1 //===-- SIMachineScheduler.cpp - SI Scheduler Interface -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// SI Machine Scheduler interface
13 //===----------------------------------------------------------------------===//
15 #include "SIMachineScheduler.h"
17 #include "SIInstrInfo.h"
18 #include "SIRegisterInfo.h"
19 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/CodeGen/LiveInterval.h"
23 #include "llvm/CodeGen/LiveIntervals.h"
24 #include "llvm/CodeGen/MachineInstr.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/MachineScheduler.h"
27 #include "llvm/CodeGen/RegisterPressure.h"
28 #include "llvm/CodeGen/SlotIndexes.h"
29 #include "llvm/CodeGen/TargetRegisterInfo.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/raw_ostream.h"
42 #define DEBUG_TYPE "machine-scheduler"
44 // This scheduler implements a different scheduling algorithm than
47 // There are several specific architecture behaviours that can't be modelled
48 // for GenericScheduler:
49 // . When accessing the result of an SGPR load instruction, you have to wait
50 // for all the SGPR load instructions before your current instruction to
52 // . When accessing the result of an VGPR load instruction, you have to wait
53 // for all the VGPR load instructions previous to the VGPR load instruction
54 // you are interested in to finish.
55 // . The less the register pressure, the best load latencies are hidden
57 // Moreover some specifities (like the fact a lot of instructions in the shader
58 // have few dependencies) makes the generic scheduler have some unpredictable
59 // behaviours. For example when register pressure becomes high, it can either
60 // manage to prevent register pressure from going too high, or it can
61 // increase register pressure even more than if it hadn't taken register
62 // pressure into account.
64 // Also some other bad behaviours are generated, like loading at the beginning
65 // of the shader a constant in VGPR you won't need until the end of the shader.
67 // The scheduling problem for SI can distinguish three main parts:
68 // . Hiding high latencies (texture sampling, etc)
69 // . Hiding low latencies (SGPR constant loading, etc)
70 // . Keeping register usage low for better latency hiding and general
73 // Some other things can also affect performance, but are hard to predict
74 // (cache usage, the fact the HW can issue several instructions from different
75 // wavefronts if different types, etc)
77 // This scheduler tries to solve the scheduling problem by dividing it into
78 // simpler sub-problems. It divides the instructions into blocks, schedules
79 // locally inside the blocks where it takes care of low latencies, and then
80 // chooses the order of the blocks by taking care of high latencies.
81 // Dividing the instructions into blocks helps control keeping register
84 // First the instructions are put into blocks.
85 // We want the blocks help control register usage and hide high latencies
86 // later. To help control register usage, we typically want all local
87 // computations, when for example you create a result that can be comsummed
88 // right away, to be contained in a block. Block inputs and outputs would
89 // typically be important results that are needed in several locations of
90 // the shader. Since we do want blocks to help hide high latencies, we want
91 // the instructions inside the block to have a minimal set of dependencies
92 // on high latencies. It will make it easy to pick blocks to hide specific
94 // The block creation algorithm is divided into several steps, and several
95 // variants can be tried during the scheduling process.
97 // Second the order of the instructions inside the blocks is chosen.
98 // At that step we do take into account only register usage and hiding
99 // low latency instructions
101 // Third the block order is chosen, there we try to hide high latencies
102 // and keep register usage low.
104 // After the third step, a pass is done to improve the hiding of low
107 // Actually when talking about 'low latency' or 'high latency' it includes
108 // both the latency to get the cache (or global mem) data go to the register,
109 // and the bandwidth limitations.
110 // Increasing the number of active wavefronts helps hide the former, but it
111 // doesn't solve the latter, thus why even if wavefront count is high, we have
112 // to try have as many instructions hiding high latencies as possible.
113 // The OpenCL doc says for example latency of 400 cycles for a global mem access,
114 // which is hidden by 10 instructions if the wavefront count is 10.
116 // Some figures taken from AMD docs:
117 // Both texture and constant L1 caches are 4-way associative with 64 bytes
119 // Constant cache is shared with 4 CUs.
120 // For texture sampling, the address generation unit receives 4 texture
121 // addresses per cycle, thus we could expect texture sampling latency to be
122 // equivalent to 4 instructions in the very best case (a VGPR is 64 work items,
123 // instructions in a wavefront group are executed every 4 cycles),
124 // or 16 instructions if the other wavefronts associated to the 3 other VALUs
125 // of the CU do texture sampling too. (Don't take these figures too seriously,
126 // as I'm not 100% sure of the computation)
127 // Data exports should get similar latency.
128 // For constant loading, the cache is shader with 4 CUs.
129 // The doc says "a throughput of 16B/cycle for each of the 4 Compute Unit"
130 // I guess if the other CU don't read the cache, it can go up to 64B/cycle.
131 // It means a simple s_buffer_load should take one instruction to hide, as
132 // well as a s_buffer_loadx2 and potentially a s_buffer_loadx8 if on the same
135 // As of today the driver doesn't preload the constants in cache, thus the
136 // first loads get extra latency. The doc says global memory access can be
137 // 300-600 cycles. We do not specially take that into account when scheduling
138 // As we expect the driver to be able to preload the constants soon.
144 static const char *getReasonStr(SIScheduleCandReason Reason) {
146 case NoCand: return "NOCAND";
147 case RegUsage: return "REGUSAGE";
148 case Latency: return "LATENCY";
149 case Successor: return "SUCCESSOR";
150 case Depth: return "DEPTH";
151 case NodeOrder: return "ORDER";
153 llvm_unreachable("Unknown reason!");
160 static bool tryLess(int TryVal, int CandVal,
161 SISchedulerCandidate &TryCand,
162 SISchedulerCandidate &Cand,
163 SIScheduleCandReason Reason) {
164 if (TryVal < CandVal) {
165 TryCand.Reason = Reason;
168 if (TryVal > CandVal) {
169 if (Cand.Reason > Reason)
170 Cand.Reason = Reason;
173 Cand.setRepeat(Reason);
177 static bool tryGreater(int TryVal, int CandVal,
178 SISchedulerCandidate &TryCand,
179 SISchedulerCandidate &Cand,
180 SIScheduleCandReason Reason) {
181 if (TryVal > CandVal) {
182 TryCand.Reason = Reason;
185 if (TryVal < CandVal) {
186 if (Cand.Reason > Reason)
187 Cand.Reason = Reason;
190 Cand.setRepeat(Reason);
193 } // end namespace SISched
194 } // end namespace llvm
196 // SIScheduleBlock //
198 void SIScheduleBlock::addUnit(SUnit *SU) {
199 NodeNum2Index[SU->NodeNum] = SUnits.size();
200 SUnits.push_back(SU);
204 void SIScheduleBlock::traceCandidate(const SISchedCandidate &Cand) {
206 dbgs() << " SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason);
211 void SIScheduleBlock::tryCandidateTopDown(SISchedCandidate &Cand,
212 SISchedCandidate &TryCand) {
213 // Initialize the candidate if needed.
214 if (!Cand.isValid()) {
215 TryCand.Reason = NodeOrder;
219 if (Cand.SGPRUsage > 60 &&
220 SISched::tryLess(TryCand.SGPRUsage, Cand.SGPRUsage,
221 TryCand, Cand, RegUsage))
224 // Schedule low latency instructions as top as possible.
225 // Order of priority is:
226 // . Low latency instructions which do not depend on other low latency
227 // instructions we haven't waited for
228 // . Other instructions which do not depend on low latency instructions
229 // we haven't waited for
231 // . All other instructions
232 // Goal is to get: low latency instructions - independent instructions
233 // - (eventually some more low latency instructions)
234 // - instructions that depend on the first low latency instructions.
235 // If in the block there is a lot of constant loads, the SGPR usage
236 // could go quite high, thus above the arbitrary limit of 60 will encourage
237 // use the already loaded constants (in order to release some SGPRs) before
239 if (SISched::tryLess(TryCand.HasLowLatencyNonWaitedParent,
240 Cand.HasLowLatencyNonWaitedParent,
241 TryCand, Cand, SIScheduleCandReason::Depth))
244 if (SISched::tryGreater(TryCand.IsLowLatency, Cand.IsLowLatency,
245 TryCand, Cand, SIScheduleCandReason::Depth))
248 if (TryCand.IsLowLatency &&
249 SISched::tryLess(TryCand.LowLatencyOffset, Cand.LowLatencyOffset,
250 TryCand, Cand, SIScheduleCandReason::Depth))
253 if (SISched::tryLess(TryCand.VGPRUsage, Cand.VGPRUsage,
254 TryCand, Cand, RegUsage))
257 // Fall through to original instruction order.
258 if (TryCand.SU->NodeNum < Cand.SU->NodeNum) {
259 TryCand.Reason = NodeOrder;
263 SUnit* SIScheduleBlock::pickNode() {
264 SISchedCandidate TopCand;
266 for (SUnit* SU : TopReadySUs) {
267 SISchedCandidate TryCand;
268 std::vector<unsigned> pressure;
269 std::vector<unsigned> MaxPressure;
270 // Predict register usage after this instruction.
272 TopRPTracker.getDownwardPressure(SU->getInstr(), pressure, MaxPressure);
273 TryCand.SGPRUsage = pressure[DAG->getSGPRSetID()];
274 TryCand.VGPRUsage = pressure[DAG->getVGPRSetID()];
275 TryCand.IsLowLatency = DAG->IsLowLatencySU[SU->NodeNum];
276 TryCand.LowLatencyOffset = DAG->LowLatencyOffset[SU->NodeNum];
277 TryCand.HasLowLatencyNonWaitedParent =
278 HasLowLatencyNonWaitedParent[NodeNum2Index[SU->NodeNum]];
279 tryCandidateTopDown(TopCand, TryCand);
280 if (TryCand.Reason != NoCand)
281 TopCand.setBest(TryCand);
288 // Schedule something valid.
289 void SIScheduleBlock::fastSchedule() {
294 for (SUnit* SU : SUnits) {
295 if (!SU->NumPredsLeft)
296 TopReadySUs.push_back(SU);
299 while (!TopReadySUs.empty()) {
300 SUnit *SU = TopReadySUs[0];
301 ScheduledSUnits.push_back(SU);
308 // Returns if the register was set between first and last.
309 static bool isDefBetween(unsigned Reg,
310 SlotIndex First, SlotIndex Last,
311 const MachineRegisterInfo *MRI,
312 const LiveIntervals *LIS) {
313 for (MachineRegisterInfo::def_instr_iterator
314 UI = MRI->def_instr_begin(Reg),
315 UE = MRI->def_instr_end(); UI != UE; ++UI) {
316 const MachineInstr* MI = &*UI;
317 if (MI->isDebugValue())
319 SlotIndex InstSlot = LIS->getInstructionIndex(*MI).getRegSlot();
320 if (InstSlot >= First && InstSlot <= Last)
326 void SIScheduleBlock::initRegPressure(MachineBasicBlock::iterator BeginBlock,
327 MachineBasicBlock::iterator EndBlock) {
328 IntervalPressure Pressure, BotPressure;
329 RegPressureTracker RPTracker(Pressure), BotRPTracker(BotPressure);
330 LiveIntervals *LIS = DAG->getLIS();
331 MachineRegisterInfo *MRI = DAG->getMRI();
332 DAG->initRPTracker(TopRPTracker);
333 DAG->initRPTracker(BotRPTracker);
334 DAG->initRPTracker(RPTracker);
336 // Goes though all SU. RPTracker captures what had to be alive for the SUs
337 // to execute, and what is still alive at the end.
338 for (SUnit* SU : ScheduledSUnits) {
339 RPTracker.setPos(SU->getInstr());
343 // Close the RPTracker to finalize live ins/outs.
344 RPTracker.closeRegion();
346 // Initialize the live ins and live outs.
347 TopRPTracker.addLiveRegs(RPTracker.getPressure().LiveInRegs);
348 BotRPTracker.addLiveRegs(RPTracker.getPressure().LiveOutRegs);
350 // Do not Track Physical Registers, because it messes up.
351 for (const auto &RegMaskPair : RPTracker.getPressure().LiveInRegs) {
352 if (TargetRegisterInfo::isVirtualRegister(RegMaskPair.RegUnit))
353 LiveInRegs.insert(RegMaskPair.RegUnit);
356 // There is several possibilities to distinguish:
357 // 1) Reg is not input to any instruction in the block, but is output of one
358 // 2) 1) + read in the block and not needed after it
359 // 3) 1) + read in the block but needed in another block
360 // 4) Reg is input of an instruction but another block will read it too
361 // 5) Reg is input of an instruction and then rewritten in the block.
362 // result is not read in the block (implies used in another block)
363 // 6) Reg is input of an instruction and then rewritten in the block.
364 // result is read in the block and not needed in another block
365 // 7) Reg is input of an instruction and then rewritten in the block.
366 // result is read in the block but also needed in another block
367 // LiveInRegs will contains all the regs in situation 4, 5, 6, 7
368 // We want LiveOutRegs to contain only Regs whose content will be read after
369 // in another block, and whose content was written in the current block,
370 // that is we want it to get 1, 3, 5, 7
371 // Since we made the MIs of a block to be packed all together before
372 // scheduling, then the LiveIntervals were correct, and the RPTracker was
373 // able to correctly handle 5 vs 6, 2 vs 3.
374 // (Note: This is not sufficient for RPTracker to not do mistakes for case 4)
375 // The RPTracker's LiveOutRegs has 1, 3, (some correct or incorrect)4, 5, 7
376 // Comparing to LiveInRegs is not sufficient to differenciate 4 vs 5, 7
377 // The use of findDefBetween removes the case 4.
378 for (const auto &RegMaskPair : RPTracker.getPressure().LiveOutRegs) {
379 unsigned Reg = RegMaskPair.RegUnit;
380 if (TargetRegisterInfo::isVirtualRegister(Reg) &&
381 isDefBetween(Reg, LIS->getInstructionIndex(*BeginBlock).getRegSlot(),
382 LIS->getInstructionIndex(*EndBlock).getRegSlot(), MRI,
384 LiveOutRegs.insert(Reg);
388 // Pressure = sum_alive_registers register size
389 // Internally llvm will represent some registers as big 128 bits registers
390 // for example, but they actually correspond to 4 actual 32 bits registers.
391 // Thus Pressure is not equal to num_alive_registers * constant.
392 LiveInPressure = TopPressure.MaxSetPressure;
393 LiveOutPressure = BotPressure.MaxSetPressure;
395 // Prepares TopRPTracker for top down scheduling.
396 TopRPTracker.closeTop();
399 void SIScheduleBlock::schedule(MachineBasicBlock::iterator BeginBlock,
400 MachineBasicBlock::iterator EndBlock) {
404 // PreScheduling phase to set LiveIn and LiveOut.
405 initRegPressure(BeginBlock, EndBlock);
408 // Schedule for real now.
412 for (SUnit* SU : SUnits) {
413 if (!SU->NumPredsLeft)
414 TopReadySUs.push_back(SU);
417 while (!TopReadySUs.empty()) {
418 SUnit *SU = pickNode();
419 ScheduledSUnits.push_back(SU);
420 TopRPTracker.setPos(SU->getInstr());
421 TopRPTracker.advance();
425 // TODO: compute InternalAdditionnalPressure.
426 InternalAdditionnalPressure.resize(TopPressure.MaxSetPressure.size());
428 // Check everything is right.
430 assert(SUnits.size() == ScheduledSUnits.size() &&
431 TopReadySUs.empty());
432 for (SUnit* SU : SUnits) {
433 assert(SU->isScheduled &&
434 SU->NumPredsLeft == 0);
441 void SIScheduleBlock::undoSchedule() {
442 for (SUnit* SU : SUnits) {
443 SU->isScheduled = false;
444 for (SDep& Succ : SU->Succs) {
445 if (BC->isSUInBlock(Succ.getSUnit(), ID))
446 undoReleaseSucc(SU, &Succ);
449 HasLowLatencyNonWaitedParent.assign(SUnits.size(), 0);
450 ScheduledSUnits.clear();
454 void SIScheduleBlock::undoReleaseSucc(SUnit *SU, SDep *SuccEdge) {
455 SUnit *SuccSU = SuccEdge->getSUnit();
457 if (SuccEdge->isWeak()) {
458 ++SuccSU->WeakPredsLeft;
461 ++SuccSU->NumPredsLeft;
464 void SIScheduleBlock::releaseSucc(SUnit *SU, SDep *SuccEdge) {
465 SUnit *SuccSU = SuccEdge->getSUnit();
467 if (SuccEdge->isWeak()) {
468 --SuccSU->WeakPredsLeft;
472 if (SuccSU->NumPredsLeft == 0) {
473 dbgs() << "*** Scheduling failed! ***\n";
475 dbgs() << " has been released too many times!\n";
476 llvm_unreachable(nullptr);
480 --SuccSU->NumPredsLeft;
483 /// Release Successors of the SU that are in the block or not.
484 void SIScheduleBlock::releaseSuccessors(SUnit *SU, bool InOrOutBlock) {
485 for (SDep& Succ : SU->Succs) {
486 SUnit *SuccSU = Succ.getSUnit();
488 if (SuccSU->NodeNum >= DAG->SUnits.size())
491 if (BC->isSUInBlock(SuccSU, ID) != InOrOutBlock)
494 releaseSucc(SU, &Succ);
495 if (SuccSU->NumPredsLeft == 0 && InOrOutBlock)
496 TopReadySUs.push_back(SuccSU);
500 void SIScheduleBlock::nodeScheduled(SUnit *SU) {
502 assert (!SU->NumPredsLeft);
503 std::vector<SUnit *>::iterator I = llvm::find(TopReadySUs, SU);
504 if (I == TopReadySUs.end()) {
505 dbgs() << "Data Structure Bug in SI Scheduler\n";
506 llvm_unreachable(nullptr);
508 TopReadySUs.erase(I);
510 releaseSuccessors(SU, true);
511 // Scheduling this node will trigger a wait,
512 // thus propagate to other instructions that they do not need to wait either.
513 if (HasLowLatencyNonWaitedParent[NodeNum2Index[SU->NodeNum]])
514 HasLowLatencyNonWaitedParent.assign(SUnits.size(), 0);
516 if (DAG->IsLowLatencySU[SU->NodeNum]) {
517 for (SDep& Succ : SU->Succs) {
518 std::map<unsigned, unsigned>::iterator I =
519 NodeNum2Index.find(Succ.getSUnit()->NodeNum);
520 if (I != NodeNum2Index.end())
521 HasLowLatencyNonWaitedParent[I->second] = 1;
524 SU->isScheduled = true;
527 void SIScheduleBlock::finalizeUnits() {
528 // We remove links from outside blocks to enable scheduling inside the block.
529 for (SUnit* SU : SUnits) {
530 releaseSuccessors(SU, false);
531 if (DAG->IsHighLatencySU[SU->NodeNum])
532 HighLatencyBlock = true;
534 HasLowLatencyNonWaitedParent.resize(SUnits.size(), 0);
537 // we maintain ascending order of IDs
538 void SIScheduleBlock::addPred(SIScheduleBlock *Pred) {
539 unsigned PredID = Pred->getID();
541 // Check if not already predecessor.
542 for (SIScheduleBlock* P : Preds) {
543 if (PredID == P->getID())
546 Preds.push_back(Pred);
548 assert(none_of(Succs,
549 [=](std::pair<SIScheduleBlock*,
550 SIScheduleBlockLinkKind> S) {
551 return PredID == S.first->getID();
553 "Loop in the Block Graph!");
556 void SIScheduleBlock::addSucc(SIScheduleBlock *Succ,
557 SIScheduleBlockLinkKind Kind) {
558 unsigned SuccID = Succ->getID();
560 // Check if not already predecessor.
561 for (std::pair<SIScheduleBlock*, SIScheduleBlockLinkKind> &S : Succs) {
562 if (SuccID == S.first->getID()) {
563 if (S.second == SIScheduleBlockLinkKind::NoData &&
564 Kind == SIScheduleBlockLinkKind::Data)
569 if (Succ->isHighLatencyBlock())
570 ++NumHighLatencySuccessors;
571 Succs.push_back(std::make_pair(Succ, Kind));
573 assert(none_of(Preds,
574 [=](SIScheduleBlock *P) { return SuccID == P->getID(); }) &&
575 "Loop in the Block Graph!");
579 void SIScheduleBlock::printDebug(bool full) {
580 dbgs() << "Block (" << ID << ")\n";
584 dbgs() << "\nContains High Latency Instruction: "
585 << HighLatencyBlock << '\n';
586 dbgs() << "\nDepends On:\n";
587 for (SIScheduleBlock* P : Preds) {
588 P->printDebug(false);
591 dbgs() << "\nSuccessors:\n";
592 for (std::pair<SIScheduleBlock*, SIScheduleBlockLinkKind> S : Succs) {
593 if (S.second == SIScheduleBlockLinkKind::Data)
594 dbgs() << "(Data Dep) ";
595 S.first->printDebug(false);
599 dbgs() << "LiveInPressure " << LiveInPressure[DAG->getSGPRSetID()] << ' '
600 << LiveInPressure[DAG->getVGPRSetID()] << '\n';
601 dbgs() << "LiveOutPressure " << LiveOutPressure[DAG->getSGPRSetID()] << ' '
602 << LiveOutPressure[DAG->getVGPRSetID()] << "\n\n";
603 dbgs() << "LiveIns:\n";
604 for (unsigned Reg : LiveInRegs)
605 dbgs() << printVRegOrUnit(Reg, DAG->getTRI()) << ' ';
607 dbgs() << "\nLiveOuts:\n";
608 for (unsigned Reg : LiveOutRegs)
609 dbgs() << printVRegOrUnit(Reg, DAG->getTRI()) << ' ';
612 dbgs() << "\nInstructions:\n";
614 for (SUnit* SU : SUnits) {
618 for (SUnit* SU : SUnits) {
623 dbgs() << "///////////////////////\n";
627 // SIScheduleBlockCreator //
629 SIScheduleBlockCreator::SIScheduleBlockCreator(SIScheduleDAGMI *DAG) :
633 SIScheduleBlockCreator::~SIScheduleBlockCreator() = default;
636 SIScheduleBlockCreator::getBlocks(SISchedulerBlockCreatorVariant BlockVariant) {
637 std::map<SISchedulerBlockCreatorVariant, SIScheduleBlocks>::iterator B =
638 Blocks.find(BlockVariant);
639 if (B == Blocks.end()) {
640 SIScheduleBlocks Res;
641 createBlocksForVariant(BlockVariant);
643 scheduleInsideBlocks();
645 Res.Blocks = CurrentBlocks;
646 Res.TopDownIndex2Block = TopDownIndex2Block;
647 Res.TopDownBlock2Index = TopDownBlock2Index;
648 Blocks[BlockVariant] = Res;
655 bool SIScheduleBlockCreator::isSUInBlock(SUnit *SU, unsigned ID) {
656 if (SU->NodeNum >= DAG->SUnits.size())
658 return CurrentBlocks[Node2CurrentBlock[SU->NodeNum]]->getID() == ID;
661 void SIScheduleBlockCreator::colorHighLatenciesAlone() {
662 unsigned DAGSize = DAG->SUnits.size();
664 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
665 SUnit *SU = &DAG->SUnits[i];
666 if (DAG->IsHighLatencySU[SU->NodeNum]) {
667 CurrentColoring[SU->NodeNum] = NextReservedID++;
673 hasDataDependencyPred(const SUnit &SU, const SUnit &FromSU) {
674 for (const auto &PredDep : SU.Preds) {
675 if (PredDep.getSUnit() == &FromSU &&
676 PredDep.getKind() == llvm::SDep::Data)
682 void SIScheduleBlockCreator::colorHighLatenciesGroups() {
683 unsigned DAGSize = DAG->SUnits.size();
684 unsigned NumHighLatencies = 0;
686 int Color = NextReservedID;
688 std::set<unsigned> FormingGroup;
690 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
691 SUnit *SU = &DAG->SUnits[i];
692 if (DAG->IsHighLatencySU[SU->NodeNum])
696 if (NumHighLatencies == 0)
699 if (NumHighLatencies <= 6)
701 else if (NumHighLatencies <= 12)
706 for (unsigned SUNum : DAG->TopDownIndex2SU) {
707 const SUnit &SU = DAG->SUnits[SUNum];
708 if (DAG->IsHighLatencySU[SU.NodeNum]) {
709 unsigned CompatibleGroup = true;
710 int ProposedColor = Color;
711 std::vector<int> AdditionalElements;
713 // We don't want to put in the same block
714 // two high latency instructions that depend
716 // One way would be to check canAddEdge
717 // in both directions, but that currently is not
718 // enough because there the high latency order is
719 // enforced (via links).
720 // Instead, look at the dependencies between the
721 // high latency instructions and deduce if it is
722 // a data dependency or not.
723 for (unsigned j : FormingGroup) {
725 std::vector<int> SubGraph;
726 // By construction (topological order), if SU and
727 // DAG->SUnits[j] are linked, DAG->SUnits[j] is neccessary
728 // in the parent graph of SU.
730 SubGraph = DAG->GetTopo()->GetSubGraph(SU, DAG->SUnits[j],
732 assert(!HasSubGraph);
734 SubGraph = DAG->GetTopo()->GetSubGraph(DAG->SUnits[j], SU,
737 continue; // No dependencies between each other
738 else if (SubGraph.size() > 5) {
739 // Too many elements would be required to be added to the block.
740 CompatibleGroup = false;
744 // Check the type of dependency
745 for (unsigned k : SubGraph) {
746 // If in the path to join the two instructions,
747 // there is another high latency instruction,
748 // or instructions colored for another block
750 if (DAG->IsHighLatencySU[k] ||
751 (CurrentColoring[k] != ProposedColor &&
752 CurrentColoring[k] != 0)) {
753 CompatibleGroup = false;
756 // If one of the SU in the subgraph depends on the result of SU j,
757 // there'll be a data dependency.
758 if (hasDataDependencyPred(DAG->SUnits[k], DAG->SUnits[j])) {
759 CompatibleGroup = false;
763 if (!CompatibleGroup)
765 // Same check for the SU
766 if (hasDataDependencyPred(SU, DAG->SUnits[j])) {
767 CompatibleGroup = false;
770 // Add all the required instructions to the block
771 // These cannot live in another block (because they
772 // depend (order dependency) on one of the
773 // instruction in the block, and are required for the
774 // high latency instruction we add.
775 AdditionalElements.insert(AdditionalElements.end(),
776 SubGraph.begin(), SubGraph.end());
779 if (CompatibleGroup) {
780 FormingGroup.insert(SU.NodeNum);
781 for (unsigned j : AdditionalElements)
782 CurrentColoring[j] = ProposedColor;
783 CurrentColoring[SU.NodeNum] = ProposedColor;
786 // Found one incompatible instruction,
787 // or has filled a big enough group.
788 // -> start a new one.
789 if (!CompatibleGroup) {
790 FormingGroup.clear();
791 Color = ++NextReservedID;
792 ProposedColor = Color;
793 FormingGroup.insert(SU.NodeNum);
794 CurrentColoring[SU.NodeNum] = ProposedColor;
796 } else if (Count == GroupSize) {
797 FormingGroup.clear();
798 Color = ++NextReservedID;
799 ProposedColor = Color;
806 void SIScheduleBlockCreator::colorComputeReservedDependencies() {
807 unsigned DAGSize = DAG->SUnits.size();
808 std::map<std::set<unsigned>, unsigned> ColorCombinations;
810 CurrentTopDownReservedDependencyColoring.clear();
811 CurrentBottomUpReservedDependencyColoring.clear();
813 CurrentTopDownReservedDependencyColoring.resize(DAGSize, 0);
814 CurrentBottomUpReservedDependencyColoring.resize(DAGSize, 0);
816 // Traverse TopDown, and give different colors to SUs depending
817 // on which combination of High Latencies they depend on.
819 for (unsigned SUNum : DAG->TopDownIndex2SU) {
820 SUnit *SU = &DAG->SUnits[SUNum];
821 std::set<unsigned> SUColors;
824 if (CurrentColoring[SU->NodeNum]) {
825 CurrentTopDownReservedDependencyColoring[SU->NodeNum] =
826 CurrentColoring[SU->NodeNum];
830 for (SDep& PredDep : SU->Preds) {
831 SUnit *Pred = PredDep.getSUnit();
832 if (PredDep.isWeak() || Pred->NodeNum >= DAGSize)
834 if (CurrentTopDownReservedDependencyColoring[Pred->NodeNum] > 0)
835 SUColors.insert(CurrentTopDownReservedDependencyColoring[Pred->NodeNum]);
837 // Color 0 by default.
838 if (SUColors.empty())
840 // Same color than parents.
841 if (SUColors.size() == 1 && *SUColors.begin() > DAGSize)
842 CurrentTopDownReservedDependencyColoring[SU->NodeNum] =
845 std::map<std::set<unsigned>, unsigned>::iterator Pos =
846 ColorCombinations.find(SUColors);
847 if (Pos != ColorCombinations.end()) {
848 CurrentTopDownReservedDependencyColoring[SU->NodeNum] = Pos->second;
850 CurrentTopDownReservedDependencyColoring[SU->NodeNum] =
852 ColorCombinations[SUColors] = NextNonReservedID++;
857 ColorCombinations.clear();
859 // Same as before, but BottomUp.
861 for (unsigned SUNum : DAG->BottomUpIndex2SU) {
862 SUnit *SU = &DAG->SUnits[SUNum];
863 std::set<unsigned> SUColors;
866 if (CurrentColoring[SU->NodeNum]) {
867 CurrentBottomUpReservedDependencyColoring[SU->NodeNum] =
868 CurrentColoring[SU->NodeNum];
872 for (SDep& SuccDep : SU->Succs) {
873 SUnit *Succ = SuccDep.getSUnit();
874 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize)
876 if (CurrentBottomUpReservedDependencyColoring[Succ->NodeNum] > 0)
877 SUColors.insert(CurrentBottomUpReservedDependencyColoring[Succ->NodeNum]);
880 if (SUColors.empty())
882 // Same color than parents.
883 if (SUColors.size() == 1 && *SUColors.begin() > DAGSize)
884 CurrentBottomUpReservedDependencyColoring[SU->NodeNum] =
887 std::map<std::set<unsigned>, unsigned>::iterator Pos =
888 ColorCombinations.find(SUColors);
889 if (Pos != ColorCombinations.end()) {
890 CurrentBottomUpReservedDependencyColoring[SU->NodeNum] = Pos->second;
892 CurrentBottomUpReservedDependencyColoring[SU->NodeNum] =
894 ColorCombinations[SUColors] = NextNonReservedID++;
900 void SIScheduleBlockCreator::colorAccordingToReservedDependencies() {
901 unsigned DAGSize = DAG->SUnits.size();
902 std::map<std::pair<unsigned, unsigned>, unsigned> ColorCombinations;
904 // Every combination of colors given by the top down
905 // and bottom up Reserved node dependency
907 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
908 SUnit *SU = &DAG->SUnits[i];
909 std::pair<unsigned, unsigned> SUColors;
911 // High latency instructions: already given.
912 if (CurrentColoring[SU->NodeNum])
915 SUColors.first = CurrentTopDownReservedDependencyColoring[SU->NodeNum];
916 SUColors.second = CurrentBottomUpReservedDependencyColoring[SU->NodeNum];
918 std::map<std::pair<unsigned, unsigned>, unsigned>::iterator Pos =
919 ColorCombinations.find(SUColors);
920 if (Pos != ColorCombinations.end()) {
921 CurrentColoring[SU->NodeNum] = Pos->second;
923 CurrentColoring[SU->NodeNum] = NextNonReservedID;
924 ColorCombinations[SUColors] = NextNonReservedID++;
929 void SIScheduleBlockCreator::colorEndsAccordingToDependencies() {
930 unsigned DAGSize = DAG->SUnits.size();
931 std::vector<int> PendingColoring = CurrentColoring;
933 assert(DAGSize >= 1 &&
934 CurrentBottomUpReservedDependencyColoring.size() == DAGSize &&
935 CurrentTopDownReservedDependencyColoring.size() == DAGSize);
936 // If there is no reserved block at all, do nothing. We don't want
937 // everything in one block.
938 if (*std::max_element(CurrentBottomUpReservedDependencyColoring.begin(),
939 CurrentBottomUpReservedDependencyColoring.end()) == 0 &&
940 *std::max_element(CurrentTopDownReservedDependencyColoring.begin(),
941 CurrentTopDownReservedDependencyColoring.end()) == 0)
944 for (unsigned SUNum : DAG->BottomUpIndex2SU) {
945 SUnit *SU = &DAG->SUnits[SUNum];
946 std::set<unsigned> SUColors;
947 std::set<unsigned> SUColorsPending;
949 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize)
952 if (CurrentBottomUpReservedDependencyColoring[SU->NodeNum] > 0 ||
953 CurrentTopDownReservedDependencyColoring[SU->NodeNum] > 0)
956 for (SDep& SuccDep : SU->Succs) {
957 SUnit *Succ = SuccDep.getSUnit();
958 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize)
960 if (CurrentBottomUpReservedDependencyColoring[Succ->NodeNum] > 0 ||
961 CurrentTopDownReservedDependencyColoring[Succ->NodeNum] > 0)
962 SUColors.insert(CurrentColoring[Succ->NodeNum]);
963 SUColorsPending.insert(PendingColoring[Succ->NodeNum]);
965 // If there is only one child/parent block, and that block
966 // is not among the ones we are removing in this path, then
967 // merge the instruction to that block
968 if (SUColors.size() == 1 && SUColorsPending.size() == 1)
969 PendingColoring[SU->NodeNum] = *SUColors.begin();
970 else // TODO: Attribute new colors depending on color
971 // combination of children.
972 PendingColoring[SU->NodeNum] = NextNonReservedID++;
974 CurrentColoring = PendingColoring;
978 void SIScheduleBlockCreator::colorForceConsecutiveOrderInGroup() {
979 unsigned DAGSize = DAG->SUnits.size();
980 unsigned PreviousColor;
981 std::set<unsigned> SeenColors;
986 PreviousColor = CurrentColoring[0];
988 for (unsigned i = 1, e = DAGSize; i != e; ++i) {
989 SUnit *SU = &DAG->SUnits[i];
990 unsigned CurrentColor = CurrentColoring[i];
991 unsigned PreviousColorSave = PreviousColor;
992 assert(i == SU->NodeNum);
994 if (CurrentColor != PreviousColor)
995 SeenColors.insert(PreviousColor);
996 PreviousColor = CurrentColor;
998 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize)
1001 if (SeenColors.find(CurrentColor) == SeenColors.end())
1004 if (PreviousColorSave != CurrentColor)
1005 CurrentColoring[i] = NextNonReservedID++;
1007 CurrentColoring[i] = CurrentColoring[i-1];
1011 void SIScheduleBlockCreator::colorMergeConstantLoadsNextGroup() {
1012 unsigned DAGSize = DAG->SUnits.size();
1014 for (unsigned SUNum : DAG->BottomUpIndex2SU) {
1015 SUnit *SU = &DAG->SUnits[SUNum];
1016 std::set<unsigned> SUColors;
1018 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize)
1021 // No predecessor: Vgpr constant loading.
1022 // Low latency instructions usually have a predecessor (the address)
1023 if (SU->Preds.size() > 0 && !DAG->IsLowLatencySU[SU->NodeNum])
1026 for (SDep& SuccDep : SU->Succs) {
1027 SUnit *Succ = SuccDep.getSUnit();
1028 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize)
1030 SUColors.insert(CurrentColoring[Succ->NodeNum]);
1032 if (SUColors.size() == 1)
1033 CurrentColoring[SU->NodeNum] = *SUColors.begin();
1037 void SIScheduleBlockCreator::colorMergeIfPossibleNextGroup() {
1038 unsigned DAGSize = DAG->SUnits.size();
1040 for (unsigned SUNum : DAG->BottomUpIndex2SU) {
1041 SUnit *SU = &DAG->SUnits[SUNum];
1042 std::set<unsigned> SUColors;
1044 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize)
1047 for (SDep& SuccDep : SU->Succs) {
1048 SUnit *Succ = SuccDep.getSUnit();
1049 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize)
1051 SUColors.insert(CurrentColoring[Succ->NodeNum]);
1053 if (SUColors.size() == 1)
1054 CurrentColoring[SU->NodeNum] = *SUColors.begin();
1058 void SIScheduleBlockCreator::colorMergeIfPossibleNextGroupOnlyForReserved() {
1059 unsigned DAGSize = DAG->SUnits.size();
1061 for (unsigned SUNum : DAG->BottomUpIndex2SU) {
1062 SUnit *SU = &DAG->SUnits[SUNum];
1063 std::set<unsigned> SUColors;
1065 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize)
1068 for (SDep& SuccDep : SU->Succs) {
1069 SUnit *Succ = SuccDep.getSUnit();
1070 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize)
1072 SUColors.insert(CurrentColoring[Succ->NodeNum]);
1074 if (SUColors.size() == 1 && *SUColors.begin() <= DAGSize)
1075 CurrentColoring[SU->NodeNum] = *SUColors.begin();
1079 void SIScheduleBlockCreator::colorMergeIfPossibleSmallGroupsToNextGroup() {
1080 unsigned DAGSize = DAG->SUnits.size();
1081 std::map<unsigned, unsigned> ColorCount;
1083 for (unsigned SUNum : DAG->BottomUpIndex2SU) {
1084 SUnit *SU = &DAG->SUnits[SUNum];
1085 unsigned color = CurrentColoring[SU->NodeNum];
1086 ++ColorCount[color];
1089 for (unsigned SUNum : DAG->BottomUpIndex2SU) {
1090 SUnit *SU = &DAG->SUnits[SUNum];
1091 unsigned color = CurrentColoring[SU->NodeNum];
1092 std::set<unsigned> SUColors;
1094 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize)
1097 if (ColorCount[color] > 1)
1100 for (SDep& SuccDep : SU->Succs) {
1101 SUnit *Succ = SuccDep.getSUnit();
1102 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize)
1104 SUColors.insert(CurrentColoring[Succ->NodeNum]);
1106 if (SUColors.size() == 1 && *SUColors.begin() != color) {
1107 --ColorCount[color];
1108 CurrentColoring[SU->NodeNum] = *SUColors.begin();
1109 ++ColorCount[*SUColors.begin()];
1114 void SIScheduleBlockCreator::cutHugeBlocks() {
1118 void SIScheduleBlockCreator::regroupNoUserInstructions() {
1119 unsigned DAGSize = DAG->SUnits.size();
1120 int GroupID = NextNonReservedID++;
1122 for (unsigned SUNum : DAG->BottomUpIndex2SU) {
1123 SUnit *SU = &DAG->SUnits[SUNum];
1124 bool hasSuccessor = false;
1126 if (CurrentColoring[SU->NodeNum] <= (int)DAGSize)
1129 for (SDep& SuccDep : SU->Succs) {
1130 SUnit *Succ = SuccDep.getSUnit();
1131 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize)
1133 hasSuccessor = true;
1136 CurrentColoring[SU->NodeNum] = GroupID;
1140 void SIScheduleBlockCreator::colorExports() {
1141 unsigned ExportColor = NextNonReservedID++;
1142 SmallVector<unsigned, 8> ExpGroup;
1144 // Put all exports together in a block.
1145 // The block will naturally end up being scheduled last,
1146 // thus putting exports at the end of the schedule, which
1147 // is better for performance.
1148 // However we must ensure, for safety, the exports can be put
1149 // together in the same block without any other instruction.
1150 // This could happen, for example, when scheduling after regalloc
1151 // if reloading a spilled register from memory using the same
1152 // register than used in a previous export.
1153 // If that happens, do not regroup the exports.
1154 for (unsigned SUNum : DAG->TopDownIndex2SU) {
1155 const SUnit &SU = DAG->SUnits[SUNum];
1156 if (SIInstrInfo::isEXP(*SU.getInstr())) {
1157 // Check the EXP can be added to the group safely,
1158 // ie without needing any other instruction.
1159 // The EXP is allowed to depend on other EXP
1160 // (they will be in the same group).
1161 for (unsigned j : ExpGroup) {
1163 std::vector<int> SubGraph;
1164 // By construction (topological order), if SU and
1165 // DAG->SUnits[j] are linked, DAG->SUnits[j] is neccessary
1166 // in the parent graph of SU.
1168 SubGraph = DAG->GetTopo()->GetSubGraph(SU, DAG->SUnits[j],
1170 assert(!HasSubGraph);
1172 SubGraph = DAG->GetTopo()->GetSubGraph(DAG->SUnits[j], SU,
1175 continue; // No dependencies between each other
1177 // SubGraph contains all the instructions required
1178 // between EXP SUnits[j] and EXP SU.
1179 for (unsigned k : SubGraph) {
1180 if (!SIInstrInfo::isEXP(*DAG->SUnits[k].getInstr()))
1181 // Other instructions than EXP would be required in the group.
1182 // Abort the groupping.
1187 ExpGroup.push_back(SUNum);
1191 // The group can be formed. Give the color.
1192 for (unsigned j : ExpGroup)
1193 CurrentColoring[j] = ExportColor;
1196 void SIScheduleBlockCreator::createBlocksForVariant(SISchedulerBlockCreatorVariant BlockVariant) {
1197 unsigned DAGSize = DAG->SUnits.size();
1198 std::map<unsigned,unsigned> RealID;
1200 CurrentBlocks.clear();
1201 CurrentColoring.clear();
1202 CurrentColoring.resize(DAGSize, 0);
1203 Node2CurrentBlock.clear();
1205 // Restore links previous scheduling variant has overridden.
1206 DAG->restoreSULinksLeft();
1209 NextNonReservedID = DAGSize + 1;
1211 LLVM_DEBUG(dbgs() << "Coloring the graph\n");
1213 if (BlockVariant == SISchedulerBlockCreatorVariant::LatenciesGrouped)
1214 colorHighLatenciesGroups();
1216 colorHighLatenciesAlone();
1217 colorComputeReservedDependencies();
1218 colorAccordingToReservedDependencies();
1219 colorEndsAccordingToDependencies();
1220 if (BlockVariant == SISchedulerBlockCreatorVariant::LatenciesAlonePlusConsecutive)
1221 colorForceConsecutiveOrderInGroup();
1222 regroupNoUserInstructions();
1223 colorMergeConstantLoadsNextGroup();
1224 colorMergeIfPossibleNextGroupOnlyForReserved();
1227 // Put SUs of same color into same block
1228 Node2CurrentBlock.resize(DAGSize, -1);
1229 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
1230 SUnit *SU = &DAG->SUnits[i];
1231 unsigned Color = CurrentColoring[SU->NodeNum];
1232 if (RealID.find(Color) == RealID.end()) {
1233 int ID = CurrentBlocks.size();
1234 BlockPtrs.push_back(llvm::make_unique<SIScheduleBlock>(DAG, this, ID));
1235 CurrentBlocks.push_back(BlockPtrs.rbegin()->get());
1238 CurrentBlocks[RealID[Color]]->addUnit(SU);
1239 Node2CurrentBlock[SU->NodeNum] = RealID[Color];
1242 // Build dependencies between blocks.
1243 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
1244 SUnit *SU = &DAG->SUnits[i];
1245 int SUID = Node2CurrentBlock[i];
1246 for (SDep& SuccDep : SU->Succs) {
1247 SUnit *Succ = SuccDep.getSUnit();
1248 if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize)
1250 if (Node2CurrentBlock[Succ->NodeNum] != SUID)
1251 CurrentBlocks[SUID]->addSucc(CurrentBlocks[Node2CurrentBlock[Succ->NodeNum]],
1252 SuccDep.isCtrl() ? NoData : Data);
1254 for (SDep& PredDep : SU->Preds) {
1255 SUnit *Pred = PredDep.getSUnit();
1256 if (PredDep.isWeak() || Pred->NodeNum >= DAGSize)
1258 if (Node2CurrentBlock[Pred->NodeNum] != SUID)
1259 CurrentBlocks[SUID]->addPred(CurrentBlocks[Node2CurrentBlock[Pred->NodeNum]]);
1263 // Free root and leafs of all blocks to enable scheduling inside them.
1264 for (unsigned i = 0, e = CurrentBlocks.size(); i != e; ++i) {
1265 SIScheduleBlock *Block = CurrentBlocks[i];
1266 Block->finalizeUnits();
1268 LLVM_DEBUG(dbgs() << "Blocks created:\n\n";
1269 for (unsigned i = 0, e = CurrentBlocks.size(); i != e; ++i) {
1270 SIScheduleBlock *Block = CurrentBlocks[i];
1271 Block->printDebug(true);
1275 // Two functions taken from Codegen/MachineScheduler.cpp
1277 /// Non-const version.
1278 static MachineBasicBlock::iterator
1279 nextIfDebug(MachineBasicBlock::iterator I,
1280 MachineBasicBlock::const_iterator End) {
1281 for (; I != End; ++I) {
1282 if (!I->isDebugInstr())
1288 void SIScheduleBlockCreator::topologicalSort() {
1289 unsigned DAGSize = CurrentBlocks.size();
1290 std::vector<int> WorkList;
1292 LLVM_DEBUG(dbgs() << "Topological Sort\n");
1294 WorkList.reserve(DAGSize);
1295 TopDownIndex2Block.resize(DAGSize);
1296 TopDownBlock2Index.resize(DAGSize);
1297 BottomUpIndex2Block.resize(DAGSize);
1299 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
1300 SIScheduleBlock *Block = CurrentBlocks[i];
1301 unsigned Degree = Block->getSuccs().size();
1302 TopDownBlock2Index[i] = Degree;
1304 WorkList.push_back(i);
1309 while (!WorkList.empty()) {
1310 int i = WorkList.back();
1311 SIScheduleBlock *Block = CurrentBlocks[i];
1312 WorkList.pop_back();
1313 TopDownBlock2Index[i] = --Id;
1314 TopDownIndex2Block[Id] = i;
1315 for (SIScheduleBlock* Pred : Block->getPreds()) {
1316 if (!--TopDownBlock2Index[Pred->getID()])
1317 WorkList.push_back(Pred->getID());
1322 // Check correctness of the ordering.
1323 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
1324 SIScheduleBlock *Block = CurrentBlocks[i];
1325 for (SIScheduleBlock* Pred : Block->getPreds()) {
1326 assert(TopDownBlock2Index[i] > TopDownBlock2Index[Pred->getID()] &&
1327 "Wrong Top Down topological sorting");
1332 BottomUpIndex2Block = std::vector<int>(TopDownIndex2Block.rbegin(),
1333 TopDownIndex2Block.rend());
1336 void SIScheduleBlockCreator::scheduleInsideBlocks() {
1337 unsigned DAGSize = CurrentBlocks.size();
1339 LLVM_DEBUG(dbgs() << "\nScheduling Blocks\n\n");
1341 // We do schedule a valid scheduling such that a Block corresponds
1342 // to a range of instructions.
1343 LLVM_DEBUG(dbgs() << "First phase: Fast scheduling for Reg Liveness\n");
1344 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
1345 SIScheduleBlock *Block = CurrentBlocks[i];
1346 Block->fastSchedule();
1349 // Note: the following code, and the part restoring previous position
1350 // is by far the most expensive operation of the Scheduler.
1352 // Do not update CurrentTop.
1353 MachineBasicBlock::iterator CurrentTopFastSched = DAG->getCurrentTop();
1354 std::vector<MachineBasicBlock::iterator> PosOld;
1355 std::vector<MachineBasicBlock::iterator> PosNew;
1356 PosOld.reserve(DAG->SUnits.size());
1357 PosNew.reserve(DAG->SUnits.size());
1359 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
1360 int BlockIndice = TopDownIndex2Block[i];
1361 SIScheduleBlock *Block = CurrentBlocks[BlockIndice];
1362 std::vector<SUnit*> SUs = Block->getScheduledUnits();
1364 for (SUnit* SU : SUs) {
1365 MachineInstr *MI = SU->getInstr();
1366 MachineBasicBlock::iterator Pos = MI;
1367 PosOld.push_back(Pos);
1368 if (&*CurrentTopFastSched == MI) {
1369 PosNew.push_back(Pos);
1370 CurrentTopFastSched = nextIfDebug(++CurrentTopFastSched,
1371 DAG->getCurrentBottom());
1373 // Update the instruction stream.
1374 DAG->getBB()->splice(CurrentTopFastSched, DAG->getBB(), MI);
1376 // Update LiveIntervals.
1377 // Note: Moving all instructions and calling handleMove every time
1378 // is the most cpu intensive operation of the scheduler.
1379 // It would gain a lot if there was a way to recompute the
1380 // LiveIntervals for the entire scheduling region.
1381 DAG->getLIS()->handleMove(*MI, /*UpdateFlags=*/true);
1382 PosNew.push_back(CurrentTopFastSched);
1387 // Now we have Block of SUs == Block of MI.
1388 // We do the final schedule for the instructions inside the block.
1389 // The property that all the SUs of the Block are grouped together as MI
1390 // is used for correct reg usage tracking.
1391 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
1392 SIScheduleBlock *Block = CurrentBlocks[i];
1393 std::vector<SUnit*> SUs = Block->getScheduledUnits();
1394 Block->schedule((*SUs.begin())->getInstr(), (*SUs.rbegin())->getInstr());
1397 LLVM_DEBUG(dbgs() << "Restoring MI Pos\n");
1398 // Restore old ordering (which prevents a LIS->handleMove bug).
1399 for (unsigned i = PosOld.size(), e = 0; i != e; --i) {
1400 MachineBasicBlock::iterator POld = PosOld[i-1];
1401 MachineBasicBlock::iterator PNew = PosNew[i-1];
1403 // Update the instruction stream.
1404 DAG->getBB()->splice(POld, DAG->getBB(), PNew);
1406 // Update LiveIntervals.
1407 DAG->getLIS()->handleMove(*POld, /*UpdateFlags=*/true);
1411 LLVM_DEBUG(for (unsigned i = 0, e = CurrentBlocks.size(); i != e; ++i) {
1412 SIScheduleBlock *Block = CurrentBlocks[i];
1413 Block->printDebug(true);
1417 void SIScheduleBlockCreator::fillStats() {
1418 unsigned DAGSize = CurrentBlocks.size();
1420 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
1421 int BlockIndice = TopDownIndex2Block[i];
1422 SIScheduleBlock *Block = CurrentBlocks[BlockIndice];
1423 if (Block->getPreds().empty())
1427 for (SIScheduleBlock *Pred : Block->getPreds()) {
1428 if (Depth < Pred->Depth + Pred->getCost())
1429 Depth = Pred->Depth + Pred->getCost();
1431 Block->Depth = Depth;
1435 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
1436 int BlockIndice = BottomUpIndex2Block[i];
1437 SIScheduleBlock *Block = CurrentBlocks[BlockIndice];
1438 if (Block->getSuccs().empty())
1441 unsigned Height = 0;
1442 for (const auto &Succ : Block->getSuccs())
1443 Height = std::max(Height, Succ.first->Height + Succ.first->getCost());
1444 Block->Height = Height;
1449 // SIScheduleBlockScheduler //
1451 SIScheduleBlockScheduler::SIScheduleBlockScheduler(SIScheduleDAGMI *DAG,
1452 SISchedulerBlockSchedulerVariant Variant,
1453 SIScheduleBlocks BlocksStruct) :
1454 DAG(DAG), Variant(Variant), Blocks(BlocksStruct.Blocks),
1455 LastPosWaitedHighLatency(0), NumBlockScheduled(0), VregCurrentUsage(0),
1456 SregCurrentUsage(0), maxVregUsage(0), maxSregUsage(0) {
1458 // Fill the usage of every output
1459 // Warning: while by construction we always have a link between two blocks
1460 // when one needs a result from the other, the number of users of an output
1461 // is not the sum of child blocks having as input the same virtual register.
1462 // Here is an example. A produces x and y. B eats x and produces x'.
1463 // C eats x' and y. The register coalescer may have attributed the same
1464 // virtual register to x and x'.
1465 // To count accurately, we do a topological sort. In case the register is
1466 // found for several parents, we increment the usage of the one with the
1467 // highest topological index.
1468 LiveOutRegsNumUsages.resize(Blocks.size());
1469 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
1470 SIScheduleBlock *Block = Blocks[i];
1471 for (unsigned Reg : Block->getInRegs()) {
1474 for (SIScheduleBlock* Pred: Block->getPreds()) {
1475 std::set<unsigned> PredOutRegs = Pred->getOutRegs();
1476 std::set<unsigned>::iterator RegPos = PredOutRegs.find(Reg);
1478 if (RegPos != PredOutRegs.end()) {
1480 if (topoInd < BlocksStruct.TopDownBlock2Index[Pred->getID()]) {
1481 topoInd = BlocksStruct.TopDownBlock2Index[Pred->getID()];
1489 int PredID = BlocksStruct.TopDownIndex2Block[topoInd];
1490 ++LiveOutRegsNumUsages[PredID][Reg];
1494 LastPosHighLatencyParentScheduled.resize(Blocks.size(), 0);
1495 BlockNumPredsLeft.resize(Blocks.size());
1496 BlockNumSuccsLeft.resize(Blocks.size());
1498 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
1499 SIScheduleBlock *Block = Blocks[i];
1500 BlockNumPredsLeft[i] = Block->getPreds().size();
1501 BlockNumSuccsLeft[i] = Block->getSuccs().size();
1505 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
1506 SIScheduleBlock *Block = Blocks[i];
1507 assert(Block->getID() == i);
1511 std::set<unsigned> InRegs = DAG->getInRegs();
1512 addLiveRegs(InRegs);
1514 // Increase LiveOutRegsNumUsages for blocks
1515 // producing registers consumed in another
1516 // scheduling region.
1517 for (unsigned Reg : DAG->getOutRegs()) {
1518 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
1519 // Do reverse traversal
1520 int ID = BlocksStruct.TopDownIndex2Block[Blocks.size()-1-i];
1521 SIScheduleBlock *Block = Blocks[ID];
1522 const std::set<unsigned> &OutRegs = Block->getOutRegs();
1524 if (OutRegs.find(Reg) == OutRegs.end())
1527 ++LiveOutRegsNumUsages[ID][Reg];
1532 // Fill LiveRegsConsumers for regs that were already
1533 // defined before scheduling.
1534 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
1535 SIScheduleBlock *Block = Blocks[i];
1536 for (unsigned Reg : Block->getInRegs()) {
1538 for (SIScheduleBlock* Pred: Block->getPreds()) {
1539 std::set<unsigned> PredOutRegs = Pred->getOutRegs();
1540 std::set<unsigned>::iterator RegPos = PredOutRegs.find(Reg);
1542 if (RegPos != PredOutRegs.end()) {
1549 ++LiveRegsConsumers[Reg];
1553 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
1554 SIScheduleBlock *Block = Blocks[i];
1555 if (BlockNumPredsLeft[i] == 0) {
1556 ReadyBlocks.push_back(Block);
1560 while (SIScheduleBlock *Block = pickBlock()) {
1561 BlocksScheduled.push_back(Block);
1562 blockScheduled(Block);
1565 LLVM_DEBUG(dbgs() << "Block Order:"; for (SIScheduleBlock *Block
1566 : BlocksScheduled) {
1567 dbgs() << ' ' << Block->getID();
1571 bool SIScheduleBlockScheduler::tryCandidateLatency(SIBlockSchedCandidate &Cand,
1572 SIBlockSchedCandidate &TryCand) {
1573 if (!Cand.isValid()) {
1574 TryCand.Reason = NodeOrder;
1578 // Try to hide high latencies.
1579 if (SISched::tryLess(TryCand.LastPosHighLatParentScheduled,
1580 Cand.LastPosHighLatParentScheduled, TryCand, Cand, Latency))
1582 // Schedule high latencies early so you can hide them better.
1583 if (SISched::tryGreater(TryCand.IsHighLatency, Cand.IsHighLatency,
1584 TryCand, Cand, Latency))
1586 if (TryCand.IsHighLatency && SISched::tryGreater(TryCand.Height, Cand.Height,
1587 TryCand, Cand, Depth))
1589 if (SISched::tryGreater(TryCand.NumHighLatencySuccessors,
1590 Cand.NumHighLatencySuccessors,
1591 TryCand, Cand, Successor))
1596 bool SIScheduleBlockScheduler::tryCandidateRegUsage(SIBlockSchedCandidate &Cand,
1597 SIBlockSchedCandidate &TryCand) {
1598 if (!Cand.isValid()) {
1599 TryCand.Reason = NodeOrder;
1603 if (SISched::tryLess(TryCand.VGPRUsageDiff > 0, Cand.VGPRUsageDiff > 0,
1604 TryCand, Cand, RegUsage))
1606 if (SISched::tryGreater(TryCand.NumSuccessors > 0,
1607 Cand.NumSuccessors > 0,
1608 TryCand, Cand, Successor))
1610 if (SISched::tryGreater(TryCand.Height, Cand.Height, TryCand, Cand, Depth))
1612 if (SISched::tryLess(TryCand.VGPRUsageDiff, Cand.VGPRUsageDiff,
1613 TryCand, Cand, RegUsage))
1618 SIScheduleBlock *SIScheduleBlockScheduler::pickBlock() {
1619 SIBlockSchedCandidate Cand;
1620 std::vector<SIScheduleBlock*>::iterator Best;
1621 SIScheduleBlock *Block;
1622 if (ReadyBlocks.empty())
1625 DAG->fillVgprSgprCost(LiveRegs.begin(), LiveRegs.end(),
1626 VregCurrentUsage, SregCurrentUsage);
1627 if (VregCurrentUsage > maxVregUsage)
1628 maxVregUsage = VregCurrentUsage;
1629 if (SregCurrentUsage > maxSregUsage)
1630 maxSregUsage = SregCurrentUsage;
1631 LLVM_DEBUG(dbgs() << "Picking New Blocks\n"; dbgs() << "Available: ";
1632 for (SIScheduleBlock *Block
1633 : ReadyBlocks) dbgs()
1634 << Block->getID() << ' ';
1635 dbgs() << "\nCurrent Live:\n";
1638 << printVRegOrUnit(Reg, DAG->getTRI()) << ' ';
1640 dbgs() << "Current VGPRs: " << VregCurrentUsage << '\n';
1641 dbgs() << "Current SGPRs: " << SregCurrentUsage << '\n';);
1643 Cand.Block = nullptr;
1644 for (std::vector<SIScheduleBlock*>::iterator I = ReadyBlocks.begin(),
1645 E = ReadyBlocks.end(); I != E; ++I) {
1646 SIBlockSchedCandidate TryCand;
1648 TryCand.IsHighLatency = TryCand.Block->isHighLatencyBlock();
1649 TryCand.VGPRUsageDiff =
1650 checkRegUsageImpact(TryCand.Block->getInRegs(),
1651 TryCand.Block->getOutRegs())[DAG->getVGPRSetID()];
1652 TryCand.NumSuccessors = TryCand.Block->getSuccs().size();
1653 TryCand.NumHighLatencySuccessors =
1654 TryCand.Block->getNumHighLatencySuccessors();
1655 TryCand.LastPosHighLatParentScheduled =
1656 (unsigned int) std::max<int> (0,
1657 LastPosHighLatencyParentScheduled[TryCand.Block->getID()] -
1658 LastPosWaitedHighLatency);
1659 TryCand.Height = TryCand.Block->Height;
1660 // Try not to increase VGPR usage too much, else we may spill.
1661 if (VregCurrentUsage > 120 ||
1662 Variant != SISchedulerBlockSchedulerVariant::BlockLatencyRegUsage) {
1663 if (!tryCandidateRegUsage(Cand, TryCand) &&
1664 Variant != SISchedulerBlockSchedulerVariant::BlockRegUsage)
1665 tryCandidateLatency(Cand, TryCand);
1667 if (!tryCandidateLatency(Cand, TryCand))
1668 tryCandidateRegUsage(Cand, TryCand);
1670 if (TryCand.Reason != NoCand) {
1671 Cand.setBest(TryCand);
1673 LLVM_DEBUG(dbgs() << "Best Current Choice: " << Cand.Block->getID() << ' '
1674 << getReasonStr(Cand.Reason) << '\n');
1678 LLVM_DEBUG(dbgs() << "Picking: " << Cand.Block->getID() << '\n';
1679 dbgs() << "Is a block with high latency instruction: "
1680 << (Cand.IsHighLatency ? "yes\n" : "no\n");
1681 dbgs() << "Position of last high latency dependency: "
1682 << Cand.LastPosHighLatParentScheduled << '\n';
1683 dbgs() << "VGPRUsageDiff: " << Cand.VGPRUsageDiff << '\n';
1687 ReadyBlocks.erase(Best);
1691 // Tracking of currently alive registers to determine VGPR Usage.
1693 void SIScheduleBlockScheduler::addLiveRegs(std::set<unsigned> &Regs) {
1694 for (unsigned Reg : Regs) {
1695 // For now only track virtual registers.
1696 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1698 // If not already in the live set, then add it.
1699 (void) LiveRegs.insert(Reg);
1703 void SIScheduleBlockScheduler::decreaseLiveRegs(SIScheduleBlock *Block,
1704 std::set<unsigned> &Regs) {
1705 for (unsigned Reg : Regs) {
1706 // For now only track virtual registers.
1707 std::set<unsigned>::iterator Pos = LiveRegs.find(Reg);
1708 assert (Pos != LiveRegs.end() && // Reg must be live.
1709 LiveRegsConsumers.find(Reg) != LiveRegsConsumers.end() &&
1710 LiveRegsConsumers[Reg] >= 1);
1711 --LiveRegsConsumers[Reg];
1712 if (LiveRegsConsumers[Reg] == 0)
1713 LiveRegs.erase(Pos);
1717 void SIScheduleBlockScheduler::releaseBlockSuccs(SIScheduleBlock *Parent) {
1718 for (const auto &Block : Parent->getSuccs()) {
1719 if (--BlockNumPredsLeft[Block.first->getID()] == 0)
1720 ReadyBlocks.push_back(Block.first);
1722 if (Parent->isHighLatencyBlock() &&
1723 Block.second == SIScheduleBlockLinkKind::Data)
1724 LastPosHighLatencyParentScheduled[Block.first->getID()] = NumBlockScheduled;
1728 void SIScheduleBlockScheduler::blockScheduled(SIScheduleBlock *Block) {
1729 decreaseLiveRegs(Block, Block->getInRegs());
1730 addLiveRegs(Block->getOutRegs());
1731 releaseBlockSuccs(Block);
1732 for (std::map<unsigned, unsigned>::iterator RegI =
1733 LiveOutRegsNumUsages[Block->getID()].begin(),
1734 E = LiveOutRegsNumUsages[Block->getID()].end(); RegI != E; ++RegI) {
1735 std::pair<unsigned, unsigned> RegP = *RegI;
1736 // We produce this register, thus it must not be previously alive.
1737 assert(LiveRegsConsumers.find(RegP.first) == LiveRegsConsumers.end() ||
1738 LiveRegsConsumers[RegP.first] == 0);
1739 LiveRegsConsumers[RegP.first] += RegP.second;
1741 if (LastPosHighLatencyParentScheduled[Block->getID()] >
1742 (unsigned)LastPosWaitedHighLatency)
1743 LastPosWaitedHighLatency =
1744 LastPosHighLatencyParentScheduled[Block->getID()];
1745 ++NumBlockScheduled;
1749 SIScheduleBlockScheduler::checkRegUsageImpact(std::set<unsigned> &InRegs,
1750 std::set<unsigned> &OutRegs) {
1751 std::vector<int> DiffSetPressure;
1752 DiffSetPressure.assign(DAG->getTRI()->getNumRegPressureSets(), 0);
1754 for (unsigned Reg : InRegs) {
1755 // For now only track virtual registers.
1756 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1758 if (LiveRegsConsumers[Reg] > 1)
1760 PSetIterator PSetI = DAG->getMRI()->getPressureSets(Reg);
1761 for (; PSetI.isValid(); ++PSetI) {
1762 DiffSetPressure[*PSetI] -= PSetI.getWeight();
1766 for (unsigned Reg : OutRegs) {
1767 // For now only track virtual registers.
1768 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1770 PSetIterator PSetI = DAG->getMRI()->getPressureSets(Reg);
1771 for (; PSetI.isValid(); ++PSetI) {
1772 DiffSetPressure[*PSetI] += PSetI.getWeight();
1776 return DiffSetPressure;
1781 struct SIScheduleBlockResult
1782 SIScheduler::scheduleVariant(SISchedulerBlockCreatorVariant BlockVariant,
1783 SISchedulerBlockSchedulerVariant ScheduleVariant) {
1784 SIScheduleBlocks Blocks = BlockCreator.getBlocks(BlockVariant);
1785 SIScheduleBlockScheduler Scheduler(DAG, ScheduleVariant, Blocks);
1786 std::vector<SIScheduleBlock*> ScheduledBlocks;
1787 struct SIScheduleBlockResult Res;
1789 ScheduledBlocks = Scheduler.getBlocks();
1791 for (unsigned b = 0; b < ScheduledBlocks.size(); ++b) {
1792 SIScheduleBlock *Block = ScheduledBlocks[b];
1793 std::vector<SUnit*> SUs = Block->getScheduledUnits();
1795 for (SUnit* SU : SUs)
1796 Res.SUs.push_back(SU->NodeNum);
1799 Res.MaxSGPRUsage = Scheduler.getSGPRUsage();
1800 Res.MaxVGPRUsage = Scheduler.getVGPRUsage();
1804 // SIScheduleDAGMI //
1806 SIScheduleDAGMI::SIScheduleDAGMI(MachineSchedContext *C) :
1807 ScheduleDAGMILive(C, llvm::make_unique<GenericScheduler>(C)) {
1808 SITII = static_cast<const SIInstrInfo*>(TII);
1809 SITRI = static_cast<const SIRegisterInfo*>(TRI);
1811 VGPRSetID = SITRI->getVGPRPressureSet();
1812 SGPRSetID = SITRI->getSGPRPressureSet();
1815 SIScheduleDAGMI::~SIScheduleDAGMI() = default;
1817 // Code adapted from scheduleDAG.cpp
1818 // Does a topological sort over the SUs.
1819 // Both TopDown and BottomUp
1820 void SIScheduleDAGMI::topologicalSort() {
1821 Topo.InitDAGTopologicalSorting();
1823 TopDownIndex2SU = std::vector<int>(Topo.begin(), Topo.end());
1824 BottomUpIndex2SU = std::vector<int>(Topo.rbegin(), Topo.rend());
1827 // Move low latencies further from their user without
1828 // increasing SGPR usage (in general)
1829 // This is to be replaced by a better pass that would
1830 // take into account SGPR usage (based on VGPR Usage
1831 // and the corresponding wavefront count), that would
1832 // try to merge groups of loads if it make sense, etc
1833 void SIScheduleDAGMI::moveLowLatencies() {
1834 unsigned DAGSize = SUnits.size();
1835 int LastLowLatencyUser = -1;
1836 int LastLowLatencyPos = -1;
1838 for (unsigned i = 0, e = ScheduledSUnits.size(); i != e; ++i) {
1839 SUnit *SU = &SUnits[ScheduledSUnits[i]];
1840 bool IsLowLatencyUser = false;
1841 unsigned MinPos = 0;
1843 for (SDep& PredDep : SU->Preds) {
1844 SUnit *Pred = PredDep.getSUnit();
1845 if (SITII->isLowLatencyInstruction(*Pred->getInstr())) {
1846 IsLowLatencyUser = true;
1848 if (Pred->NodeNum >= DAGSize)
1850 unsigned PredPos = ScheduledSUnitsInv[Pred->NodeNum];
1851 if (PredPos >= MinPos)
1852 MinPos = PredPos + 1;
1855 if (SITII->isLowLatencyInstruction(*SU->getInstr())) {
1856 unsigned BestPos = LastLowLatencyUser + 1;
1857 if ((int)BestPos <= LastLowLatencyPos)
1858 BestPos = LastLowLatencyPos + 1;
1859 if (BestPos < MinPos)
1862 for (unsigned u = i; u > BestPos; --u) {
1863 ++ScheduledSUnitsInv[ScheduledSUnits[u-1]];
1864 ScheduledSUnits[u] = ScheduledSUnits[u-1];
1866 ScheduledSUnits[BestPos] = SU->NodeNum;
1867 ScheduledSUnitsInv[SU->NodeNum] = BestPos;
1869 LastLowLatencyPos = BestPos;
1870 if (IsLowLatencyUser)
1871 LastLowLatencyUser = BestPos;
1872 } else if (IsLowLatencyUser) {
1873 LastLowLatencyUser = i;
1874 // Moves COPY instructions on which depends
1875 // the low latency instructions too.
1876 } else if (SU->getInstr()->getOpcode() == AMDGPU::COPY) {
1877 bool CopyForLowLat = false;
1878 for (SDep& SuccDep : SU->Succs) {
1879 SUnit *Succ = SuccDep.getSUnit();
1880 if (SITII->isLowLatencyInstruction(*Succ->getInstr())) {
1881 CopyForLowLat = true;
1887 for (unsigned u = i; u > MinPos; --u) {
1888 ++ScheduledSUnitsInv[ScheduledSUnits[u-1]];
1889 ScheduledSUnits[u] = ScheduledSUnits[u-1];
1891 ScheduledSUnits[MinPos] = SU->NodeNum;
1892 ScheduledSUnitsInv[SU->NodeNum] = MinPos;
1898 void SIScheduleDAGMI::restoreSULinksLeft() {
1899 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
1900 SUnits[i].isScheduled = false;
1901 SUnits[i].WeakPredsLeft = SUnitsLinksBackup[i].WeakPredsLeft;
1902 SUnits[i].NumPredsLeft = SUnitsLinksBackup[i].NumPredsLeft;
1903 SUnits[i].WeakSuccsLeft = SUnitsLinksBackup[i].WeakSuccsLeft;
1904 SUnits[i].NumSuccsLeft = SUnitsLinksBackup[i].NumSuccsLeft;
1908 // Return the Vgpr and Sgpr usage corresponding to some virtual registers.
1909 template<typename _Iterator> void
1910 SIScheduleDAGMI::fillVgprSgprCost(_Iterator First, _Iterator End,
1911 unsigned &VgprUsage, unsigned &SgprUsage) {
1914 for (_Iterator RegI = First; RegI != End; ++RegI) {
1915 unsigned Reg = *RegI;
1916 // For now only track virtual registers
1917 if (!TargetRegisterInfo::isVirtualRegister(Reg))
1919 PSetIterator PSetI = MRI.getPressureSets(Reg);
1920 for (; PSetI.isValid(); ++PSetI) {
1921 if (*PSetI == VGPRSetID)
1922 VgprUsage += PSetI.getWeight();
1923 else if (*PSetI == SGPRSetID)
1924 SgprUsage += PSetI.getWeight();
1929 void SIScheduleDAGMI::schedule()
1931 SmallVector<SUnit*, 8> TopRoots, BotRoots;
1932 SIScheduleBlockResult Best, Temp;
1933 LLVM_DEBUG(dbgs() << "Preparing Scheduling\n");
1935 buildDAGWithRegPressure();
1936 LLVM_DEBUG(for (SUnit &SU : SUnits) SU.dumpAll(this));
1939 findRootsAndBiasEdges(TopRoots, BotRoots);
1940 // We reuse several ScheduleDAGMI and ScheduleDAGMILive
1941 // functions, but to make them happy we must initialize
1942 // the default Scheduler implementation (even if we do not
1944 SchedImpl->initialize(this);
1945 initQueues(TopRoots, BotRoots);
1947 // Fill some stats to help scheduling.
1949 SUnitsLinksBackup = SUnits;
1950 IsLowLatencySU.clear();
1951 LowLatencyOffset.clear();
1952 IsHighLatencySU.clear();
1954 IsLowLatencySU.resize(SUnits.size(), 0);
1955 LowLatencyOffset.resize(SUnits.size(), 0);
1956 IsHighLatencySU.resize(SUnits.size(), 0);
1958 for (unsigned i = 0, e = (unsigned)SUnits.size(); i != e; ++i) {
1959 SUnit *SU = &SUnits[i];
1960 unsigned BaseLatReg;
1962 if (SITII->isLowLatencyInstruction(*SU->getInstr())) {
1963 IsLowLatencySU[i] = 1;
1964 if (SITII->getMemOpBaseRegImmOfs(*SU->getInstr(), BaseLatReg, OffLatReg,
1966 LowLatencyOffset[i] = OffLatReg;
1967 } else if (SITII->isHighLatencyInstruction(*SU->getInstr()))
1968 IsHighLatencySU[i] = 1;
1971 SIScheduler Scheduler(this);
1972 Best = Scheduler.scheduleVariant(SISchedulerBlockCreatorVariant::LatenciesAlone,
1973 SISchedulerBlockSchedulerVariant::BlockLatencyRegUsage);
1975 // if VGPR usage is extremely high, try other good performing variants
1976 // which could lead to lower VGPR usage
1977 if (Best.MaxVGPRUsage > 180) {
1978 static const std::pair<SISchedulerBlockCreatorVariant,
1979 SISchedulerBlockSchedulerVariant>
1981 { LatenciesAlone, BlockRegUsageLatency },
1982 // { LatenciesAlone, BlockRegUsage },
1983 { LatenciesGrouped, BlockLatencyRegUsage },
1984 // { LatenciesGrouped, BlockRegUsageLatency },
1985 // { LatenciesGrouped, BlockRegUsage },
1986 { LatenciesAlonePlusConsecutive, BlockLatencyRegUsage },
1987 // { LatenciesAlonePlusConsecutive, BlockRegUsageLatency },
1988 // { LatenciesAlonePlusConsecutive, BlockRegUsage }
1990 for (std::pair<SISchedulerBlockCreatorVariant, SISchedulerBlockSchedulerVariant> v : Variants) {
1991 Temp = Scheduler.scheduleVariant(v.first, v.second);
1992 if (Temp.MaxVGPRUsage < Best.MaxVGPRUsage)
1996 // if VGPR usage is still extremely high, we may spill. Try other variants
1997 // which are less performing, but that could lead to lower VGPR usage.
1998 if (Best.MaxVGPRUsage > 200) {
1999 static const std::pair<SISchedulerBlockCreatorVariant,
2000 SISchedulerBlockSchedulerVariant>
2002 // { LatenciesAlone, BlockRegUsageLatency },
2003 { LatenciesAlone, BlockRegUsage },
2004 // { LatenciesGrouped, BlockLatencyRegUsage },
2005 { LatenciesGrouped, BlockRegUsageLatency },
2006 { LatenciesGrouped, BlockRegUsage },
2007 // { LatenciesAlonePlusConsecutive, BlockLatencyRegUsage },
2008 { LatenciesAlonePlusConsecutive, BlockRegUsageLatency },
2009 { LatenciesAlonePlusConsecutive, BlockRegUsage }
2011 for (std::pair<SISchedulerBlockCreatorVariant, SISchedulerBlockSchedulerVariant> v : Variants) {
2012 Temp = Scheduler.scheduleVariant(v.first, v.second);
2013 if (Temp.MaxVGPRUsage < Best.MaxVGPRUsage)
2018 ScheduledSUnits = Best.SUs;
2019 ScheduledSUnitsInv.resize(SUnits.size());
2021 for (unsigned i = 0, e = (unsigned)SUnits.size(); i != e; ++i) {
2022 ScheduledSUnitsInv[ScheduledSUnits[i]] = i;
2027 // Tell the outside world about the result of the scheduling.
2029 assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker");
2030 TopRPTracker.setPos(CurrentTop);
2032 for (std::vector<unsigned>::iterator I = ScheduledSUnits.begin(),
2033 E = ScheduledSUnits.end(); I != E; ++I) {
2034 SUnit *SU = &SUnits[*I];
2036 scheduleMI(SU, true);
2038 LLVM_DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") "
2039 << *SU->getInstr());
2042 assert(CurrentTop == CurrentBottom && "Nonempty unscheduled zone.");
2047 dbgs() << "*** Final schedule for "
2048 << printMBBReference(*begin()->getParent()) << " ***\n";