xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/ScheduleDAGInstrs.cpp (revision 525fe93dc7487a1e63a90f6a2b956abc601963c1)
1 //===---- ScheduleDAGInstrs.cpp - MachineInstr Rescheduling ---------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file This implements the ScheduleDAGInstrs class, which implements
10 /// re-scheduling of MachineInstrs.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/CodeGen/ScheduleDAGInstrs.h"
15 
16 #include "llvm/ADT/IntEqClasses.h"
17 #include "llvm/ADT/MapVector.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/SparseSet.h"
20 #include "llvm/ADT/iterator_range.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/CodeGen/LiveIntervals.h"
24 #include "llvm/CodeGen/LivePhysRegs.h"
25 #include "llvm/CodeGen/MachineBasicBlock.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineFunction.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/ScheduleDFS.h"
37 #include "llvm/CodeGen/SlotIndexes.h"
38 #include "llvm/CodeGen/TargetRegisterInfo.h"
39 #include "llvm/CodeGen/TargetSubtargetInfo.h"
40 #include "llvm/Config/llvm-config.h"
41 #include "llvm/IR/Constants.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/Type.h"
44 #include "llvm/IR/Value.h"
45 #include "llvm/MC/LaneBitmask.h"
46 #include "llvm/MC/MCRegisterInfo.h"
47 #include "llvm/Support/Casting.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Compiler.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/Format.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include <algorithm>
55 #include <cassert>
56 #include <iterator>
57 #include <utility>
58 #include <vector>
59 
60 using namespace llvm;
61 
62 #define DEBUG_TYPE "machine-scheduler"
63 
64 static cl::opt<bool>
65     EnableAASchedMI("enable-aa-sched-mi", cl::Hidden,
66                     cl::desc("Enable use of AA during MI DAG construction"));
67 
68 static cl::opt<bool> UseTBAA("use-tbaa-in-sched-mi", cl::Hidden,
69     cl::init(true), cl::desc("Enable use of TBAA during MI DAG construction"));
70 
71 // Note: the two options below might be used in tuning compile time vs
72 // output quality. Setting HugeRegion so large that it will never be
73 // reached means best-effort, but may be slow.
74 
75 // When Stores and Loads maps (or NonAliasStores and NonAliasLoads)
76 // together hold this many SUs, a reduction of maps will be done.
77 static cl::opt<unsigned> HugeRegion("dag-maps-huge-region", cl::Hidden,
78     cl::init(1000), cl::desc("The limit to use while constructing the DAG "
79                              "prior to scheduling, at which point a trade-off "
80                              "is made to avoid excessive compile time."));
81 
82 static cl::opt<unsigned> ReductionSize(
83     "dag-maps-reduction-size", cl::Hidden,
84     cl::desc("A huge scheduling region will have maps reduced by this many "
85              "nodes at a time. Defaults to HugeRegion / 2."));
86 
87 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
88 static cl::opt<bool> SchedPrintCycles(
89     "sched-print-cycles", cl::Hidden, cl::init(false),
90     cl::desc("Report top/bottom cycles when dumping SUnit instances"));
91 #endif
92 
93 static unsigned getReductionSize() {
94   // Always reduce a huge region with half of the elements, except
95   // when user sets this number explicitly.
96   if (ReductionSize.getNumOccurrences() == 0)
97     return HugeRegion / 2;
98   return ReductionSize;
99 }
100 
101 static void dumpSUList(const ScheduleDAGInstrs::SUList &L) {
102 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
103   dbgs() << "{ ";
104   for (const SUnit *SU : L) {
105     dbgs() << "SU(" << SU->NodeNum << ")";
106     if (SU != L.back())
107       dbgs() << ", ";
108   }
109   dbgs() << "}\n";
110 #endif
111 }
112 
113 ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf,
114                                      const MachineLoopInfo *mli,
115                                      bool RemoveKillFlags)
116     : ScheduleDAG(mf), MLI(mli), MFI(mf.getFrameInfo()),
117       RemoveKillFlags(RemoveKillFlags),
118       UnknownValue(UndefValue::get(
119                              Type::getVoidTy(mf.getFunction().getContext()))), Topo(SUnits, &ExitSU) {
120   DbgValues.clear();
121 
122   const TargetSubtargetInfo &ST = mf.getSubtarget();
123   SchedModel.init(&ST);
124 }
125 
126 /// If this machine instr has memory reference information and it can be
127 /// tracked to a normal reference to a known object, return the Value
128 /// for that object. This function returns false the memory location is
129 /// unknown or may alias anything.
130 static bool getUnderlyingObjectsForInstr(const MachineInstr *MI,
131                                          const MachineFrameInfo &MFI,
132                                          UnderlyingObjectsVector &Objects,
133                                          const DataLayout &DL) {
134   auto AllMMOsOkay = [&]() {
135     for (const MachineMemOperand *MMO : MI->memoperands()) {
136       // TODO: Figure out whether isAtomic is really necessary (see D57601).
137       if (MMO->isVolatile() || MMO->isAtomic())
138         return false;
139 
140       if (const PseudoSourceValue *PSV = MMO->getPseudoValue()) {
141         // Function that contain tail calls don't have unique PseudoSourceValue
142         // objects. Two PseudoSourceValues might refer to the same or
143         // overlapping locations. The client code calling this function assumes
144         // this is not the case. So return a conservative answer of no known
145         // object.
146         if (MFI.hasTailCall())
147           return false;
148 
149         // For now, ignore PseudoSourceValues which may alias LLVM IR values
150         // because the code that uses this function has no way to cope with
151         // such aliases.
152         if (PSV->isAliased(&MFI))
153           return false;
154 
155         bool MayAlias = PSV->mayAlias(&MFI);
156         Objects.emplace_back(PSV, MayAlias);
157       } else if (const Value *V = MMO->getValue()) {
158         SmallVector<Value *, 4> Objs;
159         if (!getUnderlyingObjectsForCodeGen(V, Objs))
160           return false;
161 
162         for (Value *V : Objs) {
163           assert(isIdentifiedObject(V));
164           Objects.emplace_back(V, true);
165         }
166       } else
167         return false;
168     }
169     return true;
170   };
171 
172   if (!AllMMOsOkay()) {
173     Objects.clear();
174     return false;
175   }
176 
177   return true;
178 }
179 
180 void ScheduleDAGInstrs::startBlock(MachineBasicBlock *bb) {
181   BB = bb;
182 }
183 
184 void ScheduleDAGInstrs::finishBlock() {
185   // Subclasses should no longer refer to the old block.
186   BB = nullptr;
187 }
188 
189 void ScheduleDAGInstrs::enterRegion(MachineBasicBlock *bb,
190                                     MachineBasicBlock::iterator begin,
191                                     MachineBasicBlock::iterator end,
192                                     unsigned regioninstrs) {
193   assert(bb == BB && "startBlock should set BB");
194   RegionBegin = begin;
195   RegionEnd = end;
196   NumRegionInstrs = regioninstrs;
197 }
198 
199 void ScheduleDAGInstrs::exitRegion() {
200   // Nothing to do.
201 }
202 
203 void ScheduleDAGInstrs::addSchedBarrierDeps() {
204   MachineInstr *ExitMI =
205       RegionEnd != BB->end()
206           ? &*skipDebugInstructionsBackward(RegionEnd, RegionBegin)
207           : nullptr;
208   ExitSU.setInstr(ExitMI);
209   // Add dependencies on the defs and uses of the instruction.
210   if (ExitMI) {
211     for (const MachineOperand &MO : ExitMI->operands()) {
212       if (!MO.isReg() || MO.isDef()) continue;
213       Register Reg = MO.getReg();
214       if (Reg.isPhysical()) {
215         Uses.insert(PhysRegSUOper(&ExitSU, -1, Reg));
216       } else if (Reg.isVirtual() && MO.readsReg()) {
217         addVRegUseDeps(&ExitSU, ExitMI->getOperandNo(&MO));
218       }
219     }
220   }
221   if (!ExitMI || (!ExitMI->isCall() && !ExitMI->isBarrier())) {
222     // For others, e.g. fallthrough, conditional branch, assume the exit
223     // uses all the registers that are livein to the successor blocks.
224     for (const MachineBasicBlock *Succ : BB->successors()) {
225       for (const auto &LI : Succ->liveins()) {
226         if (!Uses.contains(LI.PhysReg))
227           Uses.insert(PhysRegSUOper(&ExitSU, -1, LI.PhysReg));
228       }
229     }
230   }
231 }
232 
233 /// MO is an operand of SU's instruction that defines a physical register. Adds
234 /// data dependencies from SU to any uses of the physical register.
235 void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
236   const MachineOperand &MO = SU->getInstr()->getOperand(OperIdx);
237   assert(MO.isDef() && "expect physreg def");
238 
239   // Ask the target if address-backscheduling is desirable, and if so how much.
240   const TargetSubtargetInfo &ST = MF.getSubtarget();
241 
242   // Only use any non-zero latency for real defs/uses, in contrast to
243   // "fake" operands added by regalloc.
244   const MCInstrDesc *DefMIDesc = &SU->getInstr()->getDesc();
245   bool ImplicitPseudoDef = (OperIdx >= DefMIDesc->getNumOperands() &&
246                             !DefMIDesc->hasImplicitDefOfPhysReg(MO.getReg()));
247   for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
248        Alias.isValid(); ++Alias) {
249     for (Reg2SUnitsMap::iterator I = Uses.find(*Alias); I != Uses.end(); ++I) {
250       SUnit *UseSU = I->SU;
251       if (UseSU == SU)
252         continue;
253 
254       // Adjust the dependence latency using operand def/use information,
255       // then allow the target to perform its own adjustments.
256       int UseOp = I->OpIdx;
257       MachineInstr *RegUse = nullptr;
258       SDep Dep;
259       if (UseOp < 0)
260         Dep = SDep(SU, SDep::Artificial);
261       else {
262         // Set the hasPhysRegDefs only for physreg defs that have a use within
263         // the scheduling region.
264         SU->hasPhysRegDefs = true;
265         Dep = SDep(SU, SDep::Data, *Alias);
266         RegUse = UseSU->getInstr();
267       }
268       const MCInstrDesc *UseMIDesc =
269           (RegUse ? &UseSU->getInstr()->getDesc() : nullptr);
270       bool ImplicitPseudoUse =
271           (UseMIDesc && UseOp >= ((int)UseMIDesc->getNumOperands()) &&
272            !UseMIDesc->hasImplicitUseOfPhysReg(*Alias));
273       if (!ImplicitPseudoDef && !ImplicitPseudoUse) {
274         Dep.setLatency(SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
275                                                         RegUse, UseOp));
276       } else {
277         Dep.setLatency(0);
278       }
279       ST.adjustSchedDependency(SU, OperIdx, UseSU, UseOp, Dep);
280       UseSU->addPred(Dep);
281     }
282   }
283 }
284 
285 /// Adds register dependencies (data, anti, and output) from this SUnit
286 /// to following instructions in the same scheduling region that depend the
287 /// physical register referenced at OperIdx.
288 void ScheduleDAGInstrs::addPhysRegDeps(SUnit *SU, unsigned OperIdx) {
289   MachineInstr *MI = SU->getInstr();
290   MachineOperand &MO = MI->getOperand(OperIdx);
291   Register Reg = MO.getReg();
292   // We do not need to track any dependencies for constant registers.
293   if (MRI.isConstantPhysReg(Reg))
294     return;
295 
296   const TargetSubtargetInfo &ST = MF.getSubtarget();
297 
298   // Optionally add output and anti dependencies. For anti
299   // dependencies we use a latency of 0 because for a multi-issue
300   // target we want to allow the defining instruction to issue
301   // in the same cycle as the using instruction.
302   // TODO: Using a latency of 1 here for output dependencies assumes
303   //       there's no cost for reusing registers.
304   SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
305   for (MCRegAliasIterator Alias(Reg, TRI, true); Alias.isValid(); ++Alias) {
306     if (!Defs.contains(*Alias))
307       continue;
308     for (Reg2SUnitsMap::iterator I = Defs.find(*Alias); I != Defs.end(); ++I) {
309       SUnit *DefSU = I->SU;
310       if (DefSU == &ExitSU)
311         continue;
312       if (DefSU != SU &&
313           (Kind != SDep::Output || !MO.isDead() ||
314            !DefSU->getInstr()->registerDefIsDead(*Alias))) {
315         SDep Dep(SU, Kind, /*Reg=*/*Alias);
316         if (Kind != SDep::Anti)
317           Dep.setLatency(
318             SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
319         ST.adjustSchedDependency(SU, OperIdx, DefSU, I->OpIdx, Dep);
320         DefSU->addPred(Dep);
321       }
322     }
323   }
324 
325   if (!MO.isDef()) {
326     SU->hasPhysRegUses = true;
327     // Either insert a new Reg2SUnits entry with an empty SUnits list, or
328     // retrieve the existing SUnits list for this register's uses.
329     // Push this SUnit on the use list.
330     Uses.insert(PhysRegSUOper(SU, OperIdx, Reg));
331     if (RemoveKillFlags)
332       MO.setIsKill(false);
333   } else {
334     addPhysRegDataDeps(SU, OperIdx);
335 
336     // Clear previous uses and defs of this register and its subergisters.
337     for (MCSubRegIterator SubReg(Reg, TRI, true); SubReg.isValid(); ++SubReg) {
338       if (Uses.contains(*SubReg))
339         Uses.eraseAll(*SubReg);
340       if (!MO.isDead())
341         Defs.eraseAll(*SubReg);
342     }
343     if (MO.isDead() && SU->isCall) {
344       // Calls will not be reordered because of chain dependencies (see
345       // below). Since call operands are dead, calls may continue to be added
346       // to the DefList making dependence checking quadratic in the size of
347       // the block. Instead, we leave only one call at the back of the
348       // DefList.
349       Reg2SUnitsMap::RangePair P = Defs.equal_range(Reg);
350       Reg2SUnitsMap::iterator B = P.first;
351       Reg2SUnitsMap::iterator I = P.second;
352       for (bool isBegin = I == B; !isBegin; /* empty */) {
353         isBegin = (--I) == B;
354         if (!I->SU->isCall)
355           break;
356         I = Defs.erase(I);
357       }
358     }
359 
360     // Defs are pushed in the order they are visited and never reordered.
361     Defs.insert(PhysRegSUOper(SU, OperIdx, Reg));
362   }
363 }
364 
365 LaneBitmask ScheduleDAGInstrs::getLaneMaskForMO(const MachineOperand &MO) const
366 {
367   Register Reg = MO.getReg();
368   // No point in tracking lanemasks if we don't have interesting subregisters.
369   const TargetRegisterClass &RC = *MRI.getRegClass(Reg);
370   if (!RC.HasDisjunctSubRegs)
371     return LaneBitmask::getAll();
372 
373   unsigned SubReg = MO.getSubReg();
374   if (SubReg == 0)
375     return RC.getLaneMask();
376   return TRI->getSubRegIndexLaneMask(SubReg);
377 }
378 
379 bool ScheduleDAGInstrs::deadDefHasNoUse(const MachineOperand &MO) {
380   auto RegUse = CurrentVRegUses.find(MO.getReg());
381   if (RegUse == CurrentVRegUses.end())
382     return true;
383   return (RegUse->LaneMask & getLaneMaskForMO(MO)).none();
384 }
385 
386 /// Adds register output and data dependencies from this SUnit to instructions
387 /// that occur later in the same scheduling region if they read from or write to
388 /// the virtual register defined at OperIdx.
389 ///
390 /// TODO: Hoist loop induction variable increments. This has to be
391 /// reevaluated. Generally, IV scheduling should be done before coalescing.
392 void ScheduleDAGInstrs::addVRegDefDeps(SUnit *SU, unsigned OperIdx) {
393   MachineInstr *MI = SU->getInstr();
394   MachineOperand &MO = MI->getOperand(OperIdx);
395   Register Reg = MO.getReg();
396 
397   LaneBitmask DefLaneMask;
398   LaneBitmask KillLaneMask;
399   if (TrackLaneMasks) {
400     bool IsKill = MO.getSubReg() == 0 || MO.isUndef();
401     DefLaneMask = getLaneMaskForMO(MO);
402     // If we have a <read-undef> flag, none of the lane values comes from an
403     // earlier instruction.
404     KillLaneMask = IsKill ? LaneBitmask::getAll() : DefLaneMask;
405 
406     if (MO.getSubReg() != 0 && MO.isUndef()) {
407       // There may be other subregister defs on the same instruction of the same
408       // register in later operands. The lanes of other defs will now be live
409       // after this instruction, so these should not be treated as killed by the
410       // instruction even though they appear to be killed in this one operand.
411       for (const MachineOperand &OtherMO :
412            llvm::drop_begin(MI->operands(), OperIdx + 1))
413         if (OtherMO.isReg() && OtherMO.isDef() && OtherMO.getReg() == Reg)
414           KillLaneMask &= ~getLaneMaskForMO(OtherMO);
415     }
416 
417     // Clear undef flag, we'll re-add it later once we know which subregister
418     // Def is first.
419     MO.setIsUndef(false);
420   } else {
421     DefLaneMask = LaneBitmask::getAll();
422     KillLaneMask = LaneBitmask::getAll();
423   }
424 
425   if (MO.isDead()) {
426     assert(deadDefHasNoUse(MO) && "Dead defs should have no uses");
427   } else {
428     // Add data dependence to all uses we found so far.
429     const TargetSubtargetInfo &ST = MF.getSubtarget();
430     for (VReg2SUnitOperIdxMultiMap::iterator I = CurrentVRegUses.find(Reg),
431          E = CurrentVRegUses.end(); I != E; /*empty*/) {
432       LaneBitmask LaneMask = I->LaneMask;
433       // Ignore uses of other lanes.
434       if ((LaneMask & KillLaneMask).none()) {
435         ++I;
436         continue;
437       }
438 
439       if ((LaneMask & DefLaneMask).any()) {
440         SUnit *UseSU = I->SU;
441         MachineInstr *Use = UseSU->getInstr();
442         SDep Dep(SU, SDep::Data, Reg);
443         Dep.setLatency(SchedModel.computeOperandLatency(MI, OperIdx, Use,
444                                                         I->OperandIndex));
445         ST.adjustSchedDependency(SU, OperIdx, UseSU, I->OperandIndex, Dep);
446         UseSU->addPred(Dep);
447       }
448 
449       LaneMask &= ~KillLaneMask;
450       // If we found a Def for all lanes of this use, remove it from the list.
451       if (LaneMask.any()) {
452         I->LaneMask = LaneMask;
453         ++I;
454       } else
455         I = CurrentVRegUses.erase(I);
456     }
457   }
458 
459   // Shortcut: Singly defined vregs do not have output/anti dependencies.
460   if (MRI.hasOneDef(Reg))
461     return;
462 
463   // Add output dependence to the next nearest defs of this vreg.
464   //
465   // Unless this definition is dead, the output dependence should be
466   // transitively redundant with antidependencies from this definition's
467   // uses. We're conservative for now until we have a way to guarantee the uses
468   // are not eliminated sometime during scheduling. The output dependence edge
469   // is also useful if output latency exceeds def-use latency.
470   LaneBitmask LaneMask = DefLaneMask;
471   for (VReg2SUnit &V2SU : make_range(CurrentVRegDefs.find(Reg),
472                                      CurrentVRegDefs.end())) {
473     // Ignore defs for other lanes.
474     if ((V2SU.LaneMask & LaneMask).none())
475       continue;
476     // Add an output dependence.
477     SUnit *DefSU = V2SU.SU;
478     // Ignore additional defs of the same lanes in one instruction. This can
479     // happen because lanemasks are shared for targets with too many
480     // subregisters. We also use some representration tricks/hacks where we
481     // add super-register defs/uses, to imply that although we only access parts
482     // of the reg we care about the full one.
483     if (DefSU == SU)
484       continue;
485     SDep Dep(SU, SDep::Output, Reg);
486     Dep.setLatency(
487       SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
488     DefSU->addPred(Dep);
489 
490     // Update current definition. This can get tricky if the def was about a
491     // bigger lanemask before. We then have to shrink it and create a new
492     // VReg2SUnit for the non-overlapping part.
493     LaneBitmask OverlapMask = V2SU.LaneMask & LaneMask;
494     LaneBitmask NonOverlapMask = V2SU.LaneMask & ~LaneMask;
495     V2SU.SU = SU;
496     V2SU.LaneMask = OverlapMask;
497     if (NonOverlapMask.any())
498       CurrentVRegDefs.insert(VReg2SUnit(Reg, NonOverlapMask, DefSU));
499   }
500   // If there was no CurrentVRegDefs entry for some lanes yet, create one.
501   if (LaneMask.any())
502     CurrentVRegDefs.insert(VReg2SUnit(Reg, LaneMask, SU));
503 }
504 
505 /// Adds a register data dependency if the instruction that defines the
506 /// virtual register used at OperIdx is mapped to an SUnit. Add a register
507 /// antidependency from this SUnit to instructions that occur later in the same
508 /// scheduling region if they write the virtual register.
509 ///
510 /// TODO: Handle ExitSU "uses" properly.
511 void ScheduleDAGInstrs::addVRegUseDeps(SUnit *SU, unsigned OperIdx) {
512   const MachineInstr *MI = SU->getInstr();
513   assert(!MI->isDebugOrPseudoInstr());
514 
515   const MachineOperand &MO = MI->getOperand(OperIdx);
516   Register Reg = MO.getReg();
517 
518   // Remember the use. Data dependencies will be added when we find the def.
519   LaneBitmask LaneMask = TrackLaneMasks ? getLaneMaskForMO(MO)
520                                         : LaneBitmask::getAll();
521   CurrentVRegUses.insert(VReg2SUnitOperIdx(Reg, LaneMask, OperIdx, SU));
522 
523   // Add antidependences to the following defs of the vreg.
524   for (VReg2SUnit &V2SU : make_range(CurrentVRegDefs.find(Reg),
525                                      CurrentVRegDefs.end())) {
526     // Ignore defs for unrelated lanes.
527     LaneBitmask PrevDefLaneMask = V2SU.LaneMask;
528     if ((PrevDefLaneMask & LaneMask).none())
529       continue;
530     if (V2SU.SU == SU)
531       continue;
532 
533     V2SU.SU->addPred(SDep(SU, SDep::Anti, Reg));
534   }
535 }
536 
537 /// Returns true if MI is an instruction we are unable to reason about
538 /// (like a call or something with unmodeled side effects).
539 static inline bool isGlobalMemoryObject(MachineInstr *MI) {
540   return MI->isCall() || MI->hasUnmodeledSideEffects() ||
541          (MI->hasOrderedMemoryRef() && !MI->isDereferenceableInvariantLoad());
542 }
543 
544 void ScheduleDAGInstrs::addChainDependency (SUnit *SUa, SUnit *SUb,
545                                             unsigned Latency) {
546   if (SUa->getInstr()->mayAlias(AAForDep, *SUb->getInstr(), UseTBAA)) {
547     SDep Dep(SUa, SDep::MayAliasMem);
548     Dep.setLatency(Latency);
549     SUb->addPred(Dep);
550   }
551 }
552 
553 /// Creates an SUnit for each real instruction, numbered in top-down
554 /// topological order. The instruction order A < B, implies that no edge exists
555 /// from B to A.
556 ///
557 /// Map each real instruction to its SUnit.
558 ///
559 /// After initSUnits, the SUnits vector cannot be resized and the scheduler may
560 /// hang onto SUnit pointers. We may relax this in the future by using SUnit IDs
561 /// instead of pointers.
562 ///
563 /// MachineScheduler relies on initSUnits numbering the nodes by their order in
564 /// the original instruction list.
565 void ScheduleDAGInstrs::initSUnits() {
566   // We'll be allocating one SUnit for each real instruction in the region,
567   // which is contained within a basic block.
568   SUnits.reserve(NumRegionInstrs);
569 
570   for (MachineInstr &MI : make_range(RegionBegin, RegionEnd)) {
571     if (MI.isDebugOrPseudoInstr())
572       continue;
573 
574     SUnit *SU = newSUnit(&MI);
575     MISUnitMap[&MI] = SU;
576 
577     SU->isCall = MI.isCall();
578     SU->isCommutable = MI.isCommutable();
579 
580     // Assign the Latency field of SU using target-provided information.
581     SU->Latency = SchedModel.computeInstrLatency(SU->getInstr());
582 
583     // If this SUnit uses a reserved or unbuffered resource, mark it as such.
584     //
585     // Reserved resources block an instruction from issuing and stall the
586     // entire pipeline. These are identified by BufferSize=0.
587     //
588     // Unbuffered resources prevent execution of subsequent instructions that
589     // require the same resources. This is used for in-order execution pipelines
590     // within an out-of-order core. These are identified by BufferSize=1.
591     if (SchedModel.hasInstrSchedModel()) {
592       const MCSchedClassDesc *SC = getSchedClass(SU);
593       for (const MCWriteProcResEntry &PRE :
594            make_range(SchedModel.getWriteProcResBegin(SC),
595                       SchedModel.getWriteProcResEnd(SC))) {
596         switch (SchedModel.getProcResource(PRE.ProcResourceIdx)->BufferSize) {
597         case 0:
598           SU->hasReservedResource = true;
599           break;
600         case 1:
601           SU->isUnbuffered = true;
602           break;
603         default:
604           break;
605         }
606       }
607     }
608   }
609 }
610 
611 class ScheduleDAGInstrs::Value2SUsMap : public MapVector<ValueType, SUList> {
612   /// Current total number of SUs in map.
613   unsigned NumNodes = 0;
614 
615   /// 1 for loads, 0 for stores. (see comment in SUList)
616   unsigned TrueMemOrderLatency;
617 
618 public:
619   Value2SUsMap(unsigned lat = 0) : TrueMemOrderLatency(lat) {}
620 
621   /// To keep NumNodes up to date, insert() is used instead of
622   /// this operator w/ push_back().
623   ValueType &operator[](const SUList &Key) {
624     llvm_unreachable("Don't use. Use insert() instead."); };
625 
626   /// Adds SU to the SUList of V. If Map grows huge, reduce its size by calling
627   /// reduce().
628   void inline insert(SUnit *SU, ValueType V) {
629     MapVector::operator[](V).push_back(SU);
630     NumNodes++;
631   }
632 
633   /// Clears the list of SUs mapped to V.
634   void inline clearList(ValueType V) {
635     iterator Itr = find(V);
636     if (Itr != end()) {
637       assert(NumNodes >= Itr->second.size());
638       NumNodes -= Itr->second.size();
639 
640       Itr->second.clear();
641     }
642   }
643 
644   /// Clears map from all contents.
645   void clear() {
646     MapVector<ValueType, SUList>::clear();
647     NumNodes = 0;
648   }
649 
650   unsigned inline size() const { return NumNodes; }
651 
652   /// Counts the number of SUs in this map after a reduction.
653   void reComputeSize() {
654     NumNodes = 0;
655     for (auto &I : *this)
656       NumNodes += I.second.size();
657   }
658 
659   unsigned inline getTrueMemOrderLatency() const {
660     return TrueMemOrderLatency;
661   }
662 
663   void dump();
664 };
665 
666 void ScheduleDAGInstrs::addChainDependencies(SUnit *SU,
667                                              Value2SUsMap &Val2SUsMap) {
668   for (auto &I : Val2SUsMap)
669     addChainDependencies(SU, I.second,
670                          Val2SUsMap.getTrueMemOrderLatency());
671 }
672 
673 void ScheduleDAGInstrs::addChainDependencies(SUnit *SU,
674                                              Value2SUsMap &Val2SUsMap,
675                                              ValueType V) {
676   Value2SUsMap::iterator Itr = Val2SUsMap.find(V);
677   if (Itr != Val2SUsMap.end())
678     addChainDependencies(SU, Itr->second,
679                          Val2SUsMap.getTrueMemOrderLatency());
680 }
681 
682 void ScheduleDAGInstrs::addBarrierChain(Value2SUsMap &map) {
683   assert(BarrierChain != nullptr);
684 
685   for (auto &[V, SUs] : map) {
686     (void)V;
687     for (auto *SU : SUs)
688       SU->addPredBarrier(BarrierChain);
689   }
690   map.clear();
691 }
692 
693 void ScheduleDAGInstrs::insertBarrierChain(Value2SUsMap &map) {
694   assert(BarrierChain != nullptr);
695 
696   // Go through all lists of SUs.
697   for (Value2SUsMap::iterator I = map.begin(), EE = map.end(); I != EE;) {
698     Value2SUsMap::iterator CurrItr = I++;
699     SUList &sus = CurrItr->second;
700     SUList::iterator SUItr = sus.begin(), SUEE = sus.end();
701     for (; SUItr != SUEE; ++SUItr) {
702       // Stop on BarrierChain or any instruction above it.
703       if ((*SUItr)->NodeNum <= BarrierChain->NodeNum)
704         break;
705 
706       (*SUItr)->addPredBarrier(BarrierChain);
707     }
708 
709     // Remove also the BarrierChain from list if present.
710     if (SUItr != SUEE && *SUItr == BarrierChain)
711       SUItr++;
712 
713     // Remove all SUs that are now successors of BarrierChain.
714     if (SUItr != sus.begin())
715       sus.erase(sus.begin(), SUItr);
716   }
717 
718   // Remove all entries with empty su lists.
719   map.remove_if([&](std::pair<ValueType, SUList> &mapEntry) {
720       return (mapEntry.second.empty()); });
721 
722   // Recompute the size of the map (NumNodes).
723   map.reComputeSize();
724 }
725 
726 void ScheduleDAGInstrs::buildSchedGraph(AAResults *AA,
727                                         RegPressureTracker *RPTracker,
728                                         PressureDiffs *PDiffs,
729                                         LiveIntervals *LIS,
730                                         bool TrackLaneMasks) {
731   const TargetSubtargetInfo &ST = MF.getSubtarget();
732   bool UseAA = EnableAASchedMI.getNumOccurrences() > 0 ? EnableAASchedMI
733                                                        : ST.useAA();
734   AAForDep = UseAA ? AA : nullptr;
735 
736   BarrierChain = nullptr;
737 
738   this->TrackLaneMasks = TrackLaneMasks;
739   MISUnitMap.clear();
740   ScheduleDAG::clearDAG();
741 
742   // Create an SUnit for each real instruction.
743   initSUnits();
744 
745   if (PDiffs)
746     PDiffs->init(SUnits.size());
747 
748   // We build scheduling units by walking a block's instruction list
749   // from bottom to top.
750 
751   // Each MIs' memory operand(s) is analyzed to a list of underlying
752   // objects. The SU is then inserted in the SUList(s) mapped from the
753   // Value(s). Each Value thus gets mapped to lists of SUs depending
754   // on it, stores and loads kept separately. Two SUs are trivially
755   // non-aliasing if they both depend on only identified Values and do
756   // not share any common Value.
757   Value2SUsMap Stores, Loads(1 /*TrueMemOrderLatency*/);
758 
759   // Certain memory accesses are known to not alias any SU in Stores
760   // or Loads, and have therefore their own 'NonAlias'
761   // domain. E.g. spill / reload instructions never alias LLVM I/R
762   // Values. It would be nice to assume that this type of memory
763   // accesses always have a proper memory operand modelling, and are
764   // therefore never unanalyzable, but this is conservatively not
765   // done.
766   Value2SUsMap NonAliasStores, NonAliasLoads(1 /*TrueMemOrderLatency*/);
767 
768   // Track all instructions that may raise floating-point exceptions.
769   // These do not depend on one other (or normal loads or stores), but
770   // must not be rescheduled across global barriers.  Note that we don't
771   // really need a "map" here since we don't track those MIs by value;
772   // using the same Value2SUsMap data type here is simply a matter of
773   // convenience.
774   Value2SUsMap FPExceptions;
775 
776   // Remove any stale debug info; sometimes BuildSchedGraph is called again
777   // without emitting the info from the previous call.
778   DbgValues.clear();
779   FirstDbgValue = nullptr;
780 
781   assert(Defs.empty() && Uses.empty() &&
782          "Only BuildGraph should update Defs/Uses");
783   Defs.setUniverse(TRI->getNumRegs());
784   Uses.setUniverse(TRI->getNumRegs());
785 
786   assert(CurrentVRegDefs.empty() && "nobody else should use CurrentVRegDefs");
787   assert(CurrentVRegUses.empty() && "nobody else should use CurrentVRegUses");
788   unsigned NumVirtRegs = MRI.getNumVirtRegs();
789   CurrentVRegDefs.setUniverse(NumVirtRegs);
790   CurrentVRegUses.setUniverse(NumVirtRegs);
791 
792   // Model data dependencies between instructions being scheduled and the
793   // ExitSU.
794   addSchedBarrierDeps();
795 
796   // Walk the list of instructions, from bottom moving up.
797   MachineInstr *DbgMI = nullptr;
798   for (MachineBasicBlock::iterator MII = RegionEnd, MIE = RegionBegin;
799        MII != MIE; --MII) {
800     MachineInstr &MI = *std::prev(MII);
801     if (DbgMI) {
802       DbgValues.emplace_back(DbgMI, &MI);
803       DbgMI = nullptr;
804     }
805 
806     if (MI.isDebugValue() || MI.isDebugPHI()) {
807       DbgMI = &MI;
808       continue;
809     }
810 
811     if (MI.isDebugLabel() || MI.isDebugRef() || MI.isPseudoProbe())
812       continue;
813 
814     SUnit *SU = MISUnitMap[&MI];
815     assert(SU && "No SUnit mapped to this MI");
816 
817     if (RPTracker) {
818       RegisterOperands RegOpers;
819       RegOpers.collect(MI, *TRI, MRI, TrackLaneMasks, false);
820       if (TrackLaneMasks) {
821         SlotIndex SlotIdx = LIS->getInstructionIndex(MI);
822         RegOpers.adjustLaneLiveness(*LIS, MRI, SlotIdx);
823       }
824       if (PDiffs != nullptr)
825         PDiffs->addInstruction(SU->NodeNum, RegOpers, MRI);
826 
827       if (RPTracker->getPos() == RegionEnd || &*RPTracker->getPos() != &MI)
828         RPTracker->recedeSkipDebugValues();
829       assert(&*RPTracker->getPos() == &MI && "RPTracker in sync");
830       RPTracker->recede(RegOpers);
831     }
832 
833     assert(
834         (CanHandleTerminators || (!MI.isTerminator() && !MI.isPosition())) &&
835         "Cannot schedule terminators or labels!");
836 
837     // Add register-based dependencies (data, anti, and output).
838     // For some instructions (calls, returns, inline-asm, etc.) there can
839     // be explicit uses and implicit defs, in which case the use will appear
840     // on the operand list before the def. Do two passes over the operand
841     // list to make sure that defs are processed before any uses.
842     bool HasVRegDef = false;
843     for (unsigned j = 0, n = MI.getNumOperands(); j != n; ++j) {
844       const MachineOperand &MO = MI.getOperand(j);
845       if (!MO.isReg() || !MO.isDef())
846         continue;
847       Register Reg = MO.getReg();
848       if (Reg.isPhysical()) {
849         addPhysRegDeps(SU, j);
850       } else if (Reg.isVirtual()) {
851         HasVRegDef = true;
852         addVRegDefDeps(SU, j);
853       }
854     }
855     // Now process all uses.
856     for (unsigned j = 0, n = MI.getNumOperands(); j != n; ++j) {
857       const MachineOperand &MO = MI.getOperand(j);
858       // Only look at use operands.
859       // We do not need to check for MO.readsReg() here because subsequent
860       // subregister defs will get output dependence edges and need no
861       // additional use dependencies.
862       if (!MO.isReg() || !MO.isUse())
863         continue;
864       Register Reg = MO.getReg();
865       if (Reg.isPhysical()) {
866         addPhysRegDeps(SU, j);
867       } else if (Reg.isVirtual() && MO.readsReg()) {
868         addVRegUseDeps(SU, j);
869       }
870     }
871 
872     // If we haven't seen any uses in this scheduling region, create a
873     // dependence edge to ExitSU to model the live-out latency. This is required
874     // for vreg defs with no in-region use, and prefetches with no vreg def.
875     //
876     // FIXME: NumDataSuccs would be more precise than NumSuccs here. This
877     // check currently relies on being called before adding chain deps.
878     if (SU->NumSuccs == 0 && SU->Latency > 1 && (HasVRegDef || MI.mayLoad())) {
879       SDep Dep(SU, SDep::Artificial);
880       Dep.setLatency(SU->Latency - 1);
881       ExitSU.addPred(Dep);
882     }
883 
884     // Add memory dependencies (Note: isStoreToStackSlot and
885     // isLoadFromStackSLot are not usable after stack slots are lowered to
886     // actual addresses).
887 
888     // This is a barrier event that acts as a pivotal node in the DAG.
889     if (isGlobalMemoryObject(&MI)) {
890 
891       // Become the barrier chain.
892       if (BarrierChain)
893         BarrierChain->addPredBarrier(SU);
894       BarrierChain = SU;
895 
896       LLVM_DEBUG(dbgs() << "Global memory object and new barrier chain: SU("
897                         << BarrierChain->NodeNum << ").\n";);
898 
899       // Add dependencies against everything below it and clear maps.
900       addBarrierChain(Stores);
901       addBarrierChain(Loads);
902       addBarrierChain(NonAliasStores);
903       addBarrierChain(NonAliasLoads);
904       addBarrierChain(FPExceptions);
905 
906       continue;
907     }
908 
909     // Instructions that may raise FP exceptions may not be moved
910     // across any global barriers.
911     if (MI.mayRaiseFPException()) {
912       if (BarrierChain)
913         BarrierChain->addPredBarrier(SU);
914 
915       FPExceptions.insert(SU, UnknownValue);
916 
917       if (FPExceptions.size() >= HugeRegion) {
918         LLVM_DEBUG(dbgs() << "Reducing FPExceptions map.\n";);
919         Value2SUsMap empty;
920         reduceHugeMemNodeMaps(FPExceptions, empty, getReductionSize());
921       }
922     }
923 
924     // If it's not a store or a variant load, we're done.
925     if (!MI.mayStore() &&
926         !(MI.mayLoad() && !MI.isDereferenceableInvariantLoad()))
927       continue;
928 
929     // Always add dependecy edge to BarrierChain if present.
930     if (BarrierChain)
931       BarrierChain->addPredBarrier(SU);
932 
933     // Find the underlying objects for MI. The Objs vector is either
934     // empty, or filled with the Values of memory locations which this
935     // SU depends on.
936     UnderlyingObjectsVector Objs;
937     bool ObjsFound = getUnderlyingObjectsForInstr(&MI, MFI, Objs,
938                                                   MF.getDataLayout());
939 
940     if (MI.mayStore()) {
941       if (!ObjsFound) {
942         // An unknown store depends on all stores and loads.
943         addChainDependencies(SU, Stores);
944         addChainDependencies(SU, NonAliasStores);
945         addChainDependencies(SU, Loads);
946         addChainDependencies(SU, NonAliasLoads);
947 
948         // Map this store to 'UnknownValue'.
949         Stores.insert(SU, UnknownValue);
950       } else {
951         // Add precise dependencies against all previously seen memory
952         // accesses mapped to the same Value(s).
953         for (const UnderlyingObject &UnderlObj : Objs) {
954           ValueType V = UnderlObj.getValue();
955           bool ThisMayAlias = UnderlObj.mayAlias();
956 
957           // Add dependencies to previous stores and loads mapped to V.
958           addChainDependencies(SU, (ThisMayAlias ? Stores : NonAliasStores), V);
959           addChainDependencies(SU, (ThisMayAlias ? Loads : NonAliasLoads), V);
960         }
961         // Update the store map after all chains have been added to avoid adding
962         // self-loop edge if multiple underlying objects are present.
963         for (const UnderlyingObject &UnderlObj : Objs) {
964           ValueType V = UnderlObj.getValue();
965           bool ThisMayAlias = UnderlObj.mayAlias();
966 
967           // Map this store to V.
968           (ThisMayAlias ? Stores : NonAliasStores).insert(SU, V);
969         }
970         // The store may have dependencies to unanalyzable loads and
971         // stores.
972         addChainDependencies(SU, Loads, UnknownValue);
973         addChainDependencies(SU, Stores, UnknownValue);
974       }
975     } else { // SU is a load.
976       if (!ObjsFound) {
977         // An unknown load depends on all stores.
978         addChainDependencies(SU, Stores);
979         addChainDependencies(SU, NonAliasStores);
980 
981         Loads.insert(SU, UnknownValue);
982       } else {
983         for (const UnderlyingObject &UnderlObj : Objs) {
984           ValueType V = UnderlObj.getValue();
985           bool ThisMayAlias = UnderlObj.mayAlias();
986 
987           // Add precise dependencies against all previously seen stores
988           // mapping to the same Value(s).
989           addChainDependencies(SU, (ThisMayAlias ? Stores : NonAliasStores), V);
990 
991           // Map this load to V.
992           (ThisMayAlias ? Loads : NonAliasLoads).insert(SU, V);
993         }
994         // The load may have dependencies to unanalyzable stores.
995         addChainDependencies(SU, Stores, UnknownValue);
996       }
997     }
998 
999     // Reduce maps if they grow huge.
1000     if (Stores.size() + Loads.size() >= HugeRegion) {
1001       LLVM_DEBUG(dbgs() << "Reducing Stores and Loads maps.\n";);
1002       reduceHugeMemNodeMaps(Stores, Loads, getReductionSize());
1003     }
1004     if (NonAliasStores.size() + NonAliasLoads.size() >= HugeRegion) {
1005       LLVM_DEBUG(
1006           dbgs() << "Reducing NonAliasStores and NonAliasLoads maps.\n";);
1007       reduceHugeMemNodeMaps(NonAliasStores, NonAliasLoads, getReductionSize());
1008     }
1009   }
1010 
1011   if (DbgMI)
1012     FirstDbgValue = DbgMI;
1013 
1014   Defs.clear();
1015   Uses.clear();
1016   CurrentVRegDefs.clear();
1017   CurrentVRegUses.clear();
1018 
1019   Topo.MarkDirty();
1020 }
1021 
1022 raw_ostream &llvm::operator<<(raw_ostream &OS, const PseudoSourceValue* PSV) {
1023   PSV->printCustom(OS);
1024   return OS;
1025 }
1026 
1027 void ScheduleDAGInstrs::Value2SUsMap::dump() {
1028   for (const auto &[ValType, SUs] : *this) {
1029     if (ValType.is<const Value*>()) {
1030       const Value *V = ValType.get<const Value*>();
1031       if (isa<UndefValue>(V))
1032         dbgs() << "Unknown";
1033       else
1034         V->printAsOperand(dbgs());
1035     }
1036     else if (ValType.is<const PseudoSourceValue*>())
1037       dbgs() << ValType.get<const PseudoSourceValue*>();
1038     else
1039       llvm_unreachable("Unknown Value type.");
1040 
1041     dbgs() << " : ";
1042     dumpSUList(SUs);
1043   }
1044 }
1045 
1046 void ScheduleDAGInstrs::reduceHugeMemNodeMaps(Value2SUsMap &stores,
1047                                               Value2SUsMap &loads, unsigned N) {
1048   LLVM_DEBUG(dbgs() << "Before reduction:\nStoring SUnits:\n"; stores.dump();
1049              dbgs() << "Loading SUnits:\n"; loads.dump());
1050 
1051   // Insert all SU's NodeNums into a vector and sort it.
1052   std::vector<unsigned> NodeNums;
1053   NodeNums.reserve(stores.size() + loads.size());
1054   for (const auto &[V, SUs] : stores) {
1055     (void)V;
1056     for (const auto *SU : SUs)
1057       NodeNums.push_back(SU->NodeNum);
1058   }
1059   for (const auto &[V, SUs] : loads) {
1060     (void)V;
1061     for (const auto *SU : SUs)
1062       NodeNums.push_back(SU->NodeNum);
1063   }
1064   llvm::sort(NodeNums);
1065 
1066   // The N last elements in NodeNums will be removed, and the SU with
1067   // the lowest NodeNum of them will become the new BarrierChain to
1068   // let the not yet seen SUs have a dependency to the removed SUs.
1069   assert(N <= NodeNums.size());
1070   SUnit *newBarrierChain = &SUnits[*(NodeNums.end() - N)];
1071   if (BarrierChain) {
1072     // The aliasing and non-aliasing maps reduce independently of each
1073     // other, but share a common BarrierChain. Check if the
1074     // newBarrierChain is above the former one. If it is not, it may
1075     // introduce a loop to use newBarrierChain, so keep the old one.
1076     if (newBarrierChain->NodeNum < BarrierChain->NodeNum) {
1077       BarrierChain->addPredBarrier(newBarrierChain);
1078       BarrierChain = newBarrierChain;
1079       LLVM_DEBUG(dbgs() << "Inserting new barrier chain: SU("
1080                         << BarrierChain->NodeNum << ").\n";);
1081     }
1082     else
1083       LLVM_DEBUG(dbgs() << "Keeping old barrier chain: SU("
1084                         << BarrierChain->NodeNum << ").\n";);
1085   }
1086   else
1087     BarrierChain = newBarrierChain;
1088 
1089   insertBarrierChain(stores);
1090   insertBarrierChain(loads);
1091 
1092   LLVM_DEBUG(dbgs() << "After reduction:\nStoring SUnits:\n"; stores.dump();
1093              dbgs() << "Loading SUnits:\n"; loads.dump());
1094 }
1095 
1096 static void toggleKills(const MachineRegisterInfo &MRI, LivePhysRegs &LiveRegs,
1097                         MachineInstr &MI, bool addToLiveRegs) {
1098   for (MachineOperand &MO : MI.operands()) {
1099     if (!MO.isReg() || !MO.readsReg())
1100       continue;
1101     Register Reg = MO.getReg();
1102     if (!Reg)
1103       continue;
1104 
1105     // Things that are available after the instruction are killed by it.
1106     bool IsKill = LiveRegs.available(MRI, Reg);
1107     MO.setIsKill(IsKill);
1108     if (addToLiveRegs)
1109       LiveRegs.addReg(Reg);
1110   }
1111 }
1112 
1113 void ScheduleDAGInstrs::fixupKills(MachineBasicBlock &MBB) {
1114   LLVM_DEBUG(dbgs() << "Fixup kills for " << printMBBReference(MBB) << '\n');
1115 
1116   LiveRegs.init(*TRI);
1117   LiveRegs.addLiveOuts(MBB);
1118 
1119   // Examine block from end to start...
1120   for (MachineInstr &MI : llvm::reverse(MBB)) {
1121     if (MI.isDebugOrPseudoInstr())
1122       continue;
1123 
1124     // Update liveness.  Registers that are defed but not used in this
1125     // instruction are now dead. Mark register and all subregs as they
1126     // are completely defined.
1127     for (ConstMIBundleOperands O(MI); O.isValid(); ++O) {
1128       const MachineOperand &MO = *O;
1129       if (MO.isReg()) {
1130         if (!MO.isDef())
1131           continue;
1132         Register Reg = MO.getReg();
1133         if (!Reg)
1134           continue;
1135         LiveRegs.removeReg(Reg);
1136       } else if (MO.isRegMask()) {
1137         LiveRegs.removeRegsInMask(MO);
1138       }
1139     }
1140 
1141     // If there is a bundle header fix it up first.
1142     if (!MI.isBundled()) {
1143       toggleKills(MRI, LiveRegs, MI, true);
1144     } else {
1145       MachineBasicBlock::instr_iterator Bundle = MI.getIterator();
1146       if (MI.isBundle())
1147         toggleKills(MRI, LiveRegs, MI, false);
1148 
1149       // Some targets make the (questionable) assumtion that the instructions
1150       // inside the bundle are ordered and consequently only the last use of
1151       // a register inside the bundle can kill it.
1152       MachineBasicBlock::instr_iterator I = std::next(Bundle);
1153       while (I->isBundledWithSucc())
1154         ++I;
1155       do {
1156         if (!I->isDebugOrPseudoInstr())
1157           toggleKills(MRI, LiveRegs, *I, true);
1158         --I;
1159       } while (I != Bundle);
1160     }
1161   }
1162 }
1163 
1164 void ScheduleDAGInstrs::dumpNode(const SUnit &SU) const {
1165 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1166   dumpNodeName(SU);
1167   if (SchedPrintCycles)
1168     dbgs() << " [TopReadyCycle = " << SU.TopReadyCycle
1169            << ", BottomReadyCycle = " << SU.BotReadyCycle << "]";
1170   dbgs() << ": ";
1171   SU.getInstr()->dump();
1172 #endif
1173 }
1174 
1175 void ScheduleDAGInstrs::dump() const {
1176 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1177   if (EntrySU.getInstr() != nullptr)
1178     dumpNodeAll(EntrySU);
1179   for (const SUnit &SU : SUnits)
1180     dumpNodeAll(SU);
1181   if (ExitSU.getInstr() != nullptr)
1182     dumpNodeAll(ExitSU);
1183 #endif
1184 }
1185 
1186 std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const {
1187   std::string s;
1188   raw_string_ostream oss(s);
1189   if (SU == &EntrySU)
1190     oss << "<entry>";
1191   else if (SU == &ExitSU)
1192     oss << "<exit>";
1193   else
1194     SU->getInstr()->print(oss, /*IsStandalone=*/true);
1195   return oss.str();
1196 }
1197 
1198 /// Return the basic block label. It is not necessarilly unique because a block
1199 /// contains multiple scheduling regions. But it is fine for visualization.
1200 std::string ScheduleDAGInstrs::getDAGName() const {
1201   return "dag." + BB->getFullName();
1202 }
1203 
1204 bool ScheduleDAGInstrs::canAddEdge(SUnit *SuccSU, SUnit *PredSU) {
1205   return SuccSU == &ExitSU || !Topo.IsReachable(PredSU, SuccSU);
1206 }
1207 
1208 bool ScheduleDAGInstrs::addEdge(SUnit *SuccSU, const SDep &PredDep) {
1209   if (SuccSU != &ExitSU) {
1210     // Do not use WillCreateCycle, it assumes SD scheduling.
1211     // If Pred is reachable from Succ, then the edge creates a cycle.
1212     if (Topo.IsReachable(PredDep.getSUnit(), SuccSU))
1213       return false;
1214     Topo.AddPredQueued(SuccSU, PredDep.getSUnit());
1215   }
1216   SuccSU->addPred(PredDep, /*Required=*/!PredDep.isArtificial());
1217   // Return true regardless of whether a new edge needed to be inserted.
1218   return true;
1219 }
1220 
1221 //===----------------------------------------------------------------------===//
1222 // SchedDFSResult Implementation
1223 //===----------------------------------------------------------------------===//
1224 
1225 namespace llvm {
1226 
1227 /// Internal state used to compute SchedDFSResult.
1228 class SchedDFSImpl {
1229   SchedDFSResult &R;
1230 
1231   /// Join DAG nodes into equivalence classes by their subtree.
1232   IntEqClasses SubtreeClasses;
1233   /// List PredSU, SuccSU pairs that represent data edges between subtrees.
1234   std::vector<std::pair<const SUnit *, const SUnit*>> ConnectionPairs;
1235 
1236   struct RootData {
1237     unsigned NodeID;
1238     unsigned ParentNodeID;  ///< Parent node (member of the parent subtree).
1239     unsigned SubInstrCount = 0; ///< Instr count in this tree only, not
1240                                 /// children.
1241 
1242     RootData(unsigned id): NodeID(id),
1243                            ParentNodeID(SchedDFSResult::InvalidSubtreeID) {}
1244 
1245     unsigned getSparseSetIndex() const { return NodeID; }
1246   };
1247 
1248   SparseSet<RootData> RootSet;
1249 
1250 public:
1251   SchedDFSImpl(SchedDFSResult &r): R(r), SubtreeClasses(R.DFSNodeData.size()) {
1252     RootSet.setUniverse(R.DFSNodeData.size());
1253   }
1254 
1255   /// Returns true if this node been visited by the DFS traversal.
1256   ///
1257   /// During visitPostorderNode the Node's SubtreeID is assigned to the Node
1258   /// ID. Later, SubtreeID is updated but remains valid.
1259   bool isVisited(const SUnit *SU) const {
1260     return R.DFSNodeData[SU->NodeNum].SubtreeID
1261       != SchedDFSResult::InvalidSubtreeID;
1262   }
1263 
1264   /// Initializes this node's instruction count. We don't need to flag the node
1265   /// visited until visitPostorder because the DAG cannot have cycles.
1266   void visitPreorder(const SUnit *SU) {
1267     R.DFSNodeData[SU->NodeNum].InstrCount =
1268       SU->getInstr()->isTransient() ? 0 : 1;
1269   }
1270 
1271   /// Called once for each node after all predecessors are visited. Revisit this
1272   /// node's predecessors and potentially join them now that we know the ILP of
1273   /// the other predecessors.
1274   void visitPostorderNode(const SUnit *SU) {
1275     // Mark this node as the root of a subtree. It may be joined with its
1276     // successors later.
1277     R.DFSNodeData[SU->NodeNum].SubtreeID = SU->NodeNum;
1278     RootData RData(SU->NodeNum);
1279     RData.SubInstrCount = SU->getInstr()->isTransient() ? 0 : 1;
1280 
1281     // If any predecessors are still in their own subtree, they either cannot be
1282     // joined or are large enough to remain separate. If this parent node's
1283     // total instruction count is not greater than a child subtree by at least
1284     // the subtree limit, then try to join it now since splitting subtrees is
1285     // only useful if multiple high-pressure paths are possible.
1286     unsigned InstrCount = R.DFSNodeData[SU->NodeNum].InstrCount;
1287     for (const SDep &PredDep : SU->Preds) {
1288       if (PredDep.getKind() != SDep::Data)
1289         continue;
1290       unsigned PredNum = PredDep.getSUnit()->NodeNum;
1291       if ((InstrCount - R.DFSNodeData[PredNum].InstrCount) < R.SubtreeLimit)
1292         joinPredSubtree(PredDep, SU, /*CheckLimit=*/false);
1293 
1294       // Either link or merge the TreeData entry from the child to the parent.
1295       if (R.DFSNodeData[PredNum].SubtreeID == PredNum) {
1296         // If the predecessor's parent is invalid, this is a tree edge and the
1297         // current node is the parent.
1298         if (RootSet[PredNum].ParentNodeID == SchedDFSResult::InvalidSubtreeID)
1299           RootSet[PredNum].ParentNodeID = SU->NodeNum;
1300       }
1301       else if (RootSet.count(PredNum)) {
1302         // The predecessor is not a root, but is still in the root set. This
1303         // must be the new parent that it was just joined to. Note that
1304         // RootSet[PredNum].ParentNodeID may either be invalid or may still be
1305         // set to the original parent.
1306         RData.SubInstrCount += RootSet[PredNum].SubInstrCount;
1307         RootSet.erase(PredNum);
1308       }
1309     }
1310     RootSet[SU->NodeNum] = RData;
1311   }
1312 
1313   /// Called once for each tree edge after calling visitPostOrderNode on
1314   /// the predecessor. Increment the parent node's instruction count and
1315   /// preemptively join this subtree to its parent's if it is small enough.
1316   void visitPostorderEdge(const SDep &PredDep, const SUnit *Succ) {
1317     R.DFSNodeData[Succ->NodeNum].InstrCount
1318       += R.DFSNodeData[PredDep.getSUnit()->NodeNum].InstrCount;
1319     joinPredSubtree(PredDep, Succ);
1320   }
1321 
1322   /// Adds a connection for cross edges.
1323   void visitCrossEdge(const SDep &PredDep, const SUnit *Succ) {
1324     ConnectionPairs.emplace_back(PredDep.getSUnit(), Succ);
1325   }
1326 
1327   /// Sets each node's subtree ID to the representative ID and record
1328   /// connections between trees.
1329   void finalize() {
1330     SubtreeClasses.compress();
1331     R.DFSTreeData.resize(SubtreeClasses.getNumClasses());
1332     assert(SubtreeClasses.getNumClasses() == RootSet.size()
1333            && "number of roots should match trees");
1334     for (const RootData &Root : RootSet) {
1335       unsigned TreeID = SubtreeClasses[Root.NodeID];
1336       if (Root.ParentNodeID != SchedDFSResult::InvalidSubtreeID)
1337         R.DFSTreeData[TreeID].ParentTreeID = SubtreeClasses[Root.ParentNodeID];
1338       R.DFSTreeData[TreeID].SubInstrCount = Root.SubInstrCount;
1339       // Note that SubInstrCount may be greater than InstrCount if we joined
1340       // subtrees across a cross edge. InstrCount will be attributed to the
1341       // original parent, while SubInstrCount will be attributed to the joined
1342       // parent.
1343     }
1344     R.SubtreeConnections.resize(SubtreeClasses.getNumClasses());
1345     R.SubtreeConnectLevels.resize(SubtreeClasses.getNumClasses());
1346     LLVM_DEBUG(dbgs() << R.getNumSubtrees() << " subtrees:\n");
1347     for (unsigned Idx = 0, End = R.DFSNodeData.size(); Idx != End; ++Idx) {
1348       R.DFSNodeData[Idx].SubtreeID = SubtreeClasses[Idx];
1349       LLVM_DEBUG(dbgs() << "  SU(" << Idx << ") in tree "
1350                         << R.DFSNodeData[Idx].SubtreeID << '\n');
1351     }
1352     for (const auto &[Pred, Succ] : ConnectionPairs) {
1353       unsigned PredTree = SubtreeClasses[Pred->NodeNum];
1354       unsigned SuccTree = SubtreeClasses[Succ->NodeNum];
1355       if (PredTree == SuccTree)
1356         continue;
1357       unsigned Depth = Pred->getDepth();
1358       addConnection(PredTree, SuccTree, Depth);
1359       addConnection(SuccTree, PredTree, Depth);
1360     }
1361   }
1362 
1363 protected:
1364   /// Joins the predecessor subtree with the successor that is its DFS parent.
1365   /// Applies some heuristics before joining.
1366   bool joinPredSubtree(const SDep &PredDep, const SUnit *Succ,
1367                        bool CheckLimit = true) {
1368     assert(PredDep.getKind() == SDep::Data && "Subtrees are for data edges");
1369 
1370     // Check if the predecessor is already joined.
1371     const SUnit *PredSU = PredDep.getSUnit();
1372     unsigned PredNum = PredSU->NodeNum;
1373     if (R.DFSNodeData[PredNum].SubtreeID != PredNum)
1374       return false;
1375 
1376     // Four is the magic number of successors before a node is considered a
1377     // pinch point.
1378     unsigned NumDataSucs = 0;
1379     for (const SDep &SuccDep : PredSU->Succs) {
1380       if (SuccDep.getKind() == SDep::Data) {
1381         if (++NumDataSucs >= 4)
1382           return false;
1383       }
1384     }
1385     if (CheckLimit && R.DFSNodeData[PredNum].InstrCount > R.SubtreeLimit)
1386       return false;
1387     R.DFSNodeData[PredNum].SubtreeID = Succ->NodeNum;
1388     SubtreeClasses.join(Succ->NodeNum, PredNum);
1389     return true;
1390   }
1391 
1392   /// Called by finalize() to record a connection between trees.
1393   void addConnection(unsigned FromTree, unsigned ToTree, unsigned Depth) {
1394     if (!Depth)
1395       return;
1396 
1397     do {
1398       SmallVectorImpl<SchedDFSResult::Connection> &Connections =
1399         R.SubtreeConnections[FromTree];
1400       for (SchedDFSResult::Connection &C : Connections) {
1401         if (C.TreeID == ToTree) {
1402           C.Level = std::max(C.Level, Depth);
1403           return;
1404         }
1405       }
1406       Connections.push_back(SchedDFSResult::Connection(ToTree, Depth));
1407       FromTree = R.DFSTreeData[FromTree].ParentTreeID;
1408     } while (FromTree != SchedDFSResult::InvalidSubtreeID);
1409   }
1410 };
1411 
1412 } // end namespace llvm
1413 
1414 namespace {
1415 
1416 /// Manage the stack used by a reverse depth-first search over the DAG.
1417 class SchedDAGReverseDFS {
1418   std::vector<std::pair<const SUnit *, SUnit::const_pred_iterator>> DFSStack;
1419 
1420 public:
1421   bool isComplete() const { return DFSStack.empty(); }
1422 
1423   void follow(const SUnit *SU) {
1424     DFSStack.emplace_back(SU, SU->Preds.begin());
1425   }
1426   void advance() { ++DFSStack.back().second; }
1427 
1428   const SDep *backtrack() {
1429     DFSStack.pop_back();
1430     return DFSStack.empty() ? nullptr : std::prev(DFSStack.back().second);
1431   }
1432 
1433   const SUnit *getCurr() const { return DFSStack.back().first; }
1434 
1435   SUnit::const_pred_iterator getPred() const { return DFSStack.back().second; }
1436 
1437   SUnit::const_pred_iterator getPredEnd() const {
1438     return getCurr()->Preds.end();
1439   }
1440 };
1441 
1442 } // end anonymous namespace
1443 
1444 static bool hasDataSucc(const SUnit *SU) {
1445   for (const SDep &SuccDep : SU->Succs) {
1446     if (SuccDep.getKind() == SDep::Data &&
1447         !SuccDep.getSUnit()->isBoundaryNode())
1448       return true;
1449   }
1450   return false;
1451 }
1452 
1453 /// Computes an ILP metric for all nodes in the subDAG reachable via depth-first
1454 /// search from this root.
1455 void SchedDFSResult::compute(ArrayRef<SUnit> SUnits) {
1456   if (!IsBottomUp)
1457     llvm_unreachable("Top-down ILP metric is unimplemented");
1458 
1459   SchedDFSImpl Impl(*this);
1460   for (const SUnit &SU : SUnits) {
1461     if (Impl.isVisited(&SU) || hasDataSucc(&SU))
1462       continue;
1463 
1464     SchedDAGReverseDFS DFS;
1465     Impl.visitPreorder(&SU);
1466     DFS.follow(&SU);
1467     while (true) {
1468       // Traverse the leftmost path as far as possible.
1469       while (DFS.getPred() != DFS.getPredEnd()) {
1470         const SDep &PredDep = *DFS.getPred();
1471         DFS.advance();
1472         // Ignore non-data edges.
1473         if (PredDep.getKind() != SDep::Data
1474             || PredDep.getSUnit()->isBoundaryNode()) {
1475           continue;
1476         }
1477         // An already visited edge is a cross edge, assuming an acyclic DAG.
1478         if (Impl.isVisited(PredDep.getSUnit())) {
1479           Impl.visitCrossEdge(PredDep, DFS.getCurr());
1480           continue;
1481         }
1482         Impl.visitPreorder(PredDep.getSUnit());
1483         DFS.follow(PredDep.getSUnit());
1484       }
1485       // Visit the top of the stack in postorder and backtrack.
1486       const SUnit *Child = DFS.getCurr();
1487       const SDep *PredDep = DFS.backtrack();
1488       Impl.visitPostorderNode(Child);
1489       if (PredDep)
1490         Impl.visitPostorderEdge(*PredDep, DFS.getCurr());
1491       if (DFS.isComplete())
1492         break;
1493     }
1494   }
1495   Impl.finalize();
1496 }
1497 
1498 /// The root of the given SubtreeID was just scheduled. For all subtrees
1499 /// connected to this tree, record the depth of the connection so that the
1500 /// nearest connected subtrees can be prioritized.
1501 void SchedDFSResult::scheduleTree(unsigned SubtreeID) {
1502   for (const Connection &C : SubtreeConnections[SubtreeID]) {
1503     SubtreeConnectLevels[C.TreeID] =
1504       std::max(SubtreeConnectLevels[C.TreeID], C.Level);
1505     LLVM_DEBUG(dbgs() << "  Tree: " << C.TreeID << " @"
1506                       << SubtreeConnectLevels[C.TreeID] << '\n');
1507   }
1508 }
1509 
1510 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1511 LLVM_DUMP_METHOD void ILPValue::print(raw_ostream &OS) const {
1512   OS << InstrCount << " / " << Length << " = ";
1513   if (!Length)
1514     OS << "BADILP";
1515   else
1516     OS << format("%g", ((double)InstrCount / Length));
1517 }
1518 
1519 LLVM_DUMP_METHOD void ILPValue::dump() const {
1520   dbgs() << *this << '\n';
1521 }
1522 
1523 namespace llvm {
1524 
1525 LLVM_DUMP_METHOD
1526 raw_ostream &operator<<(raw_ostream &OS, const ILPValue &Val) {
1527   Val.print(OS);
1528   return OS;
1529 }
1530 
1531 } // end namespace llvm
1532 
1533 #endif
1534