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/ADT/IntEqClasses.h"
16 #include "llvm/ADT/iterator_range.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/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
24 #include "llvm/CodeGen/LivePhysRegs.h"
25 #include "llvm/CodeGen/MachineBasicBlock.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineInstr.h"
29 #include "llvm/CodeGen/MachineInstrBundle.h"
30 #include "llvm/CodeGen/MachineMemOperand.h"
31 #include "llvm/CodeGen/MachineOperand.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/PseudoSourceValue.h"
34 #include "llvm/CodeGen/RegisterPressure.h"
35 #include "llvm/CodeGen/ScheduleDAG.h"
36 #include "llvm/CodeGen/ScheduleDAGInstrs.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 "misched"
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 /// This is the function that does the work of looking through basic
125 /// ptrtoint+arithmetic+inttoptr sequences.
126 static const Value *getUnderlyingObjectFromInt(const Value *V) {
128 if (const Operator *U = dyn_cast<Operator>(V)) {
129 // If we find a ptrtoint, we can transfer control back to the
130 // regular getUnderlyingObjectFromInt.
131 if (U->getOpcode() == Instruction::PtrToInt)
132 return U->getOperand(0);
133 // If we find an add of a constant, a multiplied value, or a phi, it's
134 // likely that the other operand will lead us to the base
135 // object. We don't have to worry about the case where the
136 // object address is somehow being computed by the multiply,
137 // because our callers only care when the result is an
138 // identifiable object.
139 if (U->getOpcode() != Instruction::Add ||
140 (!isa<ConstantInt>(U->getOperand(1)) &&
141 Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
142 !isa<PHINode>(U->getOperand(1))))
144 V = U->getOperand(0);
148 assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
152 /// This is a wrapper around GetUnderlyingObjects and adds support for basic
153 /// ptrtoint+arithmetic+inttoptr sequences.
154 static void getUnderlyingObjects(const Value *V,
155 SmallVectorImpl<Value *> &Objects,
156 const DataLayout &DL) {
157 SmallPtrSet<const Value *, 16> Visited;
158 SmallVector<const Value *, 4> Working(1, V);
160 V = Working.pop_back_val();
162 SmallVector<Value *, 4> Objs;
163 GetUnderlyingObjects(const_cast<Value *>(V), Objs, DL);
165 for (Value *V : Objs) {
166 if (!Visited.insert(V).second)
168 if (Operator::getOpcode(V) == Instruction::IntToPtr) {
170 getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
171 if (O->getType()->isPointerTy()) {
172 Working.push_back(O);
176 Objects.push_back(const_cast<Value *>(V));
178 } while (!Working.empty());
181 /// If this machine instr has memory reference information and it can be tracked
182 /// to a normal reference to a known object, return the Value for that object.
183 static void getUnderlyingObjectsForInstr(const MachineInstr *MI,
184 const MachineFrameInfo &MFI,
185 UnderlyingObjectsVector &Objects,
186 const DataLayout &DL) {
187 auto allMMOsOkay = [&]() {
188 for (const MachineMemOperand *MMO : MI->memoperands()) {
189 if (MMO->isVolatile())
192 if (const PseudoSourceValue *PSV = MMO->getPseudoValue()) {
193 // Function that contain tail calls don't have unique PseudoSourceValue
194 // objects. Two PseudoSourceValues might refer to the same or
195 // overlapping locations. The client code calling this function assumes
196 // this is not the case. So return a conservative answer of no known
198 if (MFI.hasTailCall())
201 // For now, ignore PseudoSourceValues which may alias LLVM IR values
202 // because the code that uses this function has no way to cope with
204 if (PSV->isAliased(&MFI))
207 bool MayAlias = PSV->mayAlias(&MFI);
208 Objects.push_back(UnderlyingObjectsVector::value_type(PSV, MayAlias));
209 } else if (const Value *V = MMO->getValue()) {
210 SmallVector<Value *, 4> Objs;
211 getUnderlyingObjects(V, Objs, DL);
213 for (Value *V : Objs) {
214 if (!isIdentifiedObject(V))
217 Objects.push_back(UnderlyingObjectsVector::value_type(V, true));
229 void ScheduleDAGInstrs::startBlock(MachineBasicBlock *bb) {
233 void ScheduleDAGInstrs::finishBlock() {
234 // Subclasses should no longer refer to the old block.
238 void ScheduleDAGInstrs::enterRegion(MachineBasicBlock *bb,
239 MachineBasicBlock::iterator begin,
240 MachineBasicBlock::iterator end,
241 unsigned regioninstrs) {
242 assert(bb == BB && "startBlock should set BB");
245 NumRegionInstrs = regioninstrs;
248 void ScheduleDAGInstrs::exitRegion() {
252 void ScheduleDAGInstrs::addSchedBarrierDeps() {
253 MachineInstr *ExitMI = RegionEnd != BB->end() ? &*RegionEnd : nullptr;
254 ExitSU.setInstr(ExitMI);
255 // Add dependencies on the defs and uses of the instruction.
257 for (const MachineOperand &MO : ExitMI->operands()) {
258 if (!MO.isReg() || MO.isDef()) continue;
259 unsigned Reg = MO.getReg();
260 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
261 Uses.insert(PhysRegSUOper(&ExitSU, -1, Reg));
262 } else if (TargetRegisterInfo::isVirtualRegister(Reg) && MO.readsReg()) {
263 addVRegUseDeps(&ExitSU, ExitMI->getOperandNo(&MO));
267 if (!ExitMI || (!ExitMI->isCall() && !ExitMI->isBarrier())) {
268 // For others, e.g. fallthrough, conditional branch, assume the exit
269 // uses all the registers that are livein to the successor blocks.
270 for (const MachineBasicBlock *Succ : BB->successors()) {
271 for (const auto &LI : Succ->liveins()) {
272 if (!Uses.contains(LI.PhysReg))
273 Uses.insert(PhysRegSUOper(&ExitSU, -1, LI.PhysReg));
279 /// MO is an operand of SU's instruction that defines a physical register. Adds
280 /// data dependencies from SU to any uses of the physical register.
281 void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
282 const MachineOperand &MO = SU->getInstr()->getOperand(OperIdx);
283 assert(MO.isDef() && "expect physreg def");
285 // Ask the target if address-backscheduling is desirable, and if so how much.
286 const TargetSubtargetInfo &ST = MF.getSubtarget();
288 for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
289 Alias.isValid(); ++Alias) {
290 if (!Uses.contains(*Alias))
292 for (Reg2SUnitsMap::iterator I = Uses.find(*Alias); I != Uses.end(); ++I) {
293 SUnit *UseSU = I->SU;
297 // Adjust the dependence latency using operand def/use information,
298 // then allow the target to perform its own adjustments.
299 int UseOp = I->OpIdx;
300 MachineInstr *RegUse = nullptr;
303 Dep = SDep(SU, SDep::Artificial);
305 // Set the hasPhysRegDefs only for physreg defs that have a use within
306 // the scheduling region.
307 SU->hasPhysRegDefs = true;
308 Dep = SDep(SU, SDep::Data, *Alias);
309 RegUse = UseSU->getInstr();
312 SchedModel.computeOperandLatency(SU->getInstr(), OperIdx, RegUse,
315 ST.adjustSchedDependency(SU, UseSU, Dep);
321 /// \brief Adds register dependencies (data, anti, and output) from this SUnit
322 /// to following instructions in the same scheduling region that depend the
323 /// physical register referenced at OperIdx.
324 void ScheduleDAGInstrs::addPhysRegDeps(SUnit *SU, unsigned OperIdx) {
325 MachineInstr *MI = SU->getInstr();
326 MachineOperand &MO = MI->getOperand(OperIdx);
327 unsigned Reg = MO.getReg();
328 // We do not need to track any dependencies for constant registers.
329 if (MRI.isConstantPhysReg(Reg))
332 // Optionally add output and anti dependencies. For anti
333 // dependencies we use a latency of 0 because for a multi-issue
334 // target we want to allow the defining instruction to issue
335 // in the same cycle as the using instruction.
336 // TODO: Using a latency of 1 here for output dependencies assumes
337 // there's no cost for reusing registers.
338 SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
339 for (MCRegAliasIterator Alias(Reg, TRI, true); Alias.isValid(); ++Alias) {
340 if (!Defs.contains(*Alias))
342 for (Reg2SUnitsMap::iterator I = Defs.find(*Alias); I != Defs.end(); ++I) {
343 SUnit *DefSU = I->SU;
344 if (DefSU == &ExitSU)
347 (Kind != SDep::Output || !MO.isDead() ||
348 !DefSU->getInstr()->registerDefIsDead(*Alias))) {
349 if (Kind == SDep::Anti)
350 DefSU->addPred(SDep(SU, Kind, /*Reg=*/*Alias));
352 SDep Dep(SU, Kind, /*Reg=*/*Alias);
354 SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
362 SU->hasPhysRegUses = true;
363 // Either insert a new Reg2SUnits entry with an empty SUnits list, or
364 // retrieve the existing SUnits list for this register's uses.
365 // Push this SUnit on the use list.
366 Uses.insert(PhysRegSUOper(SU, OperIdx, Reg));
370 addPhysRegDataDeps(SU, OperIdx);
372 // clear this register's use list
373 if (Uses.contains(Reg))
378 } else if (SU->isCall) {
379 // Calls will not be reordered because of chain dependencies (see
380 // below). Since call operands are dead, calls may continue to be added
381 // to the DefList making dependence checking quadratic in the size of
382 // the block. Instead, we leave only one call at the back of the
384 Reg2SUnitsMap::RangePair P = Defs.equal_range(Reg);
385 Reg2SUnitsMap::iterator B = P.first;
386 Reg2SUnitsMap::iterator I = P.second;
387 for (bool isBegin = I == B; !isBegin; /* empty */) {
388 isBegin = (--I) == B;
395 // Defs are pushed in the order they are visited and never reordered.
396 Defs.insert(PhysRegSUOper(SU, OperIdx, Reg));
400 LaneBitmask ScheduleDAGInstrs::getLaneMaskForMO(const MachineOperand &MO) const
402 unsigned Reg = MO.getReg();
403 // No point in tracking lanemasks if we don't have interesting subregisters.
404 const TargetRegisterClass &RC = *MRI.getRegClass(Reg);
405 if (!RC.HasDisjunctSubRegs)
406 return LaneBitmask::getAll();
408 unsigned SubReg = MO.getSubReg();
410 return RC.getLaneMask();
411 return TRI->getSubRegIndexLaneMask(SubReg);
414 /// Adds register output and data dependencies from this SUnit to instructions
415 /// that occur later in the same scheduling region if they read from or write to
416 /// the virtual register defined at OperIdx.
418 /// TODO: Hoist loop induction variable increments. This has to be
419 /// reevaluated. Generally, IV scheduling should be done before coalescing.
420 void ScheduleDAGInstrs::addVRegDefDeps(SUnit *SU, unsigned OperIdx) {
421 MachineInstr *MI = SU->getInstr();
422 MachineOperand &MO = MI->getOperand(OperIdx);
423 unsigned Reg = MO.getReg();
425 LaneBitmask DefLaneMask;
426 LaneBitmask KillLaneMask;
427 if (TrackLaneMasks) {
428 bool IsKill = MO.getSubReg() == 0 || MO.isUndef();
429 DefLaneMask = getLaneMaskForMO(MO);
430 // If we have a <read-undef> flag, none of the lane values comes from an
431 // earlier instruction.
432 KillLaneMask = IsKill ? LaneBitmask::getAll() : DefLaneMask;
434 // Clear undef flag, we'll re-add it later once we know which subregister
436 MO.setIsUndef(false);
438 DefLaneMask = LaneBitmask::getAll();
439 KillLaneMask = LaneBitmask::getAll();
443 assert(CurrentVRegUses.find(Reg) == CurrentVRegUses.end() &&
444 "Dead defs should have no uses");
446 // Add data dependence to all uses we found so far.
447 const TargetSubtargetInfo &ST = MF.getSubtarget();
448 for (VReg2SUnitOperIdxMultiMap::iterator I = CurrentVRegUses.find(Reg),
449 E = CurrentVRegUses.end(); I != E; /*empty*/) {
450 LaneBitmask LaneMask = I->LaneMask;
451 // Ignore uses of other lanes.
452 if ((LaneMask & KillLaneMask).none()) {
457 if ((LaneMask & DefLaneMask).any()) {
458 SUnit *UseSU = I->SU;
459 MachineInstr *Use = UseSU->getInstr();
460 SDep Dep(SU, SDep::Data, Reg);
461 Dep.setLatency(SchedModel.computeOperandLatency(MI, OperIdx, Use,
463 ST.adjustSchedDependency(SU, UseSU, Dep);
467 LaneMask &= ~KillLaneMask;
468 // If we found a Def for all lanes of this use, remove it from the list.
469 if (LaneMask.any()) {
470 I->LaneMask = LaneMask;
473 I = CurrentVRegUses.erase(I);
477 // Shortcut: Singly defined vregs do not have output/anti dependencies.
478 if (MRI.hasOneDef(Reg))
481 // Add output dependence to the next nearest defs of this vreg.
483 // Unless this definition is dead, the output dependence should be
484 // transitively redundant with antidependencies from this definition's
485 // uses. We're conservative for now until we have a way to guarantee the uses
486 // are not eliminated sometime during scheduling. The output dependence edge
487 // is also useful if output latency exceeds def-use latency.
488 LaneBitmask LaneMask = DefLaneMask;
489 for (VReg2SUnit &V2SU : make_range(CurrentVRegDefs.find(Reg),
490 CurrentVRegDefs.end())) {
491 // Ignore defs for other lanes.
492 if ((V2SU.LaneMask & LaneMask).none())
494 // Add an output dependence.
495 SUnit *DefSU = V2SU.SU;
496 // Ignore additional defs of the same lanes in one instruction. This can
497 // happen because lanemasks are shared for targets with too many
498 // subregisters. We also use some representration tricks/hacks where we
499 // add super-register defs/uses, to imply that although we only access parts
500 // of the reg we care about the full one.
503 SDep Dep(SU, SDep::Output, Reg);
505 SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
508 // Update current definition. This can get tricky if the def was about a
509 // bigger lanemask before. We then have to shrink it and create a new
510 // VReg2SUnit for the non-overlapping part.
511 LaneBitmask OverlapMask = V2SU.LaneMask & LaneMask;
512 LaneBitmask NonOverlapMask = V2SU.LaneMask & ~LaneMask;
514 V2SU.LaneMask = OverlapMask;
515 if (NonOverlapMask.any())
516 CurrentVRegDefs.insert(VReg2SUnit(Reg, NonOverlapMask, DefSU));
518 // If there was no CurrentVRegDefs entry for some lanes yet, create one.
520 CurrentVRegDefs.insert(VReg2SUnit(Reg, LaneMask, SU));
523 /// \brief Adds a register data dependency if the instruction that defines the
524 /// virtual register used at OperIdx is mapped to an SUnit. Add a register
525 /// antidependency from this SUnit to instructions that occur later in the same
526 /// scheduling region if they write the virtual register.
528 /// TODO: Handle ExitSU "uses" properly.
529 void ScheduleDAGInstrs::addVRegUseDeps(SUnit *SU, unsigned OperIdx) {
530 const MachineInstr *MI = SU->getInstr();
531 const MachineOperand &MO = MI->getOperand(OperIdx);
532 unsigned Reg = MO.getReg();
534 // Remember the use. Data dependencies will be added when we find the def.
535 LaneBitmask LaneMask = TrackLaneMasks ? getLaneMaskForMO(MO)
536 : LaneBitmask::getAll();
537 CurrentVRegUses.insert(VReg2SUnitOperIdx(Reg, LaneMask, OperIdx, SU));
539 // Add antidependences to the following defs of the vreg.
540 for (VReg2SUnit &V2SU : make_range(CurrentVRegDefs.find(Reg),
541 CurrentVRegDefs.end())) {
542 // Ignore defs for unrelated lanes.
543 LaneBitmask PrevDefLaneMask = V2SU.LaneMask;
544 if ((PrevDefLaneMask & LaneMask).none())
549 V2SU.SU->addPred(SDep(SU, SDep::Anti, Reg));
553 /// Returns true if MI is an instruction we are unable to reason about
554 /// (like a call or something with unmodeled side effects).
555 static inline bool isGlobalMemoryObject(AliasAnalysis *AA, MachineInstr *MI) {
556 return MI->isCall() || MI->hasUnmodeledSideEffects() ||
557 (MI->hasOrderedMemoryRef() && !MI->isDereferenceableInvariantLoad(AA));
560 void ScheduleDAGInstrs::addChainDependency (SUnit *SUa, SUnit *SUb,
562 if (SUa->getInstr()->mayAlias(AAForDep, *SUb->getInstr(), UseTBAA)) {
563 SDep Dep(SUa, SDep::MayAliasMem);
564 Dep.setLatency(Latency);
569 /// \brief Creates an SUnit for each real instruction, numbered in top-down
570 /// topological order. The instruction order A < B, implies that no edge exists
573 /// Map each real instruction to its SUnit.
575 /// After initSUnits, the SUnits vector cannot be resized and the scheduler may
576 /// hang onto SUnit pointers. We may relax this in the future by using SUnit IDs
577 /// instead of pointers.
579 /// MachineScheduler relies on initSUnits numbering the nodes by their order in
580 /// the original instruction list.
581 void ScheduleDAGInstrs::initSUnits() {
582 // We'll be allocating one SUnit for each real instruction in the region,
583 // which is contained within a basic block.
584 SUnits.reserve(NumRegionInstrs);
586 for (MachineInstr &MI : make_range(RegionBegin, RegionEnd)) {
587 if (MI.isDebugValue())
590 SUnit *SU = newSUnit(&MI);
591 MISUnitMap[&MI] = SU;
593 SU->isCall = MI.isCall();
594 SU->isCommutable = MI.isCommutable();
596 // Assign the Latency field of SU using target-provided information.
597 SU->Latency = SchedModel.computeInstrLatency(SU->getInstr());
599 // If this SUnit uses a reserved or unbuffered resource, mark it as such.
601 // Reserved resources block an instruction from issuing and stall the
602 // entire pipeline. These are identified by BufferSize=0.
604 // Unbuffered resources prevent execution of subsequent instructions that
605 // require the same resources. This is used for in-order execution pipelines
606 // within an out-of-order core. These are identified by BufferSize=1.
607 if (SchedModel.hasInstrSchedModel()) {
608 const MCSchedClassDesc *SC = getSchedClass(SU);
609 for (const MCWriteProcResEntry &PRE :
610 make_range(SchedModel.getWriteProcResBegin(SC),
611 SchedModel.getWriteProcResEnd(SC))) {
612 switch (SchedModel.getProcResource(PRE.ProcResourceIdx)->BufferSize) {
614 SU->hasReservedResource = true;
617 SU->isUnbuffered = true;
627 class ScheduleDAGInstrs::Value2SUsMap : public MapVector<ValueType, SUList> {
628 /// Current total number of SUs in map.
629 unsigned NumNodes = 0;
631 /// 1 for loads, 0 for stores. (see comment in SUList)
632 unsigned TrueMemOrderLatency;
635 Value2SUsMap(unsigned lat = 0) : TrueMemOrderLatency(lat) {}
637 /// To keep NumNodes up to date, insert() is used instead of
638 /// this operator w/ push_back().
639 ValueType &operator[](const SUList &Key) {
640 llvm_unreachable("Don't use. Use insert() instead."); };
642 /// Adds SU to the SUList of V. If Map grows huge, reduce its size by calling
644 void inline insert(SUnit *SU, ValueType V) {
645 MapVector::operator[](V).push_back(SU);
649 /// Clears the list of SUs mapped to V.
650 void inline clearList(ValueType V) {
651 iterator Itr = find(V);
653 assert(NumNodes >= Itr->second.size());
654 NumNodes -= Itr->second.size();
660 /// Clears map from all contents.
662 MapVector<ValueType, SUList>::clear();
666 unsigned inline size() const { return NumNodes; }
668 /// Counts the number of SUs in this map after a reduction.
669 void reComputeSize() {
671 for (auto &I : *this)
672 NumNodes += I.second.size();
675 unsigned inline getTrueMemOrderLatency() const {
676 return TrueMemOrderLatency;
682 void ScheduleDAGInstrs::addChainDependencies(SUnit *SU,
683 Value2SUsMap &Val2SUsMap) {
684 for (auto &I : Val2SUsMap)
685 addChainDependencies(SU, I.second,
686 Val2SUsMap.getTrueMemOrderLatency());
689 void ScheduleDAGInstrs::addChainDependencies(SUnit *SU,
690 Value2SUsMap &Val2SUsMap,
692 Value2SUsMap::iterator Itr = Val2SUsMap.find(V);
693 if (Itr != Val2SUsMap.end())
694 addChainDependencies(SU, Itr->second,
695 Val2SUsMap.getTrueMemOrderLatency());
698 void ScheduleDAGInstrs::addBarrierChain(Value2SUsMap &map) {
699 assert(BarrierChain != nullptr);
701 for (auto &I : map) {
702 SUList &sus = I.second;
704 SU->addPredBarrier(BarrierChain);
709 void ScheduleDAGInstrs::insertBarrierChain(Value2SUsMap &map) {
710 assert(BarrierChain != nullptr);
712 // Go through all lists of SUs.
713 for (Value2SUsMap::iterator I = map.begin(), EE = map.end(); I != EE;) {
714 Value2SUsMap::iterator CurrItr = I++;
715 SUList &sus = CurrItr->second;
716 SUList::iterator SUItr = sus.begin(), SUEE = sus.end();
717 for (; SUItr != SUEE; ++SUItr) {
718 // Stop on BarrierChain or any instruction above it.
719 if ((*SUItr)->NodeNum <= BarrierChain->NodeNum)
722 (*SUItr)->addPredBarrier(BarrierChain);
725 // Remove also the BarrierChain from list if present.
726 if (SUItr != SUEE && *SUItr == BarrierChain)
729 // Remove all SUs that are now successors of BarrierChain.
730 if (SUItr != sus.begin())
731 sus.erase(sus.begin(), SUItr);
734 // Remove all entries with empty su lists.
735 map.remove_if([&](std::pair<ValueType, SUList> &mapEntry) {
736 return (mapEntry.second.empty()); });
738 // Recompute the size of the map (NumNodes).
742 void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
743 RegPressureTracker *RPTracker,
744 PressureDiffs *PDiffs,
746 bool TrackLaneMasks) {
747 const TargetSubtargetInfo &ST = MF.getSubtarget();
748 bool UseAA = EnableAASchedMI.getNumOccurrences() > 0 ? EnableAASchedMI
750 AAForDep = UseAA ? AA : nullptr;
752 BarrierChain = nullptr;
754 this->TrackLaneMasks = TrackLaneMasks;
756 ScheduleDAG::clearDAG();
758 // Create an SUnit for each real instruction.
762 PDiffs->init(SUnits.size());
764 // We build scheduling units by walking a block's instruction list
765 // from bottom to top.
767 // Each MIs' memory operand(s) is analyzed to a list of underlying
768 // objects. The SU is then inserted in the SUList(s) mapped from the
769 // Value(s). Each Value thus gets mapped to lists of SUs depending
770 // on it, stores and loads kept separately. Two SUs are trivially
771 // non-aliasing if they both depend on only identified Values and do
772 // not share any common Value.
773 Value2SUsMap Stores, Loads(1 /*TrueMemOrderLatency*/);
775 // Certain memory accesses are known to not alias any SU in Stores
776 // or Loads, and have therefore their own 'NonAlias'
777 // domain. E.g. spill / reload instructions never alias LLVM I/R
778 // Values. It would be nice to assume that this type of memory
779 // accesses always have a proper memory operand modelling, and are
780 // therefore never unanalyzable, but this is conservatively not
782 Value2SUsMap NonAliasStores, NonAliasLoads(1 /*TrueMemOrderLatency*/);
784 // Remove any stale debug info; sometimes BuildSchedGraph is called again
785 // without emitting the info from the previous call.
787 FirstDbgValue = nullptr;
789 assert(Defs.empty() && Uses.empty() &&
790 "Only BuildGraph should update Defs/Uses");
791 Defs.setUniverse(TRI->getNumRegs());
792 Uses.setUniverse(TRI->getNumRegs());
794 assert(CurrentVRegDefs.empty() && "nobody else should use CurrentVRegDefs");
795 assert(CurrentVRegUses.empty() && "nobody else should use CurrentVRegUses");
796 unsigned NumVirtRegs = MRI.getNumVirtRegs();
797 CurrentVRegDefs.setUniverse(NumVirtRegs);
798 CurrentVRegUses.setUniverse(NumVirtRegs);
800 // Model data dependencies between instructions being scheduled and the
802 addSchedBarrierDeps();
804 // Walk the list of instructions, from bottom moving up.
805 MachineInstr *DbgMI = nullptr;
806 for (MachineBasicBlock::iterator MII = RegionEnd, MIE = RegionBegin;
808 MachineInstr &MI = *std::prev(MII);
810 DbgValues.push_back(std::make_pair(DbgMI, &MI));
814 if (MI.isDebugValue()) {
818 SUnit *SU = MISUnitMap[&MI];
819 assert(SU && "No SUnit mapped to this MI");
822 RegisterOperands RegOpers;
823 RegOpers.collect(MI, *TRI, MRI, TrackLaneMasks, false);
824 if (TrackLaneMasks) {
825 SlotIndex SlotIdx = LIS->getInstructionIndex(MI);
826 RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx);
828 if (PDiffs != nullptr)
829 PDiffs->addInstruction(SU->NodeNum, RegOpers, MRI);
831 RPTracker->recedeSkipDebugValues();
832 assert(&*RPTracker->getPos() == &MI && "RPTracker in sync");
833 RPTracker->recede(RegOpers);
837 (CanHandleTerminators || (!MI.isTerminator() && !MI.isPosition())) &&
838 "Cannot schedule terminators or labels!");
840 // Add register-based dependencies (data, anti, and output).
841 // For some instructions (calls, returns, inline-asm, etc.) there can
842 // be explicit uses and implicit defs, in which case the use will appear
843 // on the operand list before the def. Do two passes over the operand
844 // list to make sure that defs are processed before any uses.
845 bool HasVRegDef = false;
846 for (unsigned j = 0, n = MI.getNumOperands(); j != n; ++j) {
847 const MachineOperand &MO = MI.getOperand(j);
848 if (!MO.isReg() || !MO.isDef())
850 unsigned Reg = MO.getReg();
851 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
852 addPhysRegDeps(SU, j);
853 } else if (TargetRegisterInfo::isVirtualRegister(Reg)) {
855 addVRegDefDeps(SU, j);
858 // Now process all uses.
859 for (unsigned j = 0, n = MI.getNumOperands(); j != n; ++j) {
860 const MachineOperand &MO = MI.getOperand(j);
861 // Only look at use operands.
862 // We do not need to check for MO.readsReg() here because subsequent
863 // subregister defs will get output dependence edges and need no
864 // additional use dependencies.
865 if (!MO.isReg() || !MO.isUse())
867 unsigned Reg = MO.getReg();
868 if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
869 addPhysRegDeps(SU, j);
870 } else if (TargetRegisterInfo::isVirtualRegister(Reg) && MO.readsReg()) {
871 addVRegUseDeps(SU, j);
875 // If we haven't seen any uses in this scheduling region, create a
876 // dependence edge to ExitSU to model the live-out latency. This is required
877 // for vreg defs with no in-region use, and prefetches with no vreg def.
879 // FIXME: NumDataSuccs would be more precise than NumSuccs here. This
880 // check currently relies on being called before adding chain deps.
881 if (SU->NumSuccs == 0 && SU->Latency > 1 && (HasVRegDef || MI.mayLoad())) {
882 SDep Dep(SU, SDep::Artificial);
883 Dep.setLatency(SU->Latency - 1);
887 // Add memory dependencies (Note: isStoreToStackSlot and
888 // isLoadFromStackSLot are not usable after stack slots are lowered to
889 // actual addresses).
891 // This is a barrier event that acts as a pivotal node in the DAG.
892 if (isGlobalMemoryObject(AA, &MI)) {
894 // Become the barrier chain.
896 BarrierChain->addPredBarrier(SU);
899 DEBUG(dbgs() << "Global memory object and new barrier chain: SU("
900 << BarrierChain->NodeNum << ").\n";);
902 // Add dependencies against everything below it and clear maps.
903 addBarrierChain(Stores);
904 addBarrierChain(Loads);
905 addBarrierChain(NonAliasStores);
906 addBarrierChain(NonAliasLoads);
911 // If it's not a store or a variant load, we're done.
912 if (!MI.mayStore() &&
913 !(MI.mayLoad() && !MI.isDereferenceableInvariantLoad(AA)))
916 // Always add dependecy edge to BarrierChain if present.
918 BarrierChain->addPredBarrier(SU);
920 // Find the underlying objects for MI. The Objs vector is either
921 // empty, or filled with the Values of memory locations which this
922 // SU depends on. An empty vector means the memory location is
923 // unknown, and may alias anything.
924 UnderlyingObjectsVector Objs;
925 getUnderlyingObjectsForInstr(&MI, MFI, Objs, MF.getDataLayout());
929 // An unknown store depends on all stores and loads.
930 addChainDependencies(SU, Stores);
931 addChainDependencies(SU, NonAliasStores);
932 addChainDependencies(SU, Loads);
933 addChainDependencies(SU, NonAliasLoads);
935 // Map this store to 'UnknownValue'.
936 Stores.insert(SU, UnknownValue);
938 // Add precise dependencies against all previously seen memory
939 // accesses mapped to the same Value(s).
940 for (const UnderlyingObject &UnderlObj : Objs) {
941 ValueType V = UnderlObj.getValue();
942 bool ThisMayAlias = UnderlObj.mayAlias();
944 // Add dependencies to previous stores and loads mapped to V.
945 addChainDependencies(SU, (ThisMayAlias ? Stores : NonAliasStores), V);
946 addChainDependencies(SU, (ThisMayAlias ? Loads : NonAliasLoads), V);
948 // Update the store map after all chains have been added to avoid adding
949 // self-loop edge if multiple underlying objects are present.
950 for (const UnderlyingObject &UnderlObj : Objs) {
951 ValueType V = UnderlObj.getValue();
952 bool ThisMayAlias = UnderlObj.mayAlias();
954 // Map this store to V.
955 (ThisMayAlias ? Stores : NonAliasStores).insert(SU, V);
957 // The store may have dependencies to unanalyzable loads and
959 addChainDependencies(SU, Loads, UnknownValue);
960 addChainDependencies(SU, Stores, UnknownValue);
962 } else { // SU is a load.
964 // An unknown load depends on all stores.
965 addChainDependencies(SU, Stores);
966 addChainDependencies(SU, NonAliasStores);
968 Loads.insert(SU, UnknownValue);
970 for (const UnderlyingObject &UnderlObj : Objs) {
971 ValueType V = UnderlObj.getValue();
972 bool ThisMayAlias = UnderlObj.mayAlias();
974 // Add precise dependencies against all previously seen stores
975 // mapping to the same Value(s).
976 addChainDependencies(SU, (ThisMayAlias ? Stores : NonAliasStores), V);
978 // Map this load to V.
979 (ThisMayAlias ? Loads : NonAliasLoads).insert(SU, V);
981 // The load may have dependencies to unanalyzable stores.
982 addChainDependencies(SU, Stores, UnknownValue);
986 // Reduce maps if they grow huge.
987 if (Stores.size() + Loads.size() >= HugeRegion) {
988 DEBUG(dbgs() << "Reducing Stores and Loads maps.\n";);
989 reduceHugeMemNodeMaps(Stores, Loads, getReductionSize());
991 if (NonAliasStores.size() + NonAliasLoads.size() >= HugeRegion) {
992 DEBUG(dbgs() << "Reducing NonAliasStores and NonAliasLoads maps.\n";);
993 reduceHugeMemNodeMaps(NonAliasStores, NonAliasLoads, getReductionSize());
998 FirstDbgValue = DbgMI;
1002 CurrentVRegDefs.clear();
1003 CurrentVRegUses.clear();
1006 raw_ostream &llvm::operator<<(raw_ostream &OS, const PseudoSourceValue* PSV) {
1007 PSV->printCustom(OS);
1011 void ScheduleDAGInstrs::Value2SUsMap::dump() {
1012 for (auto &Itr : *this) {
1013 if (Itr.first.is<const Value*>()) {
1014 const Value *V = Itr.first.get<const Value*>();
1015 if (isa<UndefValue>(V))
1016 dbgs() << "Unknown";
1018 V->printAsOperand(dbgs());
1020 else if (Itr.first.is<const PseudoSourceValue*>())
1021 dbgs() << Itr.first.get<const PseudoSourceValue*>();
1023 llvm_unreachable("Unknown Value type.");
1026 dumpSUList(Itr.second);
1030 void ScheduleDAGInstrs::reduceHugeMemNodeMaps(Value2SUsMap &stores,
1031 Value2SUsMap &loads, unsigned N) {
1032 DEBUG(dbgs() << "Before reduction:\nStoring SUnits:\n";
1034 dbgs() << "Loading SUnits:\n";
1037 // Insert all SU's NodeNums into a vector and sort it.
1038 std::vector<unsigned> NodeNums;
1039 NodeNums.reserve(stores.size() + loads.size());
1040 for (auto &I : stores)
1041 for (auto *SU : I.second)
1042 NodeNums.push_back(SU->NodeNum);
1043 for (auto &I : loads)
1044 for (auto *SU : I.second)
1045 NodeNums.push_back(SU->NodeNum);
1046 std::sort(NodeNums.begin(), NodeNums.end());
1048 // The N last elements in NodeNums will be removed, and the SU with
1049 // the lowest NodeNum of them will become the new BarrierChain to
1050 // let the not yet seen SUs have a dependency to the removed SUs.
1051 assert(N <= NodeNums.size());
1052 SUnit *newBarrierChain = &SUnits[*(NodeNums.end() - N)];
1054 // The aliasing and non-aliasing maps reduce independently of each
1055 // other, but share a common BarrierChain. Check if the
1056 // newBarrierChain is above the former one. If it is not, it may
1057 // introduce a loop to use newBarrierChain, so keep the old one.
1058 if (newBarrierChain->NodeNum < BarrierChain->NodeNum) {
1059 BarrierChain->addPredBarrier(newBarrierChain);
1060 BarrierChain = newBarrierChain;
1061 DEBUG(dbgs() << "Inserting new barrier chain: SU("
1062 << BarrierChain->NodeNum << ").\n";);
1065 DEBUG(dbgs() << "Keeping old barrier chain: SU("
1066 << BarrierChain->NodeNum << ").\n";);
1069 BarrierChain = newBarrierChain;
1071 insertBarrierChain(stores);
1072 insertBarrierChain(loads);
1074 DEBUG(dbgs() << "After reduction:\nStoring SUnits:\n";
1076 dbgs() << "Loading SUnits:\n";
1080 static void toggleKills(const MachineRegisterInfo &MRI, LivePhysRegs &LiveRegs,
1081 MachineInstr &MI, bool addToLiveRegs) {
1082 for (MachineOperand &MO : MI.operands()) {
1083 if (!MO.isReg() || !MO.readsReg())
1085 unsigned Reg = MO.getReg();
1089 // Things that are available after the instruction are killed by it.
1090 bool IsKill = LiveRegs.available(MRI, Reg);
1091 MO.setIsKill(IsKill);
1092 if (IsKill && addToLiveRegs)
1093 LiveRegs.addReg(Reg);
1097 void ScheduleDAGInstrs::fixupKills(MachineBasicBlock &MBB) {
1098 DEBUG(dbgs() << "Fixup kills for BB#" << MBB.getNumber() << '\n');
1100 LiveRegs.init(*TRI);
1101 LiveRegs.addLiveOuts(MBB);
1103 // Examine block from end to start...
1104 for (MachineInstr &MI : make_range(MBB.rbegin(), MBB.rend())) {
1105 if (MI.isDebugValue())
1108 // Update liveness. Registers that are defed but not used in this
1109 // instruction are now dead. Mark register and all subregs as they
1110 // are completely defined.
1111 for (ConstMIBundleOperands O(MI); O.isValid(); ++O) {
1112 const MachineOperand &MO = *O;
1116 unsigned Reg = MO.getReg();
1119 LiveRegs.removeReg(Reg);
1120 } else if (MO.isRegMask()) {
1121 LiveRegs.removeRegsInMask(MO);
1125 // If there is a bundle header fix it up first.
1126 if (!MI.isBundled()) {
1127 toggleKills(MRI, LiveRegs, MI, true);
1129 MachineBasicBlock::instr_iterator First = MI.getIterator();
1130 if (MI.isBundle()) {
1131 toggleKills(MRI, LiveRegs, MI, false);
1134 // Some targets make the (questionable) assumtion that the instructions
1135 // inside the bundle are ordered and consequently only the last use of
1136 // a register inside the bundle can kill it.
1137 MachineBasicBlock::instr_iterator I = std::next(First);
1138 while (I->isBundledWithSucc())
1141 if (!I->isDebugValue())
1142 toggleKills(MRI, LiveRegs, *I, true);
1144 } while(I != First);
1149 void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const {
1150 // Cannot completely remove virtual function even in release mode.
1151 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1152 SU->getInstr()->dump();
1156 std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const {
1158 raw_string_ostream oss(s);
1161 else if (SU == &ExitSU)
1164 SU->getInstr()->print(oss, /*SkipOpers=*/true);
1168 /// Return the basic block label. It is not necessarilly unique because a block
1169 /// contains multiple scheduling regions. But it is fine for visualization.
1170 std::string ScheduleDAGInstrs::getDAGName() const {
1171 return "dag." + BB->getFullName();
1174 //===----------------------------------------------------------------------===//
1175 // SchedDFSResult Implementation
1176 //===----------------------------------------------------------------------===//
1180 /// Internal state used to compute SchedDFSResult.
1181 class SchedDFSImpl {
1184 /// Join DAG nodes into equivalence classes by their subtree.
1185 IntEqClasses SubtreeClasses;
1186 /// List PredSU, SuccSU pairs that represent data edges between subtrees.
1187 std::vector<std::pair<const SUnit *, const SUnit*>> ConnectionPairs;
1191 unsigned ParentNodeID; ///< Parent node (member of the parent subtree).
1192 unsigned SubInstrCount = 0; ///< Instr count in this tree only, not
1195 RootData(unsigned id): NodeID(id),
1196 ParentNodeID(SchedDFSResult::InvalidSubtreeID) {}
1198 unsigned getSparseSetIndex() const { return NodeID; }
1201 SparseSet<RootData> RootSet;
1204 SchedDFSImpl(SchedDFSResult &r): R(r), SubtreeClasses(R.DFSNodeData.size()) {
1205 RootSet.setUniverse(R.DFSNodeData.size());
1208 /// Returns true if this node been visited by the DFS traversal.
1210 /// During visitPostorderNode the Node's SubtreeID is assigned to the Node
1211 /// ID. Later, SubtreeID is updated but remains valid.
1212 bool isVisited(const SUnit *SU) const {
1213 return R.DFSNodeData[SU->NodeNum].SubtreeID
1214 != SchedDFSResult::InvalidSubtreeID;
1217 /// Initializes this node's instruction count. We don't need to flag the node
1218 /// visited until visitPostorder because the DAG cannot have cycles.
1219 void visitPreorder(const SUnit *SU) {
1220 R.DFSNodeData[SU->NodeNum].InstrCount =
1221 SU->getInstr()->isTransient() ? 0 : 1;
1224 /// Called once for each node after all predecessors are visited. Revisit this
1225 /// node's predecessors and potentially join them now that we know the ILP of
1226 /// the other predecessors.
1227 void visitPostorderNode(const SUnit *SU) {
1228 // Mark this node as the root of a subtree. It may be joined with its
1229 // successors later.
1230 R.DFSNodeData[SU->NodeNum].SubtreeID = SU->NodeNum;
1231 RootData RData(SU->NodeNum);
1232 RData.SubInstrCount = SU->getInstr()->isTransient() ? 0 : 1;
1234 // If any predecessors are still in their own subtree, they either cannot be
1235 // joined or are large enough to remain separate. If this parent node's
1236 // total instruction count is not greater than a child subtree by at least
1237 // the subtree limit, then try to join it now since splitting subtrees is
1238 // only useful if multiple high-pressure paths are possible.
1239 unsigned InstrCount = R.DFSNodeData[SU->NodeNum].InstrCount;
1240 for (const SDep &PredDep : SU->Preds) {
1241 if (PredDep.getKind() != SDep::Data)
1243 unsigned PredNum = PredDep.getSUnit()->NodeNum;
1244 if ((InstrCount - R.DFSNodeData[PredNum].InstrCount) < R.SubtreeLimit)
1245 joinPredSubtree(PredDep, SU, /*CheckLimit=*/false);
1247 // Either link or merge the TreeData entry from the child to the parent.
1248 if (R.DFSNodeData[PredNum].SubtreeID == PredNum) {
1249 // If the predecessor's parent is invalid, this is a tree edge and the
1250 // current node is the parent.
1251 if (RootSet[PredNum].ParentNodeID == SchedDFSResult::InvalidSubtreeID)
1252 RootSet[PredNum].ParentNodeID = SU->NodeNum;
1254 else if (RootSet.count(PredNum)) {
1255 // The predecessor is not a root, but is still in the root set. This
1256 // must be the new parent that it was just joined to. Note that
1257 // RootSet[PredNum].ParentNodeID may either be invalid or may still be
1258 // set to the original parent.
1259 RData.SubInstrCount += RootSet[PredNum].SubInstrCount;
1260 RootSet.erase(PredNum);
1263 RootSet[SU->NodeNum] = RData;
1266 /// \brief Called once for each tree edge after calling visitPostOrderNode on
1267 /// the predecessor. Increment the parent node's instruction count and
1268 /// preemptively join this subtree to its parent's if it is small enough.
1269 void visitPostorderEdge(const SDep &PredDep, const SUnit *Succ) {
1270 R.DFSNodeData[Succ->NodeNum].InstrCount
1271 += R.DFSNodeData[PredDep.getSUnit()->NodeNum].InstrCount;
1272 joinPredSubtree(PredDep, Succ);
1275 /// Adds a connection for cross edges.
1276 void visitCrossEdge(const SDep &PredDep, const SUnit *Succ) {
1277 ConnectionPairs.push_back(std::make_pair(PredDep.getSUnit(), Succ));
1280 /// Sets each node's subtree ID to the representative ID and record
1281 /// connections between trees.
1283 SubtreeClasses.compress();
1284 R.DFSTreeData.resize(SubtreeClasses.getNumClasses());
1285 assert(SubtreeClasses.getNumClasses() == RootSet.size()
1286 && "number of roots should match trees");
1287 for (const RootData &Root : RootSet) {
1288 unsigned TreeID = SubtreeClasses[Root.NodeID];
1289 if (Root.ParentNodeID != SchedDFSResult::InvalidSubtreeID)
1290 R.DFSTreeData[TreeID].ParentTreeID = SubtreeClasses[Root.ParentNodeID];
1291 R.DFSTreeData[TreeID].SubInstrCount = Root.SubInstrCount;
1292 // Note that SubInstrCount may be greater than InstrCount if we joined
1293 // subtrees across a cross edge. InstrCount will be attributed to the
1294 // original parent, while SubInstrCount will be attributed to the joined
1297 R.SubtreeConnections.resize(SubtreeClasses.getNumClasses());
1298 R.SubtreeConnectLevels.resize(SubtreeClasses.getNumClasses());
1299 DEBUG(dbgs() << R.getNumSubtrees() << " subtrees:\n");
1300 for (unsigned Idx = 0, End = R.DFSNodeData.size(); Idx != End; ++Idx) {
1301 R.DFSNodeData[Idx].SubtreeID = SubtreeClasses[Idx];
1302 DEBUG(dbgs() << " SU(" << Idx << ") in tree "
1303 << R.DFSNodeData[Idx].SubtreeID << '\n');
1305 for (const std::pair<const SUnit*, const SUnit*> &P : ConnectionPairs) {
1306 unsigned PredTree = SubtreeClasses[P.first->NodeNum];
1307 unsigned SuccTree = SubtreeClasses[P.second->NodeNum];
1308 if (PredTree == SuccTree)
1310 unsigned Depth = P.first->getDepth();
1311 addConnection(PredTree, SuccTree, Depth);
1312 addConnection(SuccTree, PredTree, Depth);
1317 /// Joins the predecessor subtree with the successor that is its DFS parent.
1318 /// Applies some heuristics before joining.
1319 bool joinPredSubtree(const SDep &PredDep, const SUnit *Succ,
1320 bool CheckLimit = true) {
1321 assert(PredDep.getKind() == SDep::Data && "Subtrees are for data edges");
1323 // Check if the predecessor is already joined.
1324 const SUnit *PredSU = PredDep.getSUnit();
1325 unsigned PredNum = PredSU->NodeNum;
1326 if (R.DFSNodeData[PredNum].SubtreeID != PredNum)
1329 // Four is the magic number of successors before a node is considered a
1331 unsigned NumDataSucs = 0;
1332 for (const SDep &SuccDep : PredSU->Succs) {
1333 if (SuccDep.getKind() == SDep::Data) {
1334 if (++NumDataSucs >= 4)
1338 if (CheckLimit && R.DFSNodeData[PredNum].InstrCount > R.SubtreeLimit)
1340 R.DFSNodeData[PredNum].SubtreeID = Succ->NodeNum;
1341 SubtreeClasses.join(Succ->NodeNum, PredNum);
1345 /// Called by finalize() to record a connection between trees.
1346 void addConnection(unsigned FromTree, unsigned ToTree, unsigned Depth) {
1351 SmallVectorImpl<SchedDFSResult::Connection> &Connections =
1352 R.SubtreeConnections[FromTree];
1353 for (SchedDFSResult::Connection &C : Connections) {
1354 if (C.TreeID == ToTree) {
1355 C.Level = std::max(C.Level, Depth);
1359 Connections.push_back(SchedDFSResult::Connection(ToTree, Depth));
1360 FromTree = R.DFSTreeData[FromTree].ParentTreeID;
1361 } while (FromTree != SchedDFSResult::InvalidSubtreeID);
1365 } // end namespace llvm
1369 /// Manage the stack used by a reverse depth-first search over the DAG.
1370 class SchedDAGReverseDFS {
1371 std::vector<std::pair<const SUnit *, SUnit::const_pred_iterator>> DFSStack;
1374 bool isComplete() const { return DFSStack.empty(); }
1376 void follow(const SUnit *SU) {
1377 DFSStack.push_back(std::make_pair(SU, SU->Preds.begin()));
1379 void advance() { ++DFSStack.back().second; }
1381 const SDep *backtrack() {
1382 DFSStack.pop_back();
1383 return DFSStack.empty() ? nullptr : std::prev(DFSStack.back().second);
1386 const SUnit *getCurr() const { return DFSStack.back().first; }
1388 SUnit::const_pred_iterator getPred() const { return DFSStack.back().second; }
1390 SUnit::const_pred_iterator getPredEnd() const {
1391 return getCurr()->Preds.end();
1395 } // end anonymous namespace
1397 static bool hasDataSucc(const SUnit *SU) {
1398 for (const SDep &SuccDep : SU->Succs) {
1399 if (SuccDep.getKind() == SDep::Data &&
1400 !SuccDep.getSUnit()->isBoundaryNode())
1406 /// Computes an ILP metric for all nodes in the subDAG reachable via depth-first
1407 /// search from this root.
1408 void SchedDFSResult::compute(ArrayRef<SUnit> SUnits) {
1410 llvm_unreachable("Top-down ILP metric is unimplemnted");
1412 SchedDFSImpl Impl(*this);
1413 for (const SUnit &SU : SUnits) {
1414 if (Impl.isVisited(&SU) || hasDataSucc(&SU))
1417 SchedDAGReverseDFS DFS;
1418 Impl.visitPreorder(&SU);
1421 // Traverse the leftmost path as far as possible.
1422 while (DFS.getPred() != DFS.getPredEnd()) {
1423 const SDep &PredDep = *DFS.getPred();
1425 // Ignore non-data edges.
1426 if (PredDep.getKind() != SDep::Data
1427 || PredDep.getSUnit()->isBoundaryNode()) {
1430 // An already visited edge is a cross edge, assuming an acyclic DAG.
1431 if (Impl.isVisited(PredDep.getSUnit())) {
1432 Impl.visitCrossEdge(PredDep, DFS.getCurr());
1435 Impl.visitPreorder(PredDep.getSUnit());
1436 DFS.follow(PredDep.getSUnit());
1438 // Visit the top of the stack in postorder and backtrack.
1439 const SUnit *Child = DFS.getCurr();
1440 const SDep *PredDep = DFS.backtrack();
1441 Impl.visitPostorderNode(Child);
1443 Impl.visitPostorderEdge(*PredDep, DFS.getCurr());
1444 if (DFS.isComplete())
1451 /// The root of the given SubtreeID was just scheduled. For all subtrees
1452 /// connected to this tree, record the depth of the connection so that the
1453 /// nearest connected subtrees can be prioritized.
1454 void SchedDFSResult::scheduleTree(unsigned SubtreeID) {
1455 for (const Connection &C : SubtreeConnections[SubtreeID]) {
1456 SubtreeConnectLevels[C.TreeID] =
1457 std::max(SubtreeConnectLevels[C.TreeID], C.Level);
1458 DEBUG(dbgs() << " Tree: " << C.TreeID
1459 << " @" << SubtreeConnectLevels[C.TreeID] << '\n');
1463 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1464 LLVM_DUMP_METHOD void ILPValue::print(raw_ostream &OS) const {
1465 OS << InstrCount << " / " << Length << " = ";
1469 OS << format("%g", ((double)InstrCount / Length));
1472 LLVM_DUMP_METHOD void ILPValue::dump() const {
1473 dbgs() << *this << '\n';
1479 raw_ostream &operator<<(raw_ostream &OS, const ILPValue &Val) {
1484 } // end namespace llvm