1 //===- HexagonMachineScheduler.cpp - MI Scheduler for Hexagon -------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // MachineScheduler schedules machine instructions after phi elimination. It
10 // preserves LiveIntervals so it can be invoked before register allocation.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "HexagonMachineScheduler.h"
15 #include "HexagonInstrInfo.h"
16 #include "HexagonSubtarget.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/CodeGen/DFAPacketizer.h"
19 #include "llvm/CodeGen/MachineBasicBlock.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/MachineInstr.h"
22 #include "llvm/CodeGen/MachineLoopInfo.h"
23 #include "llvm/CodeGen/RegisterClassInfo.h"
24 #include "llvm/CodeGen/RegisterPressure.h"
25 #include "llvm/CodeGen/ScheduleDAG.h"
26 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
27 #include "llvm/CodeGen/TargetInstrInfo.h"
28 #include "llvm/CodeGen/TargetOpcodes.h"
29 #include "llvm/CodeGen/TargetRegisterInfo.h"
30 #include "llvm/CodeGen/TargetSchedule.h"
31 #include "llvm/CodeGen/TargetSubtargetInfo.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include <algorithm>
37 #include <cassert>
38 #include <iomanip>
39 #include <limits>
40 #include <memory>
41 #include <sstream>
42 
43 using namespace llvm;
44 
45 #define DEBUG_TYPE "machine-scheduler"
46 
47 static cl::opt<bool> IgnoreBBRegPressure("ignore-bb-reg-pressure",
48     cl::Hidden, cl::ZeroOrMore, cl::init(false));
49 
50 static cl::opt<bool> UseNewerCandidate("use-newer-candidate",
51     cl::Hidden, cl::ZeroOrMore, cl::init(true));
52 
53 static cl::opt<unsigned> SchedDebugVerboseLevel("misched-verbose-level",
54     cl::Hidden, cl::ZeroOrMore, cl::init(1));
55 
56 // Check if the scheduler should penalize instructions that are available to
57 // early due to a zero-latency dependence.
58 static cl::opt<bool> CheckEarlyAvail("check-early-avail", cl::Hidden,
59     cl::ZeroOrMore, cl::init(true));
60 
61 // This value is used to determine if a register class is a high pressure set.
62 // We compute the maximum number of registers needed and divided by the total
63 // available. Then, we compare the result to this value.
64 static cl::opt<float> RPThreshold("hexagon-reg-pressure", cl::Hidden,
65     cl::init(0.75f), cl::desc("High register pressure threhold."));
66 
67 /// Return true if there is a dependence between SUd and SUu.
68 static bool hasDependence(const SUnit *SUd, const SUnit *SUu,
69                           const HexagonInstrInfo &QII) {
70   if (SUd->Succs.size() == 0)
71     return false;
72 
73   // Enable .cur formation.
74   if (QII.mayBeCurLoad(*SUd->getInstr()))
75     return false;
76 
77   if (QII.canExecuteInBundle(*SUd->getInstr(), *SUu->getInstr()))
78     return false;
79 
80   for (const auto &S : SUd->Succs) {
81     // Since we do not add pseudos to packets, might as well
82     // ignore order dependencies.
83     if (S.isCtrl())
84       continue;
85 
86     if (S.getSUnit() == SUu && S.getLatency() > 0)
87       return true;
88   }
89   return false;
90 }
91 
92 /// Check if scheduling of this SU is possible
93 /// in the current packet.
94 /// It is _not_ precise (statefull), it is more like
95 /// another heuristic. Many corner cases are figured
96 /// empirically.
97 bool VLIWResourceModel::isResourceAvailable(SUnit *SU, bool IsTop) {
98   if (!SU || !SU->getInstr())
99     return false;
100 
101   // First see if the pipeline could receive this instruction
102   // in the current cycle.
103   switch (SU->getInstr()->getOpcode()) {
104   default:
105     if (!ResourcesModel->canReserveResources(*SU->getInstr()))
106       return false;
107     break;
108   case TargetOpcode::EXTRACT_SUBREG:
109   case TargetOpcode::INSERT_SUBREG:
110   case TargetOpcode::SUBREG_TO_REG:
111   case TargetOpcode::REG_SEQUENCE:
112   case TargetOpcode::IMPLICIT_DEF:
113   case TargetOpcode::COPY:
114   case TargetOpcode::INLINEASM:
115   case TargetOpcode::INLINEASM_BR:
116     break;
117   }
118 
119   MachineBasicBlock *MBB = SU->getInstr()->getParent();
120   auto &QST = MBB->getParent()->getSubtarget<HexagonSubtarget>();
121   const auto &QII = *QST.getInstrInfo();
122 
123   // Now see if there are no other dependencies to instructions already
124   // in the packet.
125   if (IsTop) {
126     for (unsigned i = 0, e = Packet.size(); i != e; ++i)
127       if (hasDependence(Packet[i], SU, QII))
128         return false;
129   } else {
130     for (unsigned i = 0, e = Packet.size(); i != e; ++i)
131       if (hasDependence(SU, Packet[i], QII))
132         return false;
133   }
134   return true;
135 }
136 
137 /// Keep track of available resources.
138 bool VLIWResourceModel::reserveResources(SUnit *SU, bool IsTop) {
139   bool startNewCycle = false;
140   // Artificially reset state.
141   if (!SU) {
142     ResourcesModel->clearResources();
143     Packet.clear();
144     TotalPackets++;
145     return false;
146   }
147   // If this SU does not fit in the packet or the packet is now full
148   // start a new one.
149   if (!isResourceAvailable(SU, IsTop) ||
150       Packet.size() >= SchedModel->getIssueWidth()) {
151     ResourcesModel->clearResources();
152     Packet.clear();
153     TotalPackets++;
154     startNewCycle = true;
155   }
156 
157   switch (SU->getInstr()->getOpcode()) {
158   default:
159     ResourcesModel->reserveResources(*SU->getInstr());
160     break;
161   case TargetOpcode::EXTRACT_SUBREG:
162   case TargetOpcode::INSERT_SUBREG:
163   case TargetOpcode::SUBREG_TO_REG:
164   case TargetOpcode::REG_SEQUENCE:
165   case TargetOpcode::IMPLICIT_DEF:
166   case TargetOpcode::KILL:
167   case TargetOpcode::CFI_INSTRUCTION:
168   case TargetOpcode::EH_LABEL:
169   case TargetOpcode::COPY:
170   case TargetOpcode::INLINEASM:
171   case TargetOpcode::INLINEASM_BR:
172     break;
173   }
174   Packet.push_back(SU);
175 
176 #ifndef NDEBUG
177   LLVM_DEBUG(dbgs() << "Packet[" << TotalPackets << "]:\n");
178   for (unsigned i = 0, e = Packet.size(); i != e; ++i) {
179     LLVM_DEBUG(dbgs() << "\t[" << i << "] SU(");
180     LLVM_DEBUG(dbgs() << Packet[i]->NodeNum << ")\t");
181     LLVM_DEBUG(Packet[i]->getInstr()->dump());
182   }
183 #endif
184 
185   return startNewCycle;
186 }
187 
188 /// schedule - Called back from MachineScheduler::runOnMachineFunction
189 /// after setting up the current scheduling region. [RegionBegin, RegionEnd)
190 /// only includes instructions that have DAG nodes, not scheduling boundaries.
191 void VLIWMachineScheduler::schedule() {
192   LLVM_DEBUG(dbgs() << "********** MI Converging Scheduling VLIW "
193                     << printMBBReference(*BB) << " " << BB->getName()
194                     << " in_func " << BB->getParent()->getName()
195                     << " at loop depth " << MLI->getLoopDepth(BB) << " \n");
196 
197   buildDAGWithRegPressure();
198 
199   Topo.InitDAGTopologicalSorting();
200 
201   // Postprocess the DAG to add platform-specific artificial dependencies.
202   postprocessDAG();
203 
204   SmallVector<SUnit*, 8> TopRoots, BotRoots;
205   findRootsAndBiasEdges(TopRoots, BotRoots);
206 
207   // Initialize the strategy before modifying the DAG.
208   SchedImpl->initialize(this);
209 
210   LLVM_DEBUG(unsigned maxH = 0;
211              for (unsigned su = 0, e = SUnits.size(); su != e;
212                   ++su) if (SUnits[su].getHeight() > maxH) maxH =
213                  SUnits[su].getHeight();
214              dbgs() << "Max Height " << maxH << "\n";);
215   LLVM_DEBUG(unsigned maxD = 0;
216              for (unsigned su = 0, e = SUnits.size(); su != e;
217                   ++su) if (SUnits[su].getDepth() > maxD) maxD =
218                  SUnits[su].getDepth();
219              dbgs() << "Max Depth " << maxD << "\n";);
220   LLVM_DEBUG(dump());
221 
222   initQueues(TopRoots, BotRoots);
223 
224   bool IsTopNode = false;
225   while (true) {
226     LLVM_DEBUG(
227         dbgs() << "** VLIWMachineScheduler::schedule picking next node\n");
228     SUnit *SU = SchedImpl->pickNode(IsTopNode);
229     if (!SU) break;
230 
231     if (!checkSchedLimit())
232       break;
233 
234     scheduleMI(SU, IsTopNode);
235 
236     // Notify the scheduling strategy after updating the DAG.
237     SchedImpl->schedNode(SU, IsTopNode);
238 
239     updateQueues(SU, IsTopNode);
240   }
241   assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
242 
243   placeDebugValues();
244 
245   LLVM_DEBUG({
246     dbgs() << "*** Final schedule for "
247            << printMBBReference(*begin()->getParent()) << " ***\n";
248     dumpSchedule();
249     dbgs() << '\n';
250   });
251 }
252 
253 void ConvergingVLIWScheduler::initialize(ScheduleDAGMI *dag) {
254   DAG = static_cast<VLIWMachineScheduler*>(dag);
255   SchedModel = DAG->getSchedModel();
256 
257   Top.init(DAG, SchedModel);
258   Bot.init(DAG, SchedModel);
259 
260   // Initialize the HazardRecognizers. If itineraries don't exist, are empty, or
261   // are disabled, then these HazardRecs will be disabled.
262   const InstrItineraryData *Itin = DAG->getSchedModel()->getInstrItineraries();
263   const TargetSubtargetInfo &STI = DAG->MF.getSubtarget();
264   const TargetInstrInfo *TII = STI.getInstrInfo();
265   delete Top.HazardRec;
266   delete Bot.HazardRec;
267   Top.HazardRec = TII->CreateTargetMIHazardRecognizer(Itin, DAG);
268   Bot.HazardRec = TII->CreateTargetMIHazardRecognizer(Itin, DAG);
269 
270   delete Top.ResourceModel;
271   delete Bot.ResourceModel;
272   Top.ResourceModel = new VLIWResourceModel(STI, DAG->getSchedModel());
273   Bot.ResourceModel = new VLIWResourceModel(STI, DAG->getSchedModel());
274 
275   const std::vector<unsigned> &MaxPressure =
276     DAG->getRegPressure().MaxSetPressure;
277   HighPressureSets.assign(MaxPressure.size(), 0);
278   for (unsigned i = 0, e = MaxPressure.size(); i < e; ++i) {
279     unsigned Limit = DAG->getRegClassInfo()->getRegPressureSetLimit(i);
280     HighPressureSets[i] =
281       ((float) MaxPressure[i] > ((float) Limit * RPThreshold));
282   }
283 
284   assert((!ForceTopDown || !ForceBottomUp) &&
285          "-misched-topdown incompatible with -misched-bottomup");
286 }
287 
288 void ConvergingVLIWScheduler::releaseTopNode(SUnit *SU) {
289   if (SU->isScheduled)
290     return;
291 
292   for (const SDep &PI : SU->Preds) {
293     unsigned PredReadyCycle = PI.getSUnit()->TopReadyCycle;
294     unsigned MinLatency = PI.getLatency();
295 #ifndef NDEBUG
296     Top.MaxMinLatency = std::max(MinLatency, Top.MaxMinLatency);
297 #endif
298     if (SU->TopReadyCycle < PredReadyCycle + MinLatency)
299       SU->TopReadyCycle = PredReadyCycle + MinLatency;
300   }
301   Top.releaseNode(SU, SU->TopReadyCycle);
302 }
303 
304 void ConvergingVLIWScheduler::releaseBottomNode(SUnit *SU) {
305   if (SU->isScheduled)
306     return;
307 
308   assert(SU->getInstr() && "Scheduled SUnit must have instr");
309 
310   for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
311        I != E; ++I) {
312     unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
313     unsigned MinLatency = I->getLatency();
314 #ifndef NDEBUG
315     Bot.MaxMinLatency = std::max(MinLatency, Bot.MaxMinLatency);
316 #endif
317     if (SU->BotReadyCycle < SuccReadyCycle + MinLatency)
318       SU->BotReadyCycle = SuccReadyCycle + MinLatency;
319   }
320   Bot.releaseNode(SU, SU->BotReadyCycle);
321 }
322 
323 /// Does this SU have a hazard within the current instruction group.
324 ///
325 /// The scheduler supports two modes of hazard recognition. The first is the
326 /// ScheduleHazardRecognizer API. It is a fully general hazard recognizer that
327 /// supports highly complicated in-order reservation tables
328 /// (ScoreboardHazardRecognizer) and arbitrary target-specific logic.
329 ///
330 /// The second is a streamlined mechanism that checks for hazards based on
331 /// simple counters that the scheduler itself maintains. It explicitly checks
332 /// for instruction dispatch limitations, including the number of micro-ops that
333 /// can dispatch per cycle.
334 ///
335 /// TODO: Also check whether the SU must start a new group.
336 bool ConvergingVLIWScheduler::VLIWSchedBoundary::checkHazard(SUnit *SU) {
337   if (HazardRec->isEnabled())
338     return HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard;
339 
340   unsigned uops = SchedModel->getNumMicroOps(SU->getInstr());
341   if (IssueCount + uops > SchedModel->getIssueWidth())
342     return true;
343 
344   return false;
345 }
346 
347 void ConvergingVLIWScheduler::VLIWSchedBoundary::releaseNode(SUnit *SU,
348                                                      unsigned ReadyCycle) {
349   if (ReadyCycle < MinReadyCycle)
350     MinReadyCycle = ReadyCycle;
351 
352   // Check for interlocks first. For the purpose of other heuristics, an
353   // instruction that cannot issue appears as if it's not in the ReadyQueue.
354   if (ReadyCycle > CurrCycle || checkHazard(SU))
355 
356     Pending.push(SU);
357   else
358     Available.push(SU);
359 }
360 
361 /// Move the boundary of scheduled code by one cycle.
362 void ConvergingVLIWScheduler::VLIWSchedBoundary::bumpCycle() {
363   unsigned Width = SchedModel->getIssueWidth();
364   IssueCount = (IssueCount <= Width) ? 0 : IssueCount - Width;
365 
366   assert(MinReadyCycle < std::numeric_limits<unsigned>::max() &&
367          "MinReadyCycle uninitialized");
368   unsigned NextCycle = std::max(CurrCycle + 1, MinReadyCycle);
369 
370   if (!HazardRec->isEnabled()) {
371     // Bypass HazardRec virtual calls.
372     CurrCycle = NextCycle;
373   } else {
374     // Bypass getHazardType calls in case of long latency.
375     for (; CurrCycle != NextCycle; ++CurrCycle) {
376       if (isTop())
377         HazardRec->AdvanceCycle();
378       else
379         HazardRec->RecedeCycle();
380     }
381   }
382   CheckPending = true;
383 
384   LLVM_DEBUG(dbgs() << "*** Next cycle " << Available.getName() << " cycle "
385                     << CurrCycle << '\n');
386 }
387 
388 /// Move the boundary of scheduled code by one SUnit.
389 void ConvergingVLIWScheduler::VLIWSchedBoundary::bumpNode(SUnit *SU) {
390   bool startNewCycle = false;
391 
392   // Update the reservation table.
393   if (HazardRec->isEnabled()) {
394     if (!isTop() && SU->isCall) {
395       // Calls are scheduled with their preceding instructions. For bottom-up
396       // scheduling, clear the pipeline state before emitting.
397       HazardRec->Reset();
398     }
399     HazardRec->EmitInstruction(SU);
400   }
401 
402   // Update DFA model.
403   startNewCycle = ResourceModel->reserveResources(SU, isTop());
404 
405   // Check the instruction group dispatch limit.
406   // TODO: Check if this SU must end a dispatch group.
407   IssueCount += SchedModel->getNumMicroOps(SU->getInstr());
408   if (startNewCycle) {
409     LLVM_DEBUG(dbgs() << "*** Max instrs at cycle " << CurrCycle << '\n');
410     bumpCycle();
411   }
412   else
413     LLVM_DEBUG(dbgs() << "*** IssueCount " << IssueCount << " at cycle "
414                       << CurrCycle << '\n');
415 }
416 
417 /// Release pending ready nodes in to the available queue. This makes them
418 /// visible to heuristics.
419 void ConvergingVLIWScheduler::VLIWSchedBoundary::releasePending() {
420   // If the available queue is empty, it is safe to reset MinReadyCycle.
421   if (Available.empty())
422     MinReadyCycle = std::numeric_limits<unsigned>::max();
423 
424   // Check to see if any of the pending instructions are ready to issue.  If
425   // so, add them to the available queue.
426   for (unsigned i = 0, e = Pending.size(); i != e; ++i) {
427     SUnit *SU = *(Pending.begin()+i);
428     unsigned ReadyCycle = isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
429 
430     if (ReadyCycle < MinReadyCycle)
431       MinReadyCycle = ReadyCycle;
432 
433     if (ReadyCycle > CurrCycle)
434       continue;
435 
436     if (checkHazard(SU))
437       continue;
438 
439     Available.push(SU);
440     Pending.remove(Pending.begin()+i);
441     --i; --e;
442   }
443   CheckPending = false;
444 }
445 
446 /// Remove SU from the ready set for this boundary.
447 void ConvergingVLIWScheduler::VLIWSchedBoundary::removeReady(SUnit *SU) {
448   if (Available.isInQueue(SU))
449     Available.remove(Available.find(SU));
450   else {
451     assert(Pending.isInQueue(SU) && "bad ready count");
452     Pending.remove(Pending.find(SU));
453   }
454 }
455 
456 /// If this queue only has one ready candidate, return it. As a side effect,
457 /// advance the cycle until at least one node is ready. If multiple instructions
458 /// are ready, return NULL.
459 SUnit *ConvergingVLIWScheduler::VLIWSchedBoundary::pickOnlyChoice() {
460   if (CheckPending)
461     releasePending();
462 
463   auto AdvanceCycle = [this]() {
464     if (Available.empty())
465       return true;
466     if (Available.size() == 1 && Pending.size() > 0)
467       return !ResourceModel->isResourceAvailable(*Available.begin(), isTop()) ||
468         getWeakLeft(*Available.begin(), isTop()) != 0;
469     return false;
470   };
471   for (unsigned i = 0; AdvanceCycle(); ++i) {
472     assert(i <= (HazardRec->getMaxLookAhead() + MaxMinLatency) &&
473            "permanent hazard"); (void)i;
474     ResourceModel->reserveResources(nullptr, isTop());
475     bumpCycle();
476     releasePending();
477   }
478   if (Available.size() == 1)
479     return *Available.begin();
480   return nullptr;
481 }
482 
483 #ifndef NDEBUG
484 void ConvergingVLIWScheduler::traceCandidate(const char *Label,
485       const ReadyQueue &Q, SUnit *SU, int Cost, PressureChange P) {
486   dbgs() << Label << " " << Q.getName() << " ";
487   if (P.isValid())
488     dbgs() << DAG->TRI->getRegPressureSetName(P.getPSet()) << ":"
489            << P.getUnitInc() << " ";
490   else
491     dbgs() << "     ";
492   dbgs() << "cost(" << Cost << ")\t";
493   DAG->dumpNode(*SU);
494 }
495 
496 // Very detailed queue dump, to be used with higher verbosity levels.
497 void ConvergingVLIWScheduler::readyQueueVerboseDump(
498       const RegPressureTracker &RPTracker, SchedCandidate &Candidate,
499       ReadyQueue &Q) {
500   RegPressureTracker &TempTracker = const_cast<RegPressureTracker &>(RPTracker);
501 
502   dbgs() << ">>> " << Q.getName() << "\n";
503   for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
504     RegPressureDelta RPDelta;
505     TempTracker.getMaxPressureDelta((*I)->getInstr(), RPDelta,
506                                     DAG->getRegionCriticalPSets(),
507                                     DAG->getRegPressure().MaxSetPressure);
508     std::stringstream dbgstr;
509     dbgstr << "SU(" << std::setw(3) << (*I)->NodeNum << ")";
510     dbgs() << dbgstr.str();
511     SchedulingCost(Q, *I, Candidate, RPDelta, true);
512     dbgs() << "\t";
513     (*I)->getInstr()->dump();
514   }
515   dbgs() << "\n";
516 }
517 #endif
518 
519 /// isSingleUnscheduledPred - If SU2 is the only unscheduled predecessor
520 /// of SU, return true (we may have duplicates)
521 static inline bool isSingleUnscheduledPred(SUnit *SU, SUnit *SU2) {
522   if (SU->NumPredsLeft == 0)
523     return false;
524 
525   for (auto &Pred : SU->Preds) {
526     // We found an available, but not scheduled, predecessor.
527     if (!Pred.getSUnit()->isScheduled && (Pred.getSUnit() != SU2))
528       return false;
529   }
530 
531   return true;
532 }
533 
534 /// isSingleUnscheduledSucc - If SU2 is the only unscheduled successor
535 /// of SU, return true (we may have duplicates)
536 static inline bool isSingleUnscheduledSucc(SUnit *SU, SUnit *SU2) {
537   if (SU->NumSuccsLeft == 0)
538     return false;
539 
540   for (auto &Succ : SU->Succs) {
541     // We found an available, but not scheduled, successor.
542     if (!Succ.getSUnit()->isScheduled && (Succ.getSUnit() != SU2))
543       return false;
544   }
545   return true;
546 }
547 
548 /// Check if the instruction changes the register pressure of a register in the
549 /// high pressure set. The function returns a negative value if the pressure
550 /// decreases and a positive value is the pressure increases. If the instruction
551 /// doesn't use a high pressure register or doesn't change the register
552 /// pressure, then return 0.
553 int ConvergingVLIWScheduler::pressureChange(const SUnit *SU, bool isBotUp) {
554   PressureDiff &PD = DAG->getPressureDiff(SU);
555   for (auto &P : PD) {
556     if (!P.isValid())
557       continue;
558     // The pressure differences are computed bottom-up, so the comparision for
559     // an increase is positive in the bottom direction, but negative in the
560     //  top-down direction.
561     if (HighPressureSets[P.getPSet()])
562       return (isBotUp ? P.getUnitInc() : -P.getUnitInc());
563   }
564   return 0;
565 }
566 
567 // Constants used to denote relative importance of
568 // heuristic components for cost computation.
569 static const unsigned PriorityOne = 200;
570 static const unsigned PriorityTwo = 50;
571 static const unsigned PriorityThree = 75;
572 static const unsigned ScaleTwo = 10;
573 
574 /// Single point to compute overall scheduling cost.
575 /// TODO: More heuristics will be used soon.
576 int ConvergingVLIWScheduler::SchedulingCost(ReadyQueue &Q, SUnit *SU,
577                                             SchedCandidate &Candidate,
578                                             RegPressureDelta &Delta,
579                                             bool verbose) {
580   // Initial trivial priority.
581   int ResCount = 1;
582 
583   // Do not waste time on a node that is already scheduled.
584   if (!SU || SU->isScheduled)
585     return ResCount;
586 
587   LLVM_DEBUG(if (verbose) dbgs()
588              << ((Q.getID() == TopQID) ? "(top|" : "(bot|"));
589   // Forced priority is high.
590   if (SU->isScheduleHigh) {
591     ResCount += PriorityOne;
592     LLVM_DEBUG(dbgs() << "H|");
593   }
594 
595   unsigned IsAvailableAmt = 0;
596   // Critical path first.
597   if (Q.getID() == TopQID) {
598     if (Top.isLatencyBound(SU)) {
599       LLVM_DEBUG(if (verbose) dbgs() << "LB|");
600       ResCount += (SU->getHeight() * ScaleTwo);
601     }
602 
603     LLVM_DEBUG(if (verbose) {
604       std::stringstream dbgstr;
605       dbgstr << "h" << std::setw(3) << SU->getHeight() << "|";
606       dbgs() << dbgstr.str();
607     });
608 
609     // If resources are available for it, multiply the
610     // chance of scheduling.
611     if (Top.ResourceModel->isResourceAvailable(SU, true)) {
612       IsAvailableAmt = (PriorityTwo + PriorityThree);
613       ResCount += IsAvailableAmt;
614       LLVM_DEBUG(if (verbose) dbgs() << "A|");
615     } else
616       LLVM_DEBUG(if (verbose) dbgs() << " |");
617   } else {
618     if (Bot.isLatencyBound(SU)) {
619       LLVM_DEBUG(if (verbose) dbgs() << "LB|");
620       ResCount += (SU->getDepth() * ScaleTwo);
621     }
622 
623     LLVM_DEBUG(if (verbose) {
624       std::stringstream dbgstr;
625       dbgstr << "d" << std::setw(3) << SU->getDepth() << "|";
626       dbgs() << dbgstr.str();
627     });
628 
629     // If resources are available for it, multiply the
630     // chance of scheduling.
631     if (Bot.ResourceModel->isResourceAvailable(SU, false)) {
632       IsAvailableAmt = (PriorityTwo + PriorityThree);
633       ResCount += IsAvailableAmt;
634       LLVM_DEBUG(if (verbose) dbgs() << "A|");
635     } else
636       LLVM_DEBUG(if (verbose) dbgs() << " |");
637   }
638 
639   unsigned NumNodesBlocking = 0;
640   if (Q.getID() == TopQID) {
641     // How many SUs does it block from scheduling?
642     // Look at all of the successors of this node.
643     // Count the number of nodes that
644     // this node is the sole unscheduled node for.
645     if (Top.isLatencyBound(SU))
646       for (const SDep &SI : SU->Succs)
647         if (isSingleUnscheduledPred(SI.getSUnit(), SU))
648           ++NumNodesBlocking;
649   } else {
650     // How many unscheduled predecessors block this node?
651     if (Bot.isLatencyBound(SU))
652       for (const SDep &PI : SU->Preds)
653         if (isSingleUnscheduledSucc(PI.getSUnit(), SU))
654           ++NumNodesBlocking;
655   }
656   ResCount += (NumNodesBlocking * ScaleTwo);
657 
658   LLVM_DEBUG(if (verbose) {
659     std::stringstream dbgstr;
660     dbgstr << "blk " << std::setw(2) << NumNodesBlocking << ")|";
661     dbgs() << dbgstr.str();
662   });
663 
664   // Factor in reg pressure as a heuristic.
665   if (!IgnoreBBRegPressure) {
666     // Decrease priority by the amount that register pressure exceeds the limit.
667     ResCount -= (Delta.Excess.getUnitInc()*PriorityOne);
668     // Decrease priority if register pressure exceeds the limit.
669     ResCount -= (Delta.CriticalMax.getUnitInc()*PriorityOne);
670     // Decrease priority slightly if register pressure would increase over the
671     // current maximum.
672     ResCount -= (Delta.CurrentMax.getUnitInc()*PriorityTwo);
673     // If there are register pressure issues, then we remove the value added for
674     // the instruction being available. The rationale is that we really don't
675     // want to schedule an instruction that causes a spill.
676     if (IsAvailableAmt && pressureChange(SU, Q.getID() != TopQID) > 0 &&
677         (Delta.Excess.getUnitInc() || Delta.CriticalMax.getUnitInc() ||
678          Delta.CurrentMax.getUnitInc()))
679       ResCount -= IsAvailableAmt;
680     LLVM_DEBUG(if (verbose) {
681       dbgs() << "RP " << Delta.Excess.getUnitInc() << "/"
682              << Delta.CriticalMax.getUnitInc() << "/"
683              << Delta.CurrentMax.getUnitInc() << ")|";
684     });
685   }
686 
687   // Give a little extra priority to a .cur instruction if there is a resource
688   // available for it.
689   auto &QST = DAG->MF.getSubtarget<HexagonSubtarget>();
690   auto &QII = *QST.getInstrInfo();
691   if (SU->isInstr() && QII.mayBeCurLoad(*SU->getInstr())) {
692     if (Q.getID() == TopQID &&
693         Top.ResourceModel->isResourceAvailable(SU, true)) {
694       ResCount += PriorityTwo;
695       LLVM_DEBUG(if (verbose) dbgs() << "C|");
696     } else if (Q.getID() == BotQID &&
697                Bot.ResourceModel->isResourceAvailable(SU, false)) {
698       ResCount += PriorityTwo;
699       LLVM_DEBUG(if (verbose) dbgs() << "C|");
700     }
701   }
702 
703   // Give preference to a zero latency instruction if the dependent
704   // instruction is in the current packet.
705   if (Q.getID() == TopQID && getWeakLeft(SU, true) == 0) {
706     for (const SDep &PI : SU->Preds) {
707       if (!PI.getSUnit()->getInstr()->isPseudo() && PI.isAssignedRegDep() &&
708           PI.getLatency() == 0 &&
709           Top.ResourceModel->isInPacket(PI.getSUnit())) {
710         ResCount += PriorityThree;
711         LLVM_DEBUG(if (verbose) dbgs() << "Z|");
712       }
713     }
714   } else if (Q.getID() == BotQID && getWeakLeft(SU, false) == 0) {
715     for (const SDep &SI : SU->Succs) {
716       if (!SI.getSUnit()->getInstr()->isPseudo() && SI.isAssignedRegDep() &&
717           SI.getLatency() == 0 &&
718           Bot.ResourceModel->isInPacket(SI.getSUnit())) {
719         ResCount += PriorityThree;
720         LLVM_DEBUG(if (verbose) dbgs() << "Z|");
721       }
722     }
723   }
724 
725   // If the instruction has a non-zero latency dependence with an instruction in
726   // the current packet, then it should not be scheduled yet. The case occurs
727   // when the dependent instruction is scheduled in a new packet, so the
728   // scheduler updates the current cycle and pending instructions become
729   // available.
730   if (CheckEarlyAvail) {
731     if (Q.getID() == TopQID) {
732       for (const auto &PI : SU->Preds) {
733         if (PI.getLatency() > 0 &&
734             Top.ResourceModel->isInPacket(PI.getSUnit())) {
735           ResCount -= PriorityOne;
736           LLVM_DEBUG(if (verbose) dbgs() << "D|");
737         }
738       }
739     } else {
740       for (const auto &SI : SU->Succs) {
741         if (SI.getLatency() > 0 &&
742             Bot.ResourceModel->isInPacket(SI.getSUnit())) {
743           ResCount -= PriorityOne;
744           LLVM_DEBUG(if (verbose) dbgs() << "D|");
745         }
746       }
747     }
748   }
749 
750   LLVM_DEBUG(if (verbose) {
751     std::stringstream dbgstr;
752     dbgstr << "Total " << std::setw(4) << ResCount << ")";
753     dbgs() << dbgstr.str();
754   });
755 
756   return ResCount;
757 }
758 
759 /// Pick the best candidate from the top queue.
760 ///
761 /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during
762 /// DAG building. To adjust for the current scheduling location we need to
763 /// maintain the number of vreg uses remaining to be top-scheduled.
764 ConvergingVLIWScheduler::CandResult ConvergingVLIWScheduler::
765 pickNodeFromQueue(VLIWSchedBoundary &Zone, const RegPressureTracker &RPTracker,
766                   SchedCandidate &Candidate) {
767   ReadyQueue &Q = Zone.Available;
768   LLVM_DEBUG(if (SchedDebugVerboseLevel > 1)
769                  readyQueueVerboseDump(RPTracker, Candidate, Q);
770              else Q.dump(););
771 
772   // getMaxPressureDelta temporarily modifies the tracker.
773   RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
774 
775   // BestSU remains NULL if no top candidates beat the best existing candidate.
776   CandResult FoundCandidate = NoCand;
777   for (ReadyQueue::iterator I = Q.begin(), E = Q.end(); I != E; ++I) {
778     RegPressureDelta RPDelta;
779     TempTracker.getMaxPressureDelta((*I)->getInstr(), RPDelta,
780                                     DAG->getRegionCriticalPSets(),
781                                     DAG->getRegPressure().MaxSetPressure);
782 
783     int CurrentCost = SchedulingCost(Q, *I, Candidate, RPDelta, false);
784 
785     // Initialize the candidate if needed.
786     if (!Candidate.SU) {
787       LLVM_DEBUG(traceCandidate("DCAND", Q, *I, CurrentCost));
788       Candidate.SU = *I;
789       Candidate.RPDelta = RPDelta;
790       Candidate.SCost = CurrentCost;
791       FoundCandidate = NodeOrder;
792       continue;
793     }
794 
795     // Choose node order for negative cost candidates. There is no good
796     // candidate in this case.
797     if (CurrentCost < 0 && Candidate.SCost < 0) {
798       if ((Q.getID() == TopQID && (*I)->NodeNum < Candidate.SU->NodeNum)
799           || (Q.getID() == BotQID && (*I)->NodeNum > Candidate.SU->NodeNum)) {
800         LLVM_DEBUG(traceCandidate("NCAND", Q, *I, CurrentCost));
801         Candidate.SU = *I;
802         Candidate.RPDelta = RPDelta;
803         Candidate.SCost = CurrentCost;
804         FoundCandidate = NodeOrder;
805       }
806       continue;
807     }
808 
809     // Best cost.
810     if (CurrentCost > Candidate.SCost) {
811       LLVM_DEBUG(traceCandidate("CCAND", Q, *I, CurrentCost));
812       Candidate.SU = *I;
813       Candidate.RPDelta = RPDelta;
814       Candidate.SCost = CurrentCost;
815       FoundCandidate = BestCost;
816       continue;
817     }
818 
819     // Choose an instruction that does not depend on an artificial edge.
820     unsigned CurrWeak = getWeakLeft(*I, (Q.getID() == TopQID));
821     unsigned CandWeak = getWeakLeft(Candidate.SU, (Q.getID() == TopQID));
822     if (CurrWeak != CandWeak) {
823       if (CurrWeak < CandWeak) {
824         LLVM_DEBUG(traceCandidate("WCAND", Q, *I, CurrentCost));
825         Candidate.SU = *I;
826         Candidate.RPDelta = RPDelta;
827         Candidate.SCost = CurrentCost;
828         FoundCandidate = Weak;
829       }
830       continue;
831     }
832 
833     if (CurrentCost == Candidate.SCost && Zone.isLatencyBound(*I)) {
834       unsigned CurrSize, CandSize;
835       if (Q.getID() == TopQID) {
836         CurrSize = (*I)->Succs.size();
837         CandSize = Candidate.SU->Succs.size();
838       } else {
839         CurrSize = (*I)->Preds.size();
840         CandSize = Candidate.SU->Preds.size();
841       }
842       if (CurrSize > CandSize) {
843         LLVM_DEBUG(traceCandidate("SPCAND", Q, *I, CurrentCost));
844         Candidate.SU = *I;
845         Candidate.RPDelta = RPDelta;
846         Candidate.SCost = CurrentCost;
847         FoundCandidate = BestCost;
848       }
849       // Keep the old candidate if it's a better candidate. That is, don't use
850       // the subsequent tie breaker.
851       if (CurrSize != CandSize)
852         continue;
853     }
854 
855     // Tie breaker.
856     // To avoid scheduling indeterminism, we need a tie breaker
857     // for the case when cost is identical for two nodes.
858     if (UseNewerCandidate && CurrentCost == Candidate.SCost) {
859       if ((Q.getID() == TopQID && (*I)->NodeNum < Candidate.SU->NodeNum)
860           || (Q.getID() == BotQID && (*I)->NodeNum > Candidate.SU->NodeNum)) {
861         LLVM_DEBUG(traceCandidate("TCAND", Q, *I, CurrentCost));
862         Candidate.SU = *I;
863         Candidate.RPDelta = RPDelta;
864         Candidate.SCost = CurrentCost;
865         FoundCandidate = NodeOrder;
866         continue;
867       }
868     }
869 
870     // Fall through to original instruction order.
871     // Only consider node order if Candidate was chosen from this Q.
872     if (FoundCandidate == NoCand)
873       continue;
874   }
875   return FoundCandidate;
876 }
877 
878 /// Pick the best candidate node from either the top or bottom queue.
879 SUnit *ConvergingVLIWScheduler::pickNodeBidrectional(bool &IsTopNode) {
880   // Schedule as far as possible in the direction of no choice. This is most
881   // efficient, but also provides the best heuristics for CriticalPSets.
882   if (SUnit *SU = Bot.pickOnlyChoice()) {
883     LLVM_DEBUG(dbgs() << "Picked only Bottom\n");
884     IsTopNode = false;
885     return SU;
886   }
887   if (SUnit *SU = Top.pickOnlyChoice()) {
888     LLVM_DEBUG(dbgs() << "Picked only Top\n");
889     IsTopNode = true;
890     return SU;
891   }
892   SchedCandidate BotCand;
893   // Prefer bottom scheduling when heuristics are silent.
894   CandResult BotResult = pickNodeFromQueue(Bot,
895                                            DAG->getBotRPTracker(), BotCand);
896   assert(BotResult != NoCand && "failed to find the first candidate");
897 
898   // If either Q has a single candidate that provides the least increase in
899   // Excess pressure, we can immediately schedule from that Q.
900   //
901   // RegionCriticalPSets summarizes the pressure within the scheduled region and
902   // affects picking from either Q. If scheduling in one direction must
903   // increase pressure for one of the excess PSets, then schedule in that
904   // direction first to provide more freedom in the other direction.
905   if (BotResult == SingleExcess || BotResult == SingleCritical) {
906     LLVM_DEBUG(dbgs() << "Prefered Bottom Node\n");
907     IsTopNode = false;
908     return BotCand.SU;
909   }
910   // Check if the top Q has a better candidate.
911   SchedCandidate TopCand;
912   CandResult TopResult = pickNodeFromQueue(Top,
913                                            DAG->getTopRPTracker(), TopCand);
914   assert(TopResult != NoCand && "failed to find the first candidate");
915 
916   if (TopResult == SingleExcess || TopResult == SingleCritical) {
917     LLVM_DEBUG(dbgs() << "Prefered Top Node\n");
918     IsTopNode = true;
919     return TopCand.SU;
920   }
921   // If either Q has a single candidate that minimizes pressure above the
922   // original region's pressure pick it.
923   if (BotResult == SingleMax) {
924     LLVM_DEBUG(dbgs() << "Prefered Bottom Node SingleMax\n");
925     IsTopNode = false;
926     return BotCand.SU;
927   }
928   if (TopResult == SingleMax) {
929     LLVM_DEBUG(dbgs() << "Prefered Top Node SingleMax\n");
930     IsTopNode = true;
931     return TopCand.SU;
932   }
933   if (TopCand.SCost > BotCand.SCost) {
934     LLVM_DEBUG(dbgs() << "Prefered Top Node Cost\n");
935     IsTopNode = true;
936     return TopCand.SU;
937   }
938   // Otherwise prefer the bottom candidate in node order.
939   LLVM_DEBUG(dbgs() << "Prefered Bottom in Node order\n");
940   IsTopNode = false;
941   return BotCand.SU;
942 }
943 
944 /// Pick the best node to balance the schedule. Implements MachineSchedStrategy.
945 SUnit *ConvergingVLIWScheduler::pickNode(bool &IsTopNode) {
946   if (DAG->top() == DAG->bottom()) {
947     assert(Top.Available.empty() && Top.Pending.empty() &&
948            Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
949     return nullptr;
950   }
951   SUnit *SU;
952   if (ForceTopDown) {
953     SU = Top.pickOnlyChoice();
954     if (!SU) {
955       SchedCandidate TopCand;
956       CandResult TopResult =
957         pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand);
958       assert(TopResult != NoCand && "failed to find the first candidate");
959       (void)TopResult;
960       SU = TopCand.SU;
961     }
962     IsTopNode = true;
963   } else if (ForceBottomUp) {
964     SU = Bot.pickOnlyChoice();
965     if (!SU) {
966       SchedCandidate BotCand;
967       CandResult BotResult =
968         pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand);
969       assert(BotResult != NoCand && "failed to find the first candidate");
970       (void)BotResult;
971       SU = BotCand.SU;
972     }
973     IsTopNode = false;
974   } else {
975     SU = pickNodeBidrectional(IsTopNode);
976   }
977   if (SU->isTopReady())
978     Top.removeReady(SU);
979   if (SU->isBottomReady())
980     Bot.removeReady(SU);
981 
982   LLVM_DEBUG(dbgs() << "*** " << (IsTopNode ? "Top" : "Bottom")
983                     << " Scheduling instruction in cycle "
984                     << (IsTopNode ? Top.CurrCycle : Bot.CurrCycle) << " ("
985                     << reportPackets() << ")\n";
986              DAG->dumpNode(*SU));
987   return SU;
988 }
989 
990 /// Update the scheduler's state after scheduling a node. This is the same node
991 /// that was just returned by pickNode(). However, VLIWMachineScheduler needs
992 /// to update it's state based on the current cycle before MachineSchedStrategy
993 /// does.
994 void ConvergingVLIWScheduler::schedNode(SUnit *SU, bool IsTopNode) {
995   if (IsTopNode) {
996     Top.bumpNode(SU);
997     SU->TopReadyCycle = Top.CurrCycle;
998   } else {
999     Bot.bumpNode(SU);
1000     SU->BotReadyCycle = Bot.CurrCycle;
1001   }
1002 }
1003