xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/LoopInfo.cpp (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
1 //===- LoopInfo.cpp - Natural Loop Calculator -----------------------------===//
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 // This file defines the LoopInfo class that is used to identify natural loops
10 // and determine the loop depth of various nodes of the CFG.  Note that the
11 // loops identified may actually be several natural loops that share the same
12 // header node... not just a single natural loop.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "llvm/Analysis/LoopInfo.h"
17 #include "llvm/ADT/ScopeExit.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/Analysis/IVDescriptors.h"
20 #include "llvm/Analysis/LoopIterator.h"
21 #include "llvm/Analysis/LoopNestAnalysis.h"
22 #include "llvm/Analysis/MemorySSA.h"
23 #include "llvm/Analysis/MemorySSAUpdater.h"
24 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/Config/llvm-config.h"
27 #include "llvm/IR/CFG.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DebugLoc.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/IR/Metadata.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/IR/PassManager.h"
36 #include "llvm/IR/PrintPasses.h"
37 #include "llvm/InitializePasses.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/GenericLoopInfoImpl.h"
40 #include "llvm/Support/raw_ostream.h"
41 using namespace llvm;
42 
43 // Explicitly instantiate methods in LoopInfoImpl.h for IR-level Loops.
44 template class llvm::LoopBase<BasicBlock, Loop>;
45 template class llvm::LoopInfoBase<BasicBlock, Loop>;
46 
47 // Always verify loopinfo if expensive checking is enabled.
48 #ifdef EXPENSIVE_CHECKS
49 bool llvm::VerifyLoopInfo = true;
50 #else
51 bool llvm::VerifyLoopInfo = false;
52 #endif
53 static cl::opt<bool, true>
54     VerifyLoopInfoX("verify-loop-info", cl::location(VerifyLoopInfo),
55                     cl::Hidden, cl::desc("Verify loop info (time consuming)"));
56 
57 //===----------------------------------------------------------------------===//
58 // Loop implementation
59 //
60 
61 bool Loop::isLoopInvariant(const Value *V) const {
62   if (const Instruction *I = dyn_cast<Instruction>(V))
63     return !contains(I);
64   return true; // All non-instructions are loop invariant
65 }
66 
67 bool Loop::hasLoopInvariantOperands(const Instruction *I) const {
68   return all_of(I->operands(), [this](Value *V) { return isLoopInvariant(V); });
69 }
70 
71 bool Loop::makeLoopInvariant(Value *V, bool &Changed, Instruction *InsertPt,
72                              MemorySSAUpdater *MSSAU,
73                              ScalarEvolution *SE) const {
74   if (Instruction *I = dyn_cast<Instruction>(V))
75     return makeLoopInvariant(I, Changed, InsertPt, MSSAU, SE);
76   return true; // All non-instructions are loop-invariant.
77 }
78 
79 bool Loop::makeLoopInvariant(Instruction *I, bool &Changed,
80                              Instruction *InsertPt, MemorySSAUpdater *MSSAU,
81                              ScalarEvolution *SE) const {
82   // Test if the value is already loop-invariant.
83   if (isLoopInvariant(I))
84     return true;
85   if (!isSafeToSpeculativelyExecute(I))
86     return false;
87   if (I->mayReadFromMemory())
88     return false;
89   // EH block instructions are immobile.
90   if (I->isEHPad())
91     return false;
92   // Determine the insertion point, unless one was given.
93   if (!InsertPt) {
94     BasicBlock *Preheader = getLoopPreheader();
95     // Without a preheader, hoisting is not feasible.
96     if (!Preheader)
97       return false;
98     InsertPt = Preheader->getTerminator();
99   }
100   // Don't hoist instructions with loop-variant operands.
101   for (Value *Operand : I->operands())
102     if (!makeLoopInvariant(Operand, Changed, InsertPt, MSSAU, SE))
103       return false;
104 
105   // Hoist.
106   I->moveBefore(InsertPt);
107   if (MSSAU)
108     if (auto *MUD = MSSAU->getMemorySSA()->getMemoryAccess(I))
109       MSSAU->moveToPlace(MUD, InsertPt->getParent(),
110                          MemorySSA::BeforeTerminator);
111 
112   // There is possibility of hoisting this instruction above some arbitrary
113   // condition. Any metadata defined on it can be control dependent on this
114   // condition. Conservatively strip it here so that we don't give any wrong
115   // information to the optimizer.
116   I->dropUnknownNonDebugMetadata();
117 
118   if (SE)
119     SE->forgetBlockAndLoopDispositions(I);
120 
121   Changed = true;
122   return true;
123 }
124 
125 bool Loop::getIncomingAndBackEdge(BasicBlock *&Incoming,
126                                   BasicBlock *&Backedge) const {
127   BasicBlock *H = getHeader();
128 
129   Incoming = nullptr;
130   Backedge = nullptr;
131   pred_iterator PI = pred_begin(H);
132   assert(PI != pred_end(H) && "Loop must have at least one backedge!");
133   Backedge = *PI++;
134   if (PI == pred_end(H))
135     return false; // dead loop
136   Incoming = *PI++;
137   if (PI != pred_end(H))
138     return false; // multiple backedges?
139 
140   if (contains(Incoming)) {
141     if (contains(Backedge))
142       return false;
143     std::swap(Incoming, Backedge);
144   } else if (!contains(Backedge))
145     return false;
146 
147   assert(Incoming && Backedge && "expected non-null incoming and backedges");
148   return true;
149 }
150 
151 PHINode *Loop::getCanonicalInductionVariable() const {
152   BasicBlock *H = getHeader();
153 
154   BasicBlock *Incoming = nullptr, *Backedge = nullptr;
155   if (!getIncomingAndBackEdge(Incoming, Backedge))
156     return nullptr;
157 
158   // Loop over all of the PHI nodes, looking for a canonical indvar.
159   for (BasicBlock::iterator I = H->begin(); isa<PHINode>(I); ++I) {
160     PHINode *PN = cast<PHINode>(I);
161     if (ConstantInt *CI =
162             dyn_cast<ConstantInt>(PN->getIncomingValueForBlock(Incoming)))
163       if (CI->isZero())
164         if (Instruction *Inc =
165                 dyn_cast<Instruction>(PN->getIncomingValueForBlock(Backedge)))
166           if (Inc->getOpcode() == Instruction::Add && Inc->getOperand(0) == PN)
167             if (ConstantInt *CI = dyn_cast<ConstantInt>(Inc->getOperand(1)))
168               if (CI->isOne())
169                 return PN;
170   }
171   return nullptr;
172 }
173 
174 /// Get the latch condition instruction.
175 ICmpInst *Loop::getLatchCmpInst() const {
176   if (BasicBlock *Latch = getLoopLatch())
177     if (BranchInst *BI = dyn_cast_or_null<BranchInst>(Latch->getTerminator()))
178       if (BI->isConditional())
179         return dyn_cast<ICmpInst>(BI->getCondition());
180 
181   return nullptr;
182 }
183 
184 /// Return the final value of the loop induction variable if found.
185 static Value *findFinalIVValue(const Loop &L, const PHINode &IndVar,
186                                const Instruction &StepInst) {
187   ICmpInst *LatchCmpInst = L.getLatchCmpInst();
188   if (!LatchCmpInst)
189     return nullptr;
190 
191   Value *Op0 = LatchCmpInst->getOperand(0);
192   Value *Op1 = LatchCmpInst->getOperand(1);
193   if (Op0 == &IndVar || Op0 == &StepInst)
194     return Op1;
195 
196   if (Op1 == &IndVar || Op1 == &StepInst)
197     return Op0;
198 
199   return nullptr;
200 }
201 
202 std::optional<Loop::LoopBounds>
203 Loop::LoopBounds::getBounds(const Loop &L, PHINode &IndVar,
204                             ScalarEvolution &SE) {
205   InductionDescriptor IndDesc;
206   if (!InductionDescriptor::isInductionPHI(&IndVar, &L, &SE, IndDesc))
207     return std::nullopt;
208 
209   Value *InitialIVValue = IndDesc.getStartValue();
210   Instruction *StepInst = IndDesc.getInductionBinOp();
211   if (!InitialIVValue || !StepInst)
212     return std::nullopt;
213 
214   const SCEV *Step = IndDesc.getStep();
215   Value *StepInstOp1 = StepInst->getOperand(1);
216   Value *StepInstOp0 = StepInst->getOperand(0);
217   Value *StepValue = nullptr;
218   if (SE.getSCEV(StepInstOp1) == Step)
219     StepValue = StepInstOp1;
220   else if (SE.getSCEV(StepInstOp0) == Step)
221     StepValue = StepInstOp0;
222 
223   Value *FinalIVValue = findFinalIVValue(L, IndVar, *StepInst);
224   if (!FinalIVValue)
225     return std::nullopt;
226 
227   return LoopBounds(L, *InitialIVValue, *StepInst, StepValue, *FinalIVValue,
228                     SE);
229 }
230 
231 using Direction = Loop::LoopBounds::Direction;
232 
233 ICmpInst::Predicate Loop::LoopBounds::getCanonicalPredicate() const {
234   BasicBlock *Latch = L.getLoopLatch();
235   assert(Latch && "Expecting valid latch");
236 
237   BranchInst *BI = dyn_cast_or_null<BranchInst>(Latch->getTerminator());
238   assert(BI && BI->isConditional() && "Expecting conditional latch branch");
239 
240   ICmpInst *LatchCmpInst = dyn_cast<ICmpInst>(BI->getCondition());
241   assert(LatchCmpInst &&
242          "Expecting the latch compare instruction to be a CmpInst");
243 
244   // Need to inverse the predicate when first successor is not the loop
245   // header
246   ICmpInst::Predicate Pred = (BI->getSuccessor(0) == L.getHeader())
247                                  ? LatchCmpInst->getPredicate()
248                                  : LatchCmpInst->getInversePredicate();
249 
250   if (LatchCmpInst->getOperand(0) == &getFinalIVValue())
251     Pred = ICmpInst::getSwappedPredicate(Pred);
252 
253   // Need to flip strictness of the predicate when the latch compare instruction
254   // is not using StepInst
255   if (LatchCmpInst->getOperand(0) == &getStepInst() ||
256       LatchCmpInst->getOperand(1) == &getStepInst())
257     return Pred;
258 
259   // Cannot flip strictness of NE and EQ
260   if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
261     return ICmpInst::getFlippedStrictnessPredicate(Pred);
262 
263   Direction D = getDirection();
264   if (D == Direction::Increasing)
265     return ICmpInst::ICMP_SLT;
266 
267   if (D == Direction::Decreasing)
268     return ICmpInst::ICMP_SGT;
269 
270   // If cannot determine the direction, then unable to find the canonical
271   // predicate
272   return ICmpInst::BAD_ICMP_PREDICATE;
273 }
274 
275 Direction Loop::LoopBounds::getDirection() const {
276   if (const SCEVAddRecExpr *StepAddRecExpr =
277           dyn_cast<SCEVAddRecExpr>(SE.getSCEV(&getStepInst())))
278     if (const SCEV *StepRecur = StepAddRecExpr->getStepRecurrence(SE)) {
279       if (SE.isKnownPositive(StepRecur))
280         return Direction::Increasing;
281       if (SE.isKnownNegative(StepRecur))
282         return Direction::Decreasing;
283     }
284 
285   return Direction::Unknown;
286 }
287 
288 std::optional<Loop::LoopBounds> Loop::getBounds(ScalarEvolution &SE) const {
289   if (PHINode *IndVar = getInductionVariable(SE))
290     return LoopBounds::getBounds(*this, *IndVar, SE);
291 
292   return std::nullopt;
293 }
294 
295 PHINode *Loop::getInductionVariable(ScalarEvolution &SE) const {
296   if (!isLoopSimplifyForm())
297     return nullptr;
298 
299   BasicBlock *Header = getHeader();
300   assert(Header && "Expected a valid loop header");
301   ICmpInst *CmpInst = getLatchCmpInst();
302   if (!CmpInst)
303     return nullptr;
304 
305   Value *LatchCmpOp0 = CmpInst->getOperand(0);
306   Value *LatchCmpOp1 = CmpInst->getOperand(1);
307 
308   for (PHINode &IndVar : Header->phis()) {
309     InductionDescriptor IndDesc;
310     if (!InductionDescriptor::isInductionPHI(&IndVar, this, &SE, IndDesc))
311       continue;
312 
313     BasicBlock *Latch = getLoopLatch();
314     Value *StepInst = IndVar.getIncomingValueForBlock(Latch);
315 
316     // case 1:
317     // IndVar = phi[{InitialValue, preheader}, {StepInst, latch}]
318     // StepInst = IndVar + step
319     // cmp = StepInst < FinalValue
320     if (StepInst == LatchCmpOp0 || StepInst == LatchCmpOp1)
321       return &IndVar;
322 
323     // case 2:
324     // IndVar = phi[{InitialValue, preheader}, {StepInst, latch}]
325     // StepInst = IndVar + step
326     // cmp = IndVar < FinalValue
327     if (&IndVar == LatchCmpOp0 || &IndVar == LatchCmpOp1)
328       return &IndVar;
329   }
330 
331   return nullptr;
332 }
333 
334 bool Loop::getInductionDescriptor(ScalarEvolution &SE,
335                                   InductionDescriptor &IndDesc) const {
336   if (PHINode *IndVar = getInductionVariable(SE))
337     return InductionDescriptor::isInductionPHI(IndVar, this, &SE, IndDesc);
338 
339   return false;
340 }
341 
342 bool Loop::isAuxiliaryInductionVariable(PHINode &AuxIndVar,
343                                         ScalarEvolution &SE) const {
344   // Located in the loop header
345   BasicBlock *Header = getHeader();
346   if (AuxIndVar.getParent() != Header)
347     return false;
348 
349   // No uses outside of the loop
350   for (User *U : AuxIndVar.users())
351     if (const Instruction *I = dyn_cast<Instruction>(U))
352       if (!contains(I))
353         return false;
354 
355   InductionDescriptor IndDesc;
356   if (!InductionDescriptor::isInductionPHI(&AuxIndVar, this, &SE, IndDesc))
357     return false;
358 
359   // The step instruction opcode should be add or sub.
360   if (IndDesc.getInductionOpcode() != Instruction::Add &&
361       IndDesc.getInductionOpcode() != Instruction::Sub)
362     return false;
363 
364   // Incremented by a loop invariant step for each loop iteration
365   return SE.isLoopInvariant(IndDesc.getStep(), this);
366 }
367 
368 BranchInst *Loop::getLoopGuardBranch() const {
369   if (!isLoopSimplifyForm())
370     return nullptr;
371 
372   BasicBlock *Preheader = getLoopPreheader();
373   assert(Preheader && getLoopLatch() &&
374          "Expecting a loop with valid preheader and latch");
375 
376   // Loop should be in rotate form.
377   if (!isRotatedForm())
378     return nullptr;
379 
380   // Disallow loops with more than one unique exit block, as we do not verify
381   // that GuardOtherSucc post dominates all exit blocks.
382   BasicBlock *ExitFromLatch = getUniqueExitBlock();
383   if (!ExitFromLatch)
384     return nullptr;
385 
386   BasicBlock *GuardBB = Preheader->getUniquePredecessor();
387   if (!GuardBB)
388     return nullptr;
389 
390   assert(GuardBB->getTerminator() && "Expecting valid guard terminator");
391 
392   BranchInst *GuardBI = dyn_cast<BranchInst>(GuardBB->getTerminator());
393   if (!GuardBI || GuardBI->isUnconditional())
394     return nullptr;
395 
396   BasicBlock *GuardOtherSucc = (GuardBI->getSuccessor(0) == Preheader)
397                                    ? GuardBI->getSuccessor(1)
398                                    : GuardBI->getSuccessor(0);
399 
400   // Check if ExitFromLatch (or any BasicBlock which is an empty unique
401   // successor of ExitFromLatch) is equal to GuardOtherSucc. If
402   // skipEmptyBlockUntil returns GuardOtherSucc, then the guard branch for the
403   // loop is GuardBI (return GuardBI), otherwise return nullptr.
404   if (&LoopNest::skipEmptyBlockUntil(ExitFromLatch, GuardOtherSucc,
405                                      /*CheckUniquePred=*/true) ==
406       GuardOtherSucc)
407     return GuardBI;
408   else
409     return nullptr;
410 }
411 
412 bool Loop::isCanonical(ScalarEvolution &SE) const {
413   InductionDescriptor IndDesc;
414   if (!getInductionDescriptor(SE, IndDesc))
415     return false;
416 
417   ConstantInt *Init = dyn_cast_or_null<ConstantInt>(IndDesc.getStartValue());
418   if (!Init || !Init->isZero())
419     return false;
420 
421   if (IndDesc.getInductionOpcode() != Instruction::Add)
422     return false;
423 
424   ConstantInt *Step = IndDesc.getConstIntStepValue();
425   if (!Step || !Step->isOne())
426     return false;
427 
428   return true;
429 }
430 
431 // Check that 'BB' doesn't have any uses outside of the 'L'
432 static bool isBlockInLCSSAForm(const Loop &L, const BasicBlock &BB,
433                                const DominatorTree &DT, bool IgnoreTokens) {
434   for (const Instruction &I : BB) {
435     // Tokens can't be used in PHI nodes and live-out tokens prevent loop
436     // optimizations, so for the purposes of considered LCSSA form, we
437     // can ignore them.
438     if (IgnoreTokens && I.getType()->isTokenTy())
439       continue;
440 
441     for (const Use &U : I.uses()) {
442       const Instruction *UI = cast<Instruction>(U.getUser());
443       const BasicBlock *UserBB = UI->getParent();
444 
445       // For practical purposes, we consider that the use in a PHI
446       // occurs in the respective predecessor block. For more info,
447       // see the `phi` doc in LangRef and the LCSSA doc.
448       if (const PHINode *P = dyn_cast<PHINode>(UI))
449         UserBB = P->getIncomingBlock(U);
450 
451       // Check the current block, as a fast-path, before checking whether
452       // the use is anywhere in the loop.  Most values are used in the same
453       // block they are defined in.  Also, blocks not reachable from the
454       // entry are special; uses in them don't need to go through PHIs.
455       if (UserBB != &BB && !L.contains(UserBB) &&
456           DT.isReachableFromEntry(UserBB))
457         return false;
458     }
459   }
460   return true;
461 }
462 
463 bool Loop::isLCSSAForm(const DominatorTree &DT, bool IgnoreTokens) const {
464   // For each block we check that it doesn't have any uses outside of this loop.
465   return all_of(this->blocks(), [&](const BasicBlock *BB) {
466     return isBlockInLCSSAForm(*this, *BB, DT, IgnoreTokens);
467   });
468 }
469 
470 bool Loop::isRecursivelyLCSSAForm(const DominatorTree &DT, const LoopInfo &LI,
471                                   bool IgnoreTokens) const {
472   // For each block we check that it doesn't have any uses outside of its
473   // innermost loop. This process will transitively guarantee that the current
474   // loop and all of the nested loops are in LCSSA form.
475   return all_of(this->blocks(), [&](const BasicBlock *BB) {
476     return isBlockInLCSSAForm(*LI.getLoopFor(BB), *BB, DT, IgnoreTokens);
477   });
478 }
479 
480 bool Loop::isLoopSimplifyForm() const {
481   // Normal-form loops have a preheader, a single backedge, and all of their
482   // exits have all their predecessors inside the loop.
483   return getLoopPreheader() && getLoopLatch() && hasDedicatedExits();
484 }
485 
486 // Routines that reform the loop CFG and split edges often fail on indirectbr.
487 bool Loop::isSafeToClone() const {
488   // Return false if any loop blocks contain indirectbrs, or there are any calls
489   // to noduplicate functions.
490   for (BasicBlock *BB : this->blocks()) {
491     if (isa<IndirectBrInst>(BB->getTerminator()))
492       return false;
493 
494     for (Instruction &I : *BB)
495       if (auto *CB = dyn_cast<CallBase>(&I))
496         if (CB->cannotDuplicate())
497           return false;
498   }
499   return true;
500 }
501 
502 MDNode *Loop::getLoopID() const {
503   MDNode *LoopID = nullptr;
504 
505   // Go through the latch blocks and check the terminator for the metadata.
506   SmallVector<BasicBlock *, 4> LatchesBlocks;
507   getLoopLatches(LatchesBlocks);
508   for (BasicBlock *BB : LatchesBlocks) {
509     Instruction *TI = BB->getTerminator();
510     MDNode *MD = TI->getMetadata(LLVMContext::MD_loop);
511 
512     if (!MD)
513       return nullptr;
514 
515     if (!LoopID)
516       LoopID = MD;
517     else if (MD != LoopID)
518       return nullptr;
519   }
520   if (!LoopID || LoopID->getNumOperands() == 0 ||
521       LoopID->getOperand(0) != LoopID)
522     return nullptr;
523   return LoopID;
524 }
525 
526 void Loop::setLoopID(MDNode *LoopID) const {
527   assert((!LoopID || LoopID->getNumOperands() > 0) &&
528          "Loop ID needs at least one operand");
529   assert((!LoopID || LoopID->getOperand(0) == LoopID) &&
530          "Loop ID should refer to itself");
531 
532   SmallVector<BasicBlock *, 4> LoopLatches;
533   getLoopLatches(LoopLatches);
534   for (BasicBlock *BB : LoopLatches)
535     BB->getTerminator()->setMetadata(LLVMContext::MD_loop, LoopID);
536 }
537 
538 void Loop::setLoopAlreadyUnrolled() {
539   LLVMContext &Context = getHeader()->getContext();
540 
541   MDNode *DisableUnrollMD =
542       MDNode::get(Context, MDString::get(Context, "llvm.loop.unroll.disable"));
543   MDNode *LoopID = getLoopID();
544   MDNode *NewLoopID = makePostTransformationMetadata(
545       Context, LoopID, {"llvm.loop.unroll."}, {DisableUnrollMD});
546   setLoopID(NewLoopID);
547 }
548 
549 void Loop::setLoopMustProgress() {
550   LLVMContext &Context = getHeader()->getContext();
551 
552   MDNode *MustProgress = findOptionMDForLoop(this, "llvm.loop.mustprogress");
553 
554   if (MustProgress)
555     return;
556 
557   MDNode *MustProgressMD =
558       MDNode::get(Context, MDString::get(Context, "llvm.loop.mustprogress"));
559   MDNode *LoopID = getLoopID();
560   MDNode *NewLoopID =
561       makePostTransformationMetadata(Context, LoopID, {}, {MustProgressMD});
562   setLoopID(NewLoopID);
563 }
564 
565 bool Loop::isAnnotatedParallel() const {
566   MDNode *DesiredLoopIdMetadata = getLoopID();
567 
568   if (!DesiredLoopIdMetadata)
569     return false;
570 
571   MDNode *ParallelAccesses =
572       findOptionMDForLoop(this, "llvm.loop.parallel_accesses");
573   SmallPtrSet<MDNode *, 4>
574       ParallelAccessGroups; // For scalable 'contains' check.
575   if (ParallelAccesses) {
576     for (const MDOperand &MD : drop_begin(ParallelAccesses->operands())) {
577       MDNode *AccGroup = cast<MDNode>(MD.get());
578       assert(isValidAsAccessGroup(AccGroup) &&
579              "List item must be an access group");
580       ParallelAccessGroups.insert(AccGroup);
581     }
582   }
583 
584   // The loop branch contains the parallel loop metadata. In order to ensure
585   // that any parallel-loop-unaware optimization pass hasn't added loop-carried
586   // dependencies (thus converted the loop back to a sequential loop), check
587   // that all the memory instructions in the loop belong to an access group that
588   // is parallel to this loop.
589   for (BasicBlock *BB : this->blocks()) {
590     for (Instruction &I : *BB) {
591       if (!I.mayReadOrWriteMemory())
592         continue;
593 
594       if (MDNode *AccessGroup = I.getMetadata(LLVMContext::MD_access_group)) {
595         auto ContainsAccessGroup = [&ParallelAccessGroups](MDNode *AG) -> bool {
596           if (AG->getNumOperands() == 0) {
597             assert(isValidAsAccessGroup(AG) && "Item must be an access group");
598             return ParallelAccessGroups.count(AG);
599           }
600 
601           for (const MDOperand &AccessListItem : AG->operands()) {
602             MDNode *AccGroup = cast<MDNode>(AccessListItem.get());
603             assert(isValidAsAccessGroup(AccGroup) &&
604                    "List item must be an access group");
605             if (ParallelAccessGroups.count(AccGroup))
606               return true;
607           }
608           return false;
609         };
610 
611         if (ContainsAccessGroup(AccessGroup))
612           continue;
613       }
614 
615       // The memory instruction can refer to the loop identifier metadata
616       // directly or indirectly through another list metadata (in case of
617       // nested parallel loops). The loop identifier metadata refers to
618       // itself so we can check both cases with the same routine.
619       MDNode *LoopIdMD =
620           I.getMetadata(LLVMContext::MD_mem_parallel_loop_access);
621 
622       if (!LoopIdMD)
623         return false;
624 
625       if (!llvm::is_contained(LoopIdMD->operands(), DesiredLoopIdMetadata))
626         return false;
627     }
628   }
629   return true;
630 }
631 
632 DebugLoc Loop::getStartLoc() const { return getLocRange().getStart(); }
633 
634 Loop::LocRange Loop::getLocRange() const {
635   // If we have a debug location in the loop ID, then use it.
636   if (MDNode *LoopID = getLoopID()) {
637     DebugLoc Start;
638     // We use the first DebugLoc in the header as the start location of the loop
639     // and if there is a second DebugLoc in the header we use it as end location
640     // of the loop.
641     for (const MDOperand &MDO : llvm::drop_begin(LoopID->operands())) {
642       if (DILocation *L = dyn_cast<DILocation>(MDO)) {
643         if (!Start)
644           Start = DebugLoc(L);
645         else
646           return LocRange(Start, DebugLoc(L));
647       }
648     }
649 
650     if (Start)
651       return LocRange(Start);
652   }
653 
654   // Try the pre-header first.
655   if (BasicBlock *PHeadBB = getLoopPreheader())
656     if (DebugLoc DL = PHeadBB->getTerminator()->getDebugLoc())
657       return LocRange(DL);
658 
659   // If we have no pre-header or there are no instructions with debug
660   // info in it, try the header.
661   if (BasicBlock *HeadBB = getHeader())
662     return LocRange(HeadBB->getTerminator()->getDebugLoc());
663 
664   return LocRange();
665 }
666 
667 std::string Loop::getLocStr() const {
668   std::string Result;
669   raw_string_ostream OS(Result);
670   if (const DebugLoc LoopDbgLoc = getStartLoc())
671     LoopDbgLoc.print(OS);
672   else
673     // Just print the module name.
674     OS << getHeader()->getParent()->getParent()->getModuleIdentifier();
675   return Result;
676 }
677 
678 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
679 LLVM_DUMP_METHOD void Loop::dump() const { print(dbgs()); }
680 
681 LLVM_DUMP_METHOD void Loop::dumpVerbose() const {
682   print(dbgs(), /*Verbose=*/true);
683 }
684 #endif
685 
686 //===----------------------------------------------------------------------===//
687 // UnloopUpdater implementation
688 //
689 
690 namespace {
691 /// Find the new parent loop for all blocks within the "unloop" whose last
692 /// backedges has just been removed.
693 class UnloopUpdater {
694   Loop &Unloop;
695   LoopInfo *LI;
696 
697   LoopBlocksDFS DFS;
698 
699   // Map unloop's immediate subloops to their nearest reachable parents. Nested
700   // loops within these subloops will not change parents. However, an immediate
701   // subloop's new parent will be the nearest loop reachable from either its own
702   // exits *or* any of its nested loop's exits.
703   DenseMap<Loop *, Loop *> SubloopParents;
704 
705   // Flag the presence of an irreducible backedge whose destination is a block
706   // directly contained by the original unloop.
707   bool FoundIB = false;
708 
709 public:
710   UnloopUpdater(Loop *UL, LoopInfo *LInfo) : Unloop(*UL), LI(LInfo), DFS(UL) {}
711 
712   void updateBlockParents();
713 
714   void removeBlocksFromAncestors();
715 
716   void updateSubloopParents();
717 
718 protected:
719   Loop *getNearestLoop(BasicBlock *BB, Loop *BBLoop);
720 };
721 } // end anonymous namespace
722 
723 /// Update the parent loop for all blocks that are directly contained within the
724 /// original "unloop".
725 void UnloopUpdater::updateBlockParents() {
726   if (Unloop.getNumBlocks()) {
727     // Perform a post order CFG traversal of all blocks within this loop,
728     // propagating the nearest loop from successors to predecessors.
729     LoopBlocksTraversal Traversal(DFS, LI);
730     for (BasicBlock *POI : Traversal) {
731 
732       Loop *L = LI->getLoopFor(POI);
733       Loop *NL = getNearestLoop(POI, L);
734 
735       if (NL != L) {
736         // For reducible loops, NL is now an ancestor of Unloop.
737         assert((NL != &Unloop && (!NL || NL->contains(&Unloop))) &&
738                "uninitialized successor");
739         LI->changeLoopFor(POI, NL);
740       } else {
741         // Or the current block is part of a subloop, in which case its parent
742         // is unchanged.
743         assert((FoundIB || Unloop.contains(L)) && "uninitialized successor");
744       }
745     }
746   }
747   // Each irreducible loop within the unloop induces a round of iteration using
748   // the DFS result cached by Traversal.
749   bool Changed = FoundIB;
750   for (unsigned NIters = 0; Changed; ++NIters) {
751     assert(NIters < Unloop.getNumBlocks() && "runaway iterative algorithm");
752     (void)NIters;
753 
754     // Iterate over the postorder list of blocks, propagating the nearest loop
755     // from successors to predecessors as before.
756     Changed = false;
757     for (LoopBlocksDFS::POIterator POI = DFS.beginPostorder(),
758                                    POE = DFS.endPostorder();
759          POI != POE; ++POI) {
760 
761       Loop *L = LI->getLoopFor(*POI);
762       Loop *NL = getNearestLoop(*POI, L);
763       if (NL != L) {
764         assert(NL != &Unloop && (!NL || NL->contains(&Unloop)) &&
765                "uninitialized successor");
766         LI->changeLoopFor(*POI, NL);
767         Changed = true;
768       }
769     }
770   }
771 }
772 
773 /// Remove unloop's blocks from all ancestors below their new parents.
774 void UnloopUpdater::removeBlocksFromAncestors() {
775   // Remove all unloop's blocks (including those in nested subloops) from
776   // ancestors below the new parent loop.
777   for (BasicBlock *BB : Unloop.blocks()) {
778     Loop *OuterParent = LI->getLoopFor(BB);
779     if (Unloop.contains(OuterParent)) {
780       while (OuterParent->getParentLoop() != &Unloop)
781         OuterParent = OuterParent->getParentLoop();
782       OuterParent = SubloopParents[OuterParent];
783     }
784     // Remove blocks from former Ancestors except Unloop itself which will be
785     // deleted.
786     for (Loop *OldParent = Unloop.getParentLoop(); OldParent != OuterParent;
787          OldParent = OldParent->getParentLoop()) {
788       assert(OldParent && "new loop is not an ancestor of the original");
789       OldParent->removeBlockFromLoop(BB);
790     }
791   }
792 }
793 
794 /// Update the parent loop for all subloops directly nested within unloop.
795 void UnloopUpdater::updateSubloopParents() {
796   while (!Unloop.isInnermost()) {
797     Loop *Subloop = *std::prev(Unloop.end());
798     Unloop.removeChildLoop(std::prev(Unloop.end()));
799 
800     assert(SubloopParents.count(Subloop) && "DFS failed to visit subloop");
801     if (Loop *Parent = SubloopParents[Subloop])
802       Parent->addChildLoop(Subloop);
803     else
804       LI->addTopLevelLoop(Subloop);
805   }
806 }
807 
808 /// Return the nearest parent loop among this block's successors. If a successor
809 /// is a subloop header, consider its parent to be the nearest parent of the
810 /// subloop's exits.
811 ///
812 /// For subloop blocks, simply update SubloopParents and return NULL.
813 Loop *UnloopUpdater::getNearestLoop(BasicBlock *BB, Loop *BBLoop) {
814 
815   // Initially for blocks directly contained by Unloop, NearLoop == Unloop and
816   // is considered uninitialized.
817   Loop *NearLoop = BBLoop;
818 
819   Loop *Subloop = nullptr;
820   if (NearLoop != &Unloop && Unloop.contains(NearLoop)) {
821     Subloop = NearLoop;
822     // Find the subloop ancestor that is directly contained within Unloop.
823     while (Subloop->getParentLoop() != &Unloop) {
824       Subloop = Subloop->getParentLoop();
825       assert(Subloop && "subloop is not an ancestor of the original loop");
826     }
827     // Get the current nearest parent of the Subloop exits, initially Unloop.
828     NearLoop = SubloopParents.insert({Subloop, &Unloop}).first->second;
829   }
830 
831   if (succ_empty(BB)) {
832     assert(!Subloop && "subloop blocks must have a successor");
833     NearLoop = nullptr; // unloop blocks may now exit the function.
834   }
835   for (BasicBlock *Succ : successors(BB)) {
836     if (Succ == BB)
837       continue; // self loops are uninteresting
838 
839     Loop *L = LI->getLoopFor(Succ);
840     if (L == &Unloop) {
841       // This successor has not been processed. This path must lead to an
842       // irreducible backedge.
843       assert((FoundIB || !DFS.hasPostorder(Succ)) && "should have seen IB");
844       FoundIB = true;
845     }
846     if (L != &Unloop && Unloop.contains(L)) {
847       // Successor is in a subloop.
848       if (Subloop)
849         continue; // Branching within subloops. Ignore it.
850 
851       // BB branches from the original into a subloop header.
852       assert(L->getParentLoop() == &Unloop && "cannot skip into nested loops");
853 
854       // Get the current nearest parent of the Subloop's exits.
855       L = SubloopParents[L];
856       // L could be Unloop if the only exit was an irreducible backedge.
857     }
858     if (L == &Unloop) {
859       continue;
860     }
861     // Handle critical edges from Unloop into a sibling loop.
862     if (L && !L->contains(&Unloop)) {
863       L = L->getParentLoop();
864     }
865     // Remember the nearest parent loop among successors or subloop exits.
866     if (NearLoop == &Unloop || !NearLoop || NearLoop->contains(L))
867       NearLoop = L;
868   }
869   if (Subloop) {
870     SubloopParents[Subloop] = NearLoop;
871     return BBLoop;
872   }
873   return NearLoop;
874 }
875 
876 LoopInfo::LoopInfo(const DomTreeBase<BasicBlock> &DomTree) { analyze(DomTree); }
877 
878 bool LoopInfo::invalidate(Function &F, const PreservedAnalyses &PA,
879                           FunctionAnalysisManager::Invalidator &) {
880   // Check whether the analysis, all analyses on functions, or the function's
881   // CFG have been preserved.
882   auto PAC = PA.getChecker<LoopAnalysis>();
883   return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>() ||
884            PAC.preservedSet<CFGAnalyses>());
885 }
886 
887 void LoopInfo::erase(Loop *Unloop) {
888   assert(!Unloop->isInvalid() && "Loop has already been erased!");
889 
890   auto InvalidateOnExit = make_scope_exit([&]() { destroy(Unloop); });
891 
892   // First handle the special case of no parent loop to simplify the algorithm.
893   if (Unloop->isOutermost()) {
894     // Since BBLoop had no parent, Unloop blocks are no longer in a loop.
895     for (BasicBlock *BB : Unloop->blocks()) {
896       // Don't reparent blocks in subloops.
897       if (getLoopFor(BB) != Unloop)
898         continue;
899 
900       // Blocks no longer have a parent but are still referenced by Unloop until
901       // the Unloop object is deleted.
902       changeLoopFor(BB, nullptr);
903     }
904 
905     // Remove the loop from the top-level LoopInfo object.
906     for (iterator I = begin();; ++I) {
907       assert(I != end() && "Couldn't find loop");
908       if (*I == Unloop) {
909         removeLoop(I);
910         break;
911       }
912     }
913 
914     // Move all of the subloops to the top-level.
915     while (!Unloop->isInnermost())
916       addTopLevelLoop(Unloop->removeChildLoop(std::prev(Unloop->end())));
917 
918     return;
919   }
920 
921   // Update the parent loop for all blocks within the loop. Blocks within
922   // subloops will not change parents.
923   UnloopUpdater Updater(Unloop, this);
924   Updater.updateBlockParents();
925 
926   // Remove blocks from former ancestor loops.
927   Updater.removeBlocksFromAncestors();
928 
929   // Add direct subloops as children in their new parent loop.
930   Updater.updateSubloopParents();
931 
932   // Remove unloop from its parent loop.
933   Loop *ParentLoop = Unloop->getParentLoop();
934   for (Loop::iterator I = ParentLoop->begin();; ++I) {
935     assert(I != ParentLoop->end() && "Couldn't find loop");
936     if (*I == Unloop) {
937       ParentLoop->removeChildLoop(I);
938       break;
939     }
940   }
941 }
942 
943 bool LoopInfo::wouldBeOutOfLoopUseRequiringLCSSA(
944     const Value *V, const BasicBlock *ExitBB) const {
945   if (V->getType()->isTokenTy())
946     // We can't form PHIs of token type, so the definition of LCSSA excludes
947     // values of that type.
948     return false;
949 
950   const Instruction *I = dyn_cast<Instruction>(V);
951   if (!I)
952     return false;
953   const Loop *L = getLoopFor(I->getParent());
954   if (!L)
955     return false;
956   if (L->contains(ExitBB))
957     // Could be an exit bb of a subloop and contained in defining loop
958     return false;
959 
960   // We found a (new) out-of-loop use location, for a value defined in-loop.
961   // (Note that because of LCSSA, we don't have to account for values defined
962   // in sibling loops.  Such values will have LCSSA phis of their own in the
963   // common parent loop.)
964   return true;
965 }
966 
967 AnalysisKey LoopAnalysis::Key;
968 
969 LoopInfo LoopAnalysis::run(Function &F, FunctionAnalysisManager &AM) {
970   // FIXME: Currently we create a LoopInfo from scratch for every function.
971   // This may prove to be too wasteful due to deallocating and re-allocating
972   // memory each time for the underlying map and vector datastructures. At some
973   // point it may prove worthwhile to use a freelist and recycle LoopInfo
974   // objects. I don't want to add that kind of complexity until the scope of
975   // the problem is better understood.
976   LoopInfo LI;
977   LI.analyze(AM.getResult<DominatorTreeAnalysis>(F));
978   return LI;
979 }
980 
981 PreservedAnalyses LoopPrinterPass::run(Function &F,
982                                        FunctionAnalysisManager &AM) {
983   auto &LI = AM.getResult<LoopAnalysis>(F);
984   OS << "Loop info for function '" << F.getName() << "':\n";
985   LI.print(OS);
986   return PreservedAnalyses::all();
987 }
988 
989 void llvm::printLoop(Loop &L, raw_ostream &OS, const std::string &Banner) {
990 
991   if (forcePrintModuleIR()) {
992     // handling -print-module-scope
993     OS << Banner << " (loop: ";
994     L.getHeader()->printAsOperand(OS, false);
995     OS << ")\n";
996 
997     // printing whole module
998     OS << *L.getHeader()->getModule();
999     return;
1000   }
1001 
1002   OS << Banner;
1003 
1004   auto *PreHeader = L.getLoopPreheader();
1005   if (PreHeader) {
1006     OS << "\n; Preheader:";
1007     PreHeader->print(OS);
1008     OS << "\n; Loop:";
1009   }
1010 
1011   for (auto *Block : L.blocks())
1012     if (Block)
1013       Block->print(OS);
1014     else
1015       OS << "Printing <null> block";
1016 
1017   SmallVector<BasicBlock *, 8> ExitBlocks;
1018   L.getExitBlocks(ExitBlocks);
1019   if (!ExitBlocks.empty()) {
1020     OS << "\n; Exit blocks";
1021     for (auto *Block : ExitBlocks)
1022       if (Block)
1023         Block->print(OS);
1024       else
1025         OS << "Printing <null> block";
1026   }
1027 }
1028 
1029 MDNode *llvm::findOptionMDForLoopID(MDNode *LoopID, StringRef Name) {
1030   // No loop metadata node, no loop properties.
1031   if (!LoopID)
1032     return nullptr;
1033 
1034   // First operand should refer to the metadata node itself, for legacy reasons.
1035   assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1036   assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1037 
1038   // Iterate over the metdata node operands and look for MDString metadata.
1039   for (const MDOperand &MDO : llvm::drop_begin(LoopID->operands())) {
1040     MDNode *MD = dyn_cast<MDNode>(MDO);
1041     if (!MD || MD->getNumOperands() < 1)
1042       continue;
1043     MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1044     if (!S)
1045       continue;
1046     // Return the operand node if MDString holds expected metadata.
1047     if (Name == S->getString())
1048       return MD;
1049   }
1050 
1051   // Loop property not found.
1052   return nullptr;
1053 }
1054 
1055 MDNode *llvm::findOptionMDForLoop(const Loop *TheLoop, StringRef Name) {
1056   return findOptionMDForLoopID(TheLoop->getLoopID(), Name);
1057 }
1058 
1059 /// Find string metadata for loop
1060 ///
1061 /// If it has a value (e.g. {"llvm.distribute", 1} return the value as an
1062 /// operand or null otherwise.  If the string metadata is not found return
1063 /// Optional's not-a-value.
1064 std::optional<const MDOperand *>
1065 llvm::findStringMetadataForLoop(const Loop *TheLoop, StringRef Name) {
1066   MDNode *MD = findOptionMDForLoop(TheLoop, Name);
1067   if (!MD)
1068     return std::nullopt;
1069   switch (MD->getNumOperands()) {
1070   case 1:
1071     return nullptr;
1072   case 2:
1073     return &MD->getOperand(1);
1074   default:
1075     llvm_unreachable("loop metadata has 0 or 1 operand");
1076   }
1077 }
1078 
1079 std::optional<bool> llvm::getOptionalBoolLoopAttribute(const Loop *TheLoop,
1080                                                        StringRef Name) {
1081   MDNode *MD = findOptionMDForLoop(TheLoop, Name);
1082   if (!MD)
1083     return std::nullopt;
1084   switch (MD->getNumOperands()) {
1085   case 1:
1086     // When the value is absent it is interpreted as 'attribute set'.
1087     return true;
1088   case 2:
1089     if (ConstantInt *IntMD =
1090             mdconst::extract_or_null<ConstantInt>(MD->getOperand(1).get()))
1091       return IntMD->getZExtValue();
1092     return true;
1093   }
1094   llvm_unreachable("unexpected number of options");
1095 }
1096 
1097 bool llvm::getBooleanLoopAttribute(const Loop *TheLoop, StringRef Name) {
1098   return getOptionalBoolLoopAttribute(TheLoop, Name).value_or(false);
1099 }
1100 
1101 std::optional<int> llvm::getOptionalIntLoopAttribute(const Loop *TheLoop,
1102                                                      StringRef Name) {
1103   const MDOperand *AttrMD =
1104       findStringMetadataForLoop(TheLoop, Name).value_or(nullptr);
1105   if (!AttrMD)
1106     return std::nullopt;
1107 
1108   ConstantInt *IntMD = mdconst::extract_or_null<ConstantInt>(AttrMD->get());
1109   if (!IntMD)
1110     return std::nullopt;
1111 
1112   return IntMD->getSExtValue();
1113 }
1114 
1115 int llvm::getIntLoopAttribute(const Loop *TheLoop, StringRef Name,
1116                               int Default) {
1117   return getOptionalIntLoopAttribute(TheLoop, Name).value_or(Default);
1118 }
1119 
1120 CallBase *llvm::getLoopConvergenceHeart(const Loop *TheLoop) {
1121   BasicBlock *H = TheLoop->getHeader();
1122   for (Instruction &II : *H) {
1123     if (auto *CB = dyn_cast<CallBase>(&II)) {
1124       if (!CB->isConvergent())
1125         continue;
1126       // This is the heart if it uses a token defined outside the loop. The
1127       // verifier has already checked that only the loop intrinsic can use such
1128       // a token.
1129       if (auto *Token = CB->getConvergenceControlToken()) {
1130         auto *TokenDef = cast<Instruction>(Token);
1131         if (!TheLoop->contains(TokenDef->getParent()))
1132           return CB;
1133       }
1134       return nullptr;
1135     }
1136   }
1137   return nullptr;
1138 }
1139 
1140 bool llvm::isFinite(const Loop *L) {
1141   return L->getHeader()->getParent()->willReturn();
1142 }
1143 
1144 static const char *LLVMLoopMustProgress = "llvm.loop.mustprogress";
1145 
1146 bool llvm::hasMustProgress(const Loop *L) {
1147   return getBooleanLoopAttribute(L, LLVMLoopMustProgress);
1148 }
1149 
1150 bool llvm::isMustProgress(const Loop *L) {
1151   return L->getHeader()->getParent()->mustProgress() || hasMustProgress(L);
1152 }
1153 
1154 bool llvm::isValidAsAccessGroup(MDNode *Node) {
1155   return Node->getNumOperands() == 0 && Node->isDistinct();
1156 }
1157 
1158 MDNode *llvm::makePostTransformationMetadata(LLVMContext &Context,
1159                                              MDNode *OrigLoopID,
1160                                              ArrayRef<StringRef> RemovePrefixes,
1161                                              ArrayRef<MDNode *> AddAttrs) {
1162   // First remove any existing loop metadata related to this transformation.
1163   SmallVector<Metadata *, 4> MDs;
1164 
1165   // Reserve first location for self reference to the LoopID metadata node.
1166   MDs.push_back(nullptr);
1167 
1168   // Remove metadata for the transformation that has been applied or that became
1169   // outdated.
1170   if (OrigLoopID) {
1171     for (const MDOperand &MDO : llvm::drop_begin(OrigLoopID->operands())) {
1172       bool IsVectorMetadata = false;
1173       Metadata *Op = MDO;
1174       if (MDNode *MD = dyn_cast<MDNode>(Op)) {
1175         const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1176         if (S)
1177           IsVectorMetadata =
1178               llvm::any_of(RemovePrefixes, [S](StringRef Prefix) -> bool {
1179                 return S->getString().starts_with(Prefix);
1180               });
1181       }
1182       if (!IsVectorMetadata)
1183         MDs.push_back(Op);
1184     }
1185   }
1186 
1187   // Add metadata to avoid reapplying a transformation, such as
1188   // llvm.loop.unroll.disable and llvm.loop.isvectorized.
1189   MDs.append(AddAttrs.begin(), AddAttrs.end());
1190 
1191   MDNode *NewLoopID = MDNode::getDistinct(Context, MDs);
1192   // Replace the temporary node with a self-reference.
1193   NewLoopID->replaceOperandWith(0, NewLoopID);
1194   return NewLoopID;
1195 }
1196 
1197 //===----------------------------------------------------------------------===//
1198 // LoopInfo implementation
1199 //
1200 
1201 LoopInfoWrapperPass::LoopInfoWrapperPass() : FunctionPass(ID) {
1202   initializeLoopInfoWrapperPassPass(*PassRegistry::getPassRegistry());
1203 }
1204 
1205 char LoopInfoWrapperPass::ID = 0;
1206 INITIALIZE_PASS_BEGIN(LoopInfoWrapperPass, "loops", "Natural Loop Information",
1207                       true, true)
1208 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1209 INITIALIZE_PASS_END(LoopInfoWrapperPass, "loops", "Natural Loop Information",
1210                     true, true)
1211 
1212 bool LoopInfoWrapperPass::runOnFunction(Function &) {
1213   releaseMemory();
1214   LI.analyze(getAnalysis<DominatorTreeWrapperPass>().getDomTree());
1215   return false;
1216 }
1217 
1218 void LoopInfoWrapperPass::verifyAnalysis() const {
1219   // LoopInfoWrapperPass is a FunctionPass, but verifying every loop in the
1220   // function each time verifyAnalysis is called is very expensive. The
1221   // -verify-loop-info option can enable this. In order to perform some
1222   // checking by default, LoopPass has been taught to call verifyLoop manually
1223   // during loop pass sequences.
1224   if (VerifyLoopInfo) {
1225     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1226     LI.verify(DT);
1227   }
1228 }
1229 
1230 void LoopInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1231   AU.setPreservesAll();
1232   AU.addRequiredTransitive<DominatorTreeWrapperPass>();
1233 }
1234 
1235 void LoopInfoWrapperPass::print(raw_ostream &OS, const Module *) const {
1236   LI.print(OS);
1237 }
1238 
1239 PreservedAnalyses LoopVerifierPass::run(Function &F,
1240                                         FunctionAnalysisManager &AM) {
1241   LoopInfo &LI = AM.getResult<LoopAnalysis>(F);
1242   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1243   LI.verify(DT);
1244   return PreservedAnalyses::all();
1245 }
1246 
1247 //===----------------------------------------------------------------------===//
1248 // LoopBlocksDFS implementation
1249 //
1250 
1251 /// Traverse the loop blocks and store the DFS result.
1252 /// Useful for clients that just want the final DFS result and don't need to
1253 /// visit blocks during the initial traversal.
1254 void LoopBlocksDFS::perform(const LoopInfo *LI) {
1255   LoopBlocksTraversal Traversal(*this, LI);
1256   for (LoopBlocksTraversal::POTIterator POI = Traversal.begin(),
1257                                         POE = Traversal.end();
1258        POI != POE; ++POI)
1259     ;
1260 }
1261