xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/GCNIterativeScheduler.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
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 /// \file
10 /// This file implements the class GCNIterativeScheduler.
11 ///
12 //===----------------------------------------------------------------------===//
13 
14 #include "GCNIterativeScheduler.h"
15 #include "GCNSchedStrategy.h"
16 #include "SIMachineFunctionInfo.h"
17 
18 using namespace llvm;
19 
20 #define DEBUG_TYPE "machine-scheduler"
21 
22 namespace llvm {
23 
24 std::vector<const SUnit *> makeMinRegSchedule(ArrayRef<const SUnit *> TopRoots,
25                                               const ScheduleDAG &DAG);
26 
27 std::vector<const SUnit *> makeGCNILPScheduler(ArrayRef<const SUnit *> BotRoots,
28                                                const ScheduleDAG &DAG);
29 } // namespace llvm
30 
31 // shim accessors for different order containers
getMachineInstr(MachineInstr * MI)32 static inline MachineInstr *getMachineInstr(MachineInstr *MI) {
33   return MI;
34 }
getMachineInstr(const SUnit * SU)35 static inline MachineInstr *getMachineInstr(const SUnit *SU) {
36   return SU->getInstr();
37 }
getMachineInstr(const SUnit & SU)38 static inline MachineInstr *getMachineInstr(const SUnit &SU) {
39   return SU.getInstr();
40 }
41 
42 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
43 LLVM_DUMP_METHOD
printRegion(raw_ostream & OS,MachineBasicBlock::iterator Begin,MachineBasicBlock::iterator End,const LiveIntervals * LIS,unsigned MaxInstNum=std::numeric_limits<unsigned>::max ())44 static void printRegion(raw_ostream &OS,
45                         MachineBasicBlock::iterator Begin,
46                         MachineBasicBlock::iterator End,
47                         const LiveIntervals *LIS,
48                         unsigned MaxInstNum =
49                           std::numeric_limits<unsigned>::max()) {
50   auto BB = Begin->getParent();
51   OS << BB->getParent()->getName() << ":" << printMBBReference(*BB) << ' '
52      << BB->getName() << ":\n";
53   auto I = Begin;
54   MaxInstNum = std::max(MaxInstNum, 1u);
55   for (; I != End && MaxInstNum; ++I, --MaxInstNum) {
56     if (!I->isDebugInstr() && LIS)
57       OS << LIS->getInstructionIndex(*I);
58     OS << '\t' << *I;
59   }
60   if (I != End) {
61     OS << "\t...\n";
62     I = std::prev(End);
63     if (!I->isDebugInstr() && LIS)
64       OS << LIS->getInstructionIndex(*I);
65     OS << '\t' << *I;
66   }
67   if (End != BB->end()) { // print boundary inst if present
68     OS << "----\n";
69     if (LIS) OS << LIS->getInstructionIndex(*End) << '\t';
70     OS << *End;
71   }
72 }
73 
74 LLVM_DUMP_METHOD
printLivenessInfo(raw_ostream & OS,MachineBasicBlock::iterator Begin,MachineBasicBlock::iterator End,const LiveIntervals * LIS)75 static void printLivenessInfo(raw_ostream &OS,
76                               MachineBasicBlock::iterator Begin,
77                               MachineBasicBlock::iterator End,
78                               const LiveIntervals *LIS) {
79   const auto BB = Begin->getParent();
80   const auto &MRI = BB->getParent()->getRegInfo();
81 
82   const auto LiveIns = getLiveRegsBefore(*Begin, *LIS);
83   OS << "LIn RP: " << print(getRegPressure(MRI, LiveIns));
84 
85   const auto BottomMI = End == BB->end() ? std::prev(End) : End;
86   const auto LiveOuts = getLiveRegsAfter(*BottomMI, *LIS);
87   OS << "LOt RP: " << print(getRegPressure(MRI, LiveOuts));
88 }
89 
90 LLVM_DUMP_METHOD
printRegions(raw_ostream & OS) const91 void GCNIterativeScheduler::printRegions(raw_ostream &OS) const {
92   const auto &ST = MF.getSubtarget<GCNSubtarget>();
93   for (const auto R : Regions) {
94     OS << "Region to schedule ";
95     printRegion(OS, R->Begin, R->End, LIS, 1);
96     printLivenessInfo(OS, R->Begin, R->End, LIS);
97     OS << "Max RP: " << print(R->MaxPressure, &ST);
98   }
99 }
100 
101 LLVM_DUMP_METHOD
printSchedResult(raw_ostream & OS,const Region * R,const GCNRegPressure & RP) const102 void GCNIterativeScheduler::printSchedResult(raw_ostream &OS,
103                                              const Region *R,
104                                              const GCNRegPressure &RP) const {
105   OS << "\nAfter scheduling ";
106   printRegion(OS, R->Begin, R->End, LIS);
107   printSchedRP(OS, R->MaxPressure, RP);
108   OS << '\n';
109 }
110 
111 LLVM_DUMP_METHOD
printSchedRP(raw_ostream & OS,const GCNRegPressure & Before,const GCNRegPressure & After) const112 void GCNIterativeScheduler::printSchedRP(raw_ostream &OS,
113                                          const GCNRegPressure &Before,
114                                          const GCNRegPressure &After) const {
115   const auto &ST = MF.getSubtarget<GCNSubtarget>();
116   OS << "RP before: " << print(Before, &ST)
117      << "RP after:  " << print(After, &ST);
118 }
119 #endif
120 
121 // DAG builder helper
122 class GCNIterativeScheduler::BuildDAG {
123   GCNIterativeScheduler &Sch;
124   SmallVector<SUnit *, 8> TopRoots;
125 
126   SmallVector<SUnit*, 8> BotRoots;
127 public:
BuildDAG(const Region & R,GCNIterativeScheduler & _Sch)128   BuildDAG(const Region &R, GCNIterativeScheduler &_Sch)
129     : Sch(_Sch) {
130     auto BB = R.Begin->getParent();
131     Sch.BaseClass::startBlock(BB);
132     Sch.BaseClass::enterRegion(BB, R.Begin, R.End, R.NumRegionInstrs);
133 
134     Sch.buildSchedGraph(Sch.AA, nullptr, nullptr, nullptr,
135                         /*TrackLaneMask*/true);
136     Sch.Topo.InitDAGTopologicalSorting();
137     Sch.findRootsAndBiasEdges(TopRoots, BotRoots);
138   }
139 
~BuildDAG()140   ~BuildDAG() {
141     Sch.BaseClass::exitRegion();
142     Sch.BaseClass::finishBlock();
143   }
144 
getTopRoots() const145   ArrayRef<const SUnit *> getTopRoots() const {
146     return TopRoots;
147   }
getBottomRoots() const148   ArrayRef<SUnit*> getBottomRoots() const {
149     return BotRoots;
150   }
151 };
152 
153 class GCNIterativeScheduler::OverrideLegacyStrategy {
154   GCNIterativeScheduler &Sch;
155   Region &Rgn;
156   std::unique_ptr<MachineSchedStrategy> SaveSchedImpl;
157   GCNRegPressure SaveMaxRP;
158 
159 public:
OverrideLegacyStrategy(Region & R,MachineSchedStrategy & OverrideStrategy,GCNIterativeScheduler & _Sch)160   OverrideLegacyStrategy(Region &R,
161                          MachineSchedStrategy &OverrideStrategy,
162                          GCNIterativeScheduler &_Sch)
163     : Sch(_Sch)
164     , Rgn(R)
165     , SaveSchedImpl(std::move(_Sch.SchedImpl))
166     , SaveMaxRP(R.MaxPressure) {
167     Sch.SchedImpl.reset(&OverrideStrategy);
168     auto BB = R.Begin->getParent();
169     Sch.BaseClass::startBlock(BB);
170     Sch.BaseClass::enterRegion(BB, R.Begin, R.End, R.NumRegionInstrs);
171   }
172 
~OverrideLegacyStrategy()173   ~OverrideLegacyStrategy() {
174     Sch.BaseClass::exitRegion();
175     Sch.BaseClass::finishBlock();
176     Sch.SchedImpl.release();
177     Sch.SchedImpl = std::move(SaveSchedImpl);
178   }
179 
schedule()180   void schedule() {
181     assert(Sch.RegionBegin == Rgn.Begin && Sch.RegionEnd == Rgn.End);
182     LLVM_DEBUG(dbgs() << "\nScheduling ";
183                printRegion(dbgs(), Rgn.Begin, Rgn.End, Sch.LIS, 2));
184     Sch.BaseClass::schedule();
185 
186     // Unfortunately placeDebugValues incorrectly modifies RegionEnd, restore
187     Sch.RegionEnd = Rgn.End;
188     //assert(Rgn.End == Sch.RegionEnd);
189     Rgn.Begin = Sch.RegionBegin;
190     Rgn.MaxPressure.clear();
191   }
192 
restoreOrder()193   void restoreOrder() {
194     assert(Sch.RegionBegin == Rgn.Begin && Sch.RegionEnd == Rgn.End);
195     // DAG SUnits are stored using original region's order
196     // so just use SUnits as the restoring schedule
197     Sch.scheduleRegion(Rgn, Sch.SUnits, SaveMaxRP);
198   }
199 };
200 
201 namespace {
202 
203 // just a stub to make base class happy
204 class SchedStrategyStub : public MachineSchedStrategy {
205 public:
shouldTrackPressure() const206   bool shouldTrackPressure() const override { return false; }
shouldTrackLaneMasks() const207   bool shouldTrackLaneMasks() const override { return false; }
initialize(ScheduleDAGMI * DAG)208   void initialize(ScheduleDAGMI *DAG) override {}
pickNode(bool & IsTopNode)209   SUnit *pickNode(bool &IsTopNode) override { return nullptr; }
schedNode(SUnit * SU,bool IsTopNode)210   void schedNode(SUnit *SU, bool IsTopNode) override {}
releaseTopNode(SUnit * SU)211   void releaseTopNode(SUnit *SU) override {}
releaseBottomNode(SUnit * SU)212   void releaseBottomNode(SUnit *SU) override {}
213 };
214 
215 } // end anonymous namespace
216 
GCNIterativeScheduler(MachineSchedContext * C,StrategyKind S)217 GCNIterativeScheduler::GCNIterativeScheduler(MachineSchedContext *C,
218                                              StrategyKind S)
219   : BaseClass(C, std::make_unique<SchedStrategyStub>())
220   , Context(C)
221   , Strategy(S)
222   , UPTracker(*LIS) {
223 }
224 
225 // returns max pressure for a region
226 GCNRegPressure
getRegionPressure(MachineBasicBlock::iterator Begin,MachineBasicBlock::iterator End) const227 GCNIterativeScheduler::getRegionPressure(MachineBasicBlock::iterator Begin,
228                                          MachineBasicBlock::iterator End)
229   const {
230   // For the purpose of pressure tracking bottom inst of the region should
231   // be also processed. End is either BB end, BB terminator inst or sched
232   // boundary inst.
233   auto const BBEnd = Begin->getParent()->end();
234   auto const BottomMI = End == BBEnd ? std::prev(End) : End;
235 
236   // scheduleRegions walks bottom to top, so its likely we just get next
237   // instruction to track
238   auto AfterBottomMI = std::next(BottomMI);
239   if (AfterBottomMI == BBEnd ||
240       &*AfterBottomMI != UPTracker.getLastTrackedMI()) {
241     UPTracker.reset(*BottomMI);
242   } else {
243     assert(UPTracker.isValid());
244   }
245 
246   for (auto I = BottomMI; I != Begin; --I)
247     UPTracker.recede(*I);
248 
249   UPTracker.recede(*Begin);
250 
251   assert(UPTracker.isValid() ||
252          (dbgs() << "Tracked region ",
253           printRegion(dbgs(), Begin, End, LIS), false));
254   return UPTracker.getMaxPressureAndReset();
255 }
256 
257 // returns max pressure for a tentative schedule
258 template <typename Range> GCNRegPressure
getSchedulePressure(const Region & R,Range && Schedule) const259 GCNIterativeScheduler::getSchedulePressure(const Region &R,
260                                            Range &&Schedule) const {
261   auto const BBEnd = R.Begin->getParent()->end();
262   GCNUpwardRPTracker RPTracker(*LIS);
263   if (R.End != BBEnd) {
264     // R.End points to the boundary instruction but the
265     // schedule doesn't include it
266     RPTracker.reset(*R.End);
267     RPTracker.recede(*R.End);
268   } else {
269     // R.End doesn't point to the boundary instruction
270     RPTracker.reset(*std::prev(BBEnd));
271   }
272   for (auto I = Schedule.end(), B = Schedule.begin(); I != B;) {
273     RPTracker.recede(*getMachineInstr(*--I));
274   }
275   return RPTracker.getMaxPressureAndReset();
276 }
277 
enterRegion(MachineBasicBlock * BB,MachineBasicBlock::iterator Begin,MachineBasicBlock::iterator End,unsigned NumRegionInstrs)278 void GCNIterativeScheduler::enterRegion(MachineBasicBlock *BB, // overridden
279                                         MachineBasicBlock::iterator Begin,
280                                         MachineBasicBlock::iterator End,
281                                         unsigned NumRegionInstrs) {
282   BaseClass::enterRegion(BB, Begin, End, NumRegionInstrs);
283   if (NumRegionInstrs > 2) {
284     Regions.push_back(
285       new (Alloc.Allocate())
286       Region { Begin, End, NumRegionInstrs,
287                getRegionPressure(Begin, End), nullptr });
288   }
289 }
290 
schedule()291 void GCNIterativeScheduler::schedule() { // overridden
292   // do nothing
293   LLVM_DEBUG(printLivenessInfo(dbgs(), RegionBegin, RegionEnd, LIS);
294              if (!Regions.empty() && Regions.back()->Begin == RegionBegin) {
295                dbgs() << "Max RP: "
296                       << print(Regions.back()->MaxPressure,
297                                &MF.getSubtarget<GCNSubtarget>());
298              } dbgs()
299              << '\n';);
300 }
301 
finalizeSchedule()302 void GCNIterativeScheduler::finalizeSchedule() { // overridden
303   if (Regions.empty())
304     return;
305   switch (Strategy) {
306   case SCHEDULE_MINREGONLY: scheduleMinReg(); break;
307   case SCHEDULE_MINREGFORCED: scheduleMinReg(true); break;
308   case SCHEDULE_LEGACYMAXOCCUPANCY: scheduleLegacyMaxOccupancy(); break;
309   case SCHEDULE_ILP: scheduleILP(false); break;
310   }
311 }
312 
313 // Detach schedule from SUnits and interleave it with debug values.
314 // Returned schedule becomes independent of DAG state.
315 std::vector<MachineInstr*>
detachSchedule(ScheduleRef Schedule) const316 GCNIterativeScheduler::detachSchedule(ScheduleRef Schedule) const {
317   std::vector<MachineInstr*> Res;
318   Res.reserve(Schedule.size() * 2);
319 
320   if (FirstDbgValue)
321     Res.push_back(FirstDbgValue);
322 
323   const auto DbgB = DbgValues.begin(), DbgE = DbgValues.end();
324   for (const auto *SU : Schedule) {
325     Res.push_back(SU->getInstr());
326     const auto &D = std::find_if(DbgB, DbgE, [SU](decltype(*DbgB) &P) {
327       return P.second == SU->getInstr();
328     });
329     if (D != DbgE)
330       Res.push_back(D->first);
331   }
332   return Res;
333 }
334 
setBestSchedule(Region & R,ScheduleRef Schedule,const GCNRegPressure & MaxRP)335 void GCNIterativeScheduler::setBestSchedule(Region &R,
336                                             ScheduleRef Schedule,
337                                             const GCNRegPressure &MaxRP) {
338   R.BestSchedule.reset(
339     new TentativeSchedule{ detachSchedule(Schedule), MaxRP });
340 }
341 
scheduleBest(Region & R)342 void GCNIterativeScheduler::scheduleBest(Region &R) {
343   assert(R.BestSchedule.get() && "No schedule specified");
344   scheduleRegion(R, R.BestSchedule->Schedule, R.BestSchedule->MaxPressure);
345   R.BestSchedule.reset();
346 }
347 
348 // minimal required region scheduler, works for ranges of SUnits*,
349 // SUnits or MachineIntrs*
350 template <typename Range>
scheduleRegion(Region & R,Range && Schedule,const GCNRegPressure & MaxRP)351 void GCNIterativeScheduler::scheduleRegion(Region &R, Range &&Schedule,
352                                            const GCNRegPressure &MaxRP) {
353   assert(RegionBegin == R.Begin && RegionEnd == R.End);
354   assert(LIS != nullptr);
355 #ifndef NDEBUG
356   const auto SchedMaxRP = getSchedulePressure(R, Schedule);
357 #endif
358   auto BB = R.Begin->getParent();
359   auto Top = R.Begin;
360   for (const auto &I : Schedule) {
361     auto MI = getMachineInstr(I);
362     if (MI != &*Top) {
363       BB->remove(MI);
364       BB->insert(Top, MI);
365       if (!MI->isDebugInstr())
366         LIS->handleMove(*MI, true);
367     }
368     if (!MI->isDebugInstr()) {
369       // Reset read - undef flags and update them later.
370       for (auto &Op : MI->all_defs())
371         Op.setIsUndef(false);
372 
373       RegisterOperands RegOpers;
374       RegOpers.collect(*MI, *TRI, MRI, /*ShouldTrackLaneMasks*/true,
375                                        /*IgnoreDead*/false);
376       // Adjust liveness and add missing dead+read-undef flags.
377       auto SlotIdx = LIS->getInstructionIndex(*MI).getRegSlot();
378       RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx, MI);
379     }
380     Top = std::next(MI->getIterator());
381   }
382   RegionBegin = getMachineInstr(Schedule.front());
383 
384   // Schedule consisting of MachineInstr* is considered 'detached'
385   // and already interleaved with debug values
386   if (!std::is_same_v<decltype(*Schedule.begin()), MachineInstr*>) {
387     placeDebugValues();
388     // Unfortunately placeDebugValues incorrectly modifies RegionEnd, restore
389     // assert(R.End == RegionEnd);
390     RegionEnd = R.End;
391   }
392 
393   R.Begin = RegionBegin;
394   R.MaxPressure = MaxRP;
395 
396 #ifndef NDEBUG
397   const auto RegionMaxRP = getRegionPressure(R);
398   const auto &ST = MF.getSubtarget<GCNSubtarget>();
399 #endif
400   assert(
401       (SchedMaxRP == RegionMaxRP && (MaxRP.empty() || SchedMaxRP == MaxRP)) ||
402       (dbgs() << "Max RP mismatch!!!\n"
403                  "RP for schedule (calculated): "
404               << print(SchedMaxRP, &ST)
405               << "RP for schedule (reported): " << print(MaxRP, &ST)
406               << "RP after scheduling: " << print(RegionMaxRP, &ST),
407        false));
408 }
409 
410 // Sort recorded regions by pressure - highest at the front
sortRegionsByPressure(unsigned TargetOcc)411 void GCNIterativeScheduler::sortRegionsByPressure(unsigned TargetOcc) {
412   llvm::sort(Regions, [this, TargetOcc](const Region *R1, const Region *R2) {
413     return R2->MaxPressure.less(MF, R1->MaxPressure, TargetOcc);
414   });
415 }
416 
417 ///////////////////////////////////////////////////////////////////////////////
418 // Legacy MaxOccupancy Strategy
419 
420 // Tries to increase occupancy applying minreg scheduler for a sequence of
421 // most demanding regions. Obtained schedules are saved as BestSchedule for a
422 // region.
423 // TargetOcc is the best achievable occupancy for a kernel.
424 // Returns better occupancy on success or current occupancy on fail.
425 // BestSchedules aren't deleted on fail.
tryMaximizeOccupancy(unsigned TargetOcc)426 unsigned GCNIterativeScheduler::tryMaximizeOccupancy(unsigned TargetOcc) {
427   // TODO: assert Regions are sorted descending by pressure
428   const auto &ST = MF.getSubtarget<GCNSubtarget>();
429   const auto Occ = Regions.front()->MaxPressure.getOccupancy(ST);
430   LLVM_DEBUG(dbgs() << "Trying to improve occupancy, target = " << TargetOcc
431                     << ", current = " << Occ << '\n');
432 
433   auto NewOcc = TargetOcc;
434   for (auto *R : Regions) {
435     if (R->MaxPressure.getOccupancy(ST) >= NewOcc)
436       break;
437 
438     LLVM_DEBUG(printRegion(dbgs(), R->Begin, R->End, LIS, 3);
439                printLivenessInfo(dbgs(), R->Begin, R->End, LIS));
440 
441     BuildDAG DAG(*R, *this);
442     const auto MinSchedule = makeMinRegSchedule(DAG.getTopRoots(), *this);
443     const auto MaxRP = getSchedulePressure(*R, MinSchedule);
444     LLVM_DEBUG(dbgs() << "Occupancy improvement attempt:\n";
445                printSchedRP(dbgs(), R->MaxPressure, MaxRP));
446 
447     NewOcc = std::min(NewOcc, MaxRP.getOccupancy(ST));
448     if (NewOcc <= Occ)
449       break;
450 
451     setBestSchedule(*R, MinSchedule, MaxRP);
452   }
453   LLVM_DEBUG(dbgs() << "New occupancy = " << NewOcc
454                     << ", prev occupancy = " << Occ << '\n');
455   if (NewOcc > Occ) {
456     SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
457     MFI->increaseOccupancy(MF, NewOcc);
458   }
459 
460   return std::max(NewOcc, Occ);
461 }
462 
scheduleLegacyMaxOccupancy(bool TryMaximizeOccupancy)463 void GCNIterativeScheduler::scheduleLegacyMaxOccupancy(
464   bool TryMaximizeOccupancy) {
465   const auto &ST = MF.getSubtarget<GCNSubtarget>();
466   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
467   auto TgtOcc = MFI->getMinAllowedOccupancy();
468 
469   sortRegionsByPressure(TgtOcc);
470   auto Occ = Regions.front()->MaxPressure.getOccupancy(ST);
471 
472   if (TryMaximizeOccupancy && Occ < TgtOcc)
473     Occ = tryMaximizeOccupancy(TgtOcc);
474 
475   // This is really weird but for some magic scheduling regions twice
476   // gives performance improvement
477   const int NumPasses = Occ < TgtOcc ? 2 : 1;
478 
479   TgtOcc = std::min(Occ, TgtOcc);
480   LLVM_DEBUG(dbgs() << "Scheduling using default scheduler, "
481                        "target occupancy = "
482                     << TgtOcc << '\n');
483   GCNMaxOccupancySchedStrategy LStrgy(Context);
484   unsigned FinalOccupancy = std::min(Occ, MFI->getOccupancy());
485 
486   for (int I = 0; I < NumPasses; ++I) {
487     // running first pass with TargetOccupancy = 0 mimics previous scheduling
488     // approach and is a performance magic
489     LStrgy.setTargetOccupancy(I == 0 ? 0 : TgtOcc);
490     for (auto *R : Regions) {
491       OverrideLegacyStrategy Ovr(*R, LStrgy, *this);
492 
493       Ovr.schedule();
494       const auto RP = getRegionPressure(*R);
495       LLVM_DEBUG(printSchedRP(dbgs(), R->MaxPressure, RP));
496 
497       if (RP.getOccupancy(ST) < TgtOcc) {
498         LLVM_DEBUG(dbgs() << "Didn't fit into target occupancy O" << TgtOcc);
499         if (R->BestSchedule.get() &&
500             R->BestSchedule->MaxPressure.getOccupancy(ST) >= TgtOcc) {
501           LLVM_DEBUG(dbgs() << ", scheduling minimal register\n");
502           scheduleBest(*R);
503         } else {
504           LLVM_DEBUG(dbgs() << ", restoring\n");
505           Ovr.restoreOrder();
506           assert(R->MaxPressure.getOccupancy(ST) >= TgtOcc);
507         }
508       }
509       FinalOccupancy = std::min(FinalOccupancy, RP.getOccupancy(ST));
510     }
511   }
512   MFI->limitOccupancy(FinalOccupancy);
513 }
514 
515 ///////////////////////////////////////////////////////////////////////////////
516 // Minimal Register Strategy
517 
scheduleMinReg(bool force)518 void GCNIterativeScheduler::scheduleMinReg(bool force) {
519   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
520   const auto TgtOcc = MFI->getOccupancy();
521   sortRegionsByPressure(TgtOcc);
522 
523   auto MaxPressure = Regions.front()->MaxPressure;
524   for (auto *R : Regions) {
525     if (!force && R->MaxPressure.less(MF, MaxPressure, TgtOcc))
526       break;
527 
528     BuildDAG DAG(*R, *this);
529     const auto MinSchedule = makeMinRegSchedule(DAG.getTopRoots(), *this);
530 
531     const auto RP = getSchedulePressure(*R, MinSchedule);
532     LLVM_DEBUG(if (R->MaxPressure.less(MF, RP, TgtOcc)) {
533       dbgs() << "\nWarning: Pressure becomes worse after minreg!";
534       printSchedRP(dbgs(), R->MaxPressure, RP);
535     });
536 
537     if (!force && MaxPressure.less(MF, RP, TgtOcc))
538       break;
539 
540     scheduleRegion(*R, MinSchedule, RP);
541     LLVM_DEBUG(printSchedResult(dbgs(), R, RP));
542 
543     MaxPressure = RP;
544   }
545 }
546 
547 ///////////////////////////////////////////////////////////////////////////////
548 // ILP scheduler port
549 
scheduleILP(bool TryMaximizeOccupancy)550 void GCNIterativeScheduler::scheduleILP(
551   bool TryMaximizeOccupancy) {
552   const auto &ST = MF.getSubtarget<GCNSubtarget>();
553   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
554   auto TgtOcc = MFI->getMinAllowedOccupancy();
555 
556   sortRegionsByPressure(TgtOcc);
557   auto Occ = Regions.front()->MaxPressure.getOccupancy(ST);
558 
559   if (TryMaximizeOccupancy && Occ < TgtOcc)
560     Occ = tryMaximizeOccupancy(TgtOcc);
561 
562   TgtOcc = std::min(Occ, TgtOcc);
563   LLVM_DEBUG(dbgs() << "Scheduling using default scheduler, "
564                        "target occupancy = "
565                     << TgtOcc << '\n');
566 
567   unsigned FinalOccupancy = std::min(Occ, MFI->getOccupancy());
568   for (auto *R : Regions) {
569     BuildDAG DAG(*R, *this);
570     const auto ILPSchedule = makeGCNILPScheduler(DAG.getBottomRoots(), *this);
571 
572     const auto RP = getSchedulePressure(*R, ILPSchedule);
573     LLVM_DEBUG(printSchedRP(dbgs(), R->MaxPressure, RP));
574 
575     if (RP.getOccupancy(ST) < TgtOcc) {
576       LLVM_DEBUG(dbgs() << "Didn't fit into target occupancy O" << TgtOcc);
577       if (R->BestSchedule.get() &&
578         R->BestSchedule->MaxPressure.getOccupancy(ST) >= TgtOcc) {
579         LLVM_DEBUG(dbgs() << ", scheduling minimal register\n");
580         scheduleBest(*R);
581       }
582     } else {
583       scheduleRegion(*R, ILPSchedule, RP);
584       LLVM_DEBUG(printSchedResult(dbgs(), R, RP));
585       FinalOccupancy = std::min(FinalOccupancy, RP.getOccupancy(ST));
586     }
587   }
588   MFI->limitOccupancy(FinalOccupancy);
589 }
590