1 //===----- ScheduleDAGRRList.cpp - Reg pressure reduction list scheduler --===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This implements bottom-up and top-down register pressure reduction list
11 // schedulers, using standard algorithms. The basic approach uses a priority
12 // queue of available nodes to schedule. One at a time, nodes are taken from
13 // the priority queue (thus in priority order), checked for legality to
14 // schedule, and emitted if legal.
15 //
16 //===----------------------------------------------------------------------===//
17
18 #include "llvm/CodeGen/SchedulerRegistry.h"
19 #include "ScheduleDAGSDNodes.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
25 #include "llvm/CodeGen/SelectionDAGISel.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/InlineAsm.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Target/TargetInstrInfo.h"
32 #include "llvm/Target/TargetLowering.h"
33 #include "llvm/Target/TargetRegisterInfo.h"
34 #include "llvm/Target/TargetSubtargetInfo.h"
35 #include <climits>
36 using namespace llvm;
37
38 #define DEBUG_TYPE "pre-RA-sched"
39
40 STATISTIC(NumBacktracks, "Number of times scheduler backtracked");
41 STATISTIC(NumUnfolds, "Number of nodes unfolded");
42 STATISTIC(NumDups, "Number of duplicated nodes");
43 STATISTIC(NumPRCopies, "Number of physical register copies");
44
45 static RegisterScheduler
46 burrListDAGScheduler("list-burr",
47 "Bottom-up register reduction list scheduling",
48 createBURRListDAGScheduler);
49 static RegisterScheduler
50 sourceListDAGScheduler("source",
51 "Similar to list-burr but schedules in source "
52 "order when possible",
53 createSourceListDAGScheduler);
54
55 static RegisterScheduler
56 hybridListDAGScheduler("list-hybrid",
57 "Bottom-up register pressure aware list scheduling "
58 "which tries to balance latency and register pressure",
59 createHybridListDAGScheduler);
60
61 static RegisterScheduler
62 ILPListDAGScheduler("list-ilp",
63 "Bottom-up register pressure aware list scheduling "
64 "which tries to balance ILP and register pressure",
65 createILPListDAGScheduler);
66
67 static cl::opt<bool> DisableSchedCycles(
68 "disable-sched-cycles", cl::Hidden, cl::init(false),
69 cl::desc("Disable cycle-level precision during preRA scheduling"));
70
71 // Temporary sched=list-ilp flags until the heuristics are robust.
72 // Some options are also available under sched=list-hybrid.
73 static cl::opt<bool> DisableSchedRegPressure(
74 "disable-sched-reg-pressure", cl::Hidden, cl::init(false),
75 cl::desc("Disable regpressure priority in sched=list-ilp"));
76 static cl::opt<bool> DisableSchedLiveUses(
77 "disable-sched-live-uses", cl::Hidden, cl::init(true),
78 cl::desc("Disable live use priority in sched=list-ilp"));
79 static cl::opt<bool> DisableSchedVRegCycle(
80 "disable-sched-vrcycle", cl::Hidden, cl::init(false),
81 cl::desc("Disable virtual register cycle interference checks"));
82 static cl::opt<bool> DisableSchedPhysRegJoin(
83 "disable-sched-physreg-join", cl::Hidden, cl::init(false),
84 cl::desc("Disable physreg def-use affinity"));
85 static cl::opt<bool> DisableSchedStalls(
86 "disable-sched-stalls", cl::Hidden, cl::init(true),
87 cl::desc("Disable no-stall priority in sched=list-ilp"));
88 static cl::opt<bool> DisableSchedCriticalPath(
89 "disable-sched-critical-path", cl::Hidden, cl::init(false),
90 cl::desc("Disable critical path priority in sched=list-ilp"));
91 static cl::opt<bool> DisableSchedHeight(
92 "disable-sched-height", cl::Hidden, cl::init(false),
93 cl::desc("Disable scheduled-height priority in sched=list-ilp"));
94 static cl::opt<bool> Disable2AddrHack(
95 "disable-2addr-hack", cl::Hidden, cl::init(true),
96 cl::desc("Disable scheduler's two-address hack"));
97
98 static cl::opt<int> MaxReorderWindow(
99 "max-sched-reorder", cl::Hidden, cl::init(6),
100 cl::desc("Number of instructions to allow ahead of the critical path "
101 "in sched=list-ilp"));
102
103 static cl::opt<unsigned> AvgIPC(
104 "sched-avg-ipc", cl::Hidden, cl::init(1),
105 cl::desc("Average inst/cycle whan no target itinerary exists."));
106
107 namespace {
108 //===----------------------------------------------------------------------===//
109 /// ScheduleDAGRRList - The actual register reduction list scheduler
110 /// implementation. This supports both top-down and bottom-up scheduling.
111 ///
112 class ScheduleDAGRRList : public ScheduleDAGSDNodes {
113 private:
114 /// NeedLatency - True if the scheduler will make use of latency information.
115 ///
116 bool NeedLatency;
117
118 /// AvailableQueue - The priority queue to use for the available SUnits.
119 SchedulingPriorityQueue *AvailableQueue;
120
121 /// PendingQueue - This contains all of the instructions whose operands have
122 /// been issued, but their results are not ready yet (due to the latency of
123 /// the operation). Once the operands becomes available, the instruction is
124 /// added to the AvailableQueue.
125 std::vector<SUnit*> PendingQueue;
126
127 /// HazardRec - The hazard recognizer to use.
128 ScheduleHazardRecognizer *HazardRec;
129
130 /// CurCycle - The current scheduler state corresponds to this cycle.
131 unsigned CurCycle;
132
133 /// MinAvailableCycle - Cycle of the soonest available instruction.
134 unsigned MinAvailableCycle;
135
136 /// IssueCount - Count instructions issued in this cycle
137 /// Currently valid only for bottom-up scheduling.
138 unsigned IssueCount;
139
140 /// LiveRegDefs - A set of physical registers and their definition
141 /// that are "live". These nodes must be scheduled before any other nodes that
142 /// modifies the registers can be scheduled.
143 unsigned NumLiveRegs;
144 std::vector<SUnit*> LiveRegDefs;
145 std::vector<SUnit*> LiveRegGens;
146
147 // Collect interferences between physical register use/defs.
148 // Each interference is an SUnit and set of physical registers.
149 SmallVector<SUnit*, 4> Interferences;
150 typedef DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMapT;
151 LRegsMapT LRegsMap;
152
153 /// Topo - A topological ordering for SUnits which permits fast IsReachable
154 /// and similar queries.
155 ScheduleDAGTopologicalSort Topo;
156
157 // Hack to keep track of the inverse of FindCallSeqStart without more crazy
158 // DAG crawling.
159 DenseMap<SUnit*, SUnit*> CallSeqEndForStart;
160
161 public:
ScheduleDAGRRList(MachineFunction & mf,bool needlatency,SchedulingPriorityQueue * availqueue,CodeGenOpt::Level OptLevel)162 ScheduleDAGRRList(MachineFunction &mf, bool needlatency,
163 SchedulingPriorityQueue *availqueue,
164 CodeGenOpt::Level OptLevel)
165 : ScheduleDAGSDNodes(mf),
166 NeedLatency(needlatency), AvailableQueue(availqueue), CurCycle(0),
167 Topo(SUnits, nullptr) {
168
169 const TargetSubtargetInfo &STI = mf.getSubtarget();
170 if (DisableSchedCycles || !NeedLatency)
171 HazardRec = new ScheduleHazardRecognizer();
172 else
173 HazardRec = STI.getInstrInfo()->CreateTargetHazardRecognizer(&STI, this);
174 }
175
~ScheduleDAGRRList()176 ~ScheduleDAGRRList() override {
177 delete HazardRec;
178 delete AvailableQueue;
179 }
180
181 void Schedule() override;
182
getHazardRec()183 ScheduleHazardRecognizer *getHazardRec() { return HazardRec; }
184
185 /// IsReachable - Checks if SU is reachable from TargetSU.
IsReachable(const SUnit * SU,const SUnit * TargetSU)186 bool IsReachable(const SUnit *SU, const SUnit *TargetSU) {
187 return Topo.IsReachable(SU, TargetSU);
188 }
189
190 /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
191 /// create a cycle.
WillCreateCycle(SUnit * SU,SUnit * TargetSU)192 bool WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
193 return Topo.WillCreateCycle(SU, TargetSU);
194 }
195
196 /// AddPred - adds a predecessor edge to SUnit SU.
197 /// This returns true if this is a new predecessor.
198 /// Updates the topological ordering if required.
AddPred(SUnit * SU,const SDep & D)199 void AddPred(SUnit *SU, const SDep &D) {
200 Topo.AddPred(SU, D.getSUnit());
201 SU->addPred(D);
202 }
203
204 /// RemovePred - removes a predecessor edge from SUnit SU.
205 /// This returns true if an edge was removed.
206 /// Updates the topological ordering if required.
RemovePred(SUnit * SU,const SDep & D)207 void RemovePred(SUnit *SU, const SDep &D) {
208 Topo.RemovePred(SU, D.getSUnit());
209 SU->removePred(D);
210 }
211
212 private:
isReady(SUnit * SU)213 bool isReady(SUnit *SU) {
214 return DisableSchedCycles || !AvailableQueue->hasReadyFilter() ||
215 AvailableQueue->isReady(SU);
216 }
217
218 void ReleasePred(SUnit *SU, const SDep *PredEdge);
219 void ReleasePredecessors(SUnit *SU);
220 void ReleasePending();
221 void AdvanceToCycle(unsigned NextCycle);
222 void AdvancePastStalls(SUnit *SU);
223 void EmitNode(SUnit *SU);
224 void ScheduleNodeBottomUp(SUnit*);
225 void CapturePred(SDep *PredEdge);
226 void UnscheduleNodeBottomUp(SUnit*);
227 void RestoreHazardCheckerBottomUp();
228 void BacktrackBottomUp(SUnit*, SUnit*);
229 SUnit *CopyAndMoveSuccessors(SUnit*);
230 void InsertCopiesAndMoveSuccs(SUnit*, unsigned,
231 const TargetRegisterClass*,
232 const TargetRegisterClass*,
233 SmallVectorImpl<SUnit*>&);
234 bool DelayForLiveRegsBottomUp(SUnit*, SmallVectorImpl<unsigned>&);
235
236 void releaseInterferences(unsigned Reg = 0);
237
238 SUnit *PickNodeToScheduleBottomUp();
239 void ListScheduleBottomUp();
240
241 /// CreateNewSUnit - Creates a new SUnit and returns a pointer to it.
242 /// Updates the topological ordering if required.
CreateNewSUnit(SDNode * N)243 SUnit *CreateNewSUnit(SDNode *N) {
244 unsigned NumSUnits = SUnits.size();
245 SUnit *NewNode = newSUnit(N);
246 // Update the topological ordering.
247 if (NewNode->NodeNum >= NumSUnits)
248 Topo.InitDAGTopologicalSorting();
249 return NewNode;
250 }
251
252 /// CreateClone - Creates a new SUnit from an existing one.
253 /// Updates the topological ordering if required.
CreateClone(SUnit * N)254 SUnit *CreateClone(SUnit *N) {
255 unsigned NumSUnits = SUnits.size();
256 SUnit *NewNode = Clone(N);
257 // Update the topological ordering.
258 if (NewNode->NodeNum >= NumSUnits)
259 Topo.InitDAGTopologicalSorting();
260 return NewNode;
261 }
262
263 /// forceUnitLatencies - Register-pressure-reducing scheduling doesn't
264 /// need actual latency information but the hybrid scheduler does.
forceUnitLatencies() const265 bool forceUnitLatencies() const override {
266 return !NeedLatency;
267 }
268 };
269 } // end anonymous namespace
270
271 /// GetCostForDef - Looks up the register class and cost for a given definition.
272 /// Typically this just means looking up the representative register class,
273 /// but for untyped values (MVT::Untyped) it means inspecting the node's
274 /// opcode to determine what register class is being generated.
GetCostForDef(const ScheduleDAGSDNodes::RegDefIter & RegDefPos,const TargetLowering * TLI,const TargetInstrInfo * TII,const TargetRegisterInfo * TRI,unsigned & RegClass,unsigned & Cost,const MachineFunction & MF)275 static void GetCostForDef(const ScheduleDAGSDNodes::RegDefIter &RegDefPos,
276 const TargetLowering *TLI,
277 const TargetInstrInfo *TII,
278 const TargetRegisterInfo *TRI,
279 unsigned &RegClass, unsigned &Cost,
280 const MachineFunction &MF) {
281 MVT VT = RegDefPos.GetValue();
282
283 // Special handling for untyped values. These values can only come from
284 // the expansion of custom DAG-to-DAG patterns.
285 if (VT == MVT::Untyped) {
286 const SDNode *Node = RegDefPos.GetNode();
287
288 // Special handling for CopyFromReg of untyped values.
289 if (!Node->isMachineOpcode() && Node->getOpcode() == ISD::CopyFromReg) {
290 unsigned Reg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
291 const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(Reg);
292 RegClass = RC->getID();
293 Cost = 1;
294 return;
295 }
296
297 unsigned Opcode = Node->getMachineOpcode();
298 if (Opcode == TargetOpcode::REG_SEQUENCE) {
299 unsigned DstRCIdx = cast<ConstantSDNode>(Node->getOperand(0))->getZExtValue();
300 const TargetRegisterClass *RC = TRI->getRegClass(DstRCIdx);
301 RegClass = RC->getID();
302 Cost = 1;
303 return;
304 }
305
306 unsigned Idx = RegDefPos.GetIdx();
307 const MCInstrDesc Desc = TII->get(Opcode);
308 const TargetRegisterClass *RC = TII->getRegClass(Desc, Idx, TRI, MF);
309 RegClass = RC->getID();
310 // FIXME: Cost arbitrarily set to 1 because there doesn't seem to be a
311 // better way to determine it.
312 Cost = 1;
313 } else {
314 RegClass = TLI->getRepRegClassFor(VT)->getID();
315 Cost = TLI->getRepRegClassCostFor(VT);
316 }
317 }
318
319 /// Schedule - Schedule the DAG using list scheduling.
Schedule()320 void ScheduleDAGRRList::Schedule() {
321 DEBUG(dbgs()
322 << "********** List Scheduling BB#" << BB->getNumber()
323 << " '" << BB->getName() << "' **********\n");
324
325 CurCycle = 0;
326 IssueCount = 0;
327 MinAvailableCycle = DisableSchedCycles ? 0 : UINT_MAX;
328 NumLiveRegs = 0;
329 // Allocate slots for each physical register, plus one for a special register
330 // to track the virtual resource of a calling sequence.
331 LiveRegDefs.resize(TRI->getNumRegs() + 1, nullptr);
332 LiveRegGens.resize(TRI->getNumRegs() + 1, nullptr);
333 CallSeqEndForStart.clear();
334 assert(Interferences.empty() && LRegsMap.empty() && "stale Interferences");
335
336 // Build the scheduling graph.
337 BuildSchedGraph(nullptr);
338
339 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
340 SUnits[su].dumpAll(this));
341 Topo.InitDAGTopologicalSorting();
342
343 AvailableQueue->initNodes(SUnits);
344
345 HazardRec->Reset();
346
347 // Execute the actual scheduling loop.
348 ListScheduleBottomUp();
349
350 AvailableQueue->releaseState();
351
352 DEBUG({
353 dbgs() << "*** Final schedule ***\n";
354 dumpSchedule();
355 dbgs() << '\n';
356 });
357 }
358
359 //===----------------------------------------------------------------------===//
360 // Bottom-Up Scheduling
361 //===----------------------------------------------------------------------===//
362
363 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
364 /// the AvailableQueue if the count reaches zero. Also update its cycle bound.
ReleasePred(SUnit * SU,const SDep * PredEdge)365 void ScheduleDAGRRList::ReleasePred(SUnit *SU, const SDep *PredEdge) {
366 SUnit *PredSU = PredEdge->getSUnit();
367
368 #ifndef NDEBUG
369 if (PredSU->NumSuccsLeft == 0) {
370 dbgs() << "*** Scheduling failed! ***\n";
371 PredSU->dump(this);
372 dbgs() << " has been released too many times!\n";
373 llvm_unreachable(nullptr);
374 }
375 #endif
376 --PredSU->NumSuccsLeft;
377
378 if (!forceUnitLatencies()) {
379 // Updating predecessor's height. This is now the cycle when the
380 // predecessor can be scheduled without causing a pipeline stall.
381 PredSU->setHeightToAtLeast(SU->getHeight() + PredEdge->getLatency());
382 }
383
384 // If all the node's successors are scheduled, this node is ready
385 // to be scheduled. Ignore the special EntrySU node.
386 if (PredSU->NumSuccsLeft == 0 && PredSU != &EntrySU) {
387 PredSU->isAvailable = true;
388
389 unsigned Height = PredSU->getHeight();
390 if (Height < MinAvailableCycle)
391 MinAvailableCycle = Height;
392
393 if (isReady(PredSU)) {
394 AvailableQueue->push(PredSU);
395 }
396 // CapturePred and others may have left the node in the pending queue, avoid
397 // adding it twice.
398 else if (!PredSU->isPending) {
399 PredSU->isPending = true;
400 PendingQueue.push_back(PredSU);
401 }
402 }
403 }
404
405 /// IsChainDependent - Test if Outer is reachable from Inner through
406 /// chain dependencies.
IsChainDependent(SDNode * Outer,SDNode * Inner,unsigned NestLevel,const TargetInstrInfo * TII)407 static bool IsChainDependent(SDNode *Outer, SDNode *Inner,
408 unsigned NestLevel,
409 const TargetInstrInfo *TII) {
410 SDNode *N = Outer;
411 for (;;) {
412 if (N == Inner)
413 return true;
414 // For a TokenFactor, examine each operand. There may be multiple ways
415 // to get to the CALLSEQ_BEGIN, but we need to find the path with the
416 // most nesting in order to ensure that we find the corresponding match.
417 if (N->getOpcode() == ISD::TokenFactor) {
418 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
419 if (IsChainDependent(N->getOperand(i).getNode(), Inner, NestLevel, TII))
420 return true;
421 return false;
422 }
423 // Check for a lowered CALLSEQ_BEGIN or CALLSEQ_END.
424 if (N->isMachineOpcode()) {
425 if (N->getMachineOpcode() ==
426 (unsigned)TII->getCallFrameDestroyOpcode()) {
427 ++NestLevel;
428 } else if (N->getMachineOpcode() ==
429 (unsigned)TII->getCallFrameSetupOpcode()) {
430 if (NestLevel == 0)
431 return false;
432 --NestLevel;
433 }
434 }
435 // Otherwise, find the chain and continue climbing.
436 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
437 if (N->getOperand(i).getValueType() == MVT::Other) {
438 N = N->getOperand(i).getNode();
439 goto found_chain_operand;
440 }
441 return false;
442 found_chain_operand:;
443 if (N->getOpcode() == ISD::EntryToken)
444 return false;
445 }
446 }
447
448 /// FindCallSeqStart - Starting from the (lowered) CALLSEQ_END node, locate
449 /// the corresponding (lowered) CALLSEQ_BEGIN node.
450 ///
451 /// NestLevel and MaxNested are used in recursion to indcate the current level
452 /// of nesting of CALLSEQ_BEGIN and CALLSEQ_END pairs, as well as the maximum
453 /// level seen so far.
454 ///
455 /// TODO: It would be better to give CALLSEQ_END an explicit operand to point
456 /// to the corresponding CALLSEQ_BEGIN to avoid needing to search for it.
457 static SDNode *
FindCallSeqStart(SDNode * N,unsigned & NestLevel,unsigned & MaxNest,const TargetInstrInfo * TII)458 FindCallSeqStart(SDNode *N, unsigned &NestLevel, unsigned &MaxNest,
459 const TargetInstrInfo *TII) {
460 for (;;) {
461 // For a TokenFactor, examine each operand. There may be multiple ways
462 // to get to the CALLSEQ_BEGIN, but we need to find the path with the
463 // most nesting in order to ensure that we find the corresponding match.
464 if (N->getOpcode() == ISD::TokenFactor) {
465 SDNode *Best = nullptr;
466 unsigned BestMaxNest = MaxNest;
467 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
468 unsigned MyNestLevel = NestLevel;
469 unsigned MyMaxNest = MaxNest;
470 if (SDNode *New = FindCallSeqStart(N->getOperand(i).getNode(),
471 MyNestLevel, MyMaxNest, TII))
472 if (!Best || (MyMaxNest > BestMaxNest)) {
473 Best = New;
474 BestMaxNest = MyMaxNest;
475 }
476 }
477 assert(Best);
478 MaxNest = BestMaxNest;
479 return Best;
480 }
481 // Check for a lowered CALLSEQ_BEGIN or CALLSEQ_END.
482 if (N->isMachineOpcode()) {
483 if (N->getMachineOpcode() ==
484 (unsigned)TII->getCallFrameDestroyOpcode()) {
485 ++NestLevel;
486 MaxNest = std::max(MaxNest, NestLevel);
487 } else if (N->getMachineOpcode() ==
488 (unsigned)TII->getCallFrameSetupOpcode()) {
489 assert(NestLevel != 0);
490 --NestLevel;
491 if (NestLevel == 0)
492 return N;
493 }
494 }
495 // Otherwise, find the chain and continue climbing.
496 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
497 if (N->getOperand(i).getValueType() == MVT::Other) {
498 N = N->getOperand(i).getNode();
499 goto found_chain_operand;
500 }
501 return nullptr;
502 found_chain_operand:;
503 if (N->getOpcode() == ISD::EntryToken)
504 return nullptr;
505 }
506 }
507
508 /// Call ReleasePred for each predecessor, then update register live def/gen.
509 /// Always update LiveRegDefs for a register dependence even if the current SU
510 /// also defines the register. This effectively create one large live range
511 /// across a sequence of two-address node. This is important because the
512 /// entire chain must be scheduled together. Example:
513 ///
514 /// flags = (3) add
515 /// flags = (2) addc flags
516 /// flags = (1) addc flags
517 ///
518 /// results in
519 ///
520 /// LiveRegDefs[flags] = 3
521 /// LiveRegGens[flags] = 1
522 ///
523 /// If (2) addc is unscheduled, then (1) addc must also be unscheduled to avoid
524 /// interference on flags.
ReleasePredecessors(SUnit * SU)525 void ScheduleDAGRRList::ReleasePredecessors(SUnit *SU) {
526 // Bottom up: release predecessors
527 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
528 I != E; ++I) {
529 ReleasePred(SU, &*I);
530 if (I->isAssignedRegDep()) {
531 // This is a physical register dependency and it's impossible or
532 // expensive to copy the register. Make sure nothing that can
533 // clobber the register is scheduled between the predecessor and
534 // this node.
535 SUnit *RegDef = LiveRegDefs[I->getReg()]; (void)RegDef;
536 assert((!RegDef || RegDef == SU || RegDef == I->getSUnit()) &&
537 "interference on register dependence");
538 LiveRegDefs[I->getReg()] = I->getSUnit();
539 if (!LiveRegGens[I->getReg()]) {
540 ++NumLiveRegs;
541 LiveRegGens[I->getReg()] = SU;
542 }
543 }
544 }
545
546 // If we're scheduling a lowered CALLSEQ_END, find the corresponding
547 // CALLSEQ_BEGIN. Inject an artificial physical register dependence between
548 // these nodes, to prevent other calls from being interscheduled with them.
549 unsigned CallResource = TRI->getNumRegs();
550 if (!LiveRegDefs[CallResource])
551 for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode())
552 if (Node->isMachineOpcode() &&
553 Node->getMachineOpcode() == (unsigned)TII->getCallFrameDestroyOpcode()) {
554 unsigned NestLevel = 0;
555 unsigned MaxNest = 0;
556 SDNode *N = FindCallSeqStart(Node, NestLevel, MaxNest, TII);
557
558 SUnit *Def = &SUnits[N->getNodeId()];
559 CallSeqEndForStart[Def] = SU;
560
561 ++NumLiveRegs;
562 LiveRegDefs[CallResource] = Def;
563 LiveRegGens[CallResource] = SU;
564 break;
565 }
566 }
567
568 /// Check to see if any of the pending instructions are ready to issue. If
569 /// so, add them to the available queue.
ReleasePending()570 void ScheduleDAGRRList::ReleasePending() {
571 if (DisableSchedCycles) {
572 assert(PendingQueue.empty() && "pending instrs not allowed in this mode");
573 return;
574 }
575
576 // If the available queue is empty, it is safe to reset MinAvailableCycle.
577 if (AvailableQueue->empty())
578 MinAvailableCycle = UINT_MAX;
579
580 // Check to see if any of the pending instructions are ready to issue. If
581 // so, add them to the available queue.
582 for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
583 unsigned ReadyCycle = PendingQueue[i]->getHeight();
584 if (ReadyCycle < MinAvailableCycle)
585 MinAvailableCycle = ReadyCycle;
586
587 if (PendingQueue[i]->isAvailable) {
588 if (!isReady(PendingQueue[i]))
589 continue;
590 AvailableQueue->push(PendingQueue[i]);
591 }
592 PendingQueue[i]->isPending = false;
593 PendingQueue[i] = PendingQueue.back();
594 PendingQueue.pop_back();
595 --i; --e;
596 }
597 }
598
599 /// Move the scheduler state forward by the specified number of Cycles.
AdvanceToCycle(unsigned NextCycle)600 void ScheduleDAGRRList::AdvanceToCycle(unsigned NextCycle) {
601 if (NextCycle <= CurCycle)
602 return;
603
604 IssueCount = 0;
605 AvailableQueue->setCurCycle(NextCycle);
606 if (!HazardRec->isEnabled()) {
607 // Bypass lots of virtual calls in case of long latency.
608 CurCycle = NextCycle;
609 }
610 else {
611 for (; CurCycle != NextCycle; ++CurCycle) {
612 HazardRec->RecedeCycle();
613 }
614 }
615 // FIXME: Instead of visiting the pending Q each time, set a dirty flag on the
616 // available Q to release pending nodes at least once before popping.
617 ReleasePending();
618 }
619
620 /// Move the scheduler state forward until the specified node's dependents are
621 /// ready and can be scheduled with no resource conflicts.
AdvancePastStalls(SUnit * SU)622 void ScheduleDAGRRList::AdvancePastStalls(SUnit *SU) {
623 if (DisableSchedCycles)
624 return;
625
626 // FIXME: Nodes such as CopyFromReg probably should not advance the current
627 // cycle. Otherwise, we can wrongly mask real stalls. If the non-machine node
628 // has predecessors the cycle will be advanced when they are scheduled.
629 // But given the crude nature of modeling latency though such nodes, we
630 // currently need to treat these nodes like real instructions.
631 // if (!SU->getNode() || !SU->getNode()->isMachineOpcode()) return;
632
633 unsigned ReadyCycle = SU->getHeight();
634
635 // Bump CurCycle to account for latency. We assume the latency of other
636 // available instructions may be hidden by the stall (not a full pipe stall).
637 // This updates the hazard recognizer's cycle before reserving resources for
638 // this instruction.
639 AdvanceToCycle(ReadyCycle);
640
641 // Calls are scheduled in their preceding cycle, so don't conflict with
642 // hazards from instructions after the call. EmitNode will reset the
643 // scoreboard state before emitting the call.
644 if (SU->isCall)
645 return;
646
647 // FIXME: For resource conflicts in very long non-pipelined stages, we
648 // should probably skip ahead here to avoid useless scoreboard checks.
649 int Stalls = 0;
650 while (true) {
651 ScheduleHazardRecognizer::HazardType HT =
652 HazardRec->getHazardType(SU, -Stalls);
653
654 if (HT == ScheduleHazardRecognizer::NoHazard)
655 break;
656
657 ++Stalls;
658 }
659 AdvanceToCycle(CurCycle + Stalls);
660 }
661
662 /// Record this SUnit in the HazardRecognizer.
663 /// Does not update CurCycle.
EmitNode(SUnit * SU)664 void ScheduleDAGRRList::EmitNode(SUnit *SU) {
665 if (!HazardRec->isEnabled())
666 return;
667
668 // Check for phys reg copy.
669 if (!SU->getNode())
670 return;
671
672 switch (SU->getNode()->getOpcode()) {
673 default:
674 assert(SU->getNode()->isMachineOpcode() &&
675 "This target-independent node should not be scheduled.");
676 break;
677 case ISD::MERGE_VALUES:
678 case ISD::TokenFactor:
679 case ISD::LIFETIME_START:
680 case ISD::LIFETIME_END:
681 case ISD::CopyToReg:
682 case ISD::CopyFromReg:
683 case ISD::EH_LABEL:
684 // Noops don't affect the scoreboard state. Copies are likely to be
685 // removed.
686 return;
687 case ISD::INLINEASM:
688 // For inline asm, clear the pipeline state.
689 HazardRec->Reset();
690 return;
691 }
692 if (SU->isCall) {
693 // Calls are scheduled with their preceding instructions. For bottom-up
694 // scheduling, clear the pipeline state before emitting.
695 HazardRec->Reset();
696 }
697
698 HazardRec->EmitInstruction(SU);
699 }
700
701 static void resetVRegCycle(SUnit *SU);
702
703 /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
704 /// count of its predecessors. If a predecessor pending count is zero, add it to
705 /// the Available queue.
ScheduleNodeBottomUp(SUnit * SU)706 void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU) {
707 DEBUG(dbgs() << "\n*** Scheduling [" << CurCycle << "]: ");
708 DEBUG(SU->dump(this));
709
710 #ifndef NDEBUG
711 if (CurCycle < SU->getHeight())
712 DEBUG(dbgs() << " Height [" << SU->getHeight()
713 << "] pipeline stall!\n");
714 #endif
715
716 // FIXME: Do not modify node height. It may interfere with
717 // backtracking. Instead add a "ready cycle" to SUnit. Before scheduling the
718 // node its ready cycle can aid heuristics, and after scheduling it can
719 // indicate the scheduled cycle.
720 SU->setHeightToAtLeast(CurCycle);
721
722 // Reserve resources for the scheduled instruction.
723 EmitNode(SU);
724
725 Sequence.push_back(SU);
726
727 AvailableQueue->scheduledNode(SU);
728
729 // If HazardRec is disabled, and each inst counts as one cycle, then
730 // advance CurCycle before ReleasePredecessors to avoid useless pushes to
731 // PendingQueue for schedulers that implement HasReadyFilter.
732 if (!HazardRec->isEnabled() && AvgIPC < 2)
733 AdvanceToCycle(CurCycle + 1);
734
735 // Update liveness of predecessors before successors to avoid treating a
736 // two-address node as a live range def.
737 ReleasePredecessors(SU);
738
739 // Release all the implicit physical register defs that are live.
740 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
741 I != E; ++I) {
742 // LiveRegDegs[I->getReg()] != SU when SU is a two-address node.
743 if (I->isAssignedRegDep() && LiveRegDefs[I->getReg()] == SU) {
744 assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
745 --NumLiveRegs;
746 LiveRegDefs[I->getReg()] = nullptr;
747 LiveRegGens[I->getReg()] = nullptr;
748 releaseInterferences(I->getReg());
749 }
750 }
751 // Release the special call resource dependence, if this is the beginning
752 // of a call.
753 unsigned CallResource = TRI->getNumRegs();
754 if (LiveRegDefs[CallResource] == SU)
755 for (const SDNode *SUNode = SU->getNode(); SUNode;
756 SUNode = SUNode->getGluedNode()) {
757 if (SUNode->isMachineOpcode() &&
758 SUNode->getMachineOpcode() == (unsigned)TII->getCallFrameSetupOpcode()) {
759 assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
760 --NumLiveRegs;
761 LiveRegDefs[CallResource] = nullptr;
762 LiveRegGens[CallResource] = nullptr;
763 releaseInterferences(CallResource);
764 }
765 }
766
767 resetVRegCycle(SU);
768
769 SU->isScheduled = true;
770
771 // Conditions under which the scheduler should eagerly advance the cycle:
772 // (1) No available instructions
773 // (2) All pipelines full, so available instructions must have hazards.
774 //
775 // If HazardRec is disabled, the cycle was pre-advanced before calling
776 // ReleasePredecessors. In that case, IssueCount should remain 0.
777 //
778 // Check AvailableQueue after ReleasePredecessors in case of zero latency.
779 if (HazardRec->isEnabled() || AvgIPC > 1) {
780 if (SU->getNode() && SU->getNode()->isMachineOpcode())
781 ++IssueCount;
782 if ((HazardRec->isEnabled() && HazardRec->atIssueLimit())
783 || (!HazardRec->isEnabled() && IssueCount == AvgIPC))
784 AdvanceToCycle(CurCycle + 1);
785 }
786 }
787
788 /// CapturePred - This does the opposite of ReleasePred. Since SU is being
789 /// unscheduled, incrcease the succ left count of its predecessors. Remove
790 /// them from AvailableQueue if necessary.
CapturePred(SDep * PredEdge)791 void ScheduleDAGRRList::CapturePred(SDep *PredEdge) {
792 SUnit *PredSU = PredEdge->getSUnit();
793 if (PredSU->isAvailable) {
794 PredSU->isAvailable = false;
795 if (!PredSU->isPending)
796 AvailableQueue->remove(PredSU);
797 }
798
799 assert(PredSU->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!");
800 ++PredSU->NumSuccsLeft;
801 }
802
803 /// UnscheduleNodeBottomUp - Remove the node from the schedule, update its and
804 /// its predecessor states to reflect the change.
UnscheduleNodeBottomUp(SUnit * SU)805 void ScheduleDAGRRList::UnscheduleNodeBottomUp(SUnit *SU) {
806 DEBUG(dbgs() << "*** Unscheduling [" << SU->getHeight() << "]: ");
807 DEBUG(SU->dump(this));
808
809 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
810 I != E; ++I) {
811 CapturePred(&*I);
812 if (I->isAssignedRegDep() && SU == LiveRegGens[I->getReg()]){
813 assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
814 assert(LiveRegDefs[I->getReg()] == I->getSUnit() &&
815 "Physical register dependency violated?");
816 --NumLiveRegs;
817 LiveRegDefs[I->getReg()] = nullptr;
818 LiveRegGens[I->getReg()] = nullptr;
819 releaseInterferences(I->getReg());
820 }
821 }
822
823 // Reclaim the special call resource dependence, if this is the beginning
824 // of a call.
825 unsigned CallResource = TRI->getNumRegs();
826 for (const SDNode *SUNode = SU->getNode(); SUNode;
827 SUNode = SUNode->getGluedNode()) {
828 if (SUNode->isMachineOpcode() &&
829 SUNode->getMachineOpcode() == (unsigned)TII->getCallFrameSetupOpcode()) {
830 ++NumLiveRegs;
831 LiveRegDefs[CallResource] = SU;
832 LiveRegGens[CallResource] = CallSeqEndForStart[SU];
833 }
834 }
835
836 // Release the special call resource dependence, if this is the end
837 // of a call.
838 if (LiveRegGens[CallResource] == SU)
839 for (const SDNode *SUNode = SU->getNode(); SUNode;
840 SUNode = SUNode->getGluedNode()) {
841 if (SUNode->isMachineOpcode() &&
842 SUNode->getMachineOpcode() == (unsigned)TII->getCallFrameDestroyOpcode()) {
843 assert(NumLiveRegs > 0 && "NumLiveRegs is already zero!");
844 --NumLiveRegs;
845 LiveRegDefs[CallResource] = nullptr;
846 LiveRegGens[CallResource] = nullptr;
847 releaseInterferences(CallResource);
848 }
849 }
850
851 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
852 I != E; ++I) {
853 if (I->isAssignedRegDep()) {
854 if (!LiveRegDefs[I->getReg()])
855 ++NumLiveRegs;
856 // This becomes the nearest def. Note that an earlier def may still be
857 // pending if this is a two-address node.
858 LiveRegDefs[I->getReg()] = SU;
859 if (LiveRegGens[I->getReg()] == nullptr ||
860 I->getSUnit()->getHeight() < LiveRegGens[I->getReg()]->getHeight())
861 LiveRegGens[I->getReg()] = I->getSUnit();
862 }
863 }
864 if (SU->getHeight() < MinAvailableCycle)
865 MinAvailableCycle = SU->getHeight();
866
867 SU->setHeightDirty();
868 SU->isScheduled = false;
869 SU->isAvailable = true;
870 if (!DisableSchedCycles && AvailableQueue->hasReadyFilter()) {
871 // Don't make available until backtracking is complete.
872 SU->isPending = true;
873 PendingQueue.push_back(SU);
874 }
875 else {
876 AvailableQueue->push(SU);
877 }
878 AvailableQueue->unscheduledNode(SU);
879 }
880
881 /// After backtracking, the hazard checker needs to be restored to a state
882 /// corresponding the current cycle.
RestoreHazardCheckerBottomUp()883 void ScheduleDAGRRList::RestoreHazardCheckerBottomUp() {
884 HazardRec->Reset();
885
886 unsigned LookAhead = std::min((unsigned)Sequence.size(),
887 HazardRec->getMaxLookAhead());
888 if (LookAhead == 0)
889 return;
890
891 std::vector<SUnit*>::const_iterator I = (Sequence.end() - LookAhead);
892 unsigned HazardCycle = (*I)->getHeight();
893 for (std::vector<SUnit*>::const_iterator E = Sequence.end(); I != E; ++I) {
894 SUnit *SU = *I;
895 for (; SU->getHeight() > HazardCycle; ++HazardCycle) {
896 HazardRec->RecedeCycle();
897 }
898 EmitNode(SU);
899 }
900 }
901
902 /// BacktrackBottomUp - Backtrack scheduling to a previous cycle specified in
903 /// BTCycle in order to schedule a specific node.
BacktrackBottomUp(SUnit * SU,SUnit * BtSU)904 void ScheduleDAGRRList::BacktrackBottomUp(SUnit *SU, SUnit *BtSU) {
905 SUnit *OldSU = Sequence.back();
906 while (true) {
907 Sequence.pop_back();
908 // FIXME: use ready cycle instead of height
909 CurCycle = OldSU->getHeight();
910 UnscheduleNodeBottomUp(OldSU);
911 AvailableQueue->setCurCycle(CurCycle);
912 if (OldSU == BtSU)
913 break;
914 OldSU = Sequence.back();
915 }
916
917 assert(!SU->isSucc(OldSU) && "Something is wrong!");
918
919 RestoreHazardCheckerBottomUp();
920
921 ReleasePending();
922
923 ++NumBacktracks;
924 }
925
isOperandOf(const SUnit * SU,SDNode * N)926 static bool isOperandOf(const SUnit *SU, SDNode *N) {
927 for (const SDNode *SUNode = SU->getNode(); SUNode;
928 SUNode = SUNode->getGluedNode()) {
929 if (SUNode->isOperandOf(N))
930 return true;
931 }
932 return false;
933 }
934
935 /// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
936 /// successors to the newly created node.
CopyAndMoveSuccessors(SUnit * SU)937 SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) {
938 SDNode *N = SU->getNode();
939 if (!N)
940 return nullptr;
941
942 if (SU->getNode()->getGluedNode())
943 return nullptr;
944
945 SUnit *NewSU;
946 bool TryUnfold = false;
947 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
948 MVT VT = N->getSimpleValueType(i);
949 if (VT == MVT::Glue)
950 return nullptr;
951 else if (VT == MVT::Other)
952 TryUnfold = true;
953 }
954 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
955 const SDValue &Op = N->getOperand(i);
956 MVT VT = Op.getNode()->getSimpleValueType(Op.getResNo());
957 if (VT == MVT::Glue)
958 return nullptr;
959 }
960
961 if (TryUnfold) {
962 SmallVector<SDNode*, 2> NewNodes;
963 if (!TII->unfoldMemoryOperand(*DAG, N, NewNodes))
964 return nullptr;
965
966 // unfolding an x86 DEC64m operation results in store, dec, load which
967 // can't be handled here so quit
968 if (NewNodes.size() == 3)
969 return nullptr;
970
971 DEBUG(dbgs() << "Unfolding SU #" << SU->NodeNum << "\n");
972 assert(NewNodes.size() == 2 && "Expected a load folding node!");
973
974 N = NewNodes[1];
975 SDNode *LoadNode = NewNodes[0];
976 unsigned NumVals = N->getNumValues();
977 unsigned OldNumVals = SU->getNode()->getNumValues();
978 for (unsigned i = 0; i != NumVals; ++i)
979 DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), i), SDValue(N, i));
980 DAG->ReplaceAllUsesOfValueWith(SDValue(SU->getNode(), OldNumVals-1),
981 SDValue(LoadNode, 1));
982
983 // LoadNode may already exist. This can happen when there is another
984 // load from the same location and producing the same type of value
985 // but it has different alignment or volatileness.
986 bool isNewLoad = true;
987 SUnit *LoadSU;
988 if (LoadNode->getNodeId() != -1) {
989 LoadSU = &SUnits[LoadNode->getNodeId()];
990 isNewLoad = false;
991 } else {
992 LoadSU = CreateNewSUnit(LoadNode);
993 LoadNode->setNodeId(LoadSU->NodeNum);
994
995 InitNumRegDefsLeft(LoadSU);
996 computeLatency(LoadSU);
997 }
998
999 SUnit *NewSU = CreateNewSUnit(N);
1000 assert(N->getNodeId() == -1 && "Node already inserted!");
1001 N->setNodeId(NewSU->NodeNum);
1002
1003 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1004 for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
1005 if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
1006 NewSU->isTwoAddress = true;
1007 break;
1008 }
1009 }
1010 if (MCID.isCommutable())
1011 NewSU->isCommutable = true;
1012
1013 InitNumRegDefsLeft(NewSU);
1014 computeLatency(NewSU);
1015
1016 // Record all the edges to and from the old SU, by category.
1017 SmallVector<SDep, 4> ChainPreds;
1018 SmallVector<SDep, 4> ChainSuccs;
1019 SmallVector<SDep, 4> LoadPreds;
1020 SmallVector<SDep, 4> NodePreds;
1021 SmallVector<SDep, 4> NodeSuccs;
1022 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1023 I != E; ++I) {
1024 if (I->isCtrl())
1025 ChainPreds.push_back(*I);
1026 else if (isOperandOf(I->getSUnit(), LoadNode))
1027 LoadPreds.push_back(*I);
1028 else
1029 NodePreds.push_back(*I);
1030 }
1031 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1032 I != E; ++I) {
1033 if (I->isCtrl())
1034 ChainSuccs.push_back(*I);
1035 else
1036 NodeSuccs.push_back(*I);
1037 }
1038
1039 // Now assign edges to the newly-created nodes.
1040 for (unsigned i = 0, e = ChainPreds.size(); i != e; ++i) {
1041 const SDep &Pred = ChainPreds[i];
1042 RemovePred(SU, Pred);
1043 if (isNewLoad)
1044 AddPred(LoadSU, Pred);
1045 }
1046 for (unsigned i = 0, e = LoadPreds.size(); i != e; ++i) {
1047 const SDep &Pred = LoadPreds[i];
1048 RemovePred(SU, Pred);
1049 if (isNewLoad)
1050 AddPred(LoadSU, Pred);
1051 }
1052 for (unsigned i = 0, e = NodePreds.size(); i != e; ++i) {
1053 const SDep &Pred = NodePreds[i];
1054 RemovePred(SU, Pred);
1055 AddPred(NewSU, Pred);
1056 }
1057 for (unsigned i = 0, e = NodeSuccs.size(); i != e; ++i) {
1058 SDep D = NodeSuccs[i];
1059 SUnit *SuccDep = D.getSUnit();
1060 D.setSUnit(SU);
1061 RemovePred(SuccDep, D);
1062 D.setSUnit(NewSU);
1063 AddPred(SuccDep, D);
1064 // Balance register pressure.
1065 if (AvailableQueue->tracksRegPressure() && SuccDep->isScheduled
1066 && !D.isCtrl() && NewSU->NumRegDefsLeft > 0)
1067 --NewSU->NumRegDefsLeft;
1068 }
1069 for (unsigned i = 0, e = ChainSuccs.size(); i != e; ++i) {
1070 SDep D = ChainSuccs[i];
1071 SUnit *SuccDep = D.getSUnit();
1072 D.setSUnit(SU);
1073 RemovePred(SuccDep, D);
1074 if (isNewLoad) {
1075 D.setSUnit(LoadSU);
1076 AddPred(SuccDep, D);
1077 }
1078 }
1079
1080 // Add a data dependency to reflect that NewSU reads the value defined
1081 // by LoadSU.
1082 SDep D(LoadSU, SDep::Data, 0);
1083 D.setLatency(LoadSU->Latency);
1084 AddPred(NewSU, D);
1085
1086 if (isNewLoad)
1087 AvailableQueue->addNode(LoadSU);
1088 AvailableQueue->addNode(NewSU);
1089
1090 ++NumUnfolds;
1091
1092 if (NewSU->NumSuccsLeft == 0) {
1093 NewSU->isAvailable = true;
1094 return NewSU;
1095 }
1096 SU = NewSU;
1097 }
1098
1099 DEBUG(dbgs() << " Duplicating SU #" << SU->NodeNum << "\n");
1100 NewSU = CreateClone(SU);
1101
1102 // New SUnit has the exact same predecessors.
1103 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1104 I != E; ++I)
1105 if (!I->isArtificial())
1106 AddPred(NewSU, *I);
1107
1108 // Only copy scheduled successors. Cut them from old node's successor
1109 // list and move them over.
1110 SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
1111 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1112 I != E; ++I) {
1113 if (I->isArtificial())
1114 continue;
1115 SUnit *SuccSU = I->getSUnit();
1116 if (SuccSU->isScheduled) {
1117 SDep D = *I;
1118 D.setSUnit(NewSU);
1119 AddPred(SuccSU, D);
1120 D.setSUnit(SU);
1121 DelDeps.push_back(std::make_pair(SuccSU, D));
1122 }
1123 }
1124 for (unsigned i = 0, e = DelDeps.size(); i != e; ++i)
1125 RemovePred(DelDeps[i].first, DelDeps[i].second);
1126
1127 AvailableQueue->updateNode(SU);
1128 AvailableQueue->addNode(NewSU);
1129
1130 ++NumDups;
1131 return NewSU;
1132 }
1133
1134 /// InsertCopiesAndMoveSuccs - Insert register copies and move all
1135 /// scheduled successors of the given SUnit to the last copy.
InsertCopiesAndMoveSuccs(SUnit * SU,unsigned Reg,const TargetRegisterClass * DestRC,const TargetRegisterClass * SrcRC,SmallVectorImpl<SUnit * > & Copies)1136 void ScheduleDAGRRList::InsertCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
1137 const TargetRegisterClass *DestRC,
1138 const TargetRegisterClass *SrcRC,
1139 SmallVectorImpl<SUnit*> &Copies) {
1140 SUnit *CopyFromSU = CreateNewSUnit(nullptr);
1141 CopyFromSU->CopySrcRC = SrcRC;
1142 CopyFromSU->CopyDstRC = DestRC;
1143
1144 SUnit *CopyToSU = CreateNewSUnit(nullptr);
1145 CopyToSU->CopySrcRC = DestRC;
1146 CopyToSU->CopyDstRC = SrcRC;
1147
1148 // Only copy scheduled successors. Cut them from old node's successor
1149 // list and move them over.
1150 SmallVector<std::pair<SUnit *, SDep>, 4> DelDeps;
1151 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1152 I != E; ++I) {
1153 if (I->isArtificial())
1154 continue;
1155 SUnit *SuccSU = I->getSUnit();
1156 if (SuccSU->isScheduled) {
1157 SDep D = *I;
1158 D.setSUnit(CopyToSU);
1159 AddPred(SuccSU, D);
1160 DelDeps.push_back(std::make_pair(SuccSU, *I));
1161 }
1162 else {
1163 // Avoid scheduling the def-side copy before other successors. Otherwise
1164 // we could introduce another physreg interference on the copy and
1165 // continue inserting copies indefinitely.
1166 AddPred(SuccSU, SDep(CopyFromSU, SDep::Artificial));
1167 }
1168 }
1169 for (unsigned i = 0, e = DelDeps.size(); i != e; ++i)
1170 RemovePred(DelDeps[i].first, DelDeps[i].second);
1171
1172 SDep FromDep(SU, SDep::Data, Reg);
1173 FromDep.setLatency(SU->Latency);
1174 AddPred(CopyFromSU, FromDep);
1175 SDep ToDep(CopyFromSU, SDep::Data, 0);
1176 ToDep.setLatency(CopyFromSU->Latency);
1177 AddPred(CopyToSU, ToDep);
1178
1179 AvailableQueue->updateNode(SU);
1180 AvailableQueue->addNode(CopyFromSU);
1181 AvailableQueue->addNode(CopyToSU);
1182 Copies.push_back(CopyFromSU);
1183 Copies.push_back(CopyToSU);
1184
1185 ++NumPRCopies;
1186 }
1187
1188 /// getPhysicalRegisterVT - Returns the ValueType of the physical register
1189 /// definition of the specified node.
1190 /// FIXME: Move to SelectionDAG?
getPhysicalRegisterVT(SDNode * N,unsigned Reg,const TargetInstrInfo * TII)1191 static MVT getPhysicalRegisterVT(SDNode *N, unsigned Reg,
1192 const TargetInstrInfo *TII) {
1193 unsigned NumRes;
1194 if (N->getOpcode() == ISD::CopyFromReg) {
1195 // CopyFromReg has: "chain, Val, glue" so operand 1 gives the type.
1196 NumRes = 1;
1197 } else {
1198 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1199 assert(MCID.ImplicitDefs && "Physical reg def must be in implicit def list!");
1200 NumRes = MCID.getNumDefs();
1201 for (const uint16_t *ImpDef = MCID.getImplicitDefs(); *ImpDef; ++ImpDef) {
1202 if (Reg == *ImpDef)
1203 break;
1204 ++NumRes;
1205 }
1206 }
1207 return N->getSimpleValueType(NumRes);
1208 }
1209
1210 /// CheckForLiveRegDef - Return true and update live register vector if the
1211 /// specified register def of the specified SUnit clobbers any "live" registers.
CheckForLiveRegDef(SUnit * SU,unsigned Reg,std::vector<SUnit * > & LiveRegDefs,SmallSet<unsigned,4> & RegAdded,SmallVectorImpl<unsigned> & LRegs,const TargetRegisterInfo * TRI)1212 static void CheckForLiveRegDef(SUnit *SU, unsigned Reg,
1213 std::vector<SUnit*> &LiveRegDefs,
1214 SmallSet<unsigned, 4> &RegAdded,
1215 SmallVectorImpl<unsigned> &LRegs,
1216 const TargetRegisterInfo *TRI) {
1217 for (MCRegAliasIterator AliasI(Reg, TRI, true); AliasI.isValid(); ++AliasI) {
1218
1219 // Check if Ref is live.
1220 if (!LiveRegDefs[*AliasI]) continue;
1221
1222 // Allow multiple uses of the same def.
1223 if (LiveRegDefs[*AliasI] == SU) continue;
1224
1225 // Add Reg to the set of interfering live regs.
1226 if (RegAdded.insert(*AliasI).second) {
1227 LRegs.push_back(*AliasI);
1228 }
1229 }
1230 }
1231
1232 /// CheckForLiveRegDefMasked - Check for any live physregs that are clobbered
1233 /// by RegMask, and add them to LRegs.
CheckForLiveRegDefMasked(SUnit * SU,const uint32_t * RegMask,std::vector<SUnit * > & LiveRegDefs,SmallSet<unsigned,4> & RegAdded,SmallVectorImpl<unsigned> & LRegs)1234 static void CheckForLiveRegDefMasked(SUnit *SU, const uint32_t *RegMask,
1235 std::vector<SUnit*> &LiveRegDefs,
1236 SmallSet<unsigned, 4> &RegAdded,
1237 SmallVectorImpl<unsigned> &LRegs) {
1238 // Look at all live registers. Skip Reg0 and the special CallResource.
1239 for (unsigned i = 1, e = LiveRegDefs.size()-1; i != e; ++i) {
1240 if (!LiveRegDefs[i]) continue;
1241 if (LiveRegDefs[i] == SU) continue;
1242 if (!MachineOperand::clobbersPhysReg(RegMask, i)) continue;
1243 if (RegAdded.insert(i).second)
1244 LRegs.push_back(i);
1245 }
1246 }
1247
1248 /// getNodeRegMask - Returns the register mask attached to an SDNode, if any.
getNodeRegMask(const SDNode * N)1249 static const uint32_t *getNodeRegMask(const SDNode *N) {
1250 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1251 if (const RegisterMaskSDNode *Op =
1252 dyn_cast<RegisterMaskSDNode>(N->getOperand(i).getNode()))
1253 return Op->getRegMask();
1254 return nullptr;
1255 }
1256
1257 /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
1258 /// scheduling of the given node to satisfy live physical register dependencies.
1259 /// If the specific node is the last one that's available to schedule, do
1260 /// whatever is necessary (i.e. backtracking or cloning) to make it possible.
1261 bool ScheduleDAGRRList::
DelayForLiveRegsBottomUp(SUnit * SU,SmallVectorImpl<unsigned> & LRegs)1262 DelayForLiveRegsBottomUp(SUnit *SU, SmallVectorImpl<unsigned> &LRegs) {
1263 if (NumLiveRegs == 0)
1264 return false;
1265
1266 SmallSet<unsigned, 4> RegAdded;
1267 // If this node would clobber any "live" register, then it's not ready.
1268 //
1269 // If SU is the currently live definition of the same register that it uses,
1270 // then we are free to schedule it.
1271 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1272 I != E; ++I) {
1273 if (I->isAssignedRegDep() && LiveRegDefs[I->getReg()] != SU)
1274 CheckForLiveRegDef(I->getSUnit(), I->getReg(), LiveRegDefs,
1275 RegAdded, LRegs, TRI);
1276 }
1277
1278 for (SDNode *Node = SU->getNode(); Node; Node = Node->getGluedNode()) {
1279 if (Node->getOpcode() == ISD::INLINEASM) {
1280 // Inline asm can clobber physical defs.
1281 unsigned NumOps = Node->getNumOperands();
1282 if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
1283 --NumOps; // Ignore the glue operand.
1284
1285 for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
1286 unsigned Flags =
1287 cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
1288 unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
1289
1290 ++i; // Skip the ID value.
1291 if (InlineAsm::isRegDefKind(Flags) ||
1292 InlineAsm::isRegDefEarlyClobberKind(Flags) ||
1293 InlineAsm::isClobberKind(Flags)) {
1294 // Check for def of register or earlyclobber register.
1295 for (; NumVals; --NumVals, ++i) {
1296 unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
1297 if (TargetRegisterInfo::isPhysicalRegister(Reg))
1298 CheckForLiveRegDef(SU, Reg, LiveRegDefs, RegAdded, LRegs, TRI);
1299 }
1300 } else
1301 i += NumVals;
1302 }
1303 continue;
1304 }
1305
1306 if (!Node->isMachineOpcode())
1307 continue;
1308 // If we're in the middle of scheduling a call, don't begin scheduling
1309 // another call. Also, don't allow any physical registers to be live across
1310 // the call.
1311 if (Node->getMachineOpcode() == (unsigned)TII->getCallFrameDestroyOpcode()) {
1312 // Check the special calling-sequence resource.
1313 unsigned CallResource = TRI->getNumRegs();
1314 if (LiveRegDefs[CallResource]) {
1315 SDNode *Gen = LiveRegGens[CallResource]->getNode();
1316 while (SDNode *Glued = Gen->getGluedNode())
1317 Gen = Glued;
1318 if (!IsChainDependent(Gen, Node, 0, TII) &&
1319 RegAdded.insert(CallResource).second)
1320 LRegs.push_back(CallResource);
1321 }
1322 }
1323 if (const uint32_t *RegMask = getNodeRegMask(Node))
1324 CheckForLiveRegDefMasked(SU, RegMask, LiveRegDefs, RegAdded, LRegs);
1325
1326 const MCInstrDesc &MCID = TII->get(Node->getMachineOpcode());
1327 if (!MCID.ImplicitDefs)
1328 continue;
1329 for (const uint16_t *Reg = MCID.getImplicitDefs(); *Reg; ++Reg)
1330 CheckForLiveRegDef(SU, *Reg, LiveRegDefs, RegAdded, LRegs, TRI);
1331 }
1332
1333 return !LRegs.empty();
1334 }
1335
releaseInterferences(unsigned Reg)1336 void ScheduleDAGRRList::releaseInterferences(unsigned Reg) {
1337 // Add the nodes that aren't ready back onto the available list.
1338 for (unsigned i = Interferences.size(); i > 0; --i) {
1339 SUnit *SU = Interferences[i-1];
1340 LRegsMapT::iterator LRegsPos = LRegsMap.find(SU);
1341 if (Reg) {
1342 SmallVectorImpl<unsigned> &LRegs = LRegsPos->second;
1343 if (std::find(LRegs.begin(), LRegs.end(), Reg) == LRegs.end())
1344 continue;
1345 }
1346 SU->isPending = false;
1347 // The interfering node may no longer be available due to backtracking.
1348 // Furthermore, it may have been made available again, in which case it is
1349 // now already in the AvailableQueue.
1350 if (SU->isAvailable && !SU->NodeQueueId) {
1351 DEBUG(dbgs() << " Repushing SU #" << SU->NodeNum << '\n');
1352 AvailableQueue->push(SU);
1353 }
1354 if (i < Interferences.size())
1355 Interferences[i-1] = Interferences.back();
1356 Interferences.pop_back();
1357 LRegsMap.erase(LRegsPos);
1358 }
1359 }
1360
1361 /// Return a node that can be scheduled in this cycle. Requirements:
1362 /// (1) Ready: latency has been satisfied
1363 /// (2) No Hazards: resources are available
1364 /// (3) No Interferences: may unschedule to break register interferences.
PickNodeToScheduleBottomUp()1365 SUnit *ScheduleDAGRRList::PickNodeToScheduleBottomUp() {
1366 SUnit *CurSU = AvailableQueue->empty() ? nullptr : AvailableQueue->pop();
1367 while (CurSU) {
1368 SmallVector<unsigned, 4> LRegs;
1369 if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
1370 break;
1371 DEBUG(dbgs() << " Interfering reg " <<
1372 (LRegs[0] == TRI->getNumRegs() ? "CallResource"
1373 : TRI->getName(LRegs[0]))
1374 << " SU #" << CurSU->NodeNum << '\n');
1375 std::pair<LRegsMapT::iterator, bool> LRegsPair =
1376 LRegsMap.insert(std::make_pair(CurSU, LRegs));
1377 if (LRegsPair.second) {
1378 CurSU->isPending = true; // This SU is not in AvailableQueue right now.
1379 Interferences.push_back(CurSU);
1380 }
1381 else {
1382 assert(CurSU->isPending && "Interferences are pending");
1383 // Update the interference with current live regs.
1384 LRegsPair.first->second = LRegs;
1385 }
1386 CurSU = AvailableQueue->pop();
1387 }
1388 if (CurSU)
1389 return CurSU;
1390
1391 // All candidates are delayed due to live physical reg dependencies.
1392 // Try backtracking, code duplication, or inserting cross class copies
1393 // to resolve it.
1394 for (unsigned i = 0, e = Interferences.size(); i != e; ++i) {
1395 SUnit *TrySU = Interferences[i];
1396 SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
1397
1398 // Try unscheduling up to the point where it's safe to schedule
1399 // this node.
1400 SUnit *BtSU = nullptr;
1401 unsigned LiveCycle = UINT_MAX;
1402 for (unsigned j = 0, ee = LRegs.size(); j != ee; ++j) {
1403 unsigned Reg = LRegs[j];
1404 if (LiveRegGens[Reg]->getHeight() < LiveCycle) {
1405 BtSU = LiveRegGens[Reg];
1406 LiveCycle = BtSU->getHeight();
1407 }
1408 }
1409 if (!WillCreateCycle(TrySU, BtSU)) {
1410 // BacktrackBottomUp mutates Interferences!
1411 BacktrackBottomUp(TrySU, BtSU);
1412
1413 // Force the current node to be scheduled before the node that
1414 // requires the physical reg dep.
1415 if (BtSU->isAvailable) {
1416 BtSU->isAvailable = false;
1417 if (!BtSU->isPending)
1418 AvailableQueue->remove(BtSU);
1419 }
1420 DEBUG(dbgs() << "ARTIFICIAL edge from SU(" << BtSU->NodeNum << ") to SU("
1421 << TrySU->NodeNum << ")\n");
1422 AddPred(TrySU, SDep(BtSU, SDep::Artificial));
1423
1424 // If one or more successors has been unscheduled, then the current
1425 // node is no longer available.
1426 if (!TrySU->isAvailable || !TrySU->NodeQueueId)
1427 CurSU = AvailableQueue->pop();
1428 else {
1429 // Available and in AvailableQueue
1430 AvailableQueue->remove(TrySU);
1431 CurSU = TrySU;
1432 }
1433 // Interferences has been mutated. We must break.
1434 break;
1435 }
1436 }
1437
1438 if (!CurSU) {
1439 // Can't backtrack. If it's too expensive to copy the value, then try
1440 // duplicate the nodes that produces these "too expensive to copy"
1441 // values to break the dependency. In case even that doesn't work,
1442 // insert cross class copies.
1443 // If it's not too expensive, i.e. cost != -1, issue copies.
1444 SUnit *TrySU = Interferences[0];
1445 SmallVectorImpl<unsigned> &LRegs = LRegsMap[TrySU];
1446 assert(LRegs.size() == 1 && "Can't handle this yet!");
1447 unsigned Reg = LRegs[0];
1448 SUnit *LRDef = LiveRegDefs[Reg];
1449 MVT VT = getPhysicalRegisterVT(LRDef->getNode(), Reg, TII);
1450 const TargetRegisterClass *RC =
1451 TRI->getMinimalPhysRegClass(Reg, VT);
1452 const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
1453
1454 // If cross copy register class is the same as RC, then it must be possible
1455 // copy the value directly. Do not try duplicate the def.
1456 // If cross copy register class is not the same as RC, then it's possible to
1457 // copy the value but it require cross register class copies and it is
1458 // expensive.
1459 // If cross copy register class is null, then it's not possible to copy
1460 // the value at all.
1461 SUnit *NewDef = nullptr;
1462 if (DestRC != RC) {
1463 NewDef = CopyAndMoveSuccessors(LRDef);
1464 if (!DestRC && !NewDef)
1465 report_fatal_error("Can't handle live physical register dependency!");
1466 }
1467 if (!NewDef) {
1468 // Issue copies, these can be expensive cross register class copies.
1469 SmallVector<SUnit*, 2> Copies;
1470 InsertCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
1471 DEBUG(dbgs() << " Adding an edge from SU #" << TrySU->NodeNum
1472 << " to SU #" << Copies.front()->NodeNum << "\n");
1473 AddPred(TrySU, SDep(Copies.front(), SDep::Artificial));
1474 NewDef = Copies.back();
1475 }
1476
1477 DEBUG(dbgs() << " Adding an edge from SU #" << NewDef->NodeNum
1478 << " to SU #" << TrySU->NodeNum << "\n");
1479 LiveRegDefs[Reg] = NewDef;
1480 AddPred(NewDef, SDep(TrySU, SDep::Artificial));
1481 TrySU->isAvailable = false;
1482 CurSU = NewDef;
1483 }
1484 assert(CurSU && "Unable to resolve live physical register dependencies!");
1485 return CurSU;
1486 }
1487
1488 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
1489 /// schedulers.
ListScheduleBottomUp()1490 void ScheduleDAGRRList::ListScheduleBottomUp() {
1491 // Release any predecessors of the special Exit node.
1492 ReleasePredecessors(&ExitSU);
1493
1494 // Add root to Available queue.
1495 if (!SUnits.empty()) {
1496 SUnit *RootSU = &SUnits[DAG->getRoot().getNode()->getNodeId()];
1497 assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!");
1498 RootSU->isAvailable = true;
1499 AvailableQueue->push(RootSU);
1500 }
1501
1502 // While Available queue is not empty, grab the node with the highest
1503 // priority. If it is not ready put it back. Schedule the node.
1504 Sequence.reserve(SUnits.size());
1505 while (!AvailableQueue->empty() || !Interferences.empty()) {
1506 DEBUG(dbgs() << "\nExamining Available:\n";
1507 AvailableQueue->dump(this));
1508
1509 // Pick the best node to schedule taking all constraints into
1510 // consideration.
1511 SUnit *SU = PickNodeToScheduleBottomUp();
1512
1513 AdvancePastStalls(SU);
1514
1515 ScheduleNodeBottomUp(SU);
1516
1517 while (AvailableQueue->empty() && !PendingQueue.empty()) {
1518 // Advance the cycle to free resources. Skip ahead to the next ready SU.
1519 assert(MinAvailableCycle < UINT_MAX && "MinAvailableCycle uninitialized");
1520 AdvanceToCycle(std::max(CurCycle + 1, MinAvailableCycle));
1521 }
1522 }
1523
1524 // Reverse the order if it is bottom up.
1525 std::reverse(Sequence.begin(), Sequence.end());
1526
1527 #ifndef NDEBUG
1528 VerifyScheduledSequence(/*isBottomUp=*/true);
1529 #endif
1530 }
1531
1532 //===----------------------------------------------------------------------===//
1533 // RegReductionPriorityQueue Definition
1534 //===----------------------------------------------------------------------===//
1535 //
1536 // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers
1537 // to reduce register pressure.
1538 //
1539 namespace {
1540 class RegReductionPQBase;
1541
1542 struct queue_sort : public std::binary_function<SUnit*, SUnit*, bool> {
isReady__anon884f54d00211::queue_sort1543 bool isReady(SUnit* SU, unsigned CurCycle) const { return true; }
1544 };
1545
1546 #ifndef NDEBUG
1547 template<class SF>
1548 struct reverse_sort : public queue_sort {
1549 SF &SortFunc;
reverse_sort__anon884f54d00211::reverse_sort1550 reverse_sort(SF &sf) : SortFunc(sf) {}
1551
operator ()__anon884f54d00211::reverse_sort1552 bool operator()(SUnit* left, SUnit* right) const {
1553 // reverse left/right rather than simply !SortFunc(left, right)
1554 // to expose different paths in the comparison logic.
1555 return SortFunc(right, left);
1556 }
1557 };
1558 #endif // NDEBUG
1559
1560 /// bu_ls_rr_sort - Priority function for bottom up register pressure
1561 // reduction scheduler.
1562 struct bu_ls_rr_sort : public queue_sort {
1563 enum {
1564 IsBottomUp = true,
1565 HasReadyFilter = false
1566 };
1567
1568 RegReductionPQBase *SPQ;
bu_ls_rr_sort__anon884f54d00211::bu_ls_rr_sort1569 bu_ls_rr_sort(RegReductionPQBase *spq) : SPQ(spq) {}
1570
1571 bool operator()(SUnit* left, SUnit* right) const;
1572 };
1573
1574 // src_ls_rr_sort - Priority function for source order scheduler.
1575 struct src_ls_rr_sort : public queue_sort {
1576 enum {
1577 IsBottomUp = true,
1578 HasReadyFilter = false
1579 };
1580
1581 RegReductionPQBase *SPQ;
src_ls_rr_sort__anon884f54d00211::src_ls_rr_sort1582 src_ls_rr_sort(RegReductionPQBase *spq)
1583 : SPQ(spq) {}
1584
1585 bool operator()(SUnit* left, SUnit* right) const;
1586 };
1587
1588 // hybrid_ls_rr_sort - Priority function for hybrid scheduler.
1589 struct hybrid_ls_rr_sort : public queue_sort {
1590 enum {
1591 IsBottomUp = true,
1592 HasReadyFilter = false
1593 };
1594
1595 RegReductionPQBase *SPQ;
hybrid_ls_rr_sort__anon884f54d00211::hybrid_ls_rr_sort1596 hybrid_ls_rr_sort(RegReductionPQBase *spq)
1597 : SPQ(spq) {}
1598
1599 bool isReady(SUnit *SU, unsigned CurCycle) const;
1600
1601 bool operator()(SUnit* left, SUnit* right) const;
1602 };
1603
1604 // ilp_ls_rr_sort - Priority function for ILP (instruction level parallelism)
1605 // scheduler.
1606 struct ilp_ls_rr_sort : public queue_sort {
1607 enum {
1608 IsBottomUp = true,
1609 HasReadyFilter = false
1610 };
1611
1612 RegReductionPQBase *SPQ;
ilp_ls_rr_sort__anon884f54d00211::ilp_ls_rr_sort1613 ilp_ls_rr_sort(RegReductionPQBase *spq)
1614 : SPQ(spq) {}
1615
1616 bool isReady(SUnit *SU, unsigned CurCycle) const;
1617
1618 bool operator()(SUnit* left, SUnit* right) const;
1619 };
1620
1621 class RegReductionPQBase : public SchedulingPriorityQueue {
1622 protected:
1623 std::vector<SUnit*> Queue;
1624 unsigned CurQueueId;
1625 bool TracksRegPressure;
1626 bool SrcOrder;
1627
1628 // SUnits - The SUnits for the current graph.
1629 std::vector<SUnit> *SUnits;
1630
1631 MachineFunction &MF;
1632 const TargetInstrInfo *TII;
1633 const TargetRegisterInfo *TRI;
1634 const TargetLowering *TLI;
1635 ScheduleDAGRRList *scheduleDAG;
1636
1637 // SethiUllmanNumbers - The SethiUllman number for each node.
1638 std::vector<unsigned> SethiUllmanNumbers;
1639
1640 /// RegPressure - Tracking current reg pressure per register class.
1641 ///
1642 std::vector<unsigned> RegPressure;
1643
1644 /// RegLimit - Tracking the number of allocatable registers per register
1645 /// class.
1646 std::vector<unsigned> RegLimit;
1647
1648 public:
RegReductionPQBase(MachineFunction & mf,bool hasReadyFilter,bool tracksrp,bool srcorder,const TargetInstrInfo * tii,const TargetRegisterInfo * tri,const TargetLowering * tli)1649 RegReductionPQBase(MachineFunction &mf,
1650 bool hasReadyFilter,
1651 bool tracksrp,
1652 bool srcorder,
1653 const TargetInstrInfo *tii,
1654 const TargetRegisterInfo *tri,
1655 const TargetLowering *tli)
1656 : SchedulingPriorityQueue(hasReadyFilter),
1657 CurQueueId(0), TracksRegPressure(tracksrp), SrcOrder(srcorder),
1658 MF(mf), TII(tii), TRI(tri), TLI(tli), scheduleDAG(nullptr) {
1659 if (TracksRegPressure) {
1660 unsigned NumRC = TRI->getNumRegClasses();
1661 RegLimit.resize(NumRC);
1662 RegPressure.resize(NumRC);
1663 std::fill(RegLimit.begin(), RegLimit.end(), 0);
1664 std::fill(RegPressure.begin(), RegPressure.end(), 0);
1665 for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
1666 E = TRI->regclass_end(); I != E; ++I)
1667 RegLimit[(*I)->getID()] = tri->getRegPressureLimit(*I, MF);
1668 }
1669 }
1670
setScheduleDAG(ScheduleDAGRRList * scheduleDag)1671 void setScheduleDAG(ScheduleDAGRRList *scheduleDag) {
1672 scheduleDAG = scheduleDag;
1673 }
1674
getHazardRec()1675 ScheduleHazardRecognizer* getHazardRec() {
1676 return scheduleDAG->getHazardRec();
1677 }
1678
1679 void initNodes(std::vector<SUnit> &sunits) override;
1680
1681 void addNode(const SUnit *SU) override;
1682
1683 void updateNode(const SUnit *SU) override;
1684
releaseState()1685 void releaseState() override {
1686 SUnits = nullptr;
1687 SethiUllmanNumbers.clear();
1688 std::fill(RegPressure.begin(), RegPressure.end(), 0);
1689 }
1690
1691 unsigned getNodePriority(const SUnit *SU) const;
1692
getNodeOrdering(const SUnit * SU) const1693 unsigned getNodeOrdering(const SUnit *SU) const {
1694 if (!SU->getNode()) return 0;
1695
1696 return SU->getNode()->getIROrder();
1697 }
1698
empty() const1699 bool empty() const override { return Queue.empty(); }
1700
push(SUnit * U)1701 void push(SUnit *U) override {
1702 assert(!U->NodeQueueId && "Node in the queue already");
1703 U->NodeQueueId = ++CurQueueId;
1704 Queue.push_back(U);
1705 }
1706
remove(SUnit * SU)1707 void remove(SUnit *SU) override {
1708 assert(!Queue.empty() && "Queue is empty!");
1709 assert(SU->NodeQueueId != 0 && "Not in queue!");
1710 std::vector<SUnit *>::iterator I = std::find(Queue.begin(), Queue.end(),
1711 SU);
1712 if (I != std::prev(Queue.end()))
1713 std::swap(*I, Queue.back());
1714 Queue.pop_back();
1715 SU->NodeQueueId = 0;
1716 }
1717
tracksRegPressure() const1718 bool tracksRegPressure() const override { return TracksRegPressure; }
1719
1720 void dumpRegPressure() const;
1721
1722 bool HighRegPressure(const SUnit *SU) const;
1723
1724 bool MayReduceRegPressure(SUnit *SU) const;
1725
1726 int RegPressureDiff(SUnit *SU, unsigned &LiveUses) const;
1727
1728 void scheduledNode(SUnit *SU) override;
1729
1730 void unscheduledNode(SUnit *SU) override;
1731
1732 protected:
1733 bool canClobber(const SUnit *SU, const SUnit *Op);
1734 void AddPseudoTwoAddrDeps();
1735 void PrescheduleNodesWithMultipleUses();
1736 void CalculateSethiUllmanNumbers();
1737 };
1738
1739 template<class SF>
popFromQueueImpl(std::vector<SUnit * > & Q,SF & Picker)1740 static SUnit *popFromQueueImpl(std::vector<SUnit*> &Q, SF &Picker) {
1741 std::vector<SUnit *>::iterator Best = Q.begin();
1742 for (std::vector<SUnit *>::iterator I = std::next(Q.begin()),
1743 E = Q.end(); I != E; ++I)
1744 if (Picker(*Best, *I))
1745 Best = I;
1746 SUnit *V = *Best;
1747 if (Best != std::prev(Q.end()))
1748 std::swap(*Best, Q.back());
1749 Q.pop_back();
1750 return V;
1751 }
1752
1753 template<class SF>
popFromQueue(std::vector<SUnit * > & Q,SF & Picker,ScheduleDAG * DAG)1754 SUnit *popFromQueue(std::vector<SUnit*> &Q, SF &Picker, ScheduleDAG *DAG) {
1755 #ifndef NDEBUG
1756 if (DAG->StressSched) {
1757 reverse_sort<SF> RPicker(Picker);
1758 return popFromQueueImpl(Q, RPicker);
1759 }
1760 #endif
1761 (void)DAG;
1762 return popFromQueueImpl(Q, Picker);
1763 }
1764
1765 template<class SF>
1766 class RegReductionPriorityQueue : public RegReductionPQBase {
1767 SF Picker;
1768
1769 public:
RegReductionPriorityQueue(MachineFunction & mf,bool tracksrp,bool srcorder,const TargetInstrInfo * tii,const TargetRegisterInfo * tri,const TargetLowering * tli)1770 RegReductionPriorityQueue(MachineFunction &mf,
1771 bool tracksrp,
1772 bool srcorder,
1773 const TargetInstrInfo *tii,
1774 const TargetRegisterInfo *tri,
1775 const TargetLowering *tli)
1776 : RegReductionPQBase(mf, SF::HasReadyFilter, tracksrp, srcorder,
1777 tii, tri, tli),
1778 Picker(this) {}
1779
isBottomUp() const1780 bool isBottomUp() const override { return SF::IsBottomUp; }
1781
isReady(SUnit * U) const1782 bool isReady(SUnit *U) const override {
1783 return Picker.HasReadyFilter && Picker.isReady(U, getCurCycle());
1784 }
1785
pop()1786 SUnit *pop() override {
1787 if (Queue.empty()) return nullptr;
1788
1789 SUnit *V = popFromQueue(Queue, Picker, scheduleDAG);
1790 V->NodeQueueId = 0;
1791 return V;
1792 }
1793
1794 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dump(ScheduleDAG * DAG) const1795 void dump(ScheduleDAG *DAG) const override {
1796 // Emulate pop() without clobbering NodeQueueIds.
1797 std::vector<SUnit*> DumpQueue = Queue;
1798 SF DumpPicker = Picker;
1799 while (!DumpQueue.empty()) {
1800 SUnit *SU = popFromQueue(DumpQueue, DumpPicker, scheduleDAG);
1801 dbgs() << "Height " << SU->getHeight() << ": ";
1802 SU->dump(DAG);
1803 }
1804 }
1805 #endif
1806 };
1807
1808 typedef RegReductionPriorityQueue<bu_ls_rr_sort>
1809 BURegReductionPriorityQueue;
1810
1811 typedef RegReductionPriorityQueue<src_ls_rr_sort>
1812 SrcRegReductionPriorityQueue;
1813
1814 typedef RegReductionPriorityQueue<hybrid_ls_rr_sort>
1815 HybridBURRPriorityQueue;
1816
1817 typedef RegReductionPriorityQueue<ilp_ls_rr_sort>
1818 ILPBURRPriorityQueue;
1819 } // end anonymous namespace
1820
1821 //===----------------------------------------------------------------------===//
1822 // Static Node Priority for Register Pressure Reduction
1823 //===----------------------------------------------------------------------===//
1824
1825 // Check for special nodes that bypass scheduling heuristics.
1826 // Currently this pushes TokenFactor nodes down, but may be used for other
1827 // pseudo-ops as well.
1828 //
1829 // Return -1 to schedule right above left, 1 for left above right.
1830 // Return 0 if no bias exists.
checkSpecialNodes(const SUnit * left,const SUnit * right)1831 static int checkSpecialNodes(const SUnit *left, const SUnit *right) {
1832 bool LSchedLow = left->isScheduleLow;
1833 bool RSchedLow = right->isScheduleLow;
1834 if (LSchedLow != RSchedLow)
1835 return LSchedLow < RSchedLow ? 1 : -1;
1836 return 0;
1837 }
1838
1839 /// CalcNodeSethiUllmanNumber - Compute Sethi Ullman number.
1840 /// Smaller number is the higher priority.
1841 static unsigned
CalcNodeSethiUllmanNumber(const SUnit * SU,std::vector<unsigned> & SUNumbers)1842 CalcNodeSethiUllmanNumber(const SUnit *SU, std::vector<unsigned> &SUNumbers) {
1843 unsigned &SethiUllmanNumber = SUNumbers[SU->NodeNum];
1844 if (SethiUllmanNumber != 0)
1845 return SethiUllmanNumber;
1846
1847 unsigned Extra = 0;
1848 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1849 I != E; ++I) {
1850 if (I->isCtrl()) continue; // ignore chain preds
1851 SUnit *PredSU = I->getSUnit();
1852 unsigned PredSethiUllman = CalcNodeSethiUllmanNumber(PredSU, SUNumbers);
1853 if (PredSethiUllman > SethiUllmanNumber) {
1854 SethiUllmanNumber = PredSethiUllman;
1855 Extra = 0;
1856 } else if (PredSethiUllman == SethiUllmanNumber)
1857 ++Extra;
1858 }
1859
1860 SethiUllmanNumber += Extra;
1861
1862 if (SethiUllmanNumber == 0)
1863 SethiUllmanNumber = 1;
1864
1865 return SethiUllmanNumber;
1866 }
1867
1868 /// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all
1869 /// scheduling units.
CalculateSethiUllmanNumbers()1870 void RegReductionPQBase::CalculateSethiUllmanNumbers() {
1871 SethiUllmanNumbers.assign(SUnits->size(), 0);
1872
1873 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
1874 CalcNodeSethiUllmanNumber(&(*SUnits)[i], SethiUllmanNumbers);
1875 }
1876
addNode(const SUnit * SU)1877 void RegReductionPQBase::addNode(const SUnit *SU) {
1878 unsigned SUSize = SethiUllmanNumbers.size();
1879 if (SUnits->size() > SUSize)
1880 SethiUllmanNumbers.resize(SUSize*2, 0);
1881 CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
1882 }
1883
updateNode(const SUnit * SU)1884 void RegReductionPQBase::updateNode(const SUnit *SU) {
1885 SethiUllmanNumbers[SU->NodeNum] = 0;
1886 CalcNodeSethiUllmanNumber(SU, SethiUllmanNumbers);
1887 }
1888
1889 // Lower priority means schedule further down. For bottom-up scheduling, lower
1890 // priority SUs are scheduled before higher priority SUs.
getNodePriority(const SUnit * SU) const1891 unsigned RegReductionPQBase::getNodePriority(const SUnit *SU) const {
1892 assert(SU->NodeNum < SethiUllmanNumbers.size());
1893 unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0;
1894 if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
1895 // CopyToReg should be close to its uses to facilitate coalescing and
1896 // avoid spilling.
1897 return 0;
1898 if (Opc == TargetOpcode::EXTRACT_SUBREG ||
1899 Opc == TargetOpcode::SUBREG_TO_REG ||
1900 Opc == TargetOpcode::INSERT_SUBREG)
1901 // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be
1902 // close to their uses to facilitate coalescing.
1903 return 0;
1904 if (SU->NumSuccs == 0 && SU->NumPreds != 0)
1905 // If SU does not have a register use, i.e. it doesn't produce a value
1906 // that would be consumed (e.g. store), then it terminates a chain of
1907 // computation. Give it a large SethiUllman number so it will be
1908 // scheduled right before its predecessors that it doesn't lengthen
1909 // their live ranges.
1910 return 0xffff;
1911 if (SU->NumPreds == 0 && SU->NumSuccs != 0)
1912 // If SU does not have a register def, schedule it close to its uses
1913 // because it does not lengthen any live ranges.
1914 return 0;
1915 #if 1
1916 return SethiUllmanNumbers[SU->NodeNum];
1917 #else
1918 unsigned Priority = SethiUllmanNumbers[SU->NodeNum];
1919 if (SU->isCallOp) {
1920 // FIXME: This assumes all of the defs are used as call operands.
1921 int NP = (int)Priority - SU->getNode()->getNumValues();
1922 return (NP > 0) ? NP : 0;
1923 }
1924 return Priority;
1925 #endif
1926 }
1927
1928 //===----------------------------------------------------------------------===//
1929 // Register Pressure Tracking
1930 //===----------------------------------------------------------------------===//
1931
dumpRegPressure() const1932 void RegReductionPQBase::dumpRegPressure() const {
1933 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1934 for (TargetRegisterInfo::regclass_iterator I = TRI->regclass_begin(),
1935 E = TRI->regclass_end(); I != E; ++I) {
1936 const TargetRegisterClass *RC = *I;
1937 unsigned Id = RC->getID();
1938 unsigned RP = RegPressure[Id];
1939 if (!RP) continue;
1940 DEBUG(dbgs() << TRI->getRegClassName(RC) << ": " << RP << " / "
1941 << RegLimit[Id] << '\n');
1942 }
1943 #endif
1944 }
1945
HighRegPressure(const SUnit * SU) const1946 bool RegReductionPQBase::HighRegPressure(const SUnit *SU) const {
1947 if (!TLI)
1948 return false;
1949
1950 for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
1951 I != E; ++I) {
1952 if (I->isCtrl())
1953 continue;
1954 SUnit *PredSU = I->getSUnit();
1955 // NumRegDefsLeft is zero when enough uses of this node have been scheduled
1956 // to cover the number of registers defined (they are all live).
1957 if (PredSU->NumRegDefsLeft == 0) {
1958 continue;
1959 }
1960 for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
1961 RegDefPos.IsValid(); RegDefPos.Advance()) {
1962 unsigned RCId, Cost;
1963 GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
1964
1965 if ((RegPressure[RCId] + Cost) >= RegLimit[RCId])
1966 return true;
1967 }
1968 }
1969 return false;
1970 }
1971
MayReduceRegPressure(SUnit * SU) const1972 bool RegReductionPQBase::MayReduceRegPressure(SUnit *SU) const {
1973 const SDNode *N = SU->getNode();
1974
1975 if (!N->isMachineOpcode() || !SU->NumSuccs)
1976 return false;
1977
1978 unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
1979 for (unsigned i = 0; i != NumDefs; ++i) {
1980 MVT VT = N->getSimpleValueType(i);
1981 if (!N->hasAnyUseOfValue(i))
1982 continue;
1983 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
1984 if (RegPressure[RCId] >= RegLimit[RCId])
1985 return true;
1986 }
1987 return false;
1988 }
1989
1990 // Compute the register pressure contribution by this instruction by count up
1991 // for uses that are not live and down for defs. Only count register classes
1992 // that are already under high pressure. As a side effect, compute the number of
1993 // uses of registers that are already live.
1994 //
1995 // FIXME: This encompasses the logic in HighRegPressure and MayReduceRegPressure
1996 // so could probably be factored.
RegPressureDiff(SUnit * SU,unsigned & LiveUses) const1997 int RegReductionPQBase::RegPressureDiff(SUnit *SU, unsigned &LiveUses) const {
1998 LiveUses = 0;
1999 int PDiff = 0;
2000 for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
2001 I != E; ++I) {
2002 if (I->isCtrl())
2003 continue;
2004 SUnit *PredSU = I->getSUnit();
2005 // NumRegDefsLeft is zero when enough uses of this node have been scheduled
2006 // to cover the number of registers defined (they are all live).
2007 if (PredSU->NumRegDefsLeft == 0) {
2008 if (PredSU->getNode()->isMachineOpcode())
2009 ++LiveUses;
2010 continue;
2011 }
2012 for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
2013 RegDefPos.IsValid(); RegDefPos.Advance()) {
2014 MVT VT = RegDefPos.GetValue();
2015 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2016 if (RegPressure[RCId] >= RegLimit[RCId])
2017 ++PDiff;
2018 }
2019 }
2020 const SDNode *N = SU->getNode();
2021
2022 if (!N || !N->isMachineOpcode() || !SU->NumSuccs)
2023 return PDiff;
2024
2025 unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2026 for (unsigned i = 0; i != NumDefs; ++i) {
2027 MVT VT = N->getSimpleValueType(i);
2028 if (!N->hasAnyUseOfValue(i))
2029 continue;
2030 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2031 if (RegPressure[RCId] >= RegLimit[RCId])
2032 --PDiff;
2033 }
2034 return PDiff;
2035 }
2036
scheduledNode(SUnit * SU)2037 void RegReductionPQBase::scheduledNode(SUnit *SU) {
2038 if (!TracksRegPressure)
2039 return;
2040
2041 if (!SU->getNode())
2042 return;
2043
2044 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
2045 I != E; ++I) {
2046 if (I->isCtrl())
2047 continue;
2048 SUnit *PredSU = I->getSUnit();
2049 // NumRegDefsLeft is zero when enough uses of this node have been scheduled
2050 // to cover the number of registers defined (they are all live).
2051 if (PredSU->NumRegDefsLeft == 0) {
2052 continue;
2053 }
2054 // FIXME: The ScheduleDAG currently loses information about which of a
2055 // node's values is consumed by each dependence. Consequently, if the node
2056 // defines multiple register classes, we don't know which to pressurize
2057 // here. Instead the following loop consumes the register defs in an
2058 // arbitrary order. At least it handles the common case of clustered loads
2059 // to the same class. For precise liveness, each SDep needs to indicate the
2060 // result number. But that tightly couples the ScheduleDAG with the
2061 // SelectionDAG making updates tricky. A simpler hack would be to attach a
2062 // value type or register class to SDep.
2063 //
2064 // The most important aspect of register tracking is balancing the increase
2065 // here with the reduction further below. Note that this SU may use multiple
2066 // defs in PredSU. The can't be determined here, but we've already
2067 // compensated by reducing NumRegDefsLeft in PredSU during
2068 // ScheduleDAGSDNodes::AddSchedEdges.
2069 --PredSU->NumRegDefsLeft;
2070 unsigned SkipRegDefs = PredSU->NumRegDefsLeft;
2071 for (ScheduleDAGSDNodes::RegDefIter RegDefPos(PredSU, scheduleDAG);
2072 RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
2073 if (SkipRegDefs)
2074 continue;
2075
2076 unsigned RCId, Cost;
2077 GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
2078 RegPressure[RCId] += Cost;
2079 break;
2080 }
2081 }
2082
2083 // We should have this assert, but there may be dead SDNodes that never
2084 // materialize as SUnits, so they don't appear to generate liveness.
2085 //assert(SU->NumRegDefsLeft == 0 && "not all regdefs have scheduled uses");
2086 int SkipRegDefs = (int)SU->NumRegDefsLeft;
2087 for (ScheduleDAGSDNodes::RegDefIter RegDefPos(SU, scheduleDAG);
2088 RegDefPos.IsValid(); RegDefPos.Advance(), --SkipRegDefs) {
2089 if (SkipRegDefs > 0)
2090 continue;
2091 unsigned RCId, Cost;
2092 GetCostForDef(RegDefPos, TLI, TII, TRI, RCId, Cost, MF);
2093 if (RegPressure[RCId] < Cost) {
2094 // Register pressure tracking is imprecise. This can happen. But we try
2095 // hard not to let it happen because it likely results in poor scheduling.
2096 DEBUG(dbgs() << " SU(" << SU->NodeNum << ") has too many regdefs\n");
2097 RegPressure[RCId] = 0;
2098 }
2099 else {
2100 RegPressure[RCId] -= Cost;
2101 }
2102 }
2103 dumpRegPressure();
2104 }
2105
unscheduledNode(SUnit * SU)2106 void RegReductionPQBase::unscheduledNode(SUnit *SU) {
2107 if (!TracksRegPressure)
2108 return;
2109
2110 const SDNode *N = SU->getNode();
2111 if (!N) return;
2112
2113 if (!N->isMachineOpcode()) {
2114 if (N->getOpcode() != ISD::CopyToReg)
2115 return;
2116 } else {
2117 unsigned Opc = N->getMachineOpcode();
2118 if (Opc == TargetOpcode::EXTRACT_SUBREG ||
2119 Opc == TargetOpcode::INSERT_SUBREG ||
2120 Opc == TargetOpcode::SUBREG_TO_REG ||
2121 Opc == TargetOpcode::REG_SEQUENCE ||
2122 Opc == TargetOpcode::IMPLICIT_DEF)
2123 return;
2124 }
2125
2126 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
2127 I != E; ++I) {
2128 if (I->isCtrl())
2129 continue;
2130 SUnit *PredSU = I->getSUnit();
2131 // NumSuccsLeft counts all deps. Don't compare it with NumSuccs which only
2132 // counts data deps.
2133 if (PredSU->NumSuccsLeft != PredSU->Succs.size())
2134 continue;
2135 const SDNode *PN = PredSU->getNode();
2136 if (!PN->isMachineOpcode()) {
2137 if (PN->getOpcode() == ISD::CopyFromReg) {
2138 MVT VT = PN->getSimpleValueType(0);
2139 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2140 RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2141 }
2142 continue;
2143 }
2144 unsigned POpc = PN->getMachineOpcode();
2145 if (POpc == TargetOpcode::IMPLICIT_DEF)
2146 continue;
2147 if (POpc == TargetOpcode::EXTRACT_SUBREG ||
2148 POpc == TargetOpcode::INSERT_SUBREG ||
2149 POpc == TargetOpcode::SUBREG_TO_REG) {
2150 MVT VT = PN->getSimpleValueType(0);
2151 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2152 RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2153 continue;
2154 }
2155 unsigned NumDefs = TII->get(PN->getMachineOpcode()).getNumDefs();
2156 for (unsigned i = 0; i != NumDefs; ++i) {
2157 MVT VT = PN->getSimpleValueType(i);
2158 if (!PN->hasAnyUseOfValue(i))
2159 continue;
2160 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2161 if (RegPressure[RCId] < TLI->getRepRegClassCostFor(VT))
2162 // Register pressure tracking is imprecise. This can happen.
2163 RegPressure[RCId] = 0;
2164 else
2165 RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
2166 }
2167 }
2168
2169 // Check for isMachineOpcode() as PrescheduleNodesWithMultipleUses()
2170 // may transfer data dependencies to CopyToReg.
2171 if (SU->NumSuccs && N->isMachineOpcode()) {
2172 unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2173 for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
2174 MVT VT = N->getSimpleValueType(i);
2175 if (VT == MVT::Glue || VT == MVT::Other)
2176 continue;
2177 if (!N->hasAnyUseOfValue(i))
2178 continue;
2179 unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
2180 RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
2181 }
2182 }
2183
2184 dumpRegPressure();
2185 }
2186
2187 //===----------------------------------------------------------------------===//
2188 // Dynamic Node Priority for Register Pressure Reduction
2189 //===----------------------------------------------------------------------===//
2190
2191 /// closestSucc - Returns the scheduled cycle of the successor which is
2192 /// closest to the current cycle.
closestSucc(const SUnit * SU)2193 static unsigned closestSucc(const SUnit *SU) {
2194 unsigned MaxHeight = 0;
2195 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
2196 I != E; ++I) {
2197 if (I->isCtrl()) continue; // ignore chain succs
2198 unsigned Height = I->getSUnit()->getHeight();
2199 // If there are bunch of CopyToRegs stacked up, they should be considered
2200 // to be at the same position.
2201 if (I->getSUnit()->getNode() &&
2202 I->getSUnit()->getNode()->getOpcode() == ISD::CopyToReg)
2203 Height = closestSucc(I->getSUnit())+1;
2204 if (Height > MaxHeight)
2205 MaxHeight = Height;
2206 }
2207 return MaxHeight;
2208 }
2209
2210 /// calcMaxScratches - Returns an cost estimate of the worse case requirement
2211 /// for scratch registers, i.e. number of data dependencies.
calcMaxScratches(const SUnit * SU)2212 static unsigned calcMaxScratches(const SUnit *SU) {
2213 unsigned Scratches = 0;
2214 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
2215 I != E; ++I) {
2216 if (I->isCtrl()) continue; // ignore chain preds
2217 Scratches++;
2218 }
2219 return Scratches;
2220 }
2221
2222 /// hasOnlyLiveInOpers - Return true if SU has only value predecessors that are
2223 /// CopyFromReg from a virtual register.
hasOnlyLiveInOpers(const SUnit * SU)2224 static bool hasOnlyLiveInOpers(const SUnit *SU) {
2225 bool RetVal = false;
2226 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
2227 I != E; ++I) {
2228 if (I->isCtrl()) continue;
2229 const SUnit *PredSU = I->getSUnit();
2230 if (PredSU->getNode() &&
2231 PredSU->getNode()->getOpcode() == ISD::CopyFromReg) {
2232 unsigned Reg =
2233 cast<RegisterSDNode>(PredSU->getNode()->getOperand(1))->getReg();
2234 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
2235 RetVal = true;
2236 continue;
2237 }
2238 }
2239 return false;
2240 }
2241 return RetVal;
2242 }
2243
2244 /// hasOnlyLiveOutUses - Return true if SU has only value successors that are
2245 /// CopyToReg to a virtual register. This SU def is probably a liveout and
2246 /// it has no other use. It should be scheduled closer to the terminator.
hasOnlyLiveOutUses(const SUnit * SU)2247 static bool hasOnlyLiveOutUses(const SUnit *SU) {
2248 bool RetVal = false;
2249 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
2250 I != E; ++I) {
2251 if (I->isCtrl()) continue;
2252 const SUnit *SuccSU = I->getSUnit();
2253 if (SuccSU->getNode() && SuccSU->getNode()->getOpcode() == ISD::CopyToReg) {
2254 unsigned Reg =
2255 cast<RegisterSDNode>(SuccSU->getNode()->getOperand(1))->getReg();
2256 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
2257 RetVal = true;
2258 continue;
2259 }
2260 }
2261 return false;
2262 }
2263 return RetVal;
2264 }
2265
2266 // Set isVRegCycle for a node with only live in opers and live out uses. Also
2267 // set isVRegCycle for its CopyFromReg operands.
2268 //
2269 // This is only relevant for single-block loops, in which case the VRegCycle
2270 // node is likely an induction variable in which the operand and target virtual
2271 // registers should be coalesced (e.g. pre/post increment values). Setting the
2272 // isVRegCycle flag helps the scheduler prioritize other uses of the same
2273 // CopyFromReg so that this node becomes the virtual register "kill". This
2274 // avoids interference between the values live in and out of the block and
2275 // eliminates a copy inside the loop.
initVRegCycle(SUnit * SU)2276 static void initVRegCycle(SUnit *SU) {
2277 if (DisableSchedVRegCycle)
2278 return;
2279
2280 if (!hasOnlyLiveInOpers(SU) || !hasOnlyLiveOutUses(SU))
2281 return;
2282
2283 DEBUG(dbgs() << "VRegCycle: SU(" << SU->NodeNum << ")\n");
2284
2285 SU->isVRegCycle = true;
2286
2287 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
2288 I != E; ++I) {
2289 if (I->isCtrl()) continue;
2290 I->getSUnit()->isVRegCycle = true;
2291 }
2292 }
2293
2294 // After scheduling the definition of a VRegCycle, clear the isVRegCycle flag of
2295 // CopyFromReg operands. We should no longer penalize other uses of this VReg.
resetVRegCycle(SUnit * SU)2296 static void resetVRegCycle(SUnit *SU) {
2297 if (!SU->isVRegCycle)
2298 return;
2299
2300 for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
2301 I != E; ++I) {
2302 if (I->isCtrl()) continue; // ignore chain preds
2303 SUnit *PredSU = I->getSUnit();
2304 if (PredSU->isVRegCycle) {
2305 assert(PredSU->getNode()->getOpcode() == ISD::CopyFromReg &&
2306 "VRegCycle def must be CopyFromReg");
2307 I->getSUnit()->isVRegCycle = 0;
2308 }
2309 }
2310 }
2311
2312 // Return true if this SUnit uses a CopyFromReg node marked as a VRegCycle. This
2313 // means a node that defines the VRegCycle has not been scheduled yet.
hasVRegCycleUse(const SUnit * SU)2314 static bool hasVRegCycleUse(const SUnit *SU) {
2315 // If this SU also defines the VReg, don't hoist it as a "use".
2316 if (SU->isVRegCycle)
2317 return false;
2318
2319 for (SUnit::const_pred_iterator I = SU->Preds.begin(),E = SU->Preds.end();
2320 I != E; ++I) {
2321 if (I->isCtrl()) continue; // ignore chain preds
2322 if (I->getSUnit()->isVRegCycle &&
2323 I->getSUnit()->getNode()->getOpcode() == ISD::CopyFromReg) {
2324 DEBUG(dbgs() << " VReg cycle use: SU (" << SU->NodeNum << ")\n");
2325 return true;
2326 }
2327 }
2328 return false;
2329 }
2330
2331 // Check for either a dependence (latency) or resource (hazard) stall.
2332 //
2333 // Note: The ScheduleHazardRecognizer interface requires a non-const SU.
BUHasStall(SUnit * SU,int Height,RegReductionPQBase * SPQ)2334 static bool BUHasStall(SUnit *SU, int Height, RegReductionPQBase *SPQ) {
2335 if ((int)SPQ->getCurCycle() < Height) return true;
2336 if (SPQ->getHazardRec()->getHazardType(SU, 0)
2337 != ScheduleHazardRecognizer::NoHazard)
2338 return true;
2339 return false;
2340 }
2341
2342 // Return -1 if left has higher priority, 1 if right has higher priority.
2343 // Return 0 if latency-based priority is equivalent.
BUCompareLatency(SUnit * left,SUnit * right,bool checkPref,RegReductionPQBase * SPQ)2344 static int BUCompareLatency(SUnit *left, SUnit *right, bool checkPref,
2345 RegReductionPQBase *SPQ) {
2346 // Scheduling an instruction that uses a VReg whose postincrement has not yet
2347 // been scheduled will induce a copy. Model this as an extra cycle of latency.
2348 int LPenalty = hasVRegCycleUse(left) ? 1 : 0;
2349 int RPenalty = hasVRegCycleUse(right) ? 1 : 0;
2350 int LHeight = (int)left->getHeight() + LPenalty;
2351 int RHeight = (int)right->getHeight() + RPenalty;
2352
2353 bool LStall = (!checkPref || left->SchedulingPref == Sched::ILP) &&
2354 BUHasStall(left, LHeight, SPQ);
2355 bool RStall = (!checkPref || right->SchedulingPref == Sched::ILP) &&
2356 BUHasStall(right, RHeight, SPQ);
2357
2358 // If scheduling one of the node will cause a pipeline stall, delay it.
2359 // If scheduling either one of the node will cause a pipeline stall, sort
2360 // them according to their height.
2361 if (LStall) {
2362 if (!RStall)
2363 return 1;
2364 if (LHeight != RHeight)
2365 return LHeight > RHeight ? 1 : -1;
2366 } else if (RStall)
2367 return -1;
2368
2369 // If either node is scheduling for latency, sort them by height/depth
2370 // and latency.
2371 if (!checkPref || (left->SchedulingPref == Sched::ILP ||
2372 right->SchedulingPref == Sched::ILP)) {
2373 // If neither instruction stalls (!LStall && !RStall) and HazardRecognizer
2374 // is enabled, grouping instructions by cycle, then its height is already
2375 // covered so only its depth matters. We also reach this point if both stall
2376 // but have the same height.
2377 if (!SPQ->getHazardRec()->isEnabled()) {
2378 if (LHeight != RHeight)
2379 return LHeight > RHeight ? 1 : -1;
2380 }
2381 int LDepth = left->getDepth() - LPenalty;
2382 int RDepth = right->getDepth() - RPenalty;
2383 if (LDepth != RDepth) {
2384 DEBUG(dbgs() << " Comparing latency of SU (" << left->NodeNum
2385 << ") depth " << LDepth << " vs SU (" << right->NodeNum
2386 << ") depth " << RDepth << "\n");
2387 return LDepth < RDepth ? 1 : -1;
2388 }
2389 if (left->Latency != right->Latency)
2390 return left->Latency > right->Latency ? 1 : -1;
2391 }
2392 return 0;
2393 }
2394
BURRSort(SUnit * left,SUnit * right,RegReductionPQBase * SPQ)2395 static bool BURRSort(SUnit *left, SUnit *right, RegReductionPQBase *SPQ) {
2396 // Schedule physical register definitions close to their use. This is
2397 // motivated by microarchitectures that can fuse cmp+jump macro-ops. But as
2398 // long as shortening physreg live ranges is generally good, we can defer
2399 // creating a subtarget hook.
2400 if (!DisableSchedPhysRegJoin) {
2401 bool LHasPhysReg = left->hasPhysRegDefs;
2402 bool RHasPhysReg = right->hasPhysRegDefs;
2403 if (LHasPhysReg != RHasPhysReg) {
2404 #ifndef NDEBUG
2405 static const char *const PhysRegMsg[] = { " has no physreg",
2406 " defines a physreg" };
2407 #endif
2408 DEBUG(dbgs() << " SU (" << left->NodeNum << ") "
2409 << PhysRegMsg[LHasPhysReg] << " SU(" << right->NodeNum << ") "
2410 << PhysRegMsg[RHasPhysReg] << "\n");
2411 return LHasPhysReg < RHasPhysReg;
2412 }
2413 }
2414
2415 // Prioritize by Sethi-Ulmann number and push CopyToReg nodes down.
2416 unsigned LPriority = SPQ->getNodePriority(left);
2417 unsigned RPriority = SPQ->getNodePriority(right);
2418
2419 // Be really careful about hoisting call operands above previous calls.
2420 // Only allows it if it would reduce register pressure.
2421 if (left->isCall && right->isCallOp) {
2422 unsigned RNumVals = right->getNode()->getNumValues();
2423 RPriority = (RPriority > RNumVals) ? (RPriority - RNumVals) : 0;
2424 }
2425 if (right->isCall && left->isCallOp) {
2426 unsigned LNumVals = left->getNode()->getNumValues();
2427 LPriority = (LPriority > LNumVals) ? (LPriority - LNumVals) : 0;
2428 }
2429
2430 if (LPriority != RPriority)
2431 return LPriority > RPriority;
2432
2433 // One or both of the nodes are calls and their sethi-ullman numbers are the
2434 // same, then keep source order.
2435 if (left->isCall || right->isCall) {
2436 unsigned LOrder = SPQ->getNodeOrdering(left);
2437 unsigned ROrder = SPQ->getNodeOrdering(right);
2438
2439 // Prefer an ordering where the lower the non-zero order number, the higher
2440 // the preference.
2441 if ((LOrder || ROrder) && LOrder != ROrder)
2442 return LOrder != 0 && (LOrder < ROrder || ROrder == 0);
2443 }
2444
2445 // Try schedule def + use closer when Sethi-Ullman numbers are the same.
2446 // e.g.
2447 // t1 = op t2, c1
2448 // t3 = op t4, c2
2449 //
2450 // and the following instructions are both ready.
2451 // t2 = op c3
2452 // t4 = op c4
2453 //
2454 // Then schedule t2 = op first.
2455 // i.e.
2456 // t4 = op c4
2457 // t2 = op c3
2458 // t1 = op t2, c1
2459 // t3 = op t4, c2
2460 //
2461 // This creates more short live intervals.
2462 unsigned LDist = closestSucc(left);
2463 unsigned RDist = closestSucc(right);
2464 if (LDist != RDist)
2465 return LDist < RDist;
2466
2467 // How many registers becomes live when the node is scheduled.
2468 unsigned LScratch = calcMaxScratches(left);
2469 unsigned RScratch = calcMaxScratches(right);
2470 if (LScratch != RScratch)
2471 return LScratch > RScratch;
2472
2473 // Comparing latency against a call makes little sense unless the node
2474 // is register pressure-neutral.
2475 if ((left->isCall && RPriority > 0) || (right->isCall && LPriority > 0))
2476 return (left->NodeQueueId > right->NodeQueueId);
2477
2478 // Do not compare latencies when one or both of the nodes are calls.
2479 if (!DisableSchedCycles &&
2480 !(left->isCall || right->isCall)) {
2481 int result = BUCompareLatency(left, right, false /*checkPref*/, SPQ);
2482 if (result != 0)
2483 return result > 0;
2484 }
2485 else {
2486 if (left->getHeight() != right->getHeight())
2487 return left->getHeight() > right->getHeight();
2488
2489 if (left->getDepth() != right->getDepth())
2490 return left->getDepth() < right->getDepth();
2491 }
2492
2493 assert(left->NodeQueueId && right->NodeQueueId &&
2494 "NodeQueueId cannot be zero");
2495 return (left->NodeQueueId > right->NodeQueueId);
2496 }
2497
2498 // Bottom up
operator ()(SUnit * left,SUnit * right) const2499 bool bu_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2500 if (int res = checkSpecialNodes(left, right))
2501 return res > 0;
2502
2503 return BURRSort(left, right, SPQ);
2504 }
2505
2506 // Source order, otherwise bottom up.
operator ()(SUnit * left,SUnit * right) const2507 bool src_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2508 if (int res = checkSpecialNodes(left, right))
2509 return res > 0;
2510
2511 unsigned LOrder = SPQ->getNodeOrdering(left);
2512 unsigned ROrder = SPQ->getNodeOrdering(right);
2513
2514 // Prefer an ordering where the lower the non-zero order number, the higher
2515 // the preference.
2516 if ((LOrder || ROrder) && LOrder != ROrder)
2517 return LOrder != 0 && (LOrder < ROrder || ROrder == 0);
2518
2519 return BURRSort(left, right, SPQ);
2520 }
2521
2522 // If the time between now and when the instruction will be ready can cover
2523 // the spill code, then avoid adding it to the ready queue. This gives long
2524 // stalls highest priority and allows hoisting across calls. It should also
2525 // speed up processing the available queue.
isReady(SUnit * SU,unsigned CurCycle) const2526 bool hybrid_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
2527 static const unsigned ReadyDelay = 3;
2528
2529 if (SPQ->MayReduceRegPressure(SU)) return true;
2530
2531 if (SU->getHeight() > (CurCycle + ReadyDelay)) return false;
2532
2533 if (SPQ->getHazardRec()->getHazardType(SU, -ReadyDelay)
2534 != ScheduleHazardRecognizer::NoHazard)
2535 return false;
2536
2537 return true;
2538 }
2539
2540 // Return true if right should be scheduled with higher priority than left.
operator ()(SUnit * left,SUnit * right) const2541 bool hybrid_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2542 if (int res = checkSpecialNodes(left, right))
2543 return res > 0;
2544
2545 if (left->isCall || right->isCall)
2546 // No way to compute latency of calls.
2547 return BURRSort(left, right, SPQ);
2548
2549 bool LHigh = SPQ->HighRegPressure(left);
2550 bool RHigh = SPQ->HighRegPressure(right);
2551 // Avoid causing spills. If register pressure is high, schedule for
2552 // register pressure reduction.
2553 if (LHigh && !RHigh) {
2554 DEBUG(dbgs() << " pressure SU(" << left->NodeNum << ") > SU("
2555 << right->NodeNum << ")\n");
2556 return true;
2557 }
2558 else if (!LHigh && RHigh) {
2559 DEBUG(dbgs() << " pressure SU(" << right->NodeNum << ") > SU("
2560 << left->NodeNum << ")\n");
2561 return false;
2562 }
2563 if (!LHigh && !RHigh) {
2564 int result = BUCompareLatency(left, right, true /*checkPref*/, SPQ);
2565 if (result != 0)
2566 return result > 0;
2567 }
2568 return BURRSort(left, right, SPQ);
2569 }
2570
2571 // Schedule as many instructions in each cycle as possible. So don't make an
2572 // instruction available unless it is ready in the current cycle.
isReady(SUnit * SU,unsigned CurCycle) const2573 bool ilp_ls_rr_sort::isReady(SUnit *SU, unsigned CurCycle) const {
2574 if (SU->getHeight() > CurCycle) return false;
2575
2576 if (SPQ->getHazardRec()->getHazardType(SU, 0)
2577 != ScheduleHazardRecognizer::NoHazard)
2578 return false;
2579
2580 return true;
2581 }
2582
canEnableCoalescing(SUnit * SU)2583 static bool canEnableCoalescing(SUnit *SU) {
2584 unsigned Opc = SU->getNode() ? SU->getNode()->getOpcode() : 0;
2585 if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
2586 // CopyToReg should be close to its uses to facilitate coalescing and
2587 // avoid spilling.
2588 return true;
2589
2590 if (Opc == TargetOpcode::EXTRACT_SUBREG ||
2591 Opc == TargetOpcode::SUBREG_TO_REG ||
2592 Opc == TargetOpcode::INSERT_SUBREG)
2593 // EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG nodes should be
2594 // close to their uses to facilitate coalescing.
2595 return true;
2596
2597 if (SU->NumPreds == 0 && SU->NumSuccs != 0)
2598 // If SU does not have a register def, schedule it close to its uses
2599 // because it does not lengthen any live ranges.
2600 return true;
2601
2602 return false;
2603 }
2604
2605 // list-ilp is currently an experimental scheduler that allows various
2606 // heuristics to be enabled prior to the normal register reduction logic.
operator ()(SUnit * left,SUnit * right) const2607 bool ilp_ls_rr_sort::operator()(SUnit *left, SUnit *right) const {
2608 if (int res = checkSpecialNodes(left, right))
2609 return res > 0;
2610
2611 if (left->isCall || right->isCall)
2612 // No way to compute latency of calls.
2613 return BURRSort(left, right, SPQ);
2614
2615 unsigned LLiveUses = 0, RLiveUses = 0;
2616 int LPDiff = 0, RPDiff = 0;
2617 if (!DisableSchedRegPressure || !DisableSchedLiveUses) {
2618 LPDiff = SPQ->RegPressureDiff(left, LLiveUses);
2619 RPDiff = SPQ->RegPressureDiff(right, RLiveUses);
2620 }
2621 if (!DisableSchedRegPressure && LPDiff != RPDiff) {
2622 DEBUG(dbgs() << "RegPressureDiff SU(" << left->NodeNum << "): " << LPDiff
2623 << " != SU(" << right->NodeNum << "): " << RPDiff << "\n");
2624 return LPDiff > RPDiff;
2625 }
2626
2627 if (!DisableSchedRegPressure && (LPDiff > 0 || RPDiff > 0)) {
2628 bool LReduce = canEnableCoalescing(left);
2629 bool RReduce = canEnableCoalescing(right);
2630 if (LReduce && !RReduce) return false;
2631 if (RReduce && !LReduce) return true;
2632 }
2633
2634 if (!DisableSchedLiveUses && (LLiveUses != RLiveUses)) {
2635 DEBUG(dbgs() << "Live uses SU(" << left->NodeNum << "): " << LLiveUses
2636 << " != SU(" << right->NodeNum << "): " << RLiveUses << "\n");
2637 return LLiveUses < RLiveUses;
2638 }
2639
2640 if (!DisableSchedStalls) {
2641 bool LStall = BUHasStall(left, left->getHeight(), SPQ);
2642 bool RStall = BUHasStall(right, right->getHeight(), SPQ);
2643 if (LStall != RStall)
2644 return left->getHeight() > right->getHeight();
2645 }
2646
2647 if (!DisableSchedCriticalPath) {
2648 int spread = (int)left->getDepth() - (int)right->getDepth();
2649 if (std::abs(spread) > MaxReorderWindow) {
2650 DEBUG(dbgs() << "Depth of SU(" << left->NodeNum << "): "
2651 << left->getDepth() << " != SU(" << right->NodeNum << "): "
2652 << right->getDepth() << "\n");
2653 return left->getDepth() < right->getDepth();
2654 }
2655 }
2656
2657 if (!DisableSchedHeight && left->getHeight() != right->getHeight()) {
2658 int spread = (int)left->getHeight() - (int)right->getHeight();
2659 if (std::abs(spread) > MaxReorderWindow)
2660 return left->getHeight() > right->getHeight();
2661 }
2662
2663 return BURRSort(left, right, SPQ);
2664 }
2665
initNodes(std::vector<SUnit> & sunits)2666 void RegReductionPQBase::initNodes(std::vector<SUnit> &sunits) {
2667 SUnits = &sunits;
2668 // Add pseudo dependency edges for two-address nodes.
2669 if (!Disable2AddrHack)
2670 AddPseudoTwoAddrDeps();
2671 // Reroute edges to nodes with multiple uses.
2672 if (!TracksRegPressure && !SrcOrder)
2673 PrescheduleNodesWithMultipleUses();
2674 // Calculate node priorities.
2675 CalculateSethiUllmanNumbers();
2676
2677 // For single block loops, mark nodes that look like canonical IV increments.
2678 if (scheduleDAG->BB->isSuccessor(scheduleDAG->BB)) {
2679 for (unsigned i = 0, e = sunits.size(); i != e; ++i) {
2680 initVRegCycle(&sunits[i]);
2681 }
2682 }
2683 }
2684
2685 //===----------------------------------------------------------------------===//
2686 // Preschedule for Register Pressure
2687 //===----------------------------------------------------------------------===//
2688
canClobber(const SUnit * SU,const SUnit * Op)2689 bool RegReductionPQBase::canClobber(const SUnit *SU, const SUnit *Op) {
2690 if (SU->isTwoAddress) {
2691 unsigned Opc = SU->getNode()->getMachineOpcode();
2692 const MCInstrDesc &MCID = TII->get(Opc);
2693 unsigned NumRes = MCID.getNumDefs();
2694 unsigned NumOps = MCID.getNumOperands() - NumRes;
2695 for (unsigned i = 0; i != NumOps; ++i) {
2696 if (MCID.getOperandConstraint(i+NumRes, MCOI::TIED_TO) != -1) {
2697 SDNode *DU = SU->getNode()->getOperand(i).getNode();
2698 if (DU->getNodeId() != -1 &&
2699 Op->OrigNode == &(*SUnits)[DU->getNodeId()])
2700 return true;
2701 }
2702 }
2703 }
2704 return false;
2705 }
2706
2707 /// canClobberReachingPhysRegUse - True if SU would clobber one of it's
2708 /// successor's explicit physregs whose definition can reach DepSU.
2709 /// i.e. DepSU should not be scheduled above SU.
canClobberReachingPhysRegUse(const SUnit * DepSU,const SUnit * SU,ScheduleDAGRRList * scheduleDAG,const TargetInstrInfo * TII,const TargetRegisterInfo * TRI)2710 static bool canClobberReachingPhysRegUse(const SUnit *DepSU, const SUnit *SU,
2711 ScheduleDAGRRList *scheduleDAG,
2712 const TargetInstrInfo *TII,
2713 const TargetRegisterInfo *TRI) {
2714 const uint16_t *ImpDefs
2715 = TII->get(SU->getNode()->getMachineOpcode()).getImplicitDefs();
2716 const uint32_t *RegMask = getNodeRegMask(SU->getNode());
2717 if(!ImpDefs && !RegMask)
2718 return false;
2719
2720 for (SUnit::const_succ_iterator SI = SU->Succs.begin(), SE = SU->Succs.end();
2721 SI != SE; ++SI) {
2722 SUnit *SuccSU = SI->getSUnit();
2723 for (SUnit::const_pred_iterator PI = SuccSU->Preds.begin(),
2724 PE = SuccSU->Preds.end(); PI != PE; ++PI) {
2725 if (!PI->isAssignedRegDep())
2726 continue;
2727
2728 if (RegMask && MachineOperand::clobbersPhysReg(RegMask, PI->getReg()) &&
2729 scheduleDAG->IsReachable(DepSU, PI->getSUnit()))
2730 return true;
2731
2732 if (ImpDefs)
2733 for (const uint16_t *ImpDef = ImpDefs; *ImpDef; ++ImpDef)
2734 // Return true if SU clobbers this physical register use and the
2735 // definition of the register reaches from DepSU. IsReachable queries
2736 // a topological forward sort of the DAG (following the successors).
2737 if (TRI->regsOverlap(*ImpDef, PI->getReg()) &&
2738 scheduleDAG->IsReachable(DepSU, PI->getSUnit()))
2739 return true;
2740 }
2741 }
2742 return false;
2743 }
2744
2745 /// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's
2746 /// physical register defs.
canClobberPhysRegDefs(const SUnit * SuccSU,const SUnit * SU,const TargetInstrInfo * TII,const TargetRegisterInfo * TRI)2747 static bool canClobberPhysRegDefs(const SUnit *SuccSU, const SUnit *SU,
2748 const TargetInstrInfo *TII,
2749 const TargetRegisterInfo *TRI) {
2750 SDNode *N = SuccSU->getNode();
2751 unsigned NumDefs = TII->get(N->getMachineOpcode()).getNumDefs();
2752 const uint16_t *ImpDefs = TII->get(N->getMachineOpcode()).getImplicitDefs();
2753 assert(ImpDefs && "Caller should check hasPhysRegDefs");
2754 for (const SDNode *SUNode = SU->getNode(); SUNode;
2755 SUNode = SUNode->getGluedNode()) {
2756 if (!SUNode->isMachineOpcode())
2757 continue;
2758 const uint16_t *SUImpDefs =
2759 TII->get(SUNode->getMachineOpcode()).getImplicitDefs();
2760 const uint32_t *SURegMask = getNodeRegMask(SUNode);
2761 if (!SUImpDefs && !SURegMask)
2762 continue;
2763 for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
2764 MVT VT = N->getSimpleValueType(i);
2765 if (VT == MVT::Glue || VT == MVT::Other)
2766 continue;
2767 if (!N->hasAnyUseOfValue(i))
2768 continue;
2769 unsigned Reg = ImpDefs[i - NumDefs];
2770 if (SURegMask && MachineOperand::clobbersPhysReg(SURegMask, Reg))
2771 return true;
2772 if (!SUImpDefs)
2773 continue;
2774 for (;*SUImpDefs; ++SUImpDefs) {
2775 unsigned SUReg = *SUImpDefs;
2776 if (TRI->regsOverlap(Reg, SUReg))
2777 return true;
2778 }
2779 }
2780 }
2781 return false;
2782 }
2783
2784 /// PrescheduleNodesWithMultipleUses - Nodes with multiple uses
2785 /// are not handled well by the general register pressure reduction
2786 /// heuristics. When presented with code like this:
2787 ///
2788 /// N
2789 /// / |
2790 /// / |
2791 /// U store
2792 /// |
2793 /// ...
2794 ///
2795 /// the heuristics tend to push the store up, but since the
2796 /// operand of the store has another use (U), this would increase
2797 /// the length of that other use (the U->N edge).
2798 ///
2799 /// This function transforms code like the above to route U's
2800 /// dependence through the store when possible, like this:
2801 ///
2802 /// N
2803 /// ||
2804 /// ||
2805 /// store
2806 /// |
2807 /// U
2808 /// |
2809 /// ...
2810 ///
2811 /// This results in the store being scheduled immediately
2812 /// after N, which shortens the U->N live range, reducing
2813 /// register pressure.
2814 ///
PrescheduleNodesWithMultipleUses()2815 void RegReductionPQBase::PrescheduleNodesWithMultipleUses() {
2816 // Visit all the nodes in topological order, working top-down.
2817 for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
2818 SUnit *SU = &(*SUnits)[i];
2819 // For now, only look at nodes with no data successors, such as stores.
2820 // These are especially important, due to the heuristics in
2821 // getNodePriority for nodes with no data successors.
2822 if (SU->NumSuccs != 0)
2823 continue;
2824 // For now, only look at nodes with exactly one data predecessor.
2825 if (SU->NumPreds != 1)
2826 continue;
2827 // Avoid prescheduling copies to virtual registers, which don't behave
2828 // like other nodes from the perspective of scheduling heuristics.
2829 if (SDNode *N = SU->getNode())
2830 if (N->getOpcode() == ISD::CopyToReg &&
2831 TargetRegisterInfo::isVirtualRegister
2832 (cast<RegisterSDNode>(N->getOperand(1))->getReg()))
2833 continue;
2834
2835 // Locate the single data predecessor.
2836 SUnit *PredSU = nullptr;
2837 for (SUnit::const_pred_iterator II = SU->Preds.begin(),
2838 EE = SU->Preds.end(); II != EE; ++II)
2839 if (!II->isCtrl()) {
2840 PredSU = II->getSUnit();
2841 break;
2842 }
2843 assert(PredSU);
2844
2845 // Don't rewrite edges that carry physregs, because that requires additional
2846 // support infrastructure.
2847 if (PredSU->hasPhysRegDefs)
2848 continue;
2849 // Short-circuit the case where SU is PredSU's only data successor.
2850 if (PredSU->NumSuccs == 1)
2851 continue;
2852 // Avoid prescheduling to copies from virtual registers, which don't behave
2853 // like other nodes from the perspective of scheduling heuristics.
2854 if (SDNode *N = SU->getNode())
2855 if (N->getOpcode() == ISD::CopyFromReg &&
2856 TargetRegisterInfo::isVirtualRegister
2857 (cast<RegisterSDNode>(N->getOperand(1))->getReg()))
2858 continue;
2859
2860 // Perform checks on the successors of PredSU.
2861 for (SUnit::const_succ_iterator II = PredSU->Succs.begin(),
2862 EE = PredSU->Succs.end(); II != EE; ++II) {
2863 SUnit *PredSuccSU = II->getSUnit();
2864 if (PredSuccSU == SU) continue;
2865 // If PredSU has another successor with no data successors, for
2866 // now don't attempt to choose either over the other.
2867 if (PredSuccSU->NumSuccs == 0)
2868 goto outer_loop_continue;
2869 // Don't break physical register dependencies.
2870 if (SU->hasPhysRegClobbers && PredSuccSU->hasPhysRegDefs)
2871 if (canClobberPhysRegDefs(PredSuccSU, SU, TII, TRI))
2872 goto outer_loop_continue;
2873 // Don't introduce graph cycles.
2874 if (scheduleDAG->IsReachable(SU, PredSuccSU))
2875 goto outer_loop_continue;
2876 }
2877
2878 // Ok, the transformation is safe and the heuristics suggest it is
2879 // profitable. Update the graph.
2880 DEBUG(dbgs() << " Prescheduling SU #" << SU->NodeNum
2881 << " next to PredSU #" << PredSU->NodeNum
2882 << " to guide scheduling in the presence of multiple uses\n");
2883 for (unsigned i = 0; i != PredSU->Succs.size(); ++i) {
2884 SDep Edge = PredSU->Succs[i];
2885 assert(!Edge.isAssignedRegDep());
2886 SUnit *SuccSU = Edge.getSUnit();
2887 if (SuccSU != SU) {
2888 Edge.setSUnit(PredSU);
2889 scheduleDAG->RemovePred(SuccSU, Edge);
2890 scheduleDAG->AddPred(SU, Edge);
2891 Edge.setSUnit(SU);
2892 scheduleDAG->AddPred(SuccSU, Edge);
2893 --i;
2894 }
2895 }
2896 outer_loop_continue:;
2897 }
2898 }
2899
2900 /// AddPseudoTwoAddrDeps - If two nodes share an operand and one of them uses
2901 /// it as a def&use operand. Add a pseudo control edge from it to the other
2902 /// node (if it won't create a cycle) so the two-address one will be scheduled
2903 /// first (lower in the schedule). If both nodes are two-address, favor the
2904 /// one that has a CopyToReg use (more likely to be a loop induction update).
2905 /// If both are two-address, but one is commutable while the other is not
2906 /// commutable, favor the one that's not commutable.
AddPseudoTwoAddrDeps()2907 void RegReductionPQBase::AddPseudoTwoAddrDeps() {
2908 for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
2909 SUnit *SU = &(*SUnits)[i];
2910 if (!SU->isTwoAddress)
2911 continue;
2912
2913 SDNode *Node = SU->getNode();
2914 if (!Node || !Node->isMachineOpcode() || SU->getNode()->getGluedNode())
2915 continue;
2916
2917 bool isLiveOut = hasOnlyLiveOutUses(SU);
2918 unsigned Opc = Node->getMachineOpcode();
2919 const MCInstrDesc &MCID = TII->get(Opc);
2920 unsigned NumRes = MCID.getNumDefs();
2921 unsigned NumOps = MCID.getNumOperands() - NumRes;
2922 for (unsigned j = 0; j != NumOps; ++j) {
2923 if (MCID.getOperandConstraint(j+NumRes, MCOI::TIED_TO) == -1)
2924 continue;
2925 SDNode *DU = SU->getNode()->getOperand(j).getNode();
2926 if (DU->getNodeId() == -1)
2927 continue;
2928 const SUnit *DUSU = &(*SUnits)[DU->getNodeId()];
2929 if (!DUSU) continue;
2930 for (SUnit::const_succ_iterator I = DUSU->Succs.begin(),
2931 E = DUSU->Succs.end(); I != E; ++I) {
2932 if (I->isCtrl()) continue;
2933 SUnit *SuccSU = I->getSUnit();
2934 if (SuccSU == SU)
2935 continue;
2936 // Be conservative. Ignore if nodes aren't at roughly the same
2937 // depth and height.
2938 if (SuccSU->getHeight() < SU->getHeight() &&
2939 (SU->getHeight() - SuccSU->getHeight()) > 1)
2940 continue;
2941 // Skip past COPY_TO_REGCLASS nodes, so that the pseudo edge
2942 // constrains whatever is using the copy, instead of the copy
2943 // itself. In the case that the copy is coalesced, this
2944 // preserves the intent of the pseudo two-address heurietics.
2945 while (SuccSU->Succs.size() == 1 &&
2946 SuccSU->getNode()->isMachineOpcode() &&
2947 SuccSU->getNode()->getMachineOpcode() ==
2948 TargetOpcode::COPY_TO_REGCLASS)
2949 SuccSU = SuccSU->Succs.front().getSUnit();
2950 // Don't constrain non-instruction nodes.
2951 if (!SuccSU->getNode() || !SuccSU->getNode()->isMachineOpcode())
2952 continue;
2953 // Don't constrain nodes with physical register defs if the
2954 // predecessor can clobber them.
2955 if (SuccSU->hasPhysRegDefs && SU->hasPhysRegClobbers) {
2956 if (canClobberPhysRegDefs(SuccSU, SU, TII, TRI))
2957 continue;
2958 }
2959 // Don't constrain EXTRACT_SUBREG, INSERT_SUBREG, and SUBREG_TO_REG;
2960 // these may be coalesced away. We want them close to their uses.
2961 unsigned SuccOpc = SuccSU->getNode()->getMachineOpcode();
2962 if (SuccOpc == TargetOpcode::EXTRACT_SUBREG ||
2963 SuccOpc == TargetOpcode::INSERT_SUBREG ||
2964 SuccOpc == TargetOpcode::SUBREG_TO_REG)
2965 continue;
2966 if (!canClobberReachingPhysRegUse(SuccSU, SU, scheduleDAG, TII, TRI) &&
2967 (!canClobber(SuccSU, DUSU) ||
2968 (isLiveOut && !hasOnlyLiveOutUses(SuccSU)) ||
2969 (!SU->isCommutable && SuccSU->isCommutable)) &&
2970 !scheduleDAG->IsReachable(SuccSU, SU)) {
2971 DEBUG(dbgs() << " Adding a pseudo-two-addr edge from SU #"
2972 << SU->NodeNum << " to SU #" << SuccSU->NodeNum << "\n");
2973 scheduleDAG->AddPred(SU, SDep(SuccSU, SDep::Artificial));
2974 }
2975 }
2976 }
2977 }
2978 }
2979
2980 //===----------------------------------------------------------------------===//
2981 // Public Constructor Functions
2982 //===----------------------------------------------------------------------===//
2983
2984 llvm::ScheduleDAGSDNodes *
createBURRListDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level OptLevel)2985 llvm::createBURRListDAGScheduler(SelectionDAGISel *IS,
2986 CodeGenOpt::Level OptLevel) {
2987 const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
2988 const TargetInstrInfo *TII = STI.getInstrInfo();
2989 const TargetRegisterInfo *TRI = STI.getRegisterInfo();
2990
2991 BURegReductionPriorityQueue *PQ =
2992 new BURegReductionPriorityQueue(*IS->MF, false, false, TII, TRI, nullptr);
2993 ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
2994 PQ->setScheduleDAG(SD);
2995 return SD;
2996 }
2997
2998 llvm::ScheduleDAGSDNodes *
createSourceListDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level OptLevel)2999 llvm::createSourceListDAGScheduler(SelectionDAGISel *IS,
3000 CodeGenOpt::Level OptLevel) {
3001 const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3002 const TargetInstrInfo *TII = STI.getInstrInfo();
3003 const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3004
3005 SrcRegReductionPriorityQueue *PQ =
3006 new SrcRegReductionPriorityQueue(*IS->MF, false, true, TII, TRI, nullptr);
3007 ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, false, PQ, OptLevel);
3008 PQ->setScheduleDAG(SD);
3009 return SD;
3010 }
3011
3012 llvm::ScheduleDAGSDNodes *
createHybridListDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level OptLevel)3013 llvm::createHybridListDAGScheduler(SelectionDAGISel *IS,
3014 CodeGenOpt::Level OptLevel) {
3015 const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3016 const TargetInstrInfo *TII = STI.getInstrInfo();
3017 const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3018 const TargetLowering *TLI = IS->TLI;
3019
3020 HybridBURRPriorityQueue *PQ =
3021 new HybridBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI);
3022
3023 ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
3024 PQ->setScheduleDAG(SD);
3025 return SD;
3026 }
3027
3028 llvm::ScheduleDAGSDNodes *
createILPListDAGScheduler(SelectionDAGISel * IS,CodeGenOpt::Level OptLevel)3029 llvm::createILPListDAGScheduler(SelectionDAGISel *IS,
3030 CodeGenOpt::Level OptLevel) {
3031 const TargetSubtargetInfo &STI = IS->MF->getSubtarget();
3032 const TargetInstrInfo *TII = STI.getInstrInfo();
3033 const TargetRegisterInfo *TRI = STI.getRegisterInfo();
3034 const TargetLowering *TLI = IS->TLI;
3035
3036 ILPBURRPriorityQueue *PQ =
3037 new ILPBURRPriorityQueue(*IS->MF, true, false, TII, TRI, TLI);
3038 ScheduleDAGRRList *SD = new ScheduleDAGRRList(*IS->MF, true, PQ, OptLevel);
3039 PQ->setScheduleDAG(SD);
3040 return SD;
3041 }
3042