1 //===---- ScheduleDAGInstrs.cpp - MachineInstr Rescheduling ---------------===//
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 //===----------------------------------------------------------------------===//
10 /// \file This implements the ScheduleDAGInstrs class, which implements
11 /// re-scheduling of MachineInstrs.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/CodeGen/ScheduleDAGInstrs.h"
16 #include "llvm/ADT/IntEqClasses.h"
17 #include "llvm/ADT/MapVector.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/SparseSet.h"
21 #include "llvm/ADT/iterator_range.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
25 #include "llvm/CodeGen/LivePhysRegs.h"
26 #include "llvm/CodeGen/MachineBasicBlock.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineInstr.h"
30 #include "llvm/CodeGen/MachineInstrBundle.h"
31 #include "llvm/CodeGen/MachineMemOperand.h"
32 #include "llvm/CodeGen/MachineOperand.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/PseudoSourceValue.h"
35 #include "llvm/CodeGen/RegisterPressure.h"
36 #include "llvm/CodeGen/ScheduleDAG.h"
37 #include "llvm/CodeGen/ScheduleDFS.h"
38 #include "llvm/CodeGen/SlotIndexes.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/Instruction.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/Operator.h"
44 #include "llvm/IR/Type.h"
45 #include "llvm/IR/Value.h"
46 #include "llvm/MC/LaneBitmask.h"
47 #include "llvm/MC/MCRegisterInfo.h"
48 #include "llvm/Support/Casting.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/Format.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/Target/TargetRegisterInfo.h"
56 #include "llvm/Target/TargetSubtargetInfo.h"
66 #define DEBUG_TYPE "machine-scheduler"
68 static cl::opt<bool> EnableAASchedMI("enable-aa-sched-mi", cl::Hidden,
69 cl::ZeroOrMore, cl::init(false),
70 cl::desc("Enable use of AA during MI DAG construction"));
72 static cl::opt<bool> UseTBAA("use-tbaa-in-sched-mi", cl::Hidden,
73 cl::init(true), cl::desc("Enable use of TBAA during MI DAG construction"));
75 // Note: the two options below might be used in tuning compile time vs
76 // output quality. Setting HugeRegion so large that it will never be
77 // reached means best-effort, but may be slow.
79 // When Stores and Loads maps (or NonAliasStores and NonAliasLoads)
80 // together hold this many SUs, a reduction of maps will be done.
81 static cl::opt<unsigned> HugeRegion("dag-maps-huge-region", cl::Hidden,
82 cl::init(1000), cl::desc("The limit to use while constructing the DAG "
83 "prior to scheduling, at which point a trade-off "
84 "is made to avoid excessive compile time."));
86 static cl::opt<unsigned> ReductionSize(
87 "dag-maps-reduction-size", cl::Hidden,
88 cl::desc("A huge scheduling region will have maps reduced by this many "
89 "nodes at a time. Defaults to HugeRegion / 2."));
91 static unsigned getReductionSize() {
92 // Always reduce a huge region with half of the elements, except
93 // when user sets this number explicitly.
94 if (ReductionSize.getNumOccurrences() == 0)
95 return HugeRegion / 2;
99 static void dumpSUList(ScheduleDAGInstrs::SUList &L) {
100 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
102 for (const SUnit *su : L) {
103 dbgs() << "SU(" << su->NodeNum << ")";
111 ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf,
112 const MachineLoopInfo *mli,
113 bool RemoveKillFlags)
114 : ScheduleDAG(mf), MLI(mli), MFI(mf.getFrameInfo()),
115 RemoveKillFlags(RemoveKillFlags),
116 UnknownValue(UndefValue::get(
117 Type::getVoidTy(mf.getFunction()->getContext()))) {
120 const TargetSubtargetInfo &ST = mf.getSubtarget();
121 SchedModel.init(ST.getSchedModel(), &ST, TII);
124 /// If this machine instr has memory reference information and it can be tracked
125 /// to a normal reference to a known object, return the Value for that object.
126 static void getUnderlyingObjectsForInstr(const MachineInstr *MI,
127 const MachineFrameInfo &MFI,
128 UnderlyingObjectsVector &Objects,
129 const DataLayout &DL) {
130 auto allMMOsOkay = [&]() {
131 for (const MachineMemOperand *MMO : MI->memoperands()) {
132 if (MMO->isVolatile())
135 if (const PseudoSourceValue *PSV = MMO->getPseudoValue()) {
136 // Function that contain tail calls don't have unique PseudoSourceValue
137 // objects. Two PseudoSourceValues might refer to the same or
138 // overlapping locations. The client code calling this function assumes
139 // this is not the case. So return a conservative answer of no known
141 if (MFI.hasTailCall())
144 // For now, ignore PseudoSourceValues which may alias LLVM IR values
145 // because the code that uses this function has no way to cope with
147 if (PSV->isAliased(&MFI))
150 bool MayAlias = PSV->mayAlias(&MFI);
151 Objects.push_back(UnderlyingObjectsVector::value_type(PSV, MayAlias));
152 } else if (const Value *V = MMO->getValue()) {
153 SmallVector<Value *, 4> Objs;
154 getUnderlyingObjectsForCodeGen(V, Objs, DL);
156 for (Value *V : Objs) {
157 assert(isIdentifiedObject(V));
158 Objects.push_back(UnderlyingObjectsVector::value_type(V, true));
170 void ScheduleDAGInstrs::startBlock(MachineBasicBlock *bb) {
174 void ScheduleDAGInstrs::finishBlock() {
175 // Subclasses should no longer refer to the old block.
179 void ScheduleDAGInstrs::enterRegion(MachineBasicBlock *bb,
180 MachineBasicBlock::iterator begin,
181 MachineBasicBlock::iterator end,
182 unsigned regioninstrs) {
183 assert(bb == BB && "startBlock should set BB");
186 NumRegionInstrs = regioninstrs;
189 void ScheduleDAGInstrs::exitRegion() {
193 void ScheduleDAGInstrs::addSchedBarrierDeps() {
194 MachineInstr *ExitMI = RegionEnd != BB->end() ? &*RegionEnd : nullptr;
195 ExitSU.setInstr(ExitMI);
196 // Add dependencies on the defs and uses of the instruction.
198 for (const MachineOperand &MO : ExitMI->operands()) {
199 if (!MO.isReg() || MO.isDef()) continue;
200 unsigned Reg = MO.getReg();
201 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
202 Uses.insert(PhysRegSUOper(&ExitSU, -1, Reg));
203 } else if (TargetRegisterInfo::isVirtualRegister(Reg) && MO.readsReg()) {
204 addVRegUseDeps(&ExitSU, ExitMI->getOperandNo(&MO));
208 if (!ExitMI || (!ExitMI->isCall() && !ExitMI->isBarrier())) {
209 // For others, e.g. fallthrough, conditional branch, assume the exit
210 // uses all the registers that are livein to the successor blocks.
211 for (const MachineBasicBlock *Succ : BB->successors()) {
212 for (const auto &LI : Succ->liveins()) {
213 if (!Uses.contains(LI.PhysReg))
214 Uses.insert(PhysRegSUOper(&ExitSU, -1, LI.PhysReg));
220 /// MO is an operand of SU's instruction that defines a physical register. Adds
221 /// data dependencies from SU to any uses of the physical register.
222 void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
223 const MachineOperand &MO = SU->getInstr()->getOperand(OperIdx);
224 assert(MO.isDef() && "expect physreg def");
226 // Ask the target if address-backscheduling is desirable, and if so how much.
227 const TargetSubtargetInfo &ST = MF.getSubtarget();
229 for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
230 Alias.isValid(); ++Alias) {
231 if (!Uses.contains(*Alias))
233 for (Reg2SUnitsMap::iterator I = Uses.find(*Alias); I != Uses.end(); ++I) {
234 SUnit *UseSU = I->SU;
238 // Adjust the dependence latency using operand def/use information,
239 // then allow the target to perform its own adjustments.
240 int UseOp = I->OpIdx;
241 MachineInstr *RegUse = nullptr;
244 Dep = SDep(SU, SDep::Artificial);
246 // Set the hasPhysRegDefs only for physreg defs that have a use within
247 // the scheduling region.
248 SU->hasPhysRegDefs = true;
249 Dep = SDep(SU, SDep::Data, *Alias);
250 RegUse = UseSU->getInstr();
253 SchedModel.computeOperandLatency(SU->getInstr(), OperIdx, RegUse,
256 ST.adjustSchedDependency(SU, UseSU, Dep);
262 /// \brief Adds register dependencies (data, anti, and output) from this SUnit
263 /// to following instructions in the same scheduling region that depend the
264 /// physical register referenced at OperIdx.
265 void ScheduleDAGInstrs::addPhysRegDeps(SUnit *SU, unsigned OperIdx) {
266 MachineInstr *MI = SU->getInstr();
267 MachineOperand &MO = MI->getOperand(OperIdx);
268 unsigned Reg = MO.getReg();
269 // We do not need to track any dependencies for constant registers.
270 if (MRI.isConstantPhysReg(Reg))
273 // Optionally add output and anti dependencies. For anti
274 // dependencies we use a latency of 0 because for a multi-issue
275 // target we want to allow the defining instruction to issue
276 // in the same cycle as the using instruction.
277 // TODO: Using a latency of 1 here for output dependencies assumes
278 // there's no cost for reusing registers.
279 SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
280 for (MCRegAliasIterator Alias(Reg, TRI, true); Alias.isValid(); ++Alias) {
281 if (!Defs.contains(*Alias))
283 for (Reg2SUnitsMap::iterator I = Defs.find(*Alias); I != Defs.end(); ++I) {
284 SUnit *DefSU = I->SU;
285 if (DefSU == &ExitSU)
288 (Kind != SDep::Output || !MO.isDead() ||
289 !DefSU->getInstr()->registerDefIsDead(*Alias))) {
290 if (Kind == SDep::Anti)
291 DefSU->addPred(SDep(SU, Kind, /*Reg=*/*Alias));
293 SDep Dep(SU, Kind, /*Reg=*/*Alias);
295 SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
303 SU->hasPhysRegUses = true;
304 // Either insert a new Reg2SUnits entry with an empty SUnits list, or
305 // retrieve the existing SUnits list for this register's uses.
306 // Push this SUnit on the use list.
307 Uses.insert(PhysRegSUOper(SU, OperIdx, Reg));
311 addPhysRegDataDeps(SU, OperIdx);
313 // clear this register's use list
314 if (Uses.contains(Reg))
319 } else if (SU->isCall) {
320 // Calls will not be reordered because of chain dependencies (see
321 // below). Since call operands are dead, calls may continue to be added
322 // to the DefList making dependence checking quadratic in the size of
323 // the block. Instead, we leave only one call at the back of the
325 Reg2SUnitsMap::RangePair P = Defs.equal_range(Reg);
326 Reg2SUnitsMap::iterator B = P.first;
327 Reg2SUnitsMap::iterator I = P.second;
328 for (bool isBegin = I == B; !isBegin; /* empty */) {
329 isBegin = (--I) == B;
336 // Defs are pushed in the order they are visited and never reordered.
337 Defs.insert(PhysRegSUOper(SU, OperIdx, Reg));
341 LaneBitmask ScheduleDAGInstrs::getLaneMaskForMO(const MachineOperand &MO) const
343 unsigned Reg = MO.getReg();
344 // No point in tracking lanemasks if we don't have interesting subregisters.
345 const TargetRegisterClass &RC = *MRI.getRegClass(Reg);
346 if (!RC.HasDisjunctSubRegs)
347 return LaneBitmask::getAll();
349 unsigned SubReg = MO.getSubReg();
351 return RC.getLaneMask();
352 return TRI->getSubRegIndexLaneMask(SubReg);
355 /// Adds register output and data dependencies from this SUnit to instructions
356 /// that occur later in the same scheduling region if they read from or write to
357 /// the virtual register defined at OperIdx.
359 /// TODO: Hoist loop induction variable increments. This has to be
360 /// reevaluated. Generally, IV scheduling should be done before coalescing.
361 void ScheduleDAGInstrs::addVRegDefDeps(SUnit *SU, unsigned OperIdx) {
362 MachineInstr *MI = SU->getInstr();
363 MachineOperand &MO = MI->getOperand(OperIdx);
364 unsigned Reg = MO.getReg();
366 LaneBitmask DefLaneMask;
367 LaneBitmask KillLaneMask;
368 if (TrackLaneMasks) {
369 bool IsKill = MO.getSubReg() == 0 || MO.isUndef();
370 DefLaneMask = getLaneMaskForMO(MO);
371 // If we have a <read-undef> flag, none of the lane values comes from an
372 // earlier instruction.
373 KillLaneMask = IsKill ? LaneBitmask::getAll() : DefLaneMask;
375 // Clear undef flag, we'll re-add it later once we know which subregister
377 MO.setIsUndef(false);
379 DefLaneMask = LaneBitmask::getAll();
380 KillLaneMask = LaneBitmask::getAll();
384 assert(CurrentVRegUses.find(Reg) == CurrentVRegUses.end() &&
385 "Dead defs should have no uses");
387 // Add data dependence to all uses we found so far.
388 const TargetSubtargetInfo &ST = MF.getSubtarget();
389 for (VReg2SUnitOperIdxMultiMap::iterator I = CurrentVRegUses.find(Reg),
390 E = CurrentVRegUses.end(); I != E; /*empty*/) {
391 LaneBitmask LaneMask = I->LaneMask;
392 // Ignore uses of other lanes.
393 if ((LaneMask & KillLaneMask).none()) {
398 if ((LaneMask & DefLaneMask).any()) {
399 SUnit *UseSU = I->SU;
400 MachineInstr *Use = UseSU->getInstr();
401 SDep Dep(SU, SDep::Data, Reg);
402 Dep.setLatency(SchedModel.computeOperandLatency(MI, OperIdx, Use,
404 ST.adjustSchedDependency(SU, UseSU, Dep);
408 LaneMask &= ~KillLaneMask;
409 // If we found a Def for all lanes of this use, remove it from the list.
410 if (LaneMask.any()) {
411 I->LaneMask = LaneMask;
414 I = CurrentVRegUses.erase(I);
418 // Shortcut: Singly defined vregs do not have output/anti dependencies.
419 if (MRI.hasOneDef(Reg))
422 // Add output dependence to the next nearest defs of this vreg.
424 // Unless this definition is dead, the output dependence should be
425 // transitively redundant with antidependencies from this definition's
426 // uses. We're conservative for now until we have a way to guarantee the uses
427 // are not eliminated sometime during scheduling. The output dependence edge
428 // is also useful if output latency exceeds def-use latency.
429 LaneBitmask LaneMask = DefLaneMask;
430 for (VReg2SUnit &V2SU : make_range(CurrentVRegDefs.find(Reg),
431 CurrentVRegDefs.end())) {
432 // Ignore defs for other lanes.
433 if ((V2SU.LaneMask & LaneMask).none())
435 // Add an output dependence.
436 SUnit *DefSU = V2SU.SU;
437 // Ignore additional defs of the same lanes in one instruction. This can
438 // happen because lanemasks are shared for targets with too many
439 // subregisters. We also use some representration tricks/hacks where we
440 // add super-register defs/uses, to imply that although we only access parts
441 // of the reg we care about the full one.
444 SDep Dep(SU, SDep::Output, Reg);
446 SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
449 // Update current definition. This can get tricky if the def was about a
450 // bigger lanemask before. We then have to shrink it and create a new
451 // VReg2SUnit for the non-overlapping part.
452 LaneBitmask OverlapMask = V2SU.LaneMask & LaneMask;
453 LaneBitmask NonOverlapMask = V2SU.LaneMask & ~LaneMask;
455 V2SU.LaneMask = OverlapMask;
456 if (NonOverlapMask.any())
457 CurrentVRegDefs.insert(VReg2SUnit(Reg, NonOverlapMask, DefSU));
459 // If there was no CurrentVRegDefs entry for some lanes yet, create one.
461 CurrentVRegDefs.insert(VReg2SUnit(Reg, LaneMask, SU));
464 /// \brief Adds a register data dependency if the instruction that defines the
465 /// virtual register used at OperIdx is mapped to an SUnit. Add a register
466 /// antidependency from this SUnit to instructions that occur later in the same
467 /// scheduling region if they write the virtual register.
469 /// TODO: Handle ExitSU "uses" properly.
470 void ScheduleDAGInstrs::addVRegUseDeps(SUnit *SU, unsigned OperIdx) {
471 const MachineInstr *MI = SU->getInstr();
472 const MachineOperand &MO = MI->getOperand(OperIdx);
473 unsigned Reg = MO.getReg();
475 // Remember the use. Data dependencies will be added when we find the def.
476 LaneBitmask LaneMask = TrackLaneMasks ? getLaneMaskForMO(MO)
477 : LaneBitmask::getAll();
478 CurrentVRegUses.insert(VReg2SUnitOperIdx(Reg, LaneMask, OperIdx, SU));
480 // Add antidependences to the following defs of the vreg.
481 for (VReg2SUnit &V2SU : make_range(CurrentVRegDefs.find(Reg),
482 CurrentVRegDefs.end())) {
483 // Ignore defs for unrelated lanes.
484 LaneBitmask PrevDefLaneMask = V2SU.LaneMask;
485 if ((PrevDefLaneMask & LaneMask).none())
490 V2SU.SU->addPred(SDep(SU, SDep::Anti, Reg));
494 /// Returns true if MI is an instruction we are unable to reason about
495 /// (like a call or something with unmodeled side effects).
496 static inline bool isGlobalMemoryObject(AliasAnalysis *AA, MachineInstr *MI) {
497 return MI->isCall() || MI->hasUnmodeledSideEffects() ||
498 (MI->hasOrderedMemoryRef() && !MI->isDereferenceableInvariantLoad(AA));
501 void ScheduleDAGInstrs::addChainDependency (SUnit *SUa, SUnit *SUb,
503 if (SUa->getInstr()->mayAlias(AAForDep, *SUb->getInstr(), UseTBAA)) {
504 SDep Dep(SUa, SDep::MayAliasMem);
505 Dep.setLatency(Latency);
510 /// \brief Creates an SUnit for each real instruction, numbered in top-down
511 /// topological order. The instruction order A < B, implies that no edge exists
514 /// Map each real instruction to its SUnit.
516 /// After initSUnits, the SUnits vector cannot be resized and the scheduler may
517 /// hang onto SUnit pointers. We may relax this in the future by using SUnit IDs
518 /// instead of pointers.
520 /// MachineScheduler relies on initSUnits numbering the nodes by their order in
521 /// the original instruction list.
522 void ScheduleDAGInstrs::initSUnits() {
523 // We'll be allocating one SUnit for each real instruction in the region,
524 // which is contained within a basic block.
525 SUnits.reserve(NumRegionInstrs);
527 for (MachineInstr &MI : make_range(RegionBegin, RegionEnd)) {
528 if (MI.isDebugValue())
531 SUnit *SU = newSUnit(&MI);
532 MISUnitMap[&MI] = SU;
534 SU->isCall = MI.isCall();
535 SU->isCommutable = MI.isCommutable();
537 // Assign the Latency field of SU using target-provided information.
538 SU->Latency = SchedModel.computeInstrLatency(SU->getInstr());
540 // If this SUnit uses a reserved or unbuffered resource, mark it as such.
542 // Reserved resources block an instruction from issuing and stall the
543 // entire pipeline. These are identified by BufferSize=0.
545 // Unbuffered resources prevent execution of subsequent instructions that
546 // require the same resources. This is used for in-order execution pipelines
547 // within an out-of-order core. These are identified by BufferSize=1.
548 if (SchedModel.hasInstrSchedModel()) {
549 const MCSchedClassDesc *SC = getSchedClass(SU);
550 for (const MCWriteProcResEntry &PRE :
551 make_range(SchedModel.getWriteProcResBegin(SC),
552 SchedModel.getWriteProcResEnd(SC))) {
553 switch (SchedModel.getProcResource(PRE.ProcResourceIdx)->BufferSize) {
555 SU->hasReservedResource = true;
558 SU->isUnbuffered = true;
568 class ScheduleDAGInstrs::Value2SUsMap : public MapVector<ValueType, SUList> {
569 /// Current total number of SUs in map.
570 unsigned NumNodes = 0;
572 /// 1 for loads, 0 for stores. (see comment in SUList)
573 unsigned TrueMemOrderLatency;
576 Value2SUsMap(unsigned lat = 0) : TrueMemOrderLatency(lat) {}
578 /// To keep NumNodes up to date, insert() is used instead of
579 /// this operator w/ push_back().
580 ValueType &operator[](const SUList &Key) {
581 llvm_unreachable("Don't use. Use insert() instead."); };
583 /// Adds SU to the SUList of V. If Map grows huge, reduce its size by calling
585 void inline insert(SUnit *SU, ValueType V) {
586 MapVector::operator[](V).push_back(SU);
590 /// Clears the list of SUs mapped to V.
591 void inline clearList(ValueType V) {
592 iterator Itr = find(V);
594 assert(NumNodes >= Itr->second.size());
595 NumNodes -= Itr->second.size();
601 /// Clears map from all contents.
603 MapVector<ValueType, SUList>::clear();
607 unsigned inline size() const { return NumNodes; }
609 /// Counts the number of SUs in this map after a reduction.
610 void reComputeSize() {
612 for (auto &I : *this)
613 NumNodes += I.second.size();
616 unsigned inline getTrueMemOrderLatency() const {
617 return TrueMemOrderLatency;
623 void ScheduleDAGInstrs::addChainDependencies(SUnit *SU,
624 Value2SUsMap &Val2SUsMap) {
625 for (auto &I : Val2SUsMap)
626 addChainDependencies(SU, I.second,
627 Val2SUsMap.getTrueMemOrderLatency());
630 void ScheduleDAGInstrs::addChainDependencies(SUnit *SU,
631 Value2SUsMap &Val2SUsMap,
633 Value2SUsMap::iterator Itr = Val2SUsMap.find(V);
634 if (Itr != Val2SUsMap.end())
635 addChainDependencies(SU, Itr->second,
636 Val2SUsMap.getTrueMemOrderLatency());
639 void ScheduleDAGInstrs::addBarrierChain(Value2SUsMap &map) {
640 assert(BarrierChain != nullptr);
642 for (auto &I : map) {
643 SUList &sus = I.second;
645 SU->addPredBarrier(BarrierChain);
650 void ScheduleDAGInstrs::insertBarrierChain(Value2SUsMap &map) {
651 assert(BarrierChain != nullptr);
653 // Go through all lists of SUs.
654 for (Value2SUsMap::iterator I = map.begin(), EE = map.end(); I != EE;) {
655 Value2SUsMap::iterator CurrItr = I++;
656 SUList &sus = CurrItr->second;
657 SUList::iterator SUItr = sus.begin(), SUEE = sus.end();
658 for (; SUItr != SUEE; ++SUItr) {
659 // Stop on BarrierChain or any instruction above it.
660 if ((*SUItr)->NodeNum <= BarrierChain->NodeNum)
663 (*SUItr)->addPredBarrier(BarrierChain);
666 // Remove also the BarrierChain from list if present.
667 if (SUItr != SUEE && *SUItr == BarrierChain)
670 // Remove all SUs that are now successors of BarrierChain.
671 if (SUItr != sus.begin())
672 sus.erase(sus.begin(), SUItr);
675 // Remove all entries with empty su lists.
676 map.remove_if([&](std::pair<ValueType, SUList> &mapEntry) {
677 return (mapEntry.second.empty()); });
679 // Recompute the size of the map (NumNodes).
683 void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
684 RegPressureTracker *RPTracker,
685 PressureDiffs *PDiffs,
687 bool TrackLaneMasks) {
688 const TargetSubtargetInfo &ST = MF.getSubtarget();
689 bool UseAA = EnableAASchedMI.getNumOccurrences() > 0 ? EnableAASchedMI
691 AAForDep = UseAA ? AA : nullptr;
693 BarrierChain = nullptr;
695 this->TrackLaneMasks = TrackLaneMasks;
697 ScheduleDAG::clearDAG();
699 // Create an SUnit for each real instruction.
703 PDiffs->init(SUnits.size());
705 // We build scheduling units by walking a block's instruction list
706 // from bottom to top.
708 // Each MIs' memory operand(s) is analyzed to a list of underlying
709 // objects. The SU is then inserted in the SUList(s) mapped from the
710 // Value(s). Each Value thus gets mapped to lists of SUs depending
711 // on it, stores and loads kept separately. Two SUs are trivially
712 // non-aliasing if they both depend on only identified Values and do
713 // not share any common Value.
714 Value2SUsMap Stores, Loads(1 /*TrueMemOrderLatency*/);
716 // Certain memory accesses are known to not alias any SU in Stores
717 // or Loads, and have therefore their own 'NonAlias'
718 // domain. E.g. spill / reload instructions never alias LLVM I/R
719 // Values. It would be nice to assume that this type of memory
720 // accesses always have a proper memory operand modelling, and are
721 // therefore never unanalyzable, but this is conservatively not
723 Value2SUsMap NonAliasStores, NonAliasLoads(1 /*TrueMemOrderLatency*/);
725 // Remove any stale debug info; sometimes BuildSchedGraph is called again
726 // without emitting the info from the previous call.
728 FirstDbgValue = nullptr;
730 assert(Defs.empty() && Uses.empty() &&
731 "Only BuildGraph should update Defs/Uses");
732 Defs.setUniverse(TRI->getNumRegs());
733 Uses.setUniverse(TRI->getNumRegs());
735 assert(CurrentVRegDefs.empty() && "nobody else should use CurrentVRegDefs");
736 assert(CurrentVRegUses.empty() && "nobody else should use CurrentVRegUses");
737 unsigned NumVirtRegs = MRI.getNumVirtRegs();
738 CurrentVRegDefs.setUniverse(NumVirtRegs);
739 CurrentVRegUses.setUniverse(NumVirtRegs);
741 // Model data dependencies between instructions being scheduled and the
743 addSchedBarrierDeps();
745 // Walk the list of instructions, from bottom moving up.
746 MachineInstr *DbgMI = nullptr;
747 for (MachineBasicBlock::iterator MII = RegionEnd, MIE = RegionBegin;
749 MachineInstr &MI = *std::prev(MII);
751 DbgValues.push_back(std::make_pair(DbgMI, &MI));
755 if (MI.isDebugValue()) {
759 SUnit *SU = MISUnitMap[&MI];
760 assert(SU && "No SUnit mapped to this MI");
763 RegisterOperands RegOpers;
764 RegOpers.collect(MI, *TRI, MRI, TrackLaneMasks, false);
765 if (TrackLaneMasks) {
766 SlotIndex SlotIdx = LIS->getInstructionIndex(MI);
767 RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx);
769 if (PDiffs != nullptr)
770 PDiffs->addInstruction(SU->NodeNum, RegOpers, MRI);
772 RPTracker->recedeSkipDebugValues();
773 assert(&*RPTracker->getPos() == &MI && "RPTracker in sync");
774 RPTracker->recede(RegOpers);
778 (CanHandleTerminators || (!MI.isTerminator() && !MI.isPosition())) &&
779 "Cannot schedule terminators or labels!");
781 // Add register-based dependencies (data, anti, and output).
782 // For some instructions (calls, returns, inline-asm, etc.) there can
783 // be explicit uses and implicit defs, in which case the use will appear
784 // on the operand list before the def. Do two passes over the operand
785 // list to make sure that defs are processed before any uses.
786 bool HasVRegDef = false;
787 for (unsigned j = 0, n = MI.getNumOperands(); j != n; ++j) {
788 const MachineOperand &MO = MI.getOperand(j);
789 if (!MO.isReg() || !MO.isDef())
791 unsigned Reg = MO.getReg();
792 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
793 addPhysRegDeps(SU, j);
794 } else if (TargetRegisterInfo::isVirtualRegister(Reg)) {
796 addVRegDefDeps(SU, j);
799 // Now process all uses.
800 for (unsigned j = 0, n = MI.getNumOperands(); j != n; ++j) {
801 const MachineOperand &MO = MI.getOperand(j);
802 // Only look at use operands.
803 // We do not need to check for MO.readsReg() here because subsequent
804 // subregister defs will get output dependence edges and need no
805 // additional use dependencies.
806 if (!MO.isReg() || !MO.isUse())
808 unsigned Reg = MO.getReg();
809 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
810 addPhysRegDeps(SU, j);
811 } else if (TargetRegisterInfo::isVirtualRegister(Reg) && MO.readsReg()) {
812 addVRegUseDeps(SU, j);
816 // If we haven't seen any uses in this scheduling region, create a
817 // dependence edge to ExitSU to model the live-out latency. This is required
818 // for vreg defs with no in-region use, and prefetches with no vreg def.
820 // FIXME: NumDataSuccs would be more precise than NumSuccs here. This
821 // check currently relies on being called before adding chain deps.
822 if (SU->NumSuccs == 0 && SU->Latency > 1 && (HasVRegDef || MI.mayLoad())) {
823 SDep Dep(SU, SDep::Artificial);
824 Dep.setLatency(SU->Latency - 1);
828 // Add memory dependencies (Note: isStoreToStackSlot and
829 // isLoadFromStackSLot are not usable after stack slots are lowered to
830 // actual addresses).
832 // This is a barrier event that acts as a pivotal node in the DAG.
833 if (isGlobalMemoryObject(AA, &MI)) {
835 // Become the barrier chain.
837 BarrierChain->addPredBarrier(SU);
840 DEBUG(dbgs() << "Global memory object and new barrier chain: SU("
841 << BarrierChain->NodeNum << ").\n";);
843 // Add dependencies against everything below it and clear maps.
844 addBarrierChain(Stores);
845 addBarrierChain(Loads);
846 addBarrierChain(NonAliasStores);
847 addBarrierChain(NonAliasLoads);
852 // If it's not a store or a variant load, we're done.
853 if (!MI.mayStore() &&
854 !(MI.mayLoad() && !MI.isDereferenceableInvariantLoad(AA)))
857 // Always add dependecy edge to BarrierChain if present.
859 BarrierChain->addPredBarrier(SU);
861 // Find the underlying objects for MI. The Objs vector is either
862 // empty, or filled with the Values of memory locations which this
863 // SU depends on. An empty vector means the memory location is
864 // unknown, and may alias anything.
865 UnderlyingObjectsVector Objs;
866 getUnderlyingObjectsForInstr(&MI, MFI, Objs, MF.getDataLayout());
870 // An unknown store depends on all stores and loads.
871 addChainDependencies(SU, Stores);
872 addChainDependencies(SU, NonAliasStores);
873 addChainDependencies(SU, Loads);
874 addChainDependencies(SU, NonAliasLoads);
876 // Map this store to 'UnknownValue'.
877 Stores.insert(SU, UnknownValue);
879 // Add precise dependencies against all previously seen memory
880 // accesses mapped to the same Value(s).
881 for (const UnderlyingObject &UnderlObj : Objs) {
882 ValueType V = UnderlObj.getValue();
883 bool ThisMayAlias = UnderlObj.mayAlias();
885 // Add dependencies to previous stores and loads mapped to V.
886 addChainDependencies(SU, (ThisMayAlias ? Stores : NonAliasStores), V);
887 addChainDependencies(SU, (ThisMayAlias ? Loads : NonAliasLoads), V);
889 // Update the store map after all chains have been added to avoid adding
890 // self-loop edge if multiple underlying objects are present.
891 for (const UnderlyingObject &UnderlObj : Objs) {
892 ValueType V = UnderlObj.getValue();
893 bool ThisMayAlias = UnderlObj.mayAlias();
895 // Map this store to V.
896 (ThisMayAlias ? Stores : NonAliasStores).insert(SU, V);
898 // The store may have dependencies to unanalyzable loads and
900 addChainDependencies(SU, Loads, UnknownValue);
901 addChainDependencies(SU, Stores, UnknownValue);
903 } else { // SU is a load.
905 // An unknown load depends on all stores.
906 addChainDependencies(SU, Stores);
907 addChainDependencies(SU, NonAliasStores);
909 Loads.insert(SU, UnknownValue);
911 for (const UnderlyingObject &UnderlObj : Objs) {
912 ValueType V = UnderlObj.getValue();
913 bool ThisMayAlias = UnderlObj.mayAlias();
915 // Add precise dependencies against all previously seen stores
916 // mapping to the same Value(s).
917 addChainDependencies(SU, (ThisMayAlias ? Stores : NonAliasStores), V);
919 // Map this load to V.
920 (ThisMayAlias ? Loads : NonAliasLoads).insert(SU, V);
922 // The load may have dependencies to unanalyzable stores.
923 addChainDependencies(SU, Stores, UnknownValue);
927 // Reduce maps if they grow huge.
928 if (Stores.size() + Loads.size() >= HugeRegion) {
929 DEBUG(dbgs() << "Reducing Stores and Loads maps.\n";);
930 reduceHugeMemNodeMaps(Stores, Loads, getReductionSize());
932 if (NonAliasStores.size() + NonAliasLoads.size() >= HugeRegion) {
933 DEBUG(dbgs() << "Reducing NonAliasStores and NonAliasLoads maps.\n";);
934 reduceHugeMemNodeMaps(NonAliasStores, NonAliasLoads, getReductionSize());
939 FirstDbgValue = DbgMI;
943 CurrentVRegDefs.clear();
944 CurrentVRegUses.clear();
947 raw_ostream &llvm::operator<<(raw_ostream &OS, const PseudoSourceValue* PSV) {
948 PSV->printCustom(OS);
952 void ScheduleDAGInstrs::Value2SUsMap::dump() {
953 for (auto &Itr : *this) {
954 if (Itr.first.is<const Value*>()) {
955 const Value *V = Itr.first.get<const Value*>();
956 if (isa<UndefValue>(V))
959 V->printAsOperand(dbgs());
961 else if (Itr.first.is<const PseudoSourceValue*>())
962 dbgs() << Itr.first.get<const PseudoSourceValue*>();
964 llvm_unreachable("Unknown Value type.");
967 dumpSUList(Itr.second);
971 void ScheduleDAGInstrs::reduceHugeMemNodeMaps(Value2SUsMap &stores,
972 Value2SUsMap &loads, unsigned N) {
973 DEBUG(dbgs() << "Before reduction:\nStoring SUnits:\n";
975 dbgs() << "Loading SUnits:\n";
978 // Insert all SU's NodeNums into a vector and sort it.
979 std::vector<unsigned> NodeNums;
980 NodeNums.reserve(stores.size() + loads.size());
981 for (auto &I : stores)
982 for (auto *SU : I.second)
983 NodeNums.push_back(SU->NodeNum);
984 for (auto &I : loads)
985 for (auto *SU : I.second)
986 NodeNums.push_back(SU->NodeNum);
987 std::sort(NodeNums.begin(), NodeNums.end());
989 // The N last elements in NodeNums will be removed, and the SU with
990 // the lowest NodeNum of them will become the new BarrierChain to
991 // let the not yet seen SUs have a dependency to the removed SUs.
992 assert(N <= NodeNums.size());
993 SUnit *newBarrierChain = &SUnits[*(NodeNums.end() - N)];
995 // The aliasing and non-aliasing maps reduce independently of each
996 // other, but share a common BarrierChain. Check if the
997 // newBarrierChain is above the former one. If it is not, it may
998 // introduce a loop to use newBarrierChain, so keep the old one.
999 if (newBarrierChain->NodeNum < BarrierChain->NodeNum) {
1000 BarrierChain->addPredBarrier(newBarrierChain);
1001 BarrierChain = newBarrierChain;
1002 DEBUG(dbgs() << "Inserting new barrier chain: SU("
1003 << BarrierChain->NodeNum << ").\n";);
1006 DEBUG(dbgs() << "Keeping old barrier chain: SU("
1007 << BarrierChain->NodeNum << ").\n";);
1010 BarrierChain = newBarrierChain;
1012 insertBarrierChain(stores);
1013 insertBarrierChain(loads);
1015 DEBUG(dbgs() << "After reduction:\nStoring SUnits:\n";
1017 dbgs() << "Loading SUnits:\n";
1021 static void toggleKills(const MachineRegisterInfo &MRI, LivePhysRegs &LiveRegs,
1022 MachineInstr &MI, bool addToLiveRegs) {
1023 for (MachineOperand &MO : MI.operands()) {
1024 if (!MO.isReg() || !MO.readsReg())
1026 unsigned Reg = MO.getReg();
1030 // Things that are available after the instruction are killed by it.
1031 bool IsKill = LiveRegs.available(MRI, Reg);
1032 MO.setIsKill(IsKill);
1034 LiveRegs.addReg(Reg);
1038 void ScheduleDAGInstrs::fixupKills(MachineBasicBlock &MBB) {
1039 DEBUG(dbgs() << "Fixup kills for BB#" << MBB.getNumber() << '\n');
1041 LiveRegs.init(*TRI);
1042 LiveRegs.addLiveOuts(MBB);
1044 // Examine block from end to start...
1045 for (MachineInstr &MI : make_range(MBB.rbegin(), MBB.rend())) {
1046 if (MI.isDebugValue())
1049 // Update liveness. Registers that are defed but not used in this
1050 // instruction are now dead. Mark register and all subregs as they
1051 // are completely defined.
1052 for (ConstMIBundleOperands O(MI); O.isValid(); ++O) {
1053 const MachineOperand &MO = *O;
1057 unsigned Reg = MO.getReg();
1060 LiveRegs.removeReg(Reg);
1061 } else if (MO.isRegMask()) {
1062 LiveRegs.removeRegsInMask(MO);
1066 // If there is a bundle header fix it up first.
1067 if (!MI.isBundled()) {
1068 toggleKills(MRI, LiveRegs, MI, true);
1070 MachineBasicBlock::instr_iterator First = MI.getIterator();
1071 if (MI.isBundle()) {
1072 toggleKills(MRI, LiveRegs, MI, false);
1075 // Some targets make the (questionable) assumtion that the instructions
1076 // inside the bundle are ordered and consequently only the last use of
1077 // a register inside the bundle can kill it.
1078 MachineBasicBlock::instr_iterator I = std::next(First);
1079 while (I->isBundledWithSucc())
1082 if (!I->isDebugValue())
1083 toggleKills(MRI, LiveRegs, *I, true);
1085 } while(I != First);
1090 void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const {
1091 // Cannot completely remove virtual function even in release mode.
1092 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1093 SU->getInstr()->dump();
1097 std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const {
1099 raw_string_ostream oss(s);
1102 else if (SU == &ExitSU)
1105 SU->getInstr()->print(oss, /*SkipOpers=*/true);
1109 /// Return the basic block label. It is not necessarilly unique because a block
1110 /// contains multiple scheduling regions. But it is fine for visualization.
1111 std::string ScheduleDAGInstrs::getDAGName() const {
1112 return "dag." + BB->getFullName();
1115 //===----------------------------------------------------------------------===//
1116 // SchedDFSResult Implementation
1117 //===----------------------------------------------------------------------===//
1121 /// Internal state used to compute SchedDFSResult.
1122 class SchedDFSImpl {
1125 /// Join DAG nodes into equivalence classes by their subtree.
1126 IntEqClasses SubtreeClasses;
1127 /// List PredSU, SuccSU pairs that represent data edges between subtrees.
1128 std::vector<std::pair<const SUnit *, const SUnit*>> ConnectionPairs;
1132 unsigned ParentNodeID; ///< Parent node (member of the parent subtree).
1133 unsigned SubInstrCount = 0; ///< Instr count in this tree only, not
1136 RootData(unsigned id): NodeID(id),
1137 ParentNodeID(SchedDFSResult::InvalidSubtreeID) {}
1139 unsigned getSparseSetIndex() const { return NodeID; }
1142 SparseSet<RootData> RootSet;
1145 SchedDFSImpl(SchedDFSResult &r): R(r), SubtreeClasses(R.DFSNodeData.size()) {
1146 RootSet.setUniverse(R.DFSNodeData.size());
1149 /// Returns true if this node been visited by the DFS traversal.
1151 /// During visitPostorderNode the Node's SubtreeID is assigned to the Node
1152 /// ID. Later, SubtreeID is updated but remains valid.
1153 bool isVisited(const SUnit *SU) const {
1154 return R.DFSNodeData[SU->NodeNum].SubtreeID
1155 != SchedDFSResult::InvalidSubtreeID;
1158 /// Initializes this node's instruction count. We don't need to flag the node
1159 /// visited until visitPostorder because the DAG cannot have cycles.
1160 void visitPreorder(const SUnit *SU) {
1161 R.DFSNodeData[SU->NodeNum].InstrCount =
1162 SU->getInstr()->isTransient() ? 0 : 1;
1165 /// Called once for each node after all predecessors are visited. Revisit this
1166 /// node's predecessors and potentially join them now that we know the ILP of
1167 /// the other predecessors.
1168 void visitPostorderNode(const SUnit *SU) {
1169 // Mark this node as the root of a subtree. It may be joined with its
1170 // successors later.
1171 R.DFSNodeData[SU->NodeNum].SubtreeID = SU->NodeNum;
1172 RootData RData(SU->NodeNum);
1173 RData.SubInstrCount = SU->getInstr()->isTransient() ? 0 : 1;
1175 // If any predecessors are still in their own subtree, they either cannot be
1176 // joined or are large enough to remain separate. If this parent node's
1177 // total instruction count is not greater than a child subtree by at least
1178 // the subtree limit, then try to join it now since splitting subtrees is
1179 // only useful if multiple high-pressure paths are possible.
1180 unsigned InstrCount = R.DFSNodeData[SU->NodeNum].InstrCount;
1181 for (const SDep &PredDep : SU->Preds) {
1182 if (PredDep.getKind() != SDep::Data)
1184 unsigned PredNum = PredDep.getSUnit()->NodeNum;
1185 if ((InstrCount - R.DFSNodeData[PredNum].InstrCount) < R.SubtreeLimit)
1186 joinPredSubtree(PredDep, SU, /*CheckLimit=*/false);
1188 // Either link or merge the TreeData entry from the child to the parent.
1189 if (R.DFSNodeData[PredNum].SubtreeID == PredNum) {
1190 // If the predecessor's parent is invalid, this is a tree edge and the
1191 // current node is the parent.
1192 if (RootSet[PredNum].ParentNodeID == SchedDFSResult::InvalidSubtreeID)
1193 RootSet[PredNum].ParentNodeID = SU->NodeNum;
1195 else if (RootSet.count(PredNum)) {
1196 // The predecessor is not a root, but is still in the root set. This
1197 // must be the new parent that it was just joined to. Note that
1198 // RootSet[PredNum].ParentNodeID may either be invalid or may still be
1199 // set to the original parent.
1200 RData.SubInstrCount += RootSet[PredNum].SubInstrCount;
1201 RootSet.erase(PredNum);
1204 RootSet[SU->NodeNum] = RData;
1207 /// \brief Called once for each tree edge after calling visitPostOrderNode on
1208 /// the predecessor. Increment the parent node's instruction count and
1209 /// preemptively join this subtree to its parent's if it is small enough.
1210 void visitPostorderEdge(const SDep &PredDep, const SUnit *Succ) {
1211 R.DFSNodeData[Succ->NodeNum].InstrCount
1212 += R.DFSNodeData[PredDep.getSUnit()->NodeNum].InstrCount;
1213 joinPredSubtree(PredDep, Succ);
1216 /// Adds a connection for cross edges.
1217 void visitCrossEdge(const SDep &PredDep, const SUnit *Succ) {
1218 ConnectionPairs.push_back(std::make_pair(PredDep.getSUnit(), Succ));
1221 /// Sets each node's subtree ID to the representative ID and record
1222 /// connections between trees.
1224 SubtreeClasses.compress();
1225 R.DFSTreeData.resize(SubtreeClasses.getNumClasses());
1226 assert(SubtreeClasses.getNumClasses() == RootSet.size()
1227 && "number of roots should match trees");
1228 for (const RootData &Root : RootSet) {
1229 unsigned TreeID = SubtreeClasses[Root.NodeID];
1230 if (Root.ParentNodeID != SchedDFSResult::InvalidSubtreeID)
1231 R.DFSTreeData[TreeID].ParentTreeID = SubtreeClasses[Root.ParentNodeID];
1232 R.DFSTreeData[TreeID].SubInstrCount = Root.SubInstrCount;
1233 // Note that SubInstrCount may be greater than InstrCount if we joined
1234 // subtrees across a cross edge. InstrCount will be attributed to the
1235 // original parent, while SubInstrCount will be attributed to the joined
1238 R.SubtreeConnections.resize(SubtreeClasses.getNumClasses());
1239 R.SubtreeConnectLevels.resize(SubtreeClasses.getNumClasses());
1240 DEBUG(dbgs() << R.getNumSubtrees() << " subtrees:\n");
1241 for (unsigned Idx = 0, End = R.DFSNodeData.size(); Idx != End; ++Idx) {
1242 R.DFSNodeData[Idx].SubtreeID = SubtreeClasses[Idx];
1243 DEBUG(dbgs() << " SU(" << Idx << ") in tree "
1244 << R.DFSNodeData[Idx].SubtreeID << '\n');
1246 for (const std::pair<const SUnit*, const SUnit*> &P : ConnectionPairs) {
1247 unsigned PredTree = SubtreeClasses[P.first->NodeNum];
1248 unsigned SuccTree = SubtreeClasses[P.second->NodeNum];
1249 if (PredTree == SuccTree)
1251 unsigned Depth = P.first->getDepth();
1252 addConnection(PredTree, SuccTree, Depth);
1253 addConnection(SuccTree, PredTree, Depth);
1258 /// Joins the predecessor subtree with the successor that is its DFS parent.
1259 /// Applies some heuristics before joining.
1260 bool joinPredSubtree(const SDep &PredDep, const SUnit *Succ,
1261 bool CheckLimit = true) {
1262 assert(PredDep.getKind() == SDep::Data && "Subtrees are for data edges");
1264 // Check if the predecessor is already joined.
1265 const SUnit *PredSU = PredDep.getSUnit();
1266 unsigned PredNum = PredSU->NodeNum;
1267 if (R.DFSNodeData[PredNum].SubtreeID != PredNum)
1270 // Four is the magic number of successors before a node is considered a
1272 unsigned NumDataSucs = 0;
1273 for (const SDep &SuccDep : PredSU->Succs) {
1274 if (SuccDep.getKind() == SDep::Data) {
1275 if (++NumDataSucs >= 4)
1279 if (CheckLimit && R.DFSNodeData[PredNum].InstrCount > R.SubtreeLimit)
1281 R.DFSNodeData[PredNum].SubtreeID = Succ->NodeNum;
1282 SubtreeClasses.join(Succ->NodeNum, PredNum);
1286 /// Called by finalize() to record a connection between trees.
1287 void addConnection(unsigned FromTree, unsigned ToTree, unsigned Depth) {
1292 SmallVectorImpl<SchedDFSResult::Connection> &Connections =
1293 R.SubtreeConnections[FromTree];
1294 for (SchedDFSResult::Connection &C : Connections) {
1295 if (C.TreeID == ToTree) {
1296 C.Level = std::max(C.Level, Depth);
1300 Connections.push_back(SchedDFSResult::Connection(ToTree, Depth));
1301 FromTree = R.DFSTreeData[FromTree].ParentTreeID;
1302 } while (FromTree != SchedDFSResult::InvalidSubtreeID);
1306 } // end namespace llvm
1310 /// Manage the stack used by a reverse depth-first search over the DAG.
1311 class SchedDAGReverseDFS {
1312 std::vector<std::pair<const SUnit *, SUnit::const_pred_iterator>> DFSStack;
1315 bool isComplete() const { return DFSStack.empty(); }
1317 void follow(const SUnit *SU) {
1318 DFSStack.push_back(std::make_pair(SU, SU->Preds.begin()));
1320 void advance() { ++DFSStack.back().second; }
1322 const SDep *backtrack() {
1323 DFSStack.pop_back();
1324 return DFSStack.empty() ? nullptr : std::prev(DFSStack.back().second);
1327 const SUnit *getCurr() const { return DFSStack.back().first; }
1329 SUnit::const_pred_iterator getPred() const { return DFSStack.back().second; }
1331 SUnit::const_pred_iterator getPredEnd() const {
1332 return getCurr()->Preds.end();
1336 } // end anonymous namespace
1338 static bool hasDataSucc(const SUnit *SU) {
1339 for (const SDep &SuccDep : SU->Succs) {
1340 if (SuccDep.getKind() == SDep::Data &&
1341 !SuccDep.getSUnit()->isBoundaryNode())
1347 /// Computes an ILP metric for all nodes in the subDAG reachable via depth-first
1348 /// search from this root.
1349 void SchedDFSResult::compute(ArrayRef<SUnit> SUnits) {
1351 llvm_unreachable("Top-down ILP metric is unimplemnted");
1353 SchedDFSImpl Impl(*this);
1354 for (const SUnit &SU : SUnits) {
1355 if (Impl.isVisited(&SU) || hasDataSucc(&SU))
1358 SchedDAGReverseDFS DFS;
1359 Impl.visitPreorder(&SU);
1362 // Traverse the leftmost path as far as possible.
1363 while (DFS.getPred() != DFS.getPredEnd()) {
1364 const SDep &PredDep = *DFS.getPred();
1366 // Ignore non-data edges.
1367 if (PredDep.getKind() != SDep::Data
1368 || PredDep.getSUnit()->isBoundaryNode()) {
1371 // An already visited edge is a cross edge, assuming an acyclic DAG.
1372 if (Impl.isVisited(PredDep.getSUnit())) {
1373 Impl.visitCrossEdge(PredDep, DFS.getCurr());
1376 Impl.visitPreorder(PredDep.getSUnit());
1377 DFS.follow(PredDep.getSUnit());
1379 // Visit the top of the stack in postorder and backtrack.
1380 const SUnit *Child = DFS.getCurr();
1381 const SDep *PredDep = DFS.backtrack();
1382 Impl.visitPostorderNode(Child);
1384 Impl.visitPostorderEdge(*PredDep, DFS.getCurr());
1385 if (DFS.isComplete())
1392 /// The root of the given SubtreeID was just scheduled. For all subtrees
1393 /// connected to this tree, record the depth of the connection so that the
1394 /// nearest connected subtrees can be prioritized.
1395 void SchedDFSResult::scheduleTree(unsigned SubtreeID) {
1396 for (const Connection &C : SubtreeConnections[SubtreeID]) {
1397 SubtreeConnectLevels[C.TreeID] =
1398 std::max(SubtreeConnectLevels[C.TreeID], C.Level);
1399 DEBUG(dbgs() << " Tree: " << C.TreeID
1400 << " @" << SubtreeConnectLevels[C.TreeID] << '\n');
1404 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1405 LLVM_DUMP_METHOD void ILPValue::print(raw_ostream &OS) const {
1406 OS << InstrCount << " / " << Length << " = ";
1410 OS << format("%g", ((double)InstrCount / Length));
1413 LLVM_DUMP_METHOD void ILPValue::dump() const {
1414 dbgs() << *this << '\n';
1420 raw_ostream &operator<<(raw_ostream &OS, const ILPValue &Val) {
1425 } // end namespace llvm