xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp (revision 62cfcf62f627e5093fb37026a6d8c98e4d2ef04c)
1 //===-- GCNSchedStrategy.cpp - GCN Scheduler Strategy ---------------------===//
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 /// \file
10 /// This contains a MachineSchedStrategy implementation for maximizing wave
11 /// occupancy on GCN hardware.
12 //===----------------------------------------------------------------------===//
13 
14 #include "GCNSchedStrategy.h"
15 #include "AMDGPUSubtarget.h"
16 #include "SIInstrInfo.h"
17 #include "SIMachineFunctionInfo.h"
18 #include "SIRegisterInfo.h"
19 #include "Utils/AMDGPUBaseInfo.h"
20 #include "llvm/CodeGen/RegisterClassInfo.h"
21 #include "llvm/Support/MathExtras.h"
22 
23 #define DEBUG_TYPE "machine-scheduler"
24 
25 using namespace llvm;
26 
27 GCNMaxOccupancySchedStrategy::GCNMaxOccupancySchedStrategy(
28     const MachineSchedContext *C) :
29     GenericScheduler(C), TargetOccupancy(0), MF(nullptr) { }
30 
31 void GCNMaxOccupancySchedStrategy::initialize(ScheduleDAGMI *DAG) {
32   GenericScheduler::initialize(DAG);
33 
34   const SIRegisterInfo *SRI = static_cast<const SIRegisterInfo*>(TRI);
35 
36   MF = &DAG->MF;
37 
38   const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>();
39 
40   // FIXME: This is also necessary, because some passes that run after
41   // scheduling and before regalloc increase register pressure.
42   const int ErrorMargin = 3;
43 
44   SGPRExcessLimit = Context->RegClassInfo
45     ->getNumAllocatableRegs(&AMDGPU::SGPR_32RegClass) - ErrorMargin;
46   VGPRExcessLimit = Context->RegClassInfo
47     ->getNumAllocatableRegs(&AMDGPU::VGPR_32RegClass) - ErrorMargin;
48   if (TargetOccupancy) {
49     SGPRCriticalLimit = ST.getMaxNumSGPRs(TargetOccupancy, true);
50     VGPRCriticalLimit = ST.getMaxNumVGPRs(TargetOccupancy);
51   } else {
52     SGPRCriticalLimit = SRI->getRegPressureSetLimit(DAG->MF,
53                                                     SRI->getSGPRPressureSet());
54     VGPRCriticalLimit = SRI->getRegPressureSetLimit(DAG->MF,
55                                                     SRI->getVGPRPressureSet());
56   }
57 
58   SGPRCriticalLimit -= ErrorMargin;
59   VGPRCriticalLimit -= ErrorMargin;
60 }
61 
62 void GCNMaxOccupancySchedStrategy::initCandidate(SchedCandidate &Cand, SUnit *SU,
63                                      bool AtTop, const RegPressureTracker &RPTracker,
64                                      const SIRegisterInfo *SRI,
65                                      unsigned SGPRPressure,
66                                      unsigned VGPRPressure) {
67 
68   Cand.SU = SU;
69   Cand.AtTop = AtTop;
70 
71   // getDownwardPressure() and getUpwardPressure() make temporary changes to
72   // the tracker, so we need to pass those function a non-const copy.
73   RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
74 
75   Pressure.clear();
76   MaxPressure.clear();
77 
78   if (AtTop)
79     TempTracker.getDownwardPressure(SU->getInstr(), Pressure, MaxPressure);
80   else {
81     // FIXME: I think for bottom up scheduling, the register pressure is cached
82     // and can be retrieved by DAG->getPressureDif(SU).
83     TempTracker.getUpwardPressure(SU->getInstr(), Pressure, MaxPressure);
84   }
85 
86   unsigned NewSGPRPressure = Pressure[SRI->getSGPRPressureSet()];
87   unsigned NewVGPRPressure = Pressure[SRI->getVGPRPressureSet()];
88 
89   // If two instructions increase the pressure of different register sets
90   // by the same amount, the generic scheduler will prefer to schedule the
91   // instruction that increases the set with the least amount of registers,
92   // which in our case would be SGPRs.  This is rarely what we want, so
93   // when we report excess/critical register pressure, we do it either
94   // only for VGPRs or only for SGPRs.
95 
96   // FIXME: Better heuristics to determine whether to prefer SGPRs or VGPRs.
97   const unsigned MaxVGPRPressureInc = 16;
98   bool ShouldTrackVGPRs = VGPRPressure + MaxVGPRPressureInc >= VGPRExcessLimit;
99   bool ShouldTrackSGPRs = !ShouldTrackVGPRs && SGPRPressure >= SGPRExcessLimit;
100 
101 
102   // FIXME: We have to enter REG-EXCESS before we reach the actual threshold
103   // to increase the likelihood we don't go over the limits.  We should improve
104   // the analysis to look through dependencies to find the path with the least
105   // register pressure.
106 
107   // We only need to update the RPDelta for instructions that increase register
108   // pressure. Instructions that decrease or keep reg pressure the same will be
109   // marked as RegExcess in tryCandidate() when they are compared with
110   // instructions that increase the register pressure.
111   if (ShouldTrackVGPRs && NewVGPRPressure >= VGPRExcessLimit) {
112     Cand.RPDelta.Excess = PressureChange(SRI->getVGPRPressureSet());
113     Cand.RPDelta.Excess.setUnitInc(NewVGPRPressure - VGPRExcessLimit);
114   }
115 
116   if (ShouldTrackSGPRs && NewSGPRPressure >= SGPRExcessLimit) {
117     Cand.RPDelta.Excess = PressureChange(SRI->getSGPRPressureSet());
118     Cand.RPDelta.Excess.setUnitInc(NewSGPRPressure - SGPRExcessLimit);
119   }
120 
121   // Register pressure is considered 'CRITICAL' if it is approaching a value
122   // that would reduce the wave occupancy for the execution unit.  When
123   // register pressure is 'CRITICAL', increading SGPR and VGPR pressure both
124   // has the same cost, so we don't need to prefer one over the other.
125 
126   int SGPRDelta = NewSGPRPressure - SGPRCriticalLimit;
127   int VGPRDelta = NewVGPRPressure - VGPRCriticalLimit;
128 
129   if (SGPRDelta >= 0 || VGPRDelta >= 0) {
130     if (SGPRDelta > VGPRDelta) {
131       Cand.RPDelta.CriticalMax = PressureChange(SRI->getSGPRPressureSet());
132       Cand.RPDelta.CriticalMax.setUnitInc(SGPRDelta);
133     } else {
134       Cand.RPDelta.CriticalMax = PressureChange(SRI->getVGPRPressureSet());
135       Cand.RPDelta.CriticalMax.setUnitInc(VGPRDelta);
136     }
137   }
138 }
139 
140 // This function is mostly cut and pasted from
141 // GenericScheduler::pickNodeFromQueue()
142 void GCNMaxOccupancySchedStrategy::pickNodeFromQueue(SchedBoundary &Zone,
143                                          const CandPolicy &ZonePolicy,
144                                          const RegPressureTracker &RPTracker,
145                                          SchedCandidate &Cand) {
146   const SIRegisterInfo *SRI = static_cast<const SIRegisterInfo*>(TRI);
147   ArrayRef<unsigned> Pressure = RPTracker.getRegSetPressureAtPos();
148   unsigned SGPRPressure = Pressure[SRI->getSGPRPressureSet()];
149   unsigned VGPRPressure = Pressure[SRI->getVGPRPressureSet()];
150   ReadyQueue &Q = Zone.Available;
151   for (SUnit *SU : Q) {
152 
153     SchedCandidate TryCand(ZonePolicy);
154     initCandidate(TryCand, SU, Zone.isTop(), RPTracker, SRI,
155                   SGPRPressure, VGPRPressure);
156     // Pass SchedBoundary only when comparing nodes from the same boundary.
157     SchedBoundary *ZoneArg = Cand.AtTop == TryCand.AtTop ? &Zone : nullptr;
158     GenericScheduler::tryCandidate(Cand, TryCand, ZoneArg);
159     if (TryCand.Reason != NoCand) {
160       // Initialize resource delta if needed in case future heuristics query it.
161       if (TryCand.ResDelta == SchedResourceDelta())
162         TryCand.initResourceDelta(Zone.DAG, SchedModel);
163       Cand.setBest(TryCand);
164       LLVM_DEBUG(traceCandidate(Cand));
165     }
166   }
167 }
168 
169 // This function is mostly cut and pasted from
170 // GenericScheduler::pickNodeBidirectional()
171 SUnit *GCNMaxOccupancySchedStrategy::pickNodeBidirectional(bool &IsTopNode) {
172   // Schedule as far as possible in the direction of no choice. This is most
173   // efficient, but also provides the best heuristics for CriticalPSets.
174   if (SUnit *SU = Bot.pickOnlyChoice()) {
175     IsTopNode = false;
176     return SU;
177   }
178   if (SUnit *SU = Top.pickOnlyChoice()) {
179     IsTopNode = true;
180     return SU;
181   }
182   // Set the bottom-up policy based on the state of the current bottom zone and
183   // the instructions outside the zone, including the top zone.
184   CandPolicy BotPolicy;
185   setPolicy(BotPolicy, /*IsPostRA=*/false, Bot, &Top);
186   // Set the top-down policy based on the state of the current top zone and
187   // the instructions outside the zone, including the bottom zone.
188   CandPolicy TopPolicy;
189   setPolicy(TopPolicy, /*IsPostRA=*/false, Top, &Bot);
190 
191   // See if BotCand is still valid (because we previously scheduled from Top).
192   LLVM_DEBUG(dbgs() << "Picking from Bot:\n");
193   if (!BotCand.isValid() || BotCand.SU->isScheduled ||
194       BotCand.Policy != BotPolicy) {
195     BotCand.reset(CandPolicy());
196     pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), BotCand);
197     assert(BotCand.Reason != NoCand && "failed to find the first candidate");
198   } else {
199     LLVM_DEBUG(traceCandidate(BotCand));
200 #ifndef NDEBUG
201     if (VerifyScheduling) {
202       SchedCandidate TCand;
203       TCand.reset(CandPolicy());
204       pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), TCand);
205       assert(TCand.SU == BotCand.SU &&
206              "Last pick result should correspond to re-picking right now");
207     }
208 #endif
209   }
210 
211   // Check if the top Q has a better candidate.
212   LLVM_DEBUG(dbgs() << "Picking from Top:\n");
213   if (!TopCand.isValid() || TopCand.SU->isScheduled ||
214       TopCand.Policy != TopPolicy) {
215     TopCand.reset(CandPolicy());
216     pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TopCand);
217     assert(TopCand.Reason != NoCand && "failed to find the first candidate");
218   } else {
219     LLVM_DEBUG(traceCandidate(TopCand));
220 #ifndef NDEBUG
221     if (VerifyScheduling) {
222       SchedCandidate TCand;
223       TCand.reset(CandPolicy());
224       pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TCand);
225       assert(TCand.SU == TopCand.SU &&
226            "Last pick result should correspond to re-picking right now");
227     }
228 #endif
229   }
230 
231   // Pick best from BotCand and TopCand.
232   LLVM_DEBUG(dbgs() << "Top Cand: "; traceCandidate(TopCand);
233              dbgs() << "Bot Cand: "; traceCandidate(BotCand););
234   SchedCandidate Cand;
235   if (TopCand.Reason == BotCand.Reason) {
236     Cand = BotCand;
237     GenericSchedulerBase::CandReason TopReason = TopCand.Reason;
238     TopCand.Reason = NoCand;
239     GenericScheduler::tryCandidate(Cand, TopCand, nullptr);
240     if (TopCand.Reason != NoCand) {
241       Cand.setBest(TopCand);
242     } else {
243       TopCand.Reason = TopReason;
244     }
245   } else {
246     if (TopCand.Reason == RegExcess && TopCand.RPDelta.Excess.getUnitInc() <= 0) {
247       Cand = TopCand;
248     } else if (BotCand.Reason == RegExcess && BotCand.RPDelta.Excess.getUnitInc() <= 0) {
249       Cand = BotCand;
250     } else if (TopCand.Reason == RegCritical && TopCand.RPDelta.CriticalMax.getUnitInc() <= 0) {
251       Cand = TopCand;
252     } else if (BotCand.Reason == RegCritical && BotCand.RPDelta.CriticalMax.getUnitInc() <= 0) {
253       Cand = BotCand;
254     } else {
255       if (BotCand.Reason > TopCand.Reason) {
256         Cand = TopCand;
257       } else {
258         Cand = BotCand;
259       }
260     }
261   }
262   LLVM_DEBUG(dbgs() << "Picking: "; traceCandidate(Cand););
263 
264   IsTopNode = Cand.AtTop;
265   return Cand.SU;
266 }
267 
268 // This function is mostly cut and pasted from
269 // GenericScheduler::pickNode()
270 SUnit *GCNMaxOccupancySchedStrategy::pickNode(bool &IsTopNode) {
271   if (DAG->top() == DAG->bottom()) {
272     assert(Top.Available.empty() && Top.Pending.empty() &&
273            Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
274     return nullptr;
275   }
276   SUnit *SU;
277   do {
278     if (RegionPolicy.OnlyTopDown) {
279       SU = Top.pickOnlyChoice();
280       if (!SU) {
281         CandPolicy NoPolicy;
282         TopCand.reset(NoPolicy);
283         pickNodeFromQueue(Top, NoPolicy, DAG->getTopRPTracker(), TopCand);
284         assert(TopCand.Reason != NoCand && "failed to find a candidate");
285         SU = TopCand.SU;
286       }
287       IsTopNode = true;
288     } else if (RegionPolicy.OnlyBottomUp) {
289       SU = Bot.pickOnlyChoice();
290       if (!SU) {
291         CandPolicy NoPolicy;
292         BotCand.reset(NoPolicy);
293         pickNodeFromQueue(Bot, NoPolicy, DAG->getBotRPTracker(), BotCand);
294         assert(BotCand.Reason != NoCand && "failed to find a candidate");
295         SU = BotCand.SU;
296       }
297       IsTopNode = false;
298     } else {
299       SU = pickNodeBidirectional(IsTopNode);
300     }
301   } while (SU->isScheduled);
302 
303   if (SU->isTopReady())
304     Top.removeReady(SU);
305   if (SU->isBottomReady())
306     Bot.removeReady(SU);
307 
308   LLVM_DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") "
309                     << *SU->getInstr());
310   return SU;
311 }
312 
313 GCNScheduleDAGMILive::GCNScheduleDAGMILive(MachineSchedContext *C,
314                         std::unique_ptr<MachineSchedStrategy> S) :
315   ScheduleDAGMILive(C, std::move(S)),
316   ST(MF.getSubtarget<GCNSubtarget>()),
317   MFI(*MF.getInfo<SIMachineFunctionInfo>()),
318   StartingOccupancy(MFI.getOccupancy()),
319   MinOccupancy(StartingOccupancy), Stage(0), RegionIdx(0) {
320 
321   LLVM_DEBUG(dbgs() << "Starting occupancy is " << StartingOccupancy << ".\n");
322 }
323 
324 void GCNScheduleDAGMILive::schedule() {
325   if (Stage == 0) {
326     // Just record regions at the first pass.
327     Regions.push_back(std::make_pair(RegionBegin, RegionEnd));
328     return;
329   }
330 
331   std::vector<MachineInstr*> Unsched;
332   Unsched.reserve(NumRegionInstrs);
333   for (auto &I : *this) {
334     Unsched.push_back(&I);
335   }
336 
337   GCNRegPressure PressureBefore;
338   if (LIS) {
339     PressureBefore = Pressure[RegionIdx];
340 
341     LLVM_DEBUG(dbgs() << "Pressure before scheduling:\nRegion live-ins:";
342                GCNRPTracker::printLiveRegs(dbgs(), LiveIns[RegionIdx], MRI);
343                dbgs() << "Region live-in pressure:  ";
344                llvm::getRegPressure(MRI, LiveIns[RegionIdx]).print(dbgs());
345                dbgs() << "Region register pressure: ";
346                PressureBefore.print(dbgs()));
347   }
348 
349   ScheduleDAGMILive::schedule();
350   Regions[RegionIdx] = std::make_pair(RegionBegin, RegionEnd);
351 
352   if (!LIS)
353     return;
354 
355   // Check the results of scheduling.
356   GCNMaxOccupancySchedStrategy &S = (GCNMaxOccupancySchedStrategy&)*SchedImpl;
357   auto PressureAfter = getRealRegPressure();
358 
359   LLVM_DEBUG(dbgs() << "Pressure after scheduling: ";
360              PressureAfter.print(dbgs()));
361 
362   if (PressureAfter.getSGPRNum() <= S.SGPRCriticalLimit &&
363       PressureAfter.getVGPRNum() <= S.VGPRCriticalLimit) {
364     Pressure[RegionIdx] = PressureAfter;
365     LLVM_DEBUG(dbgs() << "Pressure in desired limits, done.\n");
366     return;
367   }
368   unsigned Occ = MFI.getOccupancy();
369   unsigned WavesAfter = std::min(Occ, PressureAfter.getOccupancy(ST));
370   unsigned WavesBefore = std::min(Occ, PressureBefore.getOccupancy(ST));
371   LLVM_DEBUG(dbgs() << "Occupancy before scheduling: " << WavesBefore
372                     << ", after " << WavesAfter << ".\n");
373 
374   // We could not keep current target occupancy because of the just scheduled
375   // region. Record new occupancy for next scheduling cycle.
376   unsigned NewOccupancy = std::max(WavesAfter, WavesBefore);
377   // Allow memory bound functions to drop to 4 waves if not limited by an
378   // attribute.
379   if (WavesAfter < WavesBefore && WavesAfter < MinOccupancy &&
380       WavesAfter >= MFI.getMinAllowedOccupancy()) {
381     LLVM_DEBUG(dbgs() << "Function is memory bound, allow occupancy drop up to "
382                       << MFI.getMinAllowedOccupancy() << " waves\n");
383     NewOccupancy = WavesAfter;
384   }
385   if (NewOccupancy < MinOccupancy) {
386     MinOccupancy = NewOccupancy;
387     MFI.limitOccupancy(MinOccupancy);
388     LLVM_DEBUG(dbgs() << "Occupancy lowered for the function to "
389                       << MinOccupancy << ".\n");
390   }
391 
392   if (WavesAfter >= MinOccupancy) {
393     unsigned TotalVGPRs = AMDGPU::IsaInfo::getAddressableNumVGPRs(&ST);
394     unsigned TotalSGPRs = AMDGPU::IsaInfo::getAddressableNumSGPRs(&ST);
395     if (WavesAfter > MFI.getMinWavesPerEU() ||
396         PressureAfter.less(ST, PressureBefore) ||
397         (TotalVGPRs >= PressureAfter.getVGPRNum() &&
398          TotalSGPRs >= PressureAfter.getSGPRNum())) {
399       Pressure[RegionIdx] = PressureAfter;
400       return;
401     }
402     LLVM_DEBUG(dbgs() << "New pressure will result in more spilling.\n");
403   }
404 
405   LLVM_DEBUG(dbgs() << "Attempting to revert scheduling.\n");
406   RegionEnd = RegionBegin;
407   for (MachineInstr *MI : Unsched) {
408     if (MI->isDebugInstr())
409       continue;
410 
411     if (MI->getIterator() != RegionEnd) {
412       BB->remove(MI);
413       BB->insert(RegionEnd, MI);
414       if (!MI->isDebugInstr())
415         LIS->handleMove(*MI, true);
416     }
417     // Reset read-undef flags and update them later.
418     for (auto &Op : MI->operands())
419       if (Op.isReg() && Op.isDef())
420         Op.setIsUndef(false);
421     RegisterOperands RegOpers;
422     RegOpers.collect(*MI, *TRI, MRI, ShouldTrackLaneMasks, false);
423     if (!MI->isDebugInstr()) {
424       if (ShouldTrackLaneMasks) {
425         // Adjust liveness and add missing dead+read-undef flags.
426         SlotIndex SlotIdx = LIS->getInstructionIndex(*MI).getRegSlot();
427         RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx, MI);
428       } else {
429         // Adjust for missing dead-def flags.
430         RegOpers.detectDeadDefs(*MI, *LIS);
431       }
432     }
433     RegionEnd = MI->getIterator();
434     ++RegionEnd;
435     LLVM_DEBUG(dbgs() << "Scheduling " << *MI);
436   }
437   RegionBegin = Unsched.front()->getIterator();
438   Regions[RegionIdx] = std::make_pair(RegionBegin, RegionEnd);
439 
440   placeDebugValues();
441 }
442 
443 GCNRegPressure GCNScheduleDAGMILive::getRealRegPressure() const {
444   GCNDownwardRPTracker RPTracker(*LIS);
445   RPTracker.advance(begin(), end(), &LiveIns[RegionIdx]);
446   return RPTracker.moveMaxPressure();
447 }
448 
449 void GCNScheduleDAGMILive::computeBlockPressure(const MachineBasicBlock *MBB) {
450   GCNDownwardRPTracker RPTracker(*LIS);
451 
452   // If the block has the only successor then live-ins of that successor are
453   // live-outs of the current block. We can reuse calculated live set if the
454   // successor will be sent to scheduling past current block.
455   const MachineBasicBlock *OnlySucc = nullptr;
456   if (MBB->succ_size() == 1 && !(*MBB->succ_begin())->empty()) {
457     SlotIndexes *Ind = LIS->getSlotIndexes();
458     if (Ind->getMBBStartIdx(MBB) < Ind->getMBBStartIdx(*MBB->succ_begin()))
459       OnlySucc = *MBB->succ_begin();
460   }
461 
462   // Scheduler sends regions from the end of the block upwards.
463   size_t CurRegion = RegionIdx;
464   for (size_t E = Regions.size(); CurRegion != E; ++CurRegion)
465     if (Regions[CurRegion].first->getParent() != MBB)
466       break;
467   --CurRegion;
468 
469   auto I = MBB->begin();
470   auto LiveInIt = MBBLiveIns.find(MBB);
471   if (LiveInIt != MBBLiveIns.end()) {
472     auto LiveIn = std::move(LiveInIt->second);
473     RPTracker.reset(*MBB->begin(), &LiveIn);
474     MBBLiveIns.erase(LiveInIt);
475   } else {
476     auto &Rgn = Regions[CurRegion];
477     I = Rgn.first;
478     auto *NonDbgMI = &*skipDebugInstructionsForward(Rgn.first, Rgn.second);
479     auto LRS = BBLiveInMap.lookup(NonDbgMI);
480     assert(isEqual(getLiveRegsBefore(*NonDbgMI, *LIS), LRS));
481     RPTracker.reset(*I, &LRS);
482   }
483 
484   for ( ; ; ) {
485     I = RPTracker.getNext();
486 
487     if (Regions[CurRegion].first == I) {
488       LiveIns[CurRegion] = RPTracker.getLiveRegs();
489       RPTracker.clearMaxPressure();
490     }
491 
492     if (Regions[CurRegion].second == I) {
493       Pressure[CurRegion] = RPTracker.moveMaxPressure();
494       if (CurRegion-- == RegionIdx)
495         break;
496     }
497     RPTracker.advanceToNext();
498     RPTracker.advanceBeforeNext();
499   }
500 
501   if (OnlySucc) {
502     if (I != MBB->end()) {
503       RPTracker.advanceToNext();
504       RPTracker.advance(MBB->end());
505     }
506     RPTracker.reset(*OnlySucc->begin(), &RPTracker.getLiveRegs());
507     RPTracker.advanceBeforeNext();
508     MBBLiveIns[OnlySucc] = RPTracker.moveLiveRegs();
509   }
510 }
511 
512 DenseMap<MachineInstr *, GCNRPTracker::LiveRegSet>
513 GCNScheduleDAGMILive::getBBLiveInMap() const {
514   assert(!Regions.empty());
515   std::vector<MachineInstr *> BBStarters;
516   BBStarters.reserve(Regions.size());
517   auto I = Regions.rbegin(), E = Regions.rend();
518   auto *BB = I->first->getParent();
519   do {
520     auto *MI = &*skipDebugInstructionsForward(I->first, I->second);
521     BBStarters.push_back(MI);
522     do {
523       ++I;
524     } while (I != E && I->first->getParent() == BB);
525   } while (I != E);
526   return getLiveRegMap(BBStarters, false /*After*/, *LIS);
527 }
528 
529 void GCNScheduleDAGMILive::finalizeSchedule() {
530   GCNMaxOccupancySchedStrategy &S = (GCNMaxOccupancySchedStrategy&)*SchedImpl;
531   LLVM_DEBUG(dbgs() << "All regions recorded, starting actual scheduling.\n");
532 
533   LiveIns.resize(Regions.size());
534   Pressure.resize(Regions.size());
535 
536   if (!Regions.empty())
537     BBLiveInMap = getBBLiveInMap();
538 
539   do {
540     Stage++;
541     RegionIdx = 0;
542     MachineBasicBlock *MBB = nullptr;
543 
544     if (Stage > 1) {
545       // Retry function scheduling if we found resulting occupancy and it is
546       // lower than used for first pass scheduling. This will give more freedom
547       // to schedule low register pressure blocks.
548       // Code is partially copied from MachineSchedulerBase::scheduleRegions().
549 
550       if (!LIS || StartingOccupancy <= MinOccupancy)
551         break;
552 
553       LLVM_DEBUG(
554           dbgs()
555           << "Retrying function scheduling with lowest recorded occupancy "
556           << MinOccupancy << ".\n");
557 
558       S.setTargetOccupancy(MinOccupancy);
559     }
560 
561     for (auto Region : Regions) {
562       RegionBegin = Region.first;
563       RegionEnd = Region.second;
564 
565       if (RegionBegin->getParent() != MBB) {
566         if (MBB) finishBlock();
567         MBB = RegionBegin->getParent();
568         startBlock(MBB);
569         if (Stage == 1)
570           computeBlockPressure(MBB);
571       }
572 
573       unsigned NumRegionInstrs = std::distance(begin(), end());
574       enterRegion(MBB, begin(), end(), NumRegionInstrs);
575 
576       // Skip empty scheduling regions (0 or 1 schedulable instructions).
577       if (begin() == end() || begin() == std::prev(end())) {
578         exitRegion();
579         continue;
580       }
581 
582       LLVM_DEBUG(dbgs() << "********** MI Scheduling **********\n");
583       LLVM_DEBUG(dbgs() << MF.getName() << ":" << printMBBReference(*MBB) << " "
584                         << MBB->getName() << "\n  From: " << *begin()
585                         << "    To: ";
586                  if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
587                  else dbgs() << "End";
588                  dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n');
589 
590       schedule();
591 
592       exitRegion();
593       ++RegionIdx;
594     }
595     finishBlock();
596 
597   } while (Stage < 2);
598 }
599