1 //===----- ScheduleDAGFast.cpp - Fast poor list scheduler -----------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This implements a fast scheduler.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "InstrEmitter.h"
14 #include "SDNodeDbgValue.h"
15 #include "ScheduleDAGSDNodes.h"
16 #include "llvm/ADT/SmallSet.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/CodeGen/SchedulerRegistry.h"
19 #include "llvm/CodeGen/SelectionDAGISel.h"
20 #include "llvm/CodeGen/TargetInstrInfo.h"
21 #include "llvm/CodeGen/TargetRegisterInfo.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 using namespace llvm;
27 
28 #define DEBUG_TYPE "pre-RA-sched"
29 
30 STATISTIC(NumUnfolds,    "Number of nodes unfolded");
31 STATISTIC(NumDups,       "Number of duplicated nodes");
32 STATISTIC(NumPRCopies,   "Number of physical copies");
33 
34 static RegisterScheduler
35   fastDAGScheduler("fast", "Fast suboptimal list scheduling",
36                    createFastDAGScheduler);
37 static RegisterScheduler
38   linearizeDAGScheduler("linearize", "Linearize DAG, no scheduling",
39                         createDAGLinearizer);
40 
41 
42 namespace {
43   /// FastPriorityQueue - A degenerate priority queue that considers
44   /// all nodes to have the same priority.
45   ///
46   struct FastPriorityQueue {
47     SmallVector<SUnit *, 16> Queue;
48 
empty__anonacfa8bc40111::FastPriorityQueue49     bool empty() const { return Queue.empty(); }
50 
push__anonacfa8bc40111::FastPriorityQueue51     void push(SUnit *U) {
52       Queue.push_back(U);
53     }
54 
pop__anonacfa8bc40111::FastPriorityQueue55     SUnit *pop() {
56       if (empty()) return nullptr;
57       return Queue.pop_back_val();
58     }
59   };
60 
61 //===----------------------------------------------------------------------===//
62 /// ScheduleDAGFast - The actual "fast" list scheduler implementation.
63 ///
64 class ScheduleDAGFast : public ScheduleDAGSDNodes {
65 private:
66   /// AvailableQueue - The priority queue to use for the available SUnits.
67   FastPriorityQueue AvailableQueue;
68 
69   /// LiveRegDefs - A set of physical registers and their definition
70   /// that are "live". These nodes must be scheduled before any other nodes that
71   /// modifies the registers can be scheduled.
72   unsigned NumLiveRegs = 0u;
73   std::vector<SUnit*> LiveRegDefs;
74   std::vector<unsigned> LiveRegCycles;
75 
76 public:
ScheduleDAGFast(MachineFunction & mf)77   ScheduleDAGFast(MachineFunction &mf)
78     : ScheduleDAGSDNodes(mf) {}
79 
80   void Schedule() override;
81 
82   /// AddPred - adds a predecessor edge to SUnit SU.
83   /// This returns true if this is a new predecessor.
AddPred(SUnit * SU,const SDep & D)84   void AddPred(SUnit *SU, const SDep &D) {
85     SU->addPred(D);
86   }
87 
88   /// RemovePred - removes a predecessor edge from SUnit SU.
89   /// This returns true if an edge was removed.
RemovePred(SUnit * SU,const SDep & D)90   void RemovePred(SUnit *SU, const SDep &D) {
91     SU->removePred(D);
92   }
93 
94 private:
95   void ReleasePred(SUnit *SU, SDep *PredEdge);
96   void ReleasePredecessors(SUnit *SU, unsigned CurCycle);
97   void ScheduleNodeBottomUp(SUnit*, unsigned);
98   SUnit *CopyAndMoveSuccessors(SUnit*);
99   void InsertCopiesAndMoveSuccs(SUnit*, unsigned,
100                                 const TargetRegisterClass*,
101                                 const TargetRegisterClass*,
102                                 SmallVectorImpl<SUnit*>&);
103   bool DelayForLiveRegsBottomUp(SUnit*, SmallVectorImpl<unsigned>&);
104   void ListScheduleBottomUp();
105 
106   /// forceUnitLatencies - The fast scheduler doesn't care about real latencies.
forceUnitLatencies() const107   bool forceUnitLatencies() const override { return true; }
108 };
109 }  // end anonymous namespace
110 
111 
112 /// Schedule - Schedule the DAG using list scheduling.
Schedule()113 void ScheduleDAGFast::Schedule() {
114   LLVM_DEBUG(dbgs() << "********** List Scheduling **********\n");
115 
116   NumLiveRegs = 0;
117   LiveRegDefs.resize(TRI->getNumRegs(), nullptr);
118   LiveRegCycles.resize(TRI->getNumRegs(), 0);
119 
120   // Build the scheduling graph.
121   BuildSchedGraph(nullptr);
122 
123   LLVM_DEBUG(dump());
124 
125   // Execute the actual scheduling loop.
126   ListScheduleBottomUp();
127 }
128 
129 //===----------------------------------------------------------------------===//
130 //  Bottom-Up Scheduling
131 //===----------------------------------------------------------------------===//
132 
133 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
134 /// the AvailableQueue if the count reaches zero. Also update its cycle bound.
ReleasePred(SUnit * SU,SDep * PredEdge)135 void ScheduleDAGFast::ReleasePred(SUnit *SU, SDep *PredEdge) {
136   SUnit *PredSU = PredEdge->getSUnit();
137 
138 #ifndef NDEBUG
139   if (PredSU->NumSuccsLeft == 0) {
140     dbgs() << "*** Scheduling failed! ***\n";
141     dumpNode(*PredSU);
142     dbgs() << " has been released too many times!\n";
143     llvm_unreachable(nullptr);
144   }
145 #endif
146   --PredSU->NumSuccsLeft;
147 
148   // If all the node's successors are scheduled, this node is ready
149   // to be scheduled. Ignore the special EntrySU node.
150   if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU) {
151     PredSU->isAvailable = true;
152     AvailableQueue.push(PredSU);
153   }
154 }
155 
ReleasePredecessors(SUnit * SU,unsigned CurCycle)156 void ScheduleDAGFast::ReleasePredecessors(SUnit *SU, unsigned CurCycle) {
157   // Bottom up: release predecessors
158   for (SDep &Pred : SU->Preds) {
159     ReleasePred(SU, &Pred);
160     if (Pred.isAssignedRegDep()) {
161       // This is a physical register dependency and it's impossible or
162       // expensive to copy the register. Make sure nothing that can
163       // clobber the register is scheduled between the predecessor and
164       // this node.
165       if (!LiveRegDefs[Pred.getReg()]) {
166         ++NumLiveRegs;
167         LiveRegDefs[Pred.getReg()] = Pred.getSUnit();
168         LiveRegCycles[Pred.getReg()] = CurCycle;
169       }
170     }
171   }
172 }
173 
174 /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
175 /// count of its predecessors. If a predecessor pending count is zero, add it to
176 /// the Available queue.
ScheduleNodeBottomUp(SUnit * SU,unsigned CurCycle)177 void ScheduleDAGFast::ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle) {
178   LLVM_DEBUG(dbgs() << "*** Scheduling [" << CurCycle << "]: ");
179   LLVM_DEBUG(dumpNode(*SU));
180 
181   assert(CurCycle >= SU->getHeight() && "Node scheduled below its height!");
182   SU->setHeightToAtLeast(CurCycle);
183   Sequence.push_back(SU);
184 
185   ReleasePredecessors(SU, CurCycle);
186 
187   // Release all the implicit physical register defs that are live.
188   for (SDep &Succ : SU->Succs) {
189     if (Succ.isAssignedRegDep()) {
190       if (LiveRegCycles[Succ.getReg()] == Succ.getSUnit()->getHeight()) {
191         assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
192         assert(LiveRegDefs[Succ.getReg()] == SU &&
193                "Physical register dependency violated?");
194         --NumLiveRegs;
195         LiveRegDefs[Succ.getReg()] = nullptr;
196         LiveRegCycles[Succ.getReg()] = 0;
197       }
198     }
199   }
200 
201   SU->isScheduled = true;
202 }
203 
204 /// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
205 /// successors to the newly created node.
CopyAndMoveSuccessors(SUnit * SU)206 SUnit *ScheduleDAGFast::CopyAndMoveSuccessors(SUnit *SU) {
207   if (SU->getNode()->getGluedNode())
208     return nullptr;
209 
210   SDNode *N = SU->getNode();
211   if (!N)
212     return nullptr;
213 
214   SUnit *NewSU;
215   bool TryUnfold = false;
216   for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
217     MVT VT = N->getSimpleValueType(i);
218     if (VT == MVT::Glue)
219       return nullptr;
220     else if (VT == MVT::Other)
221       TryUnfold = true;
222   }
223   for (const SDValue &Op : N->op_values()) {
224     MVT VT = Op.getNode()->getSimpleValueType(Op.getResNo());
225     if (VT == MVT::Glue)
226       return nullptr;
227   }
228 
229   if (TryUnfold) {
230     SmallVector<SDNode*, 2> NewNodes;
231     if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
232       return nullptr;
233 
234     LLVM_DEBUG(dbgs() << "Unfolding SU # " << SU->NodeNum << "\n");
235     assert(NewNodes.size() == 2 && "Expected a load folding node!");
236 
237     N = NewNodes[1];
238     SDNode *LoadNode = NewNodes[0];
239     unsigned NumVals = N->getNumValues();
240     unsigned OldNumVals = SU->getNode()->getNumValues();
241     for (unsigned i = 0; i != NumVals; ++i)
242       DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i));
243     DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals-1),
244                                    SDValue(LoadNode, 1));
245 
246     SUnit *NewSU = newSUnit(N);
247     assert(N->getNodeId() == -1 && "Node already inserted!");
248     N->setNodeId(NewSU->NodeNum);
249 
250     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
251     for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
252       if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
253         NewSU->isTwoAddress = true;
254         break;
255       }
256     }
257     if (MCID.isCommutable())
258       NewSU->isCommutable = true;
259 
260     // LoadNode may already exist. This can happen when there is another
261     // load from the same location and producing the same type of value
262     // but it has different alignment or volatileness.
263     bool isNewLoad = true;
264     SUnit *LoadSU;
265     if (LoadNode->getNodeId() != -1) {
266       LoadSU = &SUnits[LoadNode->getNodeId()];
267       isNewLoad = false;
268     } else {
269       LoadSU = newSUnit(LoadNode);
270       LoadNode->setNodeId(LoadSU->NodeNum);
271     }
272 
273     SDep ChainPred;
274     SmallVector<SDep, 4> ChainSuccs;
275     SmallVector<SDep, 4> LoadPreds;
276     SmallVector<SDep, 4> NodePreds;
277     SmallVector<SDep, 4> NodeSuccs;
278     for (SDep &Pred : SU->Preds) {
279       if (Pred.isCtrl())
280         ChainPred = Pred;
281       else if (Pred.getSUnit()->getNode() &&
282                Pred.getSUnit()->getNode()->isOperandOf(LoadNode))
283         LoadPreds.push_back(Pred);
284       else
285         NodePreds.push_back(Pred);
286     }
287     for (SDep &Succ : SU->Succs) {
288       if (Succ.isCtrl())
289         ChainSuccs.push_back(Succ);
290       else
291         NodeSuccs.push_back(Succ);
292     }
293 
294     if (ChainPred.getSUnit()) {
295       RemovePred(SU, ChainPred);
296       if (isNewLoad)
297         AddPred(LoadSU, ChainPred);
298     }
299     for (const SDep &Pred : LoadPreds) {
300       RemovePred(SU, Pred);
301       if (isNewLoad) {
302         AddPred(LoadSU, Pred);
303       }
304     }
305     for (const SDep &Pred : NodePreds) {
306       RemovePred(SU, Pred);
307       AddPred(NewSU, Pred);
308     }
309     for (SDep D : NodeSuccs) {
310       SUnit *SuccDep = D.getSUnit();
311       D.setSUnit(SU);
312       RemovePred(SuccDep, D);
313       D.setSUnit(NewSU);
314       AddPred(SuccDep, D);
315     }
316     for (SDep D : ChainSuccs) {
317       SUnit *SuccDep = D.getSUnit();
318       D.setSUnit(SU);
319       RemovePred(SuccDep, D);
320       if (isNewLoad) {
321         D.setSUnit(LoadSU);
322         AddPred(SuccDep, D);
323       }
324     }
325     if (isNewLoad) {
326       SDep D(LoadSU, SDep::Barrier);
327       D.setLatency(LoadSU->Latency);
328       AddPred(NewSU, D);
329     }
330 
331     ++NumUnfolds;
332 
333     if (NewSU->NumSuccsLeft == 0) {
334       NewSU->isAvailable = true;
335       return NewSU;
336     }
337     SU = NewSU;
338   }
339 
340   LLVM_DEBUG(dbgs() << "Duplicating SU # " << SU->NodeNum << "\n");
341   NewSU = Clone(SU);
342 
343   // New SUnit has the exact same predecessors.
344   for (SDep &Pred : SU->Preds)
345     if (!Pred.isArtificial())
346       AddPred(NewSU, Pred);
347 
348   // Only copy scheduled successors. Cut them from old node's successor
349   // list and move them over.
350   SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
351   for (SDep &Succ : SU->Succs) {
352     if (Succ.isArtificial())
353       continue;
354     SUnit *SuccSU = Succ.getSUnit();
355     if (SuccSU->isScheduled) {
356       SDep D = Succ;
357       D.setSUnit(NewSU);
358       AddPred(SuccSU, D);
359       D.setSUnit(SU);
360       DelDeps.push_back(std::make_pair(SuccSU, D));
361     }
362   }
363   for (unsigned i = 0, e = DelDeps.size(); i != e; ++i)
364     RemovePred(DelDeps[i].first, DelDeps[i].second);
365 
366   ++NumDups;
367   return NewSU;
368 }
369 
370 /// InsertCopiesAndMoveSuccs - Insert register copies and move all
371 /// scheduled successors of the given SUnit to the last copy.
InsertCopiesAndMoveSuccs(SUnit * SU,unsigned Reg,const TargetRegisterClass * DestRC,const TargetRegisterClass * SrcRC,SmallVectorImpl<SUnit * > & Copies)372 void ScheduleDAGFast::InsertCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
373                                               const TargetRegisterClass *DestRC,
374                                               const TargetRegisterClass *SrcRC,
375                                               SmallVectorImpl<SUnit*> &Copies) {
376   SUnit *CopyFromSU = newSUnit(static_cast<SDNode *>(nullptr));
377   CopyFromSU->CopySrcRC = SrcRC;
378   CopyFromSU->CopyDstRC = DestRC;
379 
380   SUnit *CopyToSU = newSUnit(static_cast<SDNode *>(nullptr));
381   CopyToSU->CopySrcRC = DestRC;
382   CopyToSU->CopyDstRC = SrcRC;
383 
384   // Only copy scheduled successors. Cut them from old node's successor
385   // list and move them over.
386   SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
387   for (SDep &Succ : SU->Succs) {
388     if (Succ.isArtificial())
389       continue;
390     SUnit *SuccSU = Succ.getSUnit();
391     if (SuccSU->isScheduled) {
392       SDep D = Succ;
393       D.setSUnit(CopyToSU);
394       AddPred(SuccSU, D);
395       DelDeps.push_back(std::make_pair(SuccSU, Succ));
396     }
397   }
398   for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) {
399     RemovePred(DelDeps[i].first, DelDeps[i].second);
400   }
401   SDep FromDep(SU, SDep::Data, Reg);
402   FromDep.setLatency(SU->Latency);
403   AddPred(CopyFromSU, FromDep);
404   SDep ToDep(CopyFromSU, SDep::Data, 0);
405   ToDep.setLatency(CopyFromSU->Latency);
406   AddPred(CopyToSU, ToDep);
407 
408   Copies.push_back(CopyFromSU);
409   Copies.push_back(CopyToSU);
410 
411   ++NumPRCopies;
412 }
413 
414 /// getPhysicalRegisterVT - Returns the ValueType of the physical register
415 /// definition of the specified node.
416 /// FIXME: Move to SelectionDAG?
getPhysicalRegisterVT(SDNode * N,unsigned Reg,const TargetInstrInfo * TII)417 static MVT getPhysicalRegisterVT(SDNode *N, unsigned Reg,
418                                  const TargetInstrInfo *TII) {
419   unsigned NumRes;
420   if (N->getOpcode() == ISD::CopyFromReg) {
421     // CopyFromReg has: "chain, Val, glue" so operand 1 gives the type.
422     NumRes = 1;
423   } else {
424     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
425     assert(!MCID.implicit_defs().empty() &&
426            "Physical reg def must be in implicit def list!");
427     NumRes = MCID.getNumDefs();
428     for (MCPhysReg ImpDef : MCID.implicit_defs()) {
429       if (Reg == ImpDef)
430         break;
431       ++NumRes;
432     }
433   }
434   return N->getSimpleValueType(NumRes);
435 }
436 
437 /// CheckForLiveRegDef - Return true and update live register vector if the
438 /// 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,const SDNode * Node=nullptr)439 static bool CheckForLiveRegDef(SUnit *SU, unsigned Reg,
440                                std::vector<SUnit *> &LiveRegDefs,
441                                SmallSet<unsigned, 4> &RegAdded,
442                                SmallVectorImpl<unsigned> &LRegs,
443                                const TargetRegisterInfo *TRI,
444                                const SDNode *Node = nullptr) {
445   bool Added = false;
446   for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) {
447     // Check if Ref is live.
448     if (!LiveRegDefs[*AI])
449       continue;
450 
451     // Allow multiple uses of the same def.
452     if (LiveRegDefs[*AI] == SU)
453       continue;
454 
455     // Allow multiple uses of same def
456     if (Node && LiveRegDefs[*AI]->getNode() == Node)
457       continue;
458 
459     // Add Reg to the set of interfering live regs.
460     if (RegAdded.insert(*AI).second) {
461       LRegs.push_back(*AI);
462       Added = true;
463     }
464   }
465   return Added;
466 }
467 
468 /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
469 /// scheduling of the given node to satisfy live physical register dependencies.
470 /// If the specific node is the last one that's available to schedule, do
471 /// whatever is necessary (i.e. backtracking or cloning) to make it possible.
DelayForLiveRegsBottomUp(SUnit * SU,SmallVectorImpl<unsigned> & LRegs)472 bool ScheduleDAGFast::DelayForLiveRegsBottomUp(SUnit *SU,
473                                               SmallVectorImpl<unsigned> &LRegs){
474   if (NumLiveRegs == 0)
475     return false;
476 
477   SmallSet<unsigned, 4> RegAdded;
478   // If this node would clobber any "live" register, then it's not ready.
479   for (SDep &Pred : SU->Preds) {
480     if (Pred.isAssignedRegDep()) {
481       CheckForLiveRegDef(Pred.getSUnit(), Pred.getReg(), LiveRegDefs,
482                          RegAdded, LRegs, TRI);
483     }
484   }
485 
486   for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) {
487     if (Node->getOpcode() == ISD::INLINEASM ||
488         Node->getOpcode() == ISD::INLINEASM_BR) {
489       // Inline asm can clobber physical defs.
490       unsigned NumOps = Node->getNumOperands();
491       if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
492         --NumOps;  // Ignore the glue operand.
493 
494       for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
495         unsigned Flags = Node->getConstantOperandVal(i);
496         const InlineAsm::Flag F(Flags);
497         unsigned NumVals = F.getNumOperandRegisters();
498 
499         ++i; // Skip the ID value.
500         if (F.isRegDefKind() || F.isRegDefEarlyClobberKind() ||
501             F.isClobberKind()) {
502           // Check for def of register or earlyclobber register.
503           for (; NumVals; --NumVals, ++i) {
504             unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
505             if (Register::isPhysicalRegister(Reg))
506               CheckForLiveRegDef(SU, Reg, LiveRegDefs, RegAdded, LRegs, TRI);
507           }
508         } else
509           i += NumVals;
510       }
511       continue;
512     }
513 
514     if (Node->getOpcode() == ISD::CopyToReg) {
515       Register Reg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
516       if (Reg.isPhysical()) {
517         SDNode *SrcNode = Node->getOperand(2).getNode();
518         CheckForLiveRegDef(SU, Reg, LiveRegDefs, RegAdded, LRegs, TRI, SrcNode);
519       }
520     }
521 
522     if (!Node->isMachineOpcode())
523       continue;
524     const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode());
525     for (MCPhysReg Reg : MCID.implicit_defs())
526       CheckForLiveRegDef(SU, Reg, LiveRegDefs, RegAdded, LRegs, TRI);
527   }
528   return !LRegs.empty();
529 }
530 
531 
532 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
533 /// schedulers.
ListScheduleBottomUp()534 void ScheduleDAGFast::ListScheduleBottomUp() {
535   unsigned CurCycle = 0;
536 
537   // Release any predecessors of the special Exit node.
538   ReleasePredecessors(&ExitSU, CurCycle);
539 
540   // Add root to Available queue.
541   if (!SUnits.empty()) {
542     SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()];
543     assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!");
544     RootSU->isAvailable = true;
545     AvailableQueue.push(RootSU);
546   }
547 
548   // While Available queue is not empty, grab the node with the highest
549   // priority. If it is not ready put it back.  Schedule the node.
550   SmallVector<SUnit*, 4> NotReady;
551   DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMap;
552   Sequence.reserve(SUnits.size());
553   while (!AvailableQueue.empty()) {
554     bool Delayed = false;
555     LRegsMap.clear();
556     SUnit *CurSU = AvailableQueue.pop();
557     while (CurSU) {
558       SmallVector<unsigned, 4> LRegs;
559       if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
560         break;
561       Delayed = true;
562       LRegsMap.insert(std::make_pair(CurSU, LRegs));
563 
564       CurSU->isPending = true;  // This SU is not in AvailableQueue right now.
565       NotReady.push_back(CurSU);
566       CurSU = AvailableQueue.pop();
567     }
568 
569     // All candidates are delayed due to live physical reg dependencies.
570     // Try code duplication or inserting cross class copies
571     // to resolve it.
572     if (Delayed && !CurSU) {
573       if (!CurSU) {
574         // Try duplicating the nodes that produces these
575         // "expensive to copy" values to break the dependency. In case even
576         // that doesn't work, insert cross class copies.
577         SUnit *TrySU = NotReady[0];
578         SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
579         assert(LRegs.size() == 1 && "Can't handle this yet!");
580         unsigned Reg = LRegs[0];
581         SUnit *LRDef = LiveRegDefs[Reg];
582         MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
583         const TargetRegisterClass *RC =
584           TRI->getMinimalPhysRegClass(Reg, VT);
585         const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
586 
587         // If cross copy register class is the same as RC, then it must be
588         // possible copy the value directly. Do not try duplicate the def.
589         // If cross copy register class is not the same as RC, then it's
590         // possible to copy the value but it require cross register class copies
591         // and it is expensive.
592         // If cross copy register class is null, then it's not possible to copy
593         // the value at all.
594         SUnit *NewDef = nullptr;
595         if (DestRC != RC) {
596           NewDef = CopyAndMoveSuccessors(LRDef);
597           if (!DestRC && !NewDef)
598             report_fatal_error("Can't handle live physical "
599                                "register dependency!");
600         }
601         if (!NewDef) {
602           // Issue copies, these can be expensive cross register class copies.
603           SmallVector<SUnit*, 2> Copies;
604           InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
605           LLVM_DEBUG(dbgs() << "Adding an edge from SU # " << TrySU->NodeNum
606                             << " to SU #" << Copies.front()->NodeNum << "\n");
607           AddPred(TrySU, SDep(Copies.front(), SDep::Artificial));
608           NewDef = Copies.back();
609         }
610 
611         LLVM_DEBUG(dbgs() << "Adding an edge from SU # " << NewDef->NodeNum
612                           << " to SU #" << TrySU->NodeNum << "\n");
613         LiveRegDefs[Reg] = NewDef;
614         AddPred(NewDef, SDep(TrySU, SDep::Artificial));
615         TrySU->isAvailable = false;
616         CurSU = NewDef;
617       }
618 
619       if (!CurSU) {
620         llvm_unreachable("Unable to resolve live physical register dependencies!");
621       }
622     }
623 
624     // Add the nodes that aren't ready back onto the available list.
625     for (unsigned i = 0, e = NotReady.size(); i != e; ++i) {
626       NotReady[i]->isPending = false;
627       // May no longer be available due to backtracking.
628       if (NotReady[i]->isAvailable)
629         AvailableQueue.push(NotReady[i]);
630     }
631     NotReady.clear();
632 
633     if (CurSU)
634       ScheduleNodeBottomUp(CurSU, CurCycle);
635     ++CurCycle;
636   }
637 
638   // Reverse the order since it is bottom up.
639   std::reverse(Sequence.begin(), Sequence.end());
640 
641 #ifndef NDEBUG
642   VerifyScheduledSequence(/*isBottomUp=*/true);
643 #endif
644 }
645 
646 
647 namespace {
648 //===----------------------------------------------------------------------===//
649 // ScheduleDAGLinearize - No scheduling scheduler, it simply linearize the
650 // DAG in topological order.
651 // IMPORTANT: this may not work for targets with phyreg dependency.
652 //
653 class ScheduleDAGLinearize : public ScheduleDAGSDNodes {
654 public:
ScheduleDAGLinearize(MachineFunction & mf)655   ScheduleDAGLinearize(MachineFunction &mf) : ScheduleDAGSDNodes(mf) {}
656 
657   void Schedule() override;
658 
659   MachineBasicBlock *
660     EmitSchedule(MachineBasicBlock::iterator &InsertPos) override;
661 
662 private:
663   std::vector<SDNode*> Sequence;
664   DenseMap<SDNode*, SDNode*> GluedMap;  // Cache glue to its user
665 
666   void ScheduleNode(SDNode *N);
667 };
668 } // end anonymous namespace
669 
ScheduleNode(SDNode * N)670 void ScheduleDAGLinearize::ScheduleNode(SDNode *N) {
671   if (N->getNodeId() != 0)
672     llvm_unreachable(nullptr);
673 
674   if (!N->isMachineOpcode() &&
675       (N->getOpcode() == ISD::EntryToken || isPassiveNode(N)))
676     // These nodes do not need to be translated into MIs.
677     return;
678 
679   LLVM_DEBUG(dbgs() << "\n*** Scheduling: ");
680   LLVM_DEBUG(N->dump(DAG));
681   Sequence.push_back(N);
682 
683   unsigned NumOps = N->getNumOperands();
684   if (unsigned NumLeft = NumOps) {
685     SDNode *GluedOpN = nullptr;
686     do {
687       const SDValue &Op = N->getOperand(NumLeft-1);
688       SDNode *OpN = Op.getNode();
689 
690       if (NumLeft == NumOps && Op.getValueType() == MVT::Glue) {
691         // Schedule glue operand right above N.
692         GluedOpN = OpN;
693         assert(OpN->getNodeId() != 0 && "Glue operand not ready?");
694         OpN->setNodeId(0);
695         ScheduleNode(OpN);
696         continue;
697       }
698 
699       if (OpN == GluedOpN)
700         // Glue operand is already scheduled.
701         continue;
702 
703       DenseMap<SDNode*, SDNode*>::iterator DI = GluedMap.find(OpN);
704       if (DI != GluedMap.end() && DI->second != N)
705         // Users of glues are counted against the glued users.
706         OpN = DI->second;
707 
708       unsigned Degree = OpN->getNodeId();
709       assert(Degree > 0 && "Predecessor over-released!");
710       OpN->setNodeId(--Degree);
711       if (Degree == 0)
712         ScheduleNode(OpN);
713     } while (--NumLeft);
714   }
715 }
716 
717 /// findGluedUser - Find the representative use of a glue value by walking
718 /// the use chain.
findGluedUser(SDNode * N)719 static SDNode *findGluedUser(SDNode *N) {
720   while (SDNode *Glued = N->getGluedUser())
721     N = Glued;
722   return N;
723 }
724 
Schedule()725 void ScheduleDAGLinearize::Schedule() {
726   LLVM_DEBUG(dbgs() << "********** DAG Linearization **********\n");
727 
728   SmallVector<SDNode*, 8> Glues;
729   unsigned DAGSize = 0;
730   for (SDNode &Node : DAG->allnodes()) {
731     SDNode *N = &Node;
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 (const SDNode *U : Glue->uses())
761       if (U == ImmGUser)
762         --Degree;
763     GUser->setNodeId(UDegree + Degree);
764     Glue->setNodeId(1);
765   }
766 
767   Sequence.reserve(DAGSize);
768   ScheduleNode(DAG->getRoot().getNode());
769 }
770 
771 MachineBasicBlock*
EmitSchedule(MachineBasicBlock::iterator & InsertPos)772 ScheduleDAGLinearize::EmitSchedule(MachineBasicBlock::iterator &InsertPos) {
773   InstrEmitter Emitter(DAG->getTarget(), BB, InsertPos);
774   DenseMap<SDValue, Register> VRBaseMap;
775 
776   LLVM_DEBUG({ dbgs() << "\n*** Final schedule ***\n"; });
777 
778   unsigned NumNodes = Sequence.size();
779   MachineBasicBlock *BB = Emitter.getBlock();
780   for (unsigned i = 0; i != NumNodes; ++i) {
781     SDNode *N = Sequence[NumNodes-i-1];
782     LLVM_DEBUG(N->dump(DAG));
783     Emitter.EmitNode(N, false, false, VRBaseMap);
784 
785     // Emit any debug values associated with the node.
786     if (N->getHasDebugValue()) {
787       MachineBasicBlock::iterator InsertPos = Emitter.getInsertPos();
788       for (auto *DV : DAG->GetDbgValues(N)) {
789         if (!DV->isEmitted())
790           if (auto *DbgMI = Emitter.EmitDbgValue(DV, VRBaseMap))
791             BB->insert(InsertPos, DbgMI);
792       }
793     }
794   }
795 
796   LLVM_DEBUG(dbgs() << '\n');
797 
798   InsertPos = Emitter.getInsertPos();
799   return Emitter.getBlock();
800 }
801 
802 //===----------------------------------------------------------------------===//
803 //                         Public Constructor Functions
804 //===----------------------------------------------------------------------===//
805 
createFastDAGScheduler(SelectionDAGISel * IS,CodeGenOptLevel)806 llvm::ScheduleDAGSDNodes *llvm::createFastDAGScheduler(SelectionDAGISel *IS,
807                                                        CodeGenOptLevel) {
808   return new ScheduleDAGFast(*IS->MF);
809 }
810 
createDAGLinearizer(SelectionDAGISel * IS,CodeGenOptLevel)811 llvm::ScheduleDAGSDNodes *llvm::createDAGLinearizer(SelectionDAGISel *IS,
812                                                     CodeGenOptLevel) {
813   return new ScheduleDAGLinearize(*IS->MF);
814 }
815