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 "SIMachineFunctionInfo.h"
16
17 #define DEBUG_TYPE "machine-scheduler"
18
19 using namespace llvm;
20
GCNMaxOccupancySchedStrategy(const MachineSchedContext * C)21 GCNMaxOccupancySchedStrategy::GCNMaxOccupancySchedStrategy(
22 const MachineSchedContext *C) :
23 GenericScheduler(C), TargetOccupancy(0), MF(nullptr) { }
24
initialize(ScheduleDAGMI * DAG)25 void GCNMaxOccupancySchedStrategy::initialize(ScheduleDAGMI *DAG) {
26 GenericScheduler::initialize(DAG);
27
28 const SIRegisterInfo *SRI = static_cast<const SIRegisterInfo*>(TRI);
29
30 MF = &DAG->MF;
31
32 const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>();
33
34 // FIXME: This is also necessary, because some passes that run after
35 // scheduling and before regalloc increase register pressure.
36 const int ErrorMargin = 3;
37
38 SGPRExcessLimit = Context->RegClassInfo
39 ->getNumAllocatableRegs(&AMDGPU::SGPR_32RegClass) - ErrorMargin;
40 VGPRExcessLimit = Context->RegClassInfo
41 ->getNumAllocatableRegs(&AMDGPU::VGPR_32RegClass) - ErrorMargin;
42 if (TargetOccupancy) {
43 SGPRCriticalLimit = ST.getMaxNumSGPRs(TargetOccupancy, true);
44 VGPRCriticalLimit = ST.getMaxNumVGPRs(TargetOccupancy);
45 } else {
46 SGPRCriticalLimit = SRI->getRegPressureSetLimit(DAG->MF,
47 AMDGPU::RegisterPressureSets::SReg_32);
48 VGPRCriticalLimit = SRI->getRegPressureSetLimit(DAG->MF,
49 AMDGPU::RegisterPressureSets::VGPR_32);
50 }
51
52 SGPRCriticalLimit -= ErrorMargin;
53 VGPRCriticalLimit -= ErrorMargin;
54 }
55
initCandidate(SchedCandidate & Cand,SUnit * SU,bool AtTop,const RegPressureTracker & RPTracker,const SIRegisterInfo * SRI,unsigned SGPRPressure,unsigned VGPRPressure)56 void GCNMaxOccupancySchedStrategy::initCandidate(SchedCandidate &Cand, SUnit *SU,
57 bool AtTop, const RegPressureTracker &RPTracker,
58 const SIRegisterInfo *SRI,
59 unsigned SGPRPressure,
60 unsigned VGPRPressure) {
61
62 Cand.SU = SU;
63 Cand.AtTop = AtTop;
64
65 // getDownwardPressure() and getUpwardPressure() make temporary changes to
66 // the tracker, so we need to pass those function a non-const copy.
67 RegPressureTracker &TempTracker = const_cast<RegPressureTracker&>(RPTracker);
68
69 Pressure.clear();
70 MaxPressure.clear();
71
72 if (AtTop)
73 TempTracker.getDownwardPressure(SU->getInstr(), Pressure, MaxPressure);
74 else {
75 // FIXME: I think for bottom up scheduling, the register pressure is cached
76 // and can be retrieved by DAG->getPressureDif(SU).
77 TempTracker.getUpwardPressure(SU->getInstr(), Pressure, MaxPressure);
78 }
79
80 unsigned NewSGPRPressure = Pressure[AMDGPU::RegisterPressureSets::SReg_32];
81 unsigned NewVGPRPressure = Pressure[AMDGPU::RegisterPressureSets::VGPR_32];
82
83 // If two instructions increase the pressure of different register sets
84 // by the same amount, the generic scheduler will prefer to schedule the
85 // instruction that increases the set with the least amount of registers,
86 // which in our case would be SGPRs. This is rarely what we want, so
87 // when we report excess/critical register pressure, we do it either
88 // only for VGPRs or only for SGPRs.
89
90 // FIXME: Better heuristics to determine whether to prefer SGPRs or VGPRs.
91 const unsigned MaxVGPRPressureInc = 16;
92 bool ShouldTrackVGPRs = VGPRPressure + MaxVGPRPressureInc >= VGPRExcessLimit;
93 bool ShouldTrackSGPRs = !ShouldTrackVGPRs && SGPRPressure >= SGPRExcessLimit;
94
95
96 // FIXME: We have to enter REG-EXCESS before we reach the actual threshold
97 // to increase the likelihood we don't go over the limits. We should improve
98 // the analysis to look through dependencies to find the path with the least
99 // register pressure.
100
101 // We only need to update the RPDelta for instructions that increase register
102 // pressure. Instructions that decrease or keep reg pressure the same will be
103 // marked as RegExcess in tryCandidate() when they are compared with
104 // instructions that increase the register pressure.
105 if (ShouldTrackVGPRs && NewVGPRPressure >= VGPRExcessLimit) {
106 Cand.RPDelta.Excess = PressureChange(AMDGPU::RegisterPressureSets::VGPR_32);
107 Cand.RPDelta.Excess.setUnitInc(NewVGPRPressure - VGPRExcessLimit);
108 }
109
110 if (ShouldTrackSGPRs && NewSGPRPressure >= SGPRExcessLimit) {
111 Cand.RPDelta.Excess = PressureChange(AMDGPU::RegisterPressureSets::SReg_32);
112 Cand.RPDelta.Excess.setUnitInc(NewSGPRPressure - SGPRExcessLimit);
113 }
114
115 // Register pressure is considered 'CRITICAL' if it is approaching a value
116 // that would reduce the wave occupancy for the execution unit. When
117 // register pressure is 'CRITICAL', increading SGPR and VGPR pressure both
118 // has the same cost, so we don't need to prefer one over the other.
119
120 int SGPRDelta = NewSGPRPressure - SGPRCriticalLimit;
121 int VGPRDelta = NewVGPRPressure - VGPRCriticalLimit;
122
123 if (SGPRDelta >= 0 || VGPRDelta >= 0) {
124 if (SGPRDelta > VGPRDelta) {
125 Cand.RPDelta.CriticalMax =
126 PressureChange(AMDGPU::RegisterPressureSets::SReg_32);
127 Cand.RPDelta.CriticalMax.setUnitInc(SGPRDelta);
128 } else {
129 Cand.RPDelta.CriticalMax =
130 PressureChange(AMDGPU::RegisterPressureSets::VGPR_32);
131 Cand.RPDelta.CriticalMax.setUnitInc(VGPRDelta);
132 }
133 }
134 }
135
136 // This function is mostly cut and pasted from
137 // GenericScheduler::pickNodeFromQueue()
pickNodeFromQueue(SchedBoundary & Zone,const CandPolicy & ZonePolicy,const RegPressureTracker & RPTracker,SchedCandidate & Cand)138 void GCNMaxOccupancySchedStrategy::pickNodeFromQueue(SchedBoundary &Zone,
139 const CandPolicy &ZonePolicy,
140 const RegPressureTracker &RPTracker,
141 SchedCandidate &Cand) {
142 const SIRegisterInfo *SRI = static_cast<const SIRegisterInfo*>(TRI);
143 ArrayRef<unsigned> Pressure = RPTracker.getRegSetPressureAtPos();
144 unsigned SGPRPressure = Pressure[AMDGPU::RegisterPressureSets::SReg_32];
145 unsigned VGPRPressure = Pressure[AMDGPU::RegisterPressureSets::VGPR_32];
146 ReadyQueue &Q = Zone.Available;
147 for (SUnit *SU : Q) {
148
149 SchedCandidate TryCand(ZonePolicy);
150 initCandidate(TryCand, SU, Zone.isTop(), RPTracker, SRI,
151 SGPRPressure, VGPRPressure);
152 // Pass SchedBoundary only when comparing nodes from the same boundary.
153 SchedBoundary *ZoneArg = Cand.AtTop == TryCand.AtTop ? &Zone : nullptr;
154 GenericScheduler::tryCandidate(Cand, TryCand, ZoneArg);
155 if (TryCand.Reason != NoCand) {
156 // Initialize resource delta if needed in case future heuristics query it.
157 if (TryCand.ResDelta == SchedResourceDelta())
158 TryCand.initResourceDelta(Zone.DAG, SchedModel);
159 Cand.setBest(TryCand);
160 LLVM_DEBUG(traceCandidate(Cand));
161 }
162 }
163 }
164
165 // This function is mostly cut and pasted from
166 // GenericScheduler::pickNodeBidirectional()
pickNodeBidirectional(bool & IsTopNode)167 SUnit *GCNMaxOccupancySchedStrategy::pickNodeBidirectional(bool &IsTopNode) {
168 // Schedule as far as possible in the direction of no choice. This is most
169 // efficient, but also provides the best heuristics for CriticalPSets.
170 if (SUnit *SU = Bot.pickOnlyChoice()) {
171 IsTopNode = false;
172 return SU;
173 }
174 if (SUnit *SU = Top.pickOnlyChoice()) {
175 IsTopNode = true;
176 return SU;
177 }
178 // Set the bottom-up policy based on the state of the current bottom zone and
179 // the instructions outside the zone, including the top zone.
180 CandPolicy BotPolicy;
181 setPolicy(BotPolicy, /*IsPostRA=*/false, Bot, &Top);
182 // Set the top-down policy based on the state of the current top zone and
183 // the instructions outside the zone, including the bottom zone.
184 CandPolicy TopPolicy;
185 setPolicy(TopPolicy, /*IsPostRA=*/false, Top, &Bot);
186
187 // See if BotCand is still valid (because we previously scheduled from Top).
188 LLVM_DEBUG(dbgs() << "Picking from Bot:\n");
189 if (!BotCand.isValid() || BotCand.SU->isScheduled ||
190 BotCand.Policy != BotPolicy) {
191 BotCand.reset(CandPolicy());
192 pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), BotCand);
193 assert(BotCand.Reason != NoCand && "failed to find the first candidate");
194 } else {
195 LLVM_DEBUG(traceCandidate(BotCand));
196 #ifndef NDEBUG
197 if (VerifyScheduling) {
198 SchedCandidate TCand;
199 TCand.reset(CandPolicy());
200 pickNodeFromQueue(Bot, BotPolicy, DAG->getBotRPTracker(), TCand);
201 assert(TCand.SU == BotCand.SU &&
202 "Last pick result should correspond to re-picking right now");
203 }
204 #endif
205 }
206
207 // Check if the top Q has a better candidate.
208 LLVM_DEBUG(dbgs() << "Picking from Top:\n");
209 if (!TopCand.isValid() || TopCand.SU->isScheduled ||
210 TopCand.Policy != TopPolicy) {
211 TopCand.reset(CandPolicy());
212 pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TopCand);
213 assert(TopCand.Reason != NoCand && "failed to find the first candidate");
214 } else {
215 LLVM_DEBUG(traceCandidate(TopCand));
216 #ifndef NDEBUG
217 if (VerifyScheduling) {
218 SchedCandidate TCand;
219 TCand.reset(CandPolicy());
220 pickNodeFromQueue(Top, TopPolicy, DAG->getTopRPTracker(), TCand);
221 assert(TCand.SU == TopCand.SU &&
222 "Last pick result should correspond to re-picking right now");
223 }
224 #endif
225 }
226
227 // Pick best from BotCand and TopCand.
228 LLVM_DEBUG(dbgs() << "Top Cand: "; traceCandidate(TopCand);
229 dbgs() << "Bot Cand: "; traceCandidate(BotCand););
230 SchedCandidate Cand = BotCand;
231 TopCand.Reason = NoCand;
232 GenericScheduler::tryCandidate(Cand, TopCand, nullptr);
233 if (TopCand.Reason != NoCand) {
234 Cand.setBest(TopCand);
235 }
236 LLVM_DEBUG(dbgs() << "Picking: "; traceCandidate(Cand););
237
238 IsTopNode = Cand.AtTop;
239 return Cand.SU;
240 }
241
242 // This function is mostly cut and pasted from
243 // GenericScheduler::pickNode()
pickNode(bool & IsTopNode)244 SUnit *GCNMaxOccupancySchedStrategy::pickNode(bool &IsTopNode) {
245 if (DAG->top() == DAG->bottom()) {
246 assert(Top.Available.empty() && Top.Pending.empty() &&
247 Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage");
248 return nullptr;
249 }
250 SUnit *SU;
251 do {
252 if (RegionPolicy.OnlyTopDown) {
253 SU = Top.pickOnlyChoice();
254 if (!SU) {
255 CandPolicy NoPolicy;
256 TopCand.reset(NoPolicy);
257 pickNodeFromQueue(Top, NoPolicy, DAG->getTopRPTracker(), TopCand);
258 assert(TopCand.Reason != NoCand && "failed to find a candidate");
259 SU = TopCand.SU;
260 }
261 IsTopNode = true;
262 } else if (RegionPolicy.OnlyBottomUp) {
263 SU = Bot.pickOnlyChoice();
264 if (!SU) {
265 CandPolicy NoPolicy;
266 BotCand.reset(NoPolicy);
267 pickNodeFromQueue(Bot, NoPolicy, DAG->getBotRPTracker(), BotCand);
268 assert(BotCand.Reason != NoCand && "failed to find a candidate");
269 SU = BotCand.SU;
270 }
271 IsTopNode = false;
272 } else {
273 SU = pickNodeBidirectional(IsTopNode);
274 }
275 } while (SU->isScheduled);
276
277 if (SU->isTopReady())
278 Top.removeReady(SU);
279 if (SU->isBottomReady())
280 Bot.removeReady(SU);
281
282 LLVM_DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") "
283 << *SU->getInstr());
284 return SU;
285 }
286
GCNScheduleDAGMILive(MachineSchedContext * C,std::unique_ptr<MachineSchedStrategy> S)287 GCNScheduleDAGMILive::GCNScheduleDAGMILive(MachineSchedContext *C,
288 std::unique_ptr<MachineSchedStrategy> S) :
289 ScheduleDAGMILive(C, std::move(S)),
290 ST(MF.getSubtarget<GCNSubtarget>()),
291 MFI(*MF.getInfo<SIMachineFunctionInfo>()),
292 StartingOccupancy(MFI.getOccupancy()),
293 MinOccupancy(StartingOccupancy), Stage(Collect), RegionIdx(0) {
294
295 LLVM_DEBUG(dbgs() << "Starting occupancy is " << StartingOccupancy << ".\n");
296 }
297
schedule()298 void GCNScheduleDAGMILive::schedule() {
299 if (Stage == Collect) {
300 // Just record regions at the first pass.
301 Regions.push_back(std::make_pair(RegionBegin, RegionEnd));
302 return;
303 }
304
305 std::vector<MachineInstr*> Unsched;
306 Unsched.reserve(NumRegionInstrs);
307 for (auto &I : *this) {
308 Unsched.push_back(&I);
309 }
310
311 GCNRegPressure PressureBefore;
312 if (LIS) {
313 PressureBefore = Pressure[RegionIdx];
314
315 LLVM_DEBUG(dbgs() << "Pressure before scheduling:\nRegion live-ins:";
316 GCNRPTracker::printLiveRegs(dbgs(), LiveIns[RegionIdx], MRI);
317 dbgs() << "Region live-in pressure: ";
318 llvm::getRegPressure(MRI, LiveIns[RegionIdx]).print(dbgs());
319 dbgs() << "Region register pressure: ";
320 PressureBefore.print(dbgs()));
321 }
322
323 ScheduleDAGMILive::schedule();
324 Regions[RegionIdx] = std::make_pair(RegionBegin, RegionEnd);
325 RescheduleRegions[RegionIdx] = false;
326
327 if (!LIS)
328 return;
329
330 // Check the results of scheduling.
331 GCNMaxOccupancySchedStrategy &S = (GCNMaxOccupancySchedStrategy&)*SchedImpl;
332 auto PressureAfter = getRealRegPressure();
333
334 LLVM_DEBUG(dbgs() << "Pressure after scheduling: ";
335 PressureAfter.print(dbgs()));
336
337 if (PressureAfter.getSGPRNum() <= S.SGPRCriticalLimit &&
338 PressureAfter.getVGPRNum() <= S.VGPRCriticalLimit) {
339 Pressure[RegionIdx] = PressureAfter;
340 LLVM_DEBUG(dbgs() << "Pressure in desired limits, done.\n");
341 return;
342 }
343 unsigned Occ = MFI.getOccupancy();
344 unsigned WavesAfter = std::min(Occ, PressureAfter.getOccupancy(ST));
345 unsigned WavesBefore = std::min(Occ, PressureBefore.getOccupancy(ST));
346 LLVM_DEBUG(dbgs() << "Occupancy before scheduling: " << WavesBefore
347 << ", after " << WavesAfter << ".\n");
348
349 // We could not keep current target occupancy because of the just scheduled
350 // region. Record new occupancy for next scheduling cycle.
351 unsigned NewOccupancy = std::max(WavesAfter, WavesBefore);
352 // Allow memory bound functions to drop to 4 waves if not limited by an
353 // attribute.
354 if (WavesAfter < WavesBefore && WavesAfter < MinOccupancy &&
355 WavesAfter >= MFI.getMinAllowedOccupancy()) {
356 LLVM_DEBUG(dbgs() << "Function is memory bound, allow occupancy drop up to "
357 << MFI.getMinAllowedOccupancy() << " waves\n");
358 NewOccupancy = WavesAfter;
359 }
360 if (NewOccupancy < MinOccupancy) {
361 MinOccupancy = NewOccupancy;
362 MFI.limitOccupancy(MinOccupancy);
363 LLVM_DEBUG(dbgs() << "Occupancy lowered for the function to "
364 << MinOccupancy << ".\n");
365 }
366
367 unsigned MaxVGPRs = ST.getMaxNumVGPRs(MF);
368 unsigned MaxSGPRs = ST.getMaxNumSGPRs(MF);
369 if (PressureAfter.getVGPRNum() > MaxVGPRs ||
370 PressureAfter.getSGPRNum() > MaxSGPRs)
371 RescheduleRegions[RegionIdx] = true;
372
373 if (WavesAfter >= MinOccupancy) {
374 if (Stage == UnclusteredReschedule &&
375 !PressureAfter.less(ST, PressureBefore)) {
376 LLVM_DEBUG(dbgs() << "Unclustered reschedule did not help.\n");
377 } else if (WavesAfter > MFI.getMinWavesPerEU() ||
378 PressureAfter.less(ST, PressureBefore) ||
379 !RescheduleRegions[RegionIdx]) {
380 Pressure[RegionIdx] = PressureAfter;
381 return;
382 } else {
383 LLVM_DEBUG(dbgs() << "New pressure will result in more spilling.\n");
384 }
385 }
386
387 LLVM_DEBUG(dbgs() << "Attempting to revert scheduling.\n");
388 RescheduleRegions[RegionIdx] = true;
389 RegionEnd = RegionBegin;
390 for (MachineInstr *MI : Unsched) {
391 if (MI->isDebugInstr())
392 continue;
393
394 if (MI->getIterator() != RegionEnd) {
395 BB->remove(MI);
396 BB->insert(RegionEnd, MI);
397 if (!MI->isDebugInstr())
398 LIS->handleMove(*MI, true);
399 }
400 // Reset read-undef flags and update them later.
401 for (auto &Op : MI->operands())
402 if (Op.isReg() && Op.isDef())
403 Op.setIsUndef(false);
404 RegisterOperands RegOpers;
405 RegOpers.collect(*MI, *TRI, MRI, ShouldTrackLaneMasks, false);
406 if (!MI->isDebugInstr()) {
407 if (ShouldTrackLaneMasks) {
408 // Adjust liveness and add missing dead+read-undef flags.
409 SlotIndex SlotIdx = LIS->getInstructionIndex(*MI).getRegSlot();
410 RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx, MI);
411 } else {
412 // Adjust for missing dead-def flags.
413 RegOpers.detectDeadDefs(*MI, *LIS);
414 }
415 }
416 RegionEnd = MI->getIterator();
417 ++RegionEnd;
418 LLVM_DEBUG(dbgs() << "Scheduling " << *MI);
419 }
420 RegionBegin = Unsched.front()->getIterator();
421 Regions[RegionIdx] = std::make_pair(RegionBegin, RegionEnd);
422
423 placeDebugValues();
424 }
425
getRealRegPressure() const426 GCNRegPressure GCNScheduleDAGMILive::getRealRegPressure() const {
427 GCNDownwardRPTracker RPTracker(*LIS);
428 RPTracker.advance(begin(), end(), &LiveIns[RegionIdx]);
429 return RPTracker.moveMaxPressure();
430 }
431
computeBlockPressure(const MachineBasicBlock * MBB)432 void GCNScheduleDAGMILive::computeBlockPressure(const MachineBasicBlock *MBB) {
433 GCNDownwardRPTracker RPTracker(*LIS);
434
435 // If the block has the only successor then live-ins of that successor are
436 // live-outs of the current block. We can reuse calculated live set if the
437 // successor will be sent to scheduling past current block.
438 const MachineBasicBlock *OnlySucc = nullptr;
439 if (MBB->succ_size() == 1 && !(*MBB->succ_begin())->empty()) {
440 SlotIndexes *Ind = LIS->getSlotIndexes();
441 if (Ind->getMBBStartIdx(MBB) < Ind->getMBBStartIdx(*MBB->succ_begin()))
442 OnlySucc = *MBB->succ_begin();
443 }
444
445 // Scheduler sends regions from the end of the block upwards.
446 size_t CurRegion = RegionIdx;
447 for (size_t E = Regions.size(); CurRegion != E; ++CurRegion)
448 if (Regions[CurRegion].first->getParent() != MBB)
449 break;
450 --CurRegion;
451
452 auto I = MBB->begin();
453 auto LiveInIt = MBBLiveIns.find(MBB);
454 if (LiveInIt != MBBLiveIns.end()) {
455 auto LiveIn = std::move(LiveInIt->second);
456 RPTracker.reset(*MBB->begin(), &LiveIn);
457 MBBLiveIns.erase(LiveInIt);
458 } else {
459 auto &Rgn = Regions[CurRegion];
460 I = Rgn.first;
461 auto *NonDbgMI = &*skipDebugInstructionsForward(Rgn.first, Rgn.second);
462 auto LRS = BBLiveInMap.lookup(NonDbgMI);
463 assert(isEqual(getLiveRegsBefore(*NonDbgMI, *LIS), LRS));
464 RPTracker.reset(*I, &LRS);
465 }
466
467 for ( ; ; ) {
468 I = RPTracker.getNext();
469
470 if (Regions[CurRegion].first == I) {
471 LiveIns[CurRegion] = RPTracker.getLiveRegs();
472 RPTracker.clearMaxPressure();
473 }
474
475 if (Regions[CurRegion].second == I) {
476 Pressure[CurRegion] = RPTracker.moveMaxPressure();
477 if (CurRegion-- == RegionIdx)
478 break;
479 }
480 RPTracker.advanceToNext();
481 RPTracker.advanceBeforeNext();
482 }
483
484 if (OnlySucc) {
485 if (I != MBB->end()) {
486 RPTracker.advanceToNext();
487 RPTracker.advance(MBB->end());
488 }
489 RPTracker.reset(*OnlySucc->begin(), &RPTracker.getLiveRegs());
490 RPTracker.advanceBeforeNext();
491 MBBLiveIns[OnlySucc] = RPTracker.moveLiveRegs();
492 }
493 }
494
495 DenseMap<MachineInstr *, GCNRPTracker::LiveRegSet>
getBBLiveInMap() const496 GCNScheduleDAGMILive::getBBLiveInMap() const {
497 assert(!Regions.empty());
498 std::vector<MachineInstr *> BBStarters;
499 BBStarters.reserve(Regions.size());
500 auto I = Regions.rbegin(), E = Regions.rend();
501 auto *BB = I->first->getParent();
502 do {
503 auto *MI = &*skipDebugInstructionsForward(I->first, I->second);
504 BBStarters.push_back(MI);
505 do {
506 ++I;
507 } while (I != E && I->first->getParent() == BB);
508 } while (I != E);
509 return getLiveRegMap(BBStarters, false /*After*/, *LIS);
510 }
511
finalizeSchedule()512 void GCNScheduleDAGMILive::finalizeSchedule() {
513 GCNMaxOccupancySchedStrategy &S = (GCNMaxOccupancySchedStrategy&)*SchedImpl;
514 LLVM_DEBUG(dbgs() << "All regions recorded, starting actual scheduling.\n");
515
516 LiveIns.resize(Regions.size());
517 Pressure.resize(Regions.size());
518 RescheduleRegions.resize(Regions.size());
519 RescheduleRegions.set();
520
521 if (!Regions.empty())
522 BBLiveInMap = getBBLiveInMap();
523
524 std::vector<std::unique_ptr<ScheduleDAGMutation>> SavedMutations;
525
526 do {
527 Stage++;
528 RegionIdx = 0;
529 MachineBasicBlock *MBB = nullptr;
530
531 if (Stage > InitialSchedule) {
532 if (!LIS)
533 break;
534
535 // Retry function scheduling if we found resulting occupancy and it is
536 // lower than used for first pass scheduling. This will give more freedom
537 // to schedule low register pressure blocks.
538 // Code is partially copied from MachineSchedulerBase::scheduleRegions().
539
540 if (Stage == UnclusteredReschedule) {
541 if (RescheduleRegions.none())
542 continue;
543 LLVM_DEBUG(dbgs() <<
544 "Retrying function scheduling without clustering.\n");
545 }
546
547 if (Stage == ClusteredLowOccupancyReschedule) {
548 if (StartingOccupancy <= MinOccupancy)
549 break;
550
551 LLVM_DEBUG(
552 dbgs()
553 << "Retrying function scheduling with lowest recorded occupancy "
554 << MinOccupancy << ".\n");
555
556 S.setTargetOccupancy(MinOccupancy);
557 }
558 }
559
560 if (Stage == UnclusteredReschedule)
561 SavedMutations.swap(Mutations);
562
563 for (auto Region : Regions) {
564 if (Stage == UnclusteredReschedule && !RescheduleRegions[RegionIdx]) {
565 ++RegionIdx;
566 continue;
567 }
568
569 RegionBegin = Region.first;
570 RegionEnd = Region.second;
571
572 if (RegionBegin->getParent() != MBB) {
573 if (MBB) finishBlock();
574 MBB = RegionBegin->getParent();
575 startBlock(MBB);
576 if (Stage == InitialSchedule)
577 computeBlockPressure(MBB);
578 }
579
580 unsigned NumRegionInstrs = std::distance(begin(), end());
581 enterRegion(MBB, begin(), end(), NumRegionInstrs);
582
583 // Skip empty scheduling regions (0 or 1 schedulable instructions).
584 if (begin() == end() || begin() == std::prev(end())) {
585 exitRegion();
586 continue;
587 }
588
589 LLVM_DEBUG(dbgs() << "********** MI Scheduling **********\n");
590 LLVM_DEBUG(dbgs() << MF.getName() << ":" << printMBBReference(*MBB) << " "
591 << MBB->getName() << "\n From: " << *begin()
592 << " To: ";
593 if (RegionEnd != MBB->end()) dbgs() << *RegionEnd;
594 else dbgs() << "End";
595 dbgs() << " RegionInstrs: " << NumRegionInstrs << '\n');
596
597 schedule();
598
599 exitRegion();
600 ++RegionIdx;
601 }
602 finishBlock();
603
604 if (Stage == UnclusteredReschedule)
605 SavedMutations.swap(Mutations);
606 } while (Stage != LastStage);
607 }
608