSU, CurCycle);
538 
539   // Add root to Available queue.
540   if (!SUnits.empty()) {
541     SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()];
542     assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!");
543     RootSU->isAvailable = true;
544     AvailableQueue.push(RootSU);
545   }
546 
547   // While Available queue is not empty, grab the node with the highest
548   // priority. If it is not ready put it back.  Schedule the node.
549   SmallVector<SUnit*, 4> NotReady;
550   DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMap;
551   Sequence.reserve(SUnits.size());
552   while (!AvailableQueue.empty()) {
553     bool Delayed = false;
554     LRegsMap.clear();
555     SUnit *CurSU = AvailableQueue.pop();
556     while (CurSU) {
557       SmallVector<unsigned, 4> LRegs;
558       if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
559         break;
560       Delayed = true;
561       LRegsMap.insert(std::make_pair(CurSU, LRegs));
562 
563       CurSU->isPending = true;  // This SU is not in AvailableQueue right now.
564       NotReady.push_back(CurSU);
565       CurSU = AvailableQueue.pop();
566     }
567 
568     // All candidates are delayed due to live physical reg dependencies.
569     // Try code duplication or inserting cross class copies
570     // to resolve it.
571     if (Delayed && !CurSU) {
572       if (!CurSU) {
573         // Try duplicating the nodes that produces these
574         // "expensive to copy" values to break the dependency. In case even
575         // that doesn't work, insert cross class copies.
576         SUnit *TrySU = NotReady[0];
577         SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
578         assert(LRegs.size() == 1 && "Can't handle this yet!");
579         unsigned Reg = LRegs[0];
580         SUnit *LRDef = LiveRegDefs[Reg];
581         MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
582         const TargetRegisterClass *RC =
583           TRI->getMinimalPhysRegClass(Reg, VT);
584         const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
585 
586         // If cross copy register class is the same as RC, then it must be
587         // possible copy the value directly. Do not try duplicate the def.
588         // If cross copy register class is not the same as RC, then it's
589         // possible to copy the value but it require cross register class copies
590         // and it is expensive.
591         // If cross copy register class is null, then it's not possible to copy
592         // the value at all.
593         SUnit *NewDef = nullptr;
594         if (DestRC != RC) {
595           NewDef = CopyAndMoveSuccessors(LRDef);
596           if (!DestRC && !NewDef)
597             report_fatal_error("Can't handle live physical "
598                                "register dependency!");
599         }
600         if (!NewDef) {
601           // Issue copies, these can be expensive cross register class copies.
602           SmallVector<SUnit*, 2> Copies;
603           InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
604           DEBUG(dbgs() << "Adding an edge from SU # " << TrySU->NodeNum
605                        << " to SU #" << Copies.front()->NodeNum << "\n");
606           AddPred(TrySU, SDep(Copies.front(), SDep::Artificial));
607           NewDef = Copies.back();
608         }
609 
610         DEBUG(dbgs() << "Adding an edge from SU # " << NewDef->NodeNum
611                      << " to SU #" << TrySU->NodeNum << "\n");
612         LiveRegDefs[Reg] = NewDef;
613         AddPred(NewDef, SDep(TrySU, SDep::Artificial));
614         TrySU->isAvailable = false;
615         CurSU = NewDef;
616       }
617 
618       if (!CurSU) {
619         llvm_unreachable("Unable to resolve live physical register dependencies!");
620       }
621     }
622 
623     // Add the nodes that aren't ready back onto the available list.
624     for (unsigned i = 0, e = NotReady.size(); i != e; ++i) {
625       NotReady[i]->isPending = false;
626       // May no longer be available due to backtracking.
627       if (NotReady[i]->isAvailable)
628         AvailableQueue.push(NotReady[i]);
629     }
630     NotReady.clear();
631 
632     if (CurSU)
633       ScheduleNodeBottomUp(CurSU, CurCycle);
634     ++CurCycle;
635   }
636 
637   // Reverse the order since it is bottom up.
638   std::reverse(Sequence.begin(), Sequence.end());
639 
640 #ifndef NDEBUG
641   VerifyScheduledSequence(/*isBottomUp=*/true);
642 #endif
643 }
644 
645 
646 namespace {
647 //===----------------------------------------------------------------------===//
648 // ScheduleDAGLinearize - No scheduling scheduler, it simply linearize the
649 // DAG in topological order.
650 // IMPORTANT: this may not work for targets with phyreg dependency.
651 //
652 class ScheduleDAGLinearize : public ScheduleDAGSDNodes {
653 public:
ScheduleDAGLinearize(MachineFunction & mf)654   ScheduleDAGLinearize(MachineFunction &mf) : ScheduleDAGSDNodes(mf) {}
655 
656   void Schedule() override;
657 
658   MachineBasicBlock *
659     EmitSchedule(MachineBasicBlock::iterator &InsertPos) override;
660 
661 private:
662   std::vector<SDNode*> Sequence;
663   DenseMap<SDNode*, SDNode*> GluedMap;  // Cache glue to its user
664 
665   void ScheduleNode(SDNode *N);
666 };
667 } // end anonymous namespace
668 
ScheduleNode(SDNode * N)669 void ScheduleDAGLinearize::ScheduleNode(SDNode *N) {
670   if (N->getNodeId() != 0)
671     llvm_unreachable(nullptr);
672 
673   if (!N->isMachineOpcode() &&
674       (N->getOpcode() == ISD::EntryToken || isPassiveNode(N)))
675     // These nodes do not need to be translated into MIs.
676     return;
677 
678   DEBUG(dbgs() << "\n*** Scheduling: ");
679   DEBUG(N->dump(DAG));
680   Sequence.push_back(N);
681 
682   unsigned NumOps = N->getNumOperands();
683   if (unsigned NumLeft = NumOps) {
684     SDNode *GluedOpN = nullptr;
685     do {
686       const SDValue &Op = N->getOperand(NumLeft-1);
687       SDNode *OpN = Op.getNode();
688 
689       if (NumLeft == NumOps && Op.getValueType() == MVT::Glue) {
690         // Schedule glue operand right above N.
691         GluedOpN = OpN;
692         assert(OpN->getNodeId() != 0 && "Glue operand not ready?");
693         OpN->setNodeId(0);
694         ScheduleNode(OpN);
695         continue;
696       }
697 
698       if (OpN == GluedOpN)
699         // Glue operand is already scheduled.
700         continue;
701 
702       DenseMap<SDNode*, SDNode*>::iterator DI = GluedMap.find(OpN);
703       if (DI != GluedMap.end() && DI->second != N)
704         // Users of glues are counted against the glued users.
705         OpN = DI->second;
706 
707       unsigned Degree = OpN->getNodeId();
708       assert(Degree > 0 && "Predecessor over-released!");
709       OpN->setNodeId(--Degree);
710       if (Degree == 0)
711         ScheduleNode(OpN);
712     } while (--NumLeft);
713   }
714 }
715 
716 /// findGluedUser - Find the representative use of a glue value by walking
717 /// the use chain.
findGluedUser(SDNode * N)718 static SDNode *findGluedUser(SDNode *N) {
719   while (SDNode *Glued = N->getGluedUser())
720     N = Glued;
721   return N;
722 }
723 
Schedule()724 void ScheduleDAGLinearize::Schedule() {
725   DEBUG(dbgs() << "********** DAG Linearization **********\n");
726 
727   SmallVector<SDNode*, 8> Glues;
728   unsigned DAGSize = 0;
729   for (SelectionDAG::allnodes_iterator I = DAG->allnodes_begin(),
730          E = DAG->allnodes_end(); I != E; ++I) {
731     SDNode *N = I;
732 
733     // Use node id to record degree.
734     unsigned Degree = N->use_size();
735     N->setNodeId(Degree);
736     unsigned NumVals = N->getNumValues();
737     if (NumVals && N->getValueType(NumVals-1) == MVT::Glue &&
738         N->hasAnyUseOfValue(NumVals-1)) {
739       SDNode *User = findGluedUser(N);
740       if (User) {
741         Glues.push_back(N);
742         GluedMap.insert(std::make_pair(N, User));
743       }
744     }
745 
746     if (N->isMachineOpcode() ||
747         (N->getOpcode() != ISD::EntryToken && !isPassiveNode(N)))
748       ++DAGSize;
749   }
750 
751   for (unsigned i = 0, e = Glues.size(); i != e; ++i) {
752     SDNode *Glue = Glues[i];
753     SDNode *GUser = GluedMap[Glue];
754     unsigned Degree = Glue->getNodeId();
755     unsigned UDegree = GUser->getNodeId();
756 
757     // Glue user must be scheduled together with the glue operand. So other
758     // users of the glue operand must be treated as its users.
759     SDNode *ImmGUser = Glue->getGluedUser();
760     for (SDNode::use_iterator ui = Glue->use_begin(), ue = Glue->use_end();
761          ui != ue; ++ui)
762       if (*ui == ImmGUser)
763         --Degree;
764     GUser->setNodeId(UDegree + Degree);
765     Glue->setNodeId(1);
766   }
767 
768   Sequence.reserve(DAGSize);
769   ScheduleNode(DAG->getRoot().getNode());
770 }
771 
772 MachineBasicBlock*
EmitSchedule(MachineBasicBlock::iterator & InsertPos)773 ScheduleDAGLinearize::EmitSchedule(MachineBasicBlock::iterator &InsertPos) {
774   InstrEmitter Emitter(BB, InsertPos);
775   DenseMap<SDValue, unsigned> VRBaseMap;
776 
777   DEBUG({
778       dbgs() << "\n*** Final schedule ***\n";
779     });
780 
781   // FIXME: Handle dbg_values.
782   unsigned NumNodes = Sequence.size();
783   for (unsigned i = 0; i != NumNodes; ++i) {
784     SDNode *N = Sequence[NumNodes-i-1];
785     DEBUG(N->dump(DAG));
786     Emitter.EmitNode(N, false, false, VRBaseMap);
787   }
788 
789   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