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__anonf3b1b9750111::FastPriorityQueue51 bool empty() const { return Queue.empty(); }
52
push__anonf3b1b9750111::FastPriorityQueue53 void push(SUnit *U) {
54 Queue.push_back(U);
55 }
56
pop__anonf3b1b9750111::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 EVT VT = N->getValueType(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 EVT VT = Op.getNode()->getValueType(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 EVT getPhysicalRegisterVT(SDNode *N, unsigned Reg,
435 const TargetInstrInfo *TII) {
436 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
437 assert(MCID.ImplicitDefs && "Physical reg def must be in implicit def list!");
438 unsigned NumRes = MCID.getNumDefs();
439 for (const uint16_t *ImpDef = MCID.getImplicitDefs(); *ImpDef; ++ImpDef) {
440 if (Reg == *ImpDef)
441 break;
442 ++NumRes;
443 }
444 return N->getValueType(NumRes);
445 }
446
447 /// CheckForLiveRegDef - Return true and update live register vector if the
448 /// 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)449 static bool CheckForLiveRegDef(SUnit *SU, unsigned Reg,
450 std::vector<SUnit*> &LiveRegDefs,
451 SmallSet<unsigned, 4> &RegAdded,
452 SmallVectorImpl<unsigned> &LRegs,
453 const TargetRegisterInfo *TRI) {
454 bool Added = false;
455 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) {
456 if (LiveRegDefs[*AI] && LiveRegDefs[*AI] != SU) {
457 if (RegAdded.insert(*AI)) {
458 LRegs.push_back(*AI);
459 Added = true;
460 }
461 }
462 }
463 return Added;
464 }
465
466 /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
467 /// scheduling of the given node to satisfy live physical register dependencies.
468 /// If the specific node is the last one that's available to schedule, do
469 /// whatever is necessary (i.e. backtracking or cloning) to make it possible.
DelayForLiveRegsBottomUp(SUnit * SU,SmallVectorImpl<unsigned> & LRegs)470 bool ScheduleDAGFast::DelayForLiveRegsBottomUp(SUnit *SU,
471 SmallVectorImpl<unsigned> &LRegs){
472 if (NumLiveRegs == 0)
473 return false;
474
475 SmallSet<unsigned, 4> RegAdded;
476 // If this node would clobber any "live" register, then it's not ready.
477 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
478 I != E; ++I) {
479 if (I->isAssignedRegDep()) {
480 CheckForLiveRegDef(I->getSUnit(), I->getReg(), LiveRegDefs,
481 RegAdded, LRegs, TRI);
482 }
483 }
484
485 for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) {
486 if (Node->getOpcode() == ISD::INLINEASM) {
487 // Inline asm can clobber physical defs.
488 unsigned NumOps = Node->getNumOperands();
489 if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
490 --NumOps; // Ignore the glue operand.
491
492 for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
493 unsigned Flags =
494 cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
495 unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
496
497 ++i; // Skip the ID value.
498 if (InlineAsm::isRegDefKind(Flags) ||
499 InlineAsm::isRegDefEarlyClobberKind(Flags) ||
500 InlineAsm::isClobberKind(Flags)) {
501 // Check for def of register or earlyclobber register.
502 for (; NumVals; --NumVals, ++i) {
503 unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
504 if (TargetRegisterInfo::isPhysicalRegister(Reg))
505 CheckForLiveRegDef(SU, Reg, LiveRegDefs, RegAdded, LRegs, TRI);
506 }
507 } else
508 i += NumVals;
509 }
510 continue;
511 }
512 if (!Node->isMachineOpcode())
513 continue;
514 const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode());
515 if (!MCID.ImplicitDefs)
516 continue;
517 for (const uint16_t *Reg = MCID.getImplicitDefs(); *Reg; ++Reg) {
518 CheckForLiveRegDef(SU, *Reg, LiveRegDefs, RegAdded, LRegs, TRI);
519 }
520 }
521 return !LRegs.empty();
522 }
523
524
525 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
526 /// schedulers.
ListScheduleBottomUp()527 void ScheduleDAGFast::ListScheduleBottomUp() {
528 unsigned CurCycle = 0;
529
530 // Release any predecessors of the special Exit node.
531 ReleasePredecessors(&ExitSU, CurCycle);
532
533 // Add root to Available queue.
534 if (!SUnits.empty()) {
535 SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()];
536 assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!");
537 RootSU->isAvailable = true;
538 AvailableQueue.push(RootSU);
539 }
540
541 // While Available queue is not empty, grab the node with the highest
542 // priority. If it is not ready put it back. Schedule the node.
543 SmallVector<SUnit*, 4> NotReady;
544 DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMap;
545 Sequence.reserve(SUnits.size());
546 while (!AvailableQueue.empty()) {
547 bool Delayed = false;
548 LRegsMap.clear();
549 SUnit *CurSU = AvailableQueue.pop();
550 while (CurSU) {
551 SmallVector<unsigned, 4> LRegs;
552 if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
553 break;
554 Delayed = true;
555 LRegsMap.insert(std::make_pair(CurSU, LRegs));
556
557 CurSU->isPending = true; // This SU is not in AvailableQueue right now.
558 NotReady.push_back(CurSU);
559 CurSU = AvailableQueue.pop();
560 }
561
562 // All candidates are delayed due to live physical reg dependencies.
563 // Try code duplication or inserting cross class copies
564 // to resolve it.
565 if (Delayed && !CurSU) {
566 if (!CurSU) {
567 // Try duplicating the nodes that produces these
568 // "expensive to copy" values to break the dependency. In case even
569 // that doesn't work, insert cross class copies.
570 SUnit *TrySU = NotReady[0];
571 SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
572 assert(LRegs.size() == 1 && "Can't handle this yet!");
573 unsigned Reg = LRegs[0];
574 SUnit *LRDef = LiveRegDefs[Reg];
575 EVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
576 const TargetRegisterClass *RC =
577 TRI->getMinimalPhysRegClass(Reg, VT);
578 const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
579
580 // If cross copy register class is the same as RC, then it must be
581 // possible copy the value directly. Do not try duplicate the def.
582 // If cross copy register class is not the same as RC, then it's
583 // possible to copy the value but it require cross register class copies
584 // and it is expensive.
585 // If cross copy register class is null, then it's not possible to copy
586 // the value at all.
587 SUnit *NewDef = nullptr;
588 if (DestRC != RC) {
589 NewDef = CopyAndMoveSuccessors(LRDef);
590 if (!DestRC && !NewDef)
591 report_fatal_error("Can't handle live physical "
592 "register dependency!");
593 }
594 if (!NewDef) {
595 // Issue copies, these can be expensive cross register class copies.
596 SmallVector<SUnit*, 2> Copies;
597 InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
598 DEBUG(dbgs() << "Adding an edge from SU # " << TrySU->NodeNum
599 << " to SU #" << Copies.front()->NodeNum << "\n");
600 AddPred(TrySU, SDep(Copies.front(), SDep::Artificial));
601 NewDef = Copies.back();
602 }
603
604 DEBUG(dbgs() << "Adding an edge from SU # " << NewDef->NodeNum
605 << " to SU #" << TrySU->NodeNum << "\n");
606 LiveRegDefs[Reg] = NewDef;
607 AddPred(NewDef, SDep(TrySU, SDep::Artificial));
608 TrySU->isAvailable = false;
609 CurSU = NewDef;
610 }
611
612 if (!CurSU) {
613 llvm_unreachable("Unable to resolve live physical register dependencies!");
614 }
615 }
616
617 // Add the nodes that aren't ready back onto the available list.
618 for (unsigned i = 0, e = NotReady.size(); i != e; ++i) {
619 NotReady[i]->isPending = false;
620 // May no longer be available due to backtracking.
621 if (NotReady[i]->isAvailable)
622 AvailableQueue.push(NotReady[i]);
623 }
624 NotReady.clear();
625
626 if (CurSU)
627 ScheduleNodeBottomUp(CurSU, CurCycle);
628 ++CurCycle;
629 }
630
631 // Reverse the order since it is bottom up.
632 std::reverse(Sequence.begin(), Sequence.end());
633
634 #ifndef NDEBUG
635 VerifyScheduledSequence(/*isBottomUp=*/true);
636 #endif
637 }
638
639
640 namespace {
641 //===----------------------------------------------------------------------===//
642 // ScheduleDAGLinearize - No scheduling scheduler, it simply linearize the
643 // DAG in topological order.
644 // IMPORTANT: this may not work for targets with phyreg dependency.
645 //
646 class ScheduleDAGLinearize : public ScheduleDAGSDNodes {
647 public:
ScheduleDAGLinearize(MachineFunction & mf)648 ScheduleDAGLinearize(MachineFunction &mf) : ScheduleDAGSDNodes(mf) {}
649
650 void Schedule() override;
651
652 MachineBasicBlock *
653 EmitSchedule(MachineBasicBlock::iterator &InsertPos) override;
654
655 private:
656 std::vector<SDNode*> Sequence;
657 DenseMap<SDNode*, SDNode*> GluedMap; // Cache glue to its user
658
659 void ScheduleNode(SDNode *N);
660 };
661 } // end anonymous namespace
662
ScheduleNode(SDNode * N)663 void ScheduleDAGLinearize::ScheduleNode(SDNode *N) {
664 if (N->getNodeId() != 0)
665 llvm_unreachable(nullptr);
666
667 if (!N->isMachineOpcode() &&
668 (N->getOpcode() == ISD::EntryToken || isPassiveNode(N)))
669 // These nodes do not need to be translated into MIs.
670 return;
671
672 DEBUG(dbgs() << "\n*** Scheduling: ");
673 DEBUG(N->dump(DAG));
674 Sequence.push_back(N);
675
676 unsigned NumOps = N->getNumOperands();
677 if (unsigned NumLeft = NumOps) {
678 SDNode *GluedOpN = nullptr;
679 do {
680 const SDValue &Op = N->getOperand(NumLeft-1);
681 SDNode *OpN = Op.getNode();
682
683 if (NumLeft == NumOps && Op.getValueType() == MVT::Glue) {
684 // Schedule glue operand right above N.
685 GluedOpN = OpN;
686 assert(OpN->getNodeId() != 0 && "Glue operand not ready?");
687 OpN->setNodeId(0);
688 ScheduleNode(OpN);
689 continue;
690 }
691
692 if (OpN == GluedOpN)
693 // Glue operand is already scheduled.
694 continue;
695
696 DenseMap<SDNode*, SDNode*>::iterator DI = GluedMap.find(OpN);
697 if (DI != GluedMap.end() && DI->second != N)
698 // Users of glues are counted against the glued users.
699 OpN = DI->second;
700
701 unsigned Degree = OpN->getNodeId();
702 assert(Degree > 0 && "Predecessor over-released!");
703 OpN->setNodeId(--Degree);
704 if (Degree == 0)
705 ScheduleNode(OpN);
706 } while (--NumLeft);
707 }
708 }
709
710 /// findGluedUser - Find the representative use of a glue value by walking
711 /// the use chain.
findGluedUser(SDNode * N)712 static SDNode *findGluedUser(SDNode *N) {
713 while (SDNode *Glued = N->getGluedUser())
714 N = Glued;
715 return N;
716 }
717
Schedule()718 void ScheduleDAGLinearize::Schedule() {
719 DEBUG(dbgs() << "********** DAG Linearization **********\n");
720
721 SmallVector<SDNode*, 8> Glues;
722 unsigned DAGSize = 0;
723 for (SelectionDAG::allnodes_iterator I = DAG->allnodes_begin(),
724 E = DAG->allnodes_end(); I != E; ++I) {
725 SDNode *N = I;
726
727 // Use node id to record degree.
728 unsigned Degree = N->use_size();
729 N->setNodeId(Degree);
730 unsigned NumVals = N->getNumValues();
731 if (NumVals && N->getValueType(NumVals-1) == MVT::Glue &&
732 N->hasAnyUseOfValue(NumVals-1)) {
733 SDNode *User = findGluedUser(N);
734 if (User) {
735 Glues.push_back(N);
736 GluedMap.insert(std::make_pair(N, User));
737 }
738 }
739
740 if (N->isMachineOpcode() ||
741 (N->getOpcode() != ISD::EntryToken && !isPassiveNode(N)))
742 ++DAGSize;
743 }
744
745 for (unsigned i = 0, e = Glues.size(); i != e; ++i) {
746 SDNode *Glue = Glues[i];
747 SDNode *GUser = GluedMap[Glue];
748 unsigned Degree = Glue->getNodeId();
749 unsigned UDegree = GUser->getNodeId();
750
751 // Glue user must be scheduled together with the glue operand. So other
752 // users of the glue operand must be treated as its users.
753 SDNode *ImmGUser = Glue->getGluedUser();
754 for (SDNode::use_iterator ui = Glue->use_begin(), ue = Glue->use_end();
755 ui != ue; ++ui)
756 if (*ui == ImmGUser)
757 --Degree;
758 GUser->setNodeId(UDegree + Degree);
759 Glue->setNodeId(1);
760 }
761
762 Sequence.reserve(DAGSize);
763 ScheduleNode(DAG->getRoot().getNode());
764 }
765
766 MachineBasicBlock*
EmitSchedule(MachineBasicBlock::iterator & InsertPos)767 ScheduleDAGLinearize::EmitSchedule(MachineBasicBlock::iterator &InsertPos) {
768 InstrEmitter Emitter(BB, InsertPos);
769 DenseMap<SDValue, unsigned> VRBaseMap;
770
771 DEBUG({
772 dbgs() << "\n*** Final schedule ***\n";
773 });
774
775 // FIXME: Handle dbg_values.
776 unsigned NumNodes = Sequence.size();
777 for (unsigned i = 0; i != NumNodes; ++i) {
778 SDNode *N = Sequence[NumNodes-i-1];
779 DEBUG(N->dump(DAG));
780 Emitter.EmitNode(N, false, false, VRBaseMap);
781 }
782
783 DEBUG(dbgs() << '\n');
784
785 InsertPos = Emitter.getInsertPos();
786 return Emitter.getBlock();
787 }
788
789 //===----------------------------------------------------------------------===//
790 // Public Constructor Functions
791 //===----------------------------------------------------------------------===//
792
793 llvm::ScheduleDAGSDNodes *
createFastDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level)794 llvm::createFastDAGScheduler(SelectionDAGISel *IS, CodeGenOpt::Level) {
795 return new ScheduleDAGFast(*IS->MF);
796 }
797
798 llvm::ScheduleDAGSDNodes *
createDAGLinearizer(SelectionDAGISel * IS,CodeGenOpt::Level)799 llvm::createDAGLinearizer(SelectionDAGISel *IS, CodeGenOpt::Level) {
800 return new ScheduleDAGLinearize(*IS->MF);
801 }
802