1 //===- GCNIterativeScheduler.cpp ------------------------------------------===//
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 #include "GCNIterativeScheduler.h"
10 #include "AMDGPUSubtarget.h"
11 #include "GCNRegPressure.h"
12 #include "GCNSchedStrategy.h"
13 #include "SIMachineFunctionInfo.h"
14 #include "llvm/ADT/ArrayRef.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/CodeGen/LiveIntervals.h"
18 #include "llvm/CodeGen/MachineBasicBlock.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/RegisterPressure.h"
21 #include "llvm/CodeGen/ScheduleDAG.h"
22 #include "llvm/Config/llvm-config.h"
23 #include "llvm/Support/Compiler.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/raw_ostream.h"
26 #include <algorithm>
27 #include <cassert>
28 #include <iterator>
29 #include <limits>
30 #include <memory>
31 #include <type_traits>
32 #include <vector>
33
34 using namespace llvm;
35
36 #define DEBUG_TYPE "machine-scheduler"
37
38 namespace llvm {
39
40 std::vector<const SUnit *> makeMinRegSchedule(ArrayRef<const SUnit *> TopRoots,
41 const ScheduleDAG &DAG);
42
43 std::vector<const SUnit*> makeGCNILPScheduler(ArrayRef<const SUnit*> BotRoots,
44 const ScheduleDAG &DAG);
45 }
46
47 // shim accessors for different order containers
getMachineInstr(MachineInstr * MI)48 static inline MachineInstr *getMachineInstr(MachineInstr *MI) {
49 return MI;
50 }
getMachineInstr(const SUnit * SU)51 static inline MachineInstr *getMachineInstr(const SUnit *SU) {
52 return SU->getInstr();
53 }
getMachineInstr(const SUnit & SU)54 static inline MachineInstr *getMachineInstr(const SUnit &SU) {
55 return SU.getInstr();
56 }
57
58 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
59 LLVM_DUMP_METHOD
printRegion(raw_ostream & OS,MachineBasicBlock::iterator Begin,MachineBasicBlock::iterator End,const LiveIntervals * LIS,unsigned MaxInstNum=std::numeric_limits<unsigned>::max ())60 static void printRegion(raw_ostream &OS,
61 MachineBasicBlock::iterator Begin,
62 MachineBasicBlock::iterator End,
63 const LiveIntervals *LIS,
64 unsigned MaxInstNum =
65 std::numeric_limits<unsigned>::max()) {
66 auto BB = Begin->getParent();
67 OS << BB->getParent()->getName() << ":" << printMBBReference(*BB) << ' '
68 << BB->getName() << ":\n";
69 auto I = Begin;
70 MaxInstNum = std::max(MaxInstNum, 1u);
71 for (; I != End && MaxInstNum; ++I, --MaxInstNum) {
72 if (!I->isDebugInstr() && LIS)
73 OS << LIS->getInstructionIndex(*I);
74 OS << '\t' << *I;
75 }
76 if (I != End) {
77 OS << "\t...\n";
78 I = std::prev(End);
79 if (!I->isDebugInstr() && LIS)
80 OS << LIS->getInstructionIndex(*I);
81 OS << '\t' << *I;
82 }
83 if (End != BB->end()) { // print boundary inst if present
84 OS << "----\n";
85 if (LIS) OS << LIS->getInstructionIndex(*End) << '\t';
86 OS << *End;
87 }
88 }
89
90 LLVM_DUMP_METHOD
printLivenessInfo(raw_ostream & OS,MachineBasicBlock::iterator Begin,MachineBasicBlock::iterator End,const LiveIntervals * LIS)91 static void printLivenessInfo(raw_ostream &OS,
92 MachineBasicBlock::iterator Begin,
93 MachineBasicBlock::iterator End,
94 const LiveIntervals *LIS) {
95 const auto BB = Begin->getParent();
96 const auto &MRI = BB->getParent()->getRegInfo();
97
98 const auto LiveIns = getLiveRegsBefore(*Begin, *LIS);
99 OS << "LIn RP: ";
100 getRegPressure(MRI, LiveIns).print(OS);
101
102 const auto BottomMI = End == BB->end() ? std::prev(End) : End;
103 const auto LiveOuts = getLiveRegsAfter(*BottomMI, *LIS);
104 OS << "LOt RP: ";
105 getRegPressure(MRI, LiveOuts).print(OS);
106 }
107
108 LLVM_DUMP_METHOD
printRegions(raw_ostream & OS) const109 void GCNIterativeScheduler::printRegions(raw_ostream &OS) const {
110 const auto &ST = MF.getSubtarget<GCNSubtarget>();
111 for (const auto R : Regions) {
112 OS << "Region to schedule ";
113 printRegion(OS, R->Begin, R->End, LIS, 1);
114 printLivenessInfo(OS, R->Begin, R->End, LIS);
115 OS << "Max RP: ";
116 R->MaxPressure.print(OS, &ST);
117 }
118 }
119
120 LLVM_DUMP_METHOD
printSchedResult(raw_ostream & OS,const Region * R,const GCNRegPressure & RP) const121 void GCNIterativeScheduler::printSchedResult(raw_ostream &OS,
122 const Region *R,
123 const GCNRegPressure &RP) const {
124 OS << "\nAfter scheduling ";
125 printRegion(OS, R->Begin, R->End, LIS);
126 printSchedRP(OS, R->MaxPressure, RP);
127 OS << '\n';
128 }
129
130 LLVM_DUMP_METHOD
printSchedRP(raw_ostream & OS,const GCNRegPressure & Before,const GCNRegPressure & After) const131 void GCNIterativeScheduler::printSchedRP(raw_ostream &OS,
132 const GCNRegPressure &Before,
133 const GCNRegPressure &After) const {
134 const auto &ST = MF.getSubtarget<GCNSubtarget>();
135 OS << "RP before: ";
136 Before.print(OS, &ST);
137 OS << "RP after: ";
138 After.print(OS, &ST);
139 }
140 #endif
141
142 // DAG builder helper
143 class GCNIterativeScheduler::BuildDAG {
144 GCNIterativeScheduler &Sch;
145 SmallVector<SUnit *, 8> TopRoots;
146
147 SmallVector<SUnit*, 8> BotRoots;
148 public:
BuildDAG(const Region & R,GCNIterativeScheduler & _Sch)149 BuildDAG(const Region &R, GCNIterativeScheduler &_Sch)
150 : Sch(_Sch) {
151 auto BB = R.Begin->getParent();
152 Sch.BaseClass::startBlock(BB);
153 Sch.BaseClass::enterRegion(BB, R.Begin, R.End, R.NumRegionInstrs);
154
155 Sch.buildSchedGraph(Sch.AA, nullptr, nullptr, nullptr,
156 /*TrackLaneMask*/true);
157 Sch.Topo.InitDAGTopologicalSorting();
158 Sch.findRootsAndBiasEdges(TopRoots, BotRoots);
159 }
160
~BuildDAG()161 ~BuildDAG() {
162 Sch.BaseClass::exitRegion();
163 Sch.BaseClass::finishBlock();
164 }
165
getTopRoots() const166 ArrayRef<const SUnit *> getTopRoots() const {
167 return TopRoots;
168 }
getBottomRoots() const169 ArrayRef<SUnit*> getBottomRoots() const {
170 return BotRoots;
171 }
172 };
173
174 class GCNIterativeScheduler::OverrideLegacyStrategy {
175 GCNIterativeScheduler &Sch;
176 Region &Rgn;
177 std::unique_ptr<MachineSchedStrategy> SaveSchedImpl;
178 GCNRegPressure SaveMaxRP;
179
180 public:
OverrideLegacyStrategy(Region & R,MachineSchedStrategy & OverrideStrategy,GCNIterativeScheduler & _Sch)181 OverrideLegacyStrategy(Region &R,
182 MachineSchedStrategy &OverrideStrategy,
183 GCNIterativeScheduler &_Sch)
184 : Sch(_Sch)
185 , Rgn(R)
186 , SaveSchedImpl(std::move(_Sch.SchedImpl))
187 , SaveMaxRP(R.MaxPressure) {
188 Sch.SchedImpl.reset(&OverrideStrategy);
189 auto BB = R.Begin->getParent();
190 Sch.BaseClass::startBlock(BB);
191 Sch.BaseClass::enterRegion(BB, R.Begin, R.End, R.NumRegionInstrs);
192 }
193
~OverrideLegacyStrategy()194 ~OverrideLegacyStrategy() {
195 Sch.BaseClass::exitRegion();
196 Sch.BaseClass::finishBlock();
197 Sch.SchedImpl.release();
198 Sch.SchedImpl = std::move(SaveSchedImpl);
199 }
200
schedule()201 void schedule() {
202 assert(Sch.RegionBegin == Rgn.Begin && Sch.RegionEnd == Rgn.End);
203 LLVM_DEBUG(dbgs() << "\nScheduling ";
204 printRegion(dbgs(), Rgn.Begin, Rgn.End, Sch.LIS, 2));
205 Sch.BaseClass::schedule();
206
207 // Unfortunatelly placeDebugValues incorrectly modifies RegionEnd, restore
208 Sch.RegionEnd = Rgn.End;
209 //assert(Rgn.End == Sch.RegionEnd);
210 Rgn.Begin = Sch.RegionBegin;
211 Rgn.MaxPressure.clear();
212 }
213
restoreOrder()214 void restoreOrder() {
215 assert(Sch.RegionBegin == Rgn.Begin && Sch.RegionEnd == Rgn.End);
216 // DAG SUnits are stored using original region's order
217 // so just use SUnits as the restoring schedule
218 Sch.scheduleRegion(Rgn, Sch.SUnits, SaveMaxRP);
219 }
220 };
221
222 namespace {
223
224 // just a stub to make base class happy
225 class SchedStrategyStub : public MachineSchedStrategy {
226 public:
shouldTrackPressure() const227 bool shouldTrackPressure() const override { return false; }
shouldTrackLaneMasks() const228 bool shouldTrackLaneMasks() const override { return false; }
initialize(ScheduleDAGMI * DAG)229 void initialize(ScheduleDAGMI *DAG) override {}
pickNode(bool & IsTopNode)230 SUnit *pickNode(bool &IsTopNode) override { return nullptr; }
schedNode(SUnit * SU,bool IsTopNode)231 void schedNode(SUnit *SU, bool IsTopNode) override {}
releaseTopNode(SUnit * SU)232 void releaseTopNode(SUnit *SU) override {}
releaseBottomNode(SUnit * SU)233 void releaseBottomNode(SUnit *SU) override {}
234 };
235
236 } // end anonymous namespace
237
GCNIterativeScheduler(MachineSchedContext * C,StrategyKind S)238 GCNIterativeScheduler::GCNIterativeScheduler(MachineSchedContext *C,
239 StrategyKind S)
240 : BaseClass(C, std::make_unique<SchedStrategyStub>())
241 , Context(C)
242 , Strategy(S)
243 , UPTracker(*LIS) {
244 }
245
246 // returns max pressure for a region
247 GCNRegPressure
getRegionPressure(MachineBasicBlock::iterator Begin,MachineBasicBlock::iterator End) const248 GCNIterativeScheduler::getRegionPressure(MachineBasicBlock::iterator Begin,
249 MachineBasicBlock::iterator End)
250 const {
251 // For the purpose of pressure tracking bottom inst of the region should
252 // be also processed. End is either BB end, BB terminator inst or sched
253 // boundary inst.
254 auto const BBEnd = Begin->getParent()->end();
255 auto const BottomMI = End == BBEnd ? std::prev(End) : End;
256
257 // scheduleRegions walks bottom to top, so its likely we just get next
258 // instruction to track
259 auto AfterBottomMI = std::next(BottomMI);
260 if (AfterBottomMI == BBEnd ||
261 &*AfterBottomMI != UPTracker.getLastTrackedMI()) {
262 UPTracker.reset(*BottomMI);
263 } else {
264 assert(UPTracker.isValid());
265 }
266
267 for (auto I = BottomMI; I != Begin; --I)
268 UPTracker.recede(*I);
269
270 UPTracker.recede(*Begin);
271
272 assert(UPTracker.isValid() ||
273 (dbgs() << "Tracked region ",
274 printRegion(dbgs(), Begin, End, LIS), false));
275 return UPTracker.moveMaxPressure();
276 }
277
278 // returns max pressure for a tentative schedule
279 template <typename Range> GCNRegPressure
getSchedulePressure(const Region & R,Range && Schedule) const280 GCNIterativeScheduler::getSchedulePressure(const Region &R,
281 Range &&Schedule) const {
282 auto const BBEnd = R.Begin->getParent()->end();
283 GCNUpwardRPTracker RPTracker(*LIS);
284 if (R.End != BBEnd) {
285 // R.End points to the boundary instruction but the
286 // schedule doesn't include it
287 RPTracker.reset(*R.End);
288 RPTracker.recede(*R.End);
289 } else {
290 // R.End doesn't point to the boundary instruction
291 RPTracker.reset(*std::prev(BBEnd));
292 }
293 for (auto I = Schedule.end(), B = Schedule.begin(); I != B;) {
294 RPTracker.recede(*getMachineInstr(*--I));
295 }
296 return RPTracker.moveMaxPressure();
297 }
298
enterRegion(MachineBasicBlock * BB,MachineBasicBlock::iterator Begin,MachineBasicBlock::iterator End,unsigned NumRegionInstrs)299 void GCNIterativeScheduler::enterRegion(MachineBasicBlock *BB, // overriden
300 MachineBasicBlock::iterator Begin,
301 MachineBasicBlock::iterator End,
302 unsigned NumRegionInstrs) {
303 BaseClass::enterRegion(BB, Begin, End, NumRegionInstrs);
304 if (NumRegionInstrs > 2) {
305 Regions.push_back(
306 new (Alloc.Allocate())
307 Region { Begin, End, NumRegionInstrs,
308 getRegionPressure(Begin, End), nullptr });
309 }
310 }
311
schedule()312 void GCNIterativeScheduler::schedule() { // overriden
313 // do nothing
314 LLVM_DEBUG(printLivenessInfo(dbgs(), RegionBegin, RegionEnd, LIS);
315 if (!Regions.empty() && Regions.back()->Begin == RegionBegin) {
316 dbgs() << "Max RP: ";
317 Regions.back()->MaxPressure.print(
318 dbgs(), &MF.getSubtarget<GCNSubtarget>());
319 } dbgs()
320 << '\n';);
321 }
322
finalizeSchedule()323 void GCNIterativeScheduler::finalizeSchedule() { // overriden
324 if (Regions.empty())
325 return;
326 switch (Strategy) {
327 case SCHEDULE_MINREGONLY: scheduleMinReg(); break;
328 case SCHEDULE_MINREGFORCED: scheduleMinReg(true); break;
329 case SCHEDULE_LEGACYMAXOCCUPANCY: scheduleLegacyMaxOccupancy(); break;
330 case SCHEDULE_ILP: scheduleILP(false); break;
331 }
332 }
333
334 // Detach schedule from SUnits and interleave it with debug values.
335 // Returned schedule becomes independent of DAG state.
336 std::vector<MachineInstr*>
detachSchedule(ScheduleRef Schedule) const337 GCNIterativeScheduler::detachSchedule(ScheduleRef Schedule) const {
338 std::vector<MachineInstr*> Res;
339 Res.reserve(Schedule.size() * 2);
340
341 if (FirstDbgValue)
342 Res.push_back(FirstDbgValue);
343
344 const auto DbgB = DbgValues.begin(), DbgE = DbgValues.end();
345 for (auto SU : Schedule) {
346 Res.push_back(SU->getInstr());
347 const auto &D = std::find_if(DbgB, DbgE, [SU](decltype(*DbgB) &P) {
348 return P.second == SU->getInstr();
349 });
350 if (D != DbgE)
351 Res.push_back(D->first);
352 }
353 return Res;
354 }
355
setBestSchedule(Region & R,ScheduleRef Schedule,const GCNRegPressure & MaxRP)356 void GCNIterativeScheduler::setBestSchedule(Region &R,
357 ScheduleRef Schedule,
358 const GCNRegPressure &MaxRP) {
359 R.BestSchedule.reset(
360 new TentativeSchedule{ detachSchedule(Schedule), MaxRP });
361 }
362
scheduleBest(Region & R)363 void GCNIterativeScheduler::scheduleBest(Region &R) {
364 assert(R.BestSchedule.get() && "No schedule specified");
365 scheduleRegion(R, R.BestSchedule->Schedule, R.BestSchedule->MaxPressure);
366 R.BestSchedule.reset();
367 }
368
369 // minimal required region scheduler, works for ranges of SUnits*,
370 // SUnits or MachineIntrs*
371 template <typename Range>
scheduleRegion(Region & R,Range && Schedule,const GCNRegPressure & MaxRP)372 void GCNIterativeScheduler::scheduleRegion(Region &R, Range &&Schedule,
373 const GCNRegPressure &MaxRP) {
374 assert(RegionBegin == R.Begin && RegionEnd == R.End);
375 assert(LIS != nullptr);
376 #ifndef NDEBUG
377 const auto SchedMaxRP = getSchedulePressure(R, Schedule);
378 #endif
379 auto BB = R.Begin->getParent();
380 auto Top = R.Begin;
381 for (const auto &I : Schedule) {
382 auto MI = getMachineInstr(I);
383 if (MI != &*Top) {
384 BB->remove(MI);
385 BB->insert(Top, MI);
386 if (!MI->isDebugInstr())
387 LIS->handleMove(*MI, true);
388 }
389 if (!MI->isDebugInstr()) {
390 // Reset read - undef flags and update them later.
391 for (auto &Op : MI->operands())
392 if (Op.isReg() && Op.isDef())
393 Op.setIsUndef(false);
394
395 RegisterOperands RegOpers;
396 RegOpers.collect(*MI, *TRI, MRI, /*ShouldTrackLaneMasks*/true,
397 /*IgnoreDead*/false);
398 // Adjust liveness and add missing dead+read-undef flags.
399 auto SlotIdx = LIS->getInstructionIndex(*MI).getRegSlot();
400 RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx, MI);
401 }
402 Top = std::next(MI->getIterator());
403 }
404 RegionBegin = getMachineInstr(Schedule.front());
405
406 // Schedule consisting of MachineInstr* is considered 'detached'
407 // and already interleaved with debug values
408 if (!std::is_same<decltype(*Schedule.begin()), MachineInstr*>::value) {
409 placeDebugValues();
410 // Unfortunatelly placeDebugValues incorrectly modifies RegionEnd, restore
411 //assert(R.End == RegionEnd);
412 RegionEnd = R.End;
413 }
414
415 R.Begin = RegionBegin;
416 R.MaxPressure = MaxRP;
417
418 #ifndef NDEBUG
419 const auto RegionMaxRP = getRegionPressure(R);
420 const auto &ST = MF.getSubtarget<GCNSubtarget>();
421 #endif
422 assert((SchedMaxRP == RegionMaxRP && (MaxRP.empty() || SchedMaxRP == MaxRP))
423 || (dbgs() << "Max RP mismatch!!!\n"
424 "RP for schedule (calculated): ",
425 SchedMaxRP.print(dbgs(), &ST),
426 dbgs() << "RP for schedule (reported): ",
427 MaxRP.print(dbgs(), &ST),
428 dbgs() << "RP after scheduling: ",
429 RegionMaxRP.print(dbgs(), &ST),
430 false));
431 }
432
433 // Sort recorded regions by pressure - highest at the front
sortRegionsByPressure(unsigned TargetOcc)434 void GCNIterativeScheduler::sortRegionsByPressure(unsigned TargetOcc) {
435 const auto &ST = MF.getSubtarget<GCNSubtarget>();
436 llvm::sort(Regions, [&ST, TargetOcc](const Region *R1, const Region *R2) {
437 return R2->MaxPressure.less(ST, R1->MaxPressure, TargetOcc);
438 });
439 }
440
441 ///////////////////////////////////////////////////////////////////////////////
442 // Legacy MaxOccupancy Strategy
443
444 // Tries to increase occupancy applying minreg scheduler for a sequence of
445 // most demanding regions. Obtained schedules are saved as BestSchedule for a
446 // region.
447 // TargetOcc is the best achievable occupancy for a kernel.
448 // Returns better occupancy on success or current occupancy on fail.
449 // BestSchedules aren't deleted on fail.
tryMaximizeOccupancy(unsigned TargetOcc)450 unsigned GCNIterativeScheduler::tryMaximizeOccupancy(unsigned TargetOcc) {
451 // TODO: assert Regions are sorted descending by pressure
452 const auto &ST = MF.getSubtarget<GCNSubtarget>();
453 const auto Occ = Regions.front()->MaxPressure.getOccupancy(ST);
454 LLVM_DEBUG(dbgs() << "Trying to improve occupancy, target = " << TargetOcc
455 << ", current = " << Occ << '\n');
456
457 auto NewOcc = TargetOcc;
458 for (auto R : Regions) {
459 if (R->MaxPressure.getOccupancy(ST) >= NewOcc)
460 break;
461
462 LLVM_DEBUG(printRegion(dbgs(), R->Begin, R->End, LIS, 3);
463 printLivenessInfo(dbgs(), R->Begin, R->End, LIS));
464
465 BuildDAG DAG(*R, *this);
466 const auto MinSchedule = makeMinRegSchedule(DAG.getTopRoots(), *this);
467 const auto MaxRP = getSchedulePressure(*R, MinSchedule);
468 LLVM_DEBUG(dbgs() << "Occupancy improvement attempt:\n";
469 printSchedRP(dbgs(), R->MaxPressure, MaxRP));
470
471 NewOcc = std::min(NewOcc, MaxRP.getOccupancy(ST));
472 if (NewOcc <= Occ)
473 break;
474
475 setBestSchedule(*R, MinSchedule, MaxRP);
476 }
477 LLVM_DEBUG(dbgs() << "New occupancy = " << NewOcc
478 << ", prev occupancy = " << Occ << '\n');
479 if (NewOcc > Occ) {
480 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
481 MFI->increaseOccupancy(MF, NewOcc);
482 }
483
484 return std::max(NewOcc, Occ);
485 }
486
scheduleLegacyMaxOccupancy(bool TryMaximizeOccupancy)487 void GCNIterativeScheduler::scheduleLegacyMaxOccupancy(
488 bool TryMaximizeOccupancy) {
489 const auto &ST = MF.getSubtarget<GCNSubtarget>();
490 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
491 auto TgtOcc = MFI->getMinAllowedOccupancy();
492
493 sortRegionsByPressure(TgtOcc);
494 auto Occ = Regions.front()->MaxPressure.getOccupancy(ST);
495
496 if (TryMaximizeOccupancy && Occ < TgtOcc)
497 Occ = tryMaximizeOccupancy(TgtOcc);
498
499 // This is really weird but for some magic scheduling regions twice
500 // gives performance improvement
501 const int NumPasses = Occ < TgtOcc ? 2 : 1;
502
503 TgtOcc = std::min(Occ, TgtOcc);
504 LLVM_DEBUG(dbgs() << "Scheduling using default scheduler, "
505 "target occupancy = "
506 << TgtOcc << '\n');
507 GCNMaxOccupancySchedStrategy LStrgy(Context);
508 unsigned FinalOccupancy = std::min(Occ, MFI->getOccupancy());
509
510 for (int I = 0; I < NumPasses; ++I) {
511 // running first pass with TargetOccupancy = 0 mimics previous scheduling
512 // approach and is a performance magic
513 LStrgy.setTargetOccupancy(I == 0 ? 0 : TgtOcc);
514 for (auto R : Regions) {
515 OverrideLegacyStrategy Ovr(*R, LStrgy, *this);
516
517 Ovr.schedule();
518 const auto RP = getRegionPressure(*R);
519 LLVM_DEBUG(printSchedRP(dbgs(), R->MaxPressure, RP));
520
521 if (RP.getOccupancy(ST) < TgtOcc) {
522 LLVM_DEBUG(dbgs() << "Didn't fit into target occupancy O" << TgtOcc);
523 if (R->BestSchedule.get() &&
524 R->BestSchedule->MaxPressure.getOccupancy(ST) >= TgtOcc) {
525 LLVM_DEBUG(dbgs() << ", scheduling minimal register\n");
526 scheduleBest(*R);
527 } else {
528 LLVM_DEBUG(dbgs() << ", restoring\n");
529 Ovr.restoreOrder();
530 assert(R->MaxPressure.getOccupancy(ST) >= TgtOcc);
531 }
532 }
533 FinalOccupancy = std::min(FinalOccupancy, RP.getOccupancy(ST));
534 }
535 }
536 MFI->limitOccupancy(FinalOccupancy);
537 }
538
539 ///////////////////////////////////////////////////////////////////////////////
540 // Minimal Register Strategy
541
scheduleMinReg(bool force)542 void GCNIterativeScheduler::scheduleMinReg(bool force) {
543 const auto &ST = MF.getSubtarget<GCNSubtarget>();
544 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
545 const auto TgtOcc = MFI->getOccupancy();
546 sortRegionsByPressure(TgtOcc);
547
548 auto MaxPressure = Regions.front()->MaxPressure;
549 for (auto R : Regions) {
550 if (!force && R->MaxPressure.less(ST, MaxPressure, TgtOcc))
551 break;
552
553 BuildDAG DAG(*R, *this);
554 const auto MinSchedule = makeMinRegSchedule(DAG.getTopRoots(), *this);
555
556 const auto RP = getSchedulePressure(*R, MinSchedule);
557 LLVM_DEBUG(if (R->MaxPressure.less(ST, RP, TgtOcc)) {
558 dbgs() << "\nWarning: Pressure becomes worse after minreg!";
559 printSchedRP(dbgs(), R->MaxPressure, RP);
560 });
561
562 if (!force && MaxPressure.less(ST, RP, TgtOcc))
563 break;
564
565 scheduleRegion(*R, MinSchedule, RP);
566 LLVM_DEBUG(printSchedResult(dbgs(), R, RP));
567
568 MaxPressure = RP;
569 }
570 }
571
572 ///////////////////////////////////////////////////////////////////////////////
573 // ILP scheduler port
574
scheduleILP(bool TryMaximizeOccupancy)575 void GCNIterativeScheduler::scheduleILP(
576 bool TryMaximizeOccupancy) {
577 const auto &ST = MF.getSubtarget<GCNSubtarget>();
578 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
579 auto TgtOcc = MFI->getMinAllowedOccupancy();
580
581 sortRegionsByPressure(TgtOcc);
582 auto Occ = Regions.front()->MaxPressure.getOccupancy(ST);
583
584 if (TryMaximizeOccupancy && Occ < TgtOcc)
585 Occ = tryMaximizeOccupancy(TgtOcc);
586
587 TgtOcc = std::min(Occ, TgtOcc);
588 LLVM_DEBUG(dbgs() << "Scheduling using default scheduler, "
589 "target occupancy = "
590 << TgtOcc << '\n');
591
592 unsigned FinalOccupancy = std::min(Occ, MFI->getOccupancy());
593 for (auto R : Regions) {
594 BuildDAG DAG(*R, *this);
595 const auto ILPSchedule = makeGCNILPScheduler(DAG.getBottomRoots(), *this);
596
597 const auto RP = getSchedulePressure(*R, ILPSchedule);
598 LLVM_DEBUG(printSchedRP(dbgs(), R->MaxPressure, RP));
599
600 if (RP.getOccupancy(ST) < TgtOcc) {
601 LLVM_DEBUG(dbgs() << "Didn't fit into target occupancy O" << TgtOcc);
602 if (R->BestSchedule.get() &&
603 R->BestSchedule->MaxPressure.getOccupancy(ST) >= TgtOcc) {
604 LLVM_DEBUG(dbgs() << ", scheduling minimal register\n");
605 scheduleBest(*R);
606 }
607 } else {
608 scheduleRegion(*R, ILPSchedule, RP);
609 LLVM_DEBUG(printSchedResult(dbgs(), R, RP));
610 FinalOccupancy = std::min(FinalOccupancy, RP.getOccupancy(ST));
611 }
612 }
613 MFI->limitOccupancy(FinalOccupancy);
614 }
615