xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/LoopInfo.cpp (revision 4d3fc8b0570b29fb0d6ee9525f104d52176ff0d4)
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/LoopInfoImpl.h"
21 #include "llvm/Analysis/LoopIterator.h"
22 #include "llvm/Analysis/LoopNestAnalysis.h"
23 #include "llvm/Analysis/MemorySSA.h"
24 #include "llvm/Analysis/MemorySSAUpdater.h"
25 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/Config/llvm-config.h"
28 #include "llvm/IR/CFG.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DebugLoc.h"
31 #include "llvm/IR/Dominators.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/LLVMContext.h"
34 #include "llvm/IR/Metadata.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/raw_ostream.h"
40 using namespace llvm;
41 
42 // Explicitly instantiate methods in LoopInfoImpl.h for IR-level Loops.
43 template class llvm::LoopBase<BasicBlock, Loop>;
44 template class llvm::LoopInfoBase<BasicBlock, Loop>;
45 
46 // Always verify loopinfo if expensive checking is enabled.
47 #ifdef EXPENSIVE_CHECKS
48 bool llvm::VerifyLoopInfo = true;
49 #else
50 bool llvm::VerifyLoopInfo = false;
51 #endif
52 static cl::opt<bool, true>
53     VerifyLoopInfoX("verify-loop-info", cl::location(VerifyLoopInfo),
54                     cl::Hidden, cl::desc("Verify loop info (time consuming)"));
55 
56 //===----------------------------------------------------------------------===//
57 // Loop implementation
58 //
59 
60 bool Loop::isLoopInvariant(const Value *V) const {
61   if (const Instruction *I = dyn_cast<Instruction>(V))
62     return !contains(I);
63   return true; // All non-instructions are loop invariant
64 }
65 
66 bool Loop::hasLoopInvariantOperands(const Instruction *I) const {
67   return all_of(I->operands(), [this](Value *V) { return isLoopInvariant(V); });
68 }
69 
70 bool Loop::makeLoopInvariant(Value *V, bool &Changed, Instruction *InsertPt,
71                              MemorySSAUpdater *MSSAU) const {
72   if (Instruction *I = dyn_cast<Instruction>(V))
73     return makeLoopInvariant(I, Changed, InsertPt, MSSAU);
74   return true; // All non-instructions are loop-invariant.
75 }
76 
77 bool Loop::makeLoopInvariant(Instruction *I, bool &Changed,
78                              Instruction *InsertPt,
79                              MemorySSAUpdater *MSSAU) const {
80   // Test if the value is already loop-invariant.
81   if (isLoopInvariant(I))
82     return true;
83   if (!isSafeToSpeculativelyExecute(I))
84     return false;
85   if (I->mayReadFromMemory())
86     return false;
87   // EH block instructions are immobile.
88   if (I->isEHPad())
89     return false;
90   // Determine the insertion point, unless one was given.
91   if (!InsertPt) {
92     BasicBlock *Preheader = getLoopPreheader();
93     // Without a preheader, hoisting is not feasible.
94     if (!Preheader)
95       return false;
96     InsertPt = Preheader->getTerminator();
97   }
98   // Don't hoist instructions with loop-variant operands.
99   for (Value *Operand : I->operands())
100     if (!makeLoopInvariant(Operand, Changed, InsertPt, MSSAU))
101       return false;
102 
103   // Hoist.
104   I->moveBefore(InsertPt);
105   if (MSSAU)
106     if (auto *MUD = MSSAU->getMemorySSA()->getMemoryAccess(I))
107       MSSAU->moveToPlace(MUD, InsertPt->getParent(),
108                          MemorySSA::BeforeTerminator);
109 
110   // There is possibility of hoisting this instruction above some arbitrary
111   // condition. Any metadata defined on it can be control dependent on this
112   // condition. Conservatively strip it here so that we don't give any wrong
113   // information to the optimizer.
114   I->dropUnknownNonDebugMetadata();
115 
116   Changed = true;
117   return true;
118 }
119 
120 bool Loop::getIncomingAndBackEdge(BasicBlock *&Incoming,
121                                   BasicBlock *&Backedge) const {
122   BasicBlock *H = getHeader();
123 
124   Incoming = nullptr;
125   Backedge = nullptr;
126   pred_iterator PI = pred_begin(H);
127   assert(PI != pred_end(H) && "Loop must have at least one backedge!");
128   Backedge = *PI++;
129   if (PI == pred_end(H))
130     return false; // dead loop
131   Incoming = *PI++;
132   if (PI != pred_end(H))
133     return false; // multiple backedges?
134 
135   if (contains(Incoming)) {
136     if (contains(Backedge))
137       return false;
138     std::swap(Incoming, Backedge);
139   } else if (!contains(Backedge))
140     return false;
141 
142   assert(Incoming && Backedge && "expected non-null incoming and backedges");
143   return true;
144 }
145 
146 PHINode *Loop::getCanonicalInductionVariable() const {
147   BasicBlock *H = getHeader();
148 
149   BasicBlock *Incoming = nullptr, *Backedge = nullptr;
150   if (!getIncomingAndBackEdge(Incoming, Backedge))
151     return nullptr;
152 
153   // Loop over all of the PHI nodes, looking for a canonical indvar.
154   for (BasicBlock::iterator I = H->begin(); isa<PHINode>(I); ++I) {
155     PHINode *PN = cast<PHINode>(I);
156     if (ConstantInt *CI =
157             dyn_cast<ConstantInt>(PN->getIncomingValueForBlock(Incoming)))
158       if (CI->isZero())
159         if (Instruction *Inc =
160                 dyn_cast<Instruction>(PN->getIncomingValueForBlock(Backedge)))
161           if (Inc->getOpcode() == Instruction::Add && Inc->getOperand(0) == PN)
162             if (ConstantInt *CI = dyn_cast<ConstantInt>(Inc->getOperand(1)))
163               if (CI->isOne())
164                 return PN;
165   }
166   return nullptr;
167 }
168 
169 /// Get the latch condition instruction.
170 ICmpInst *Loop::getLatchCmpInst() const {
171   if (BasicBlock *Latch = getLoopLatch())
172     if (BranchInst *BI = dyn_cast_or_null<BranchInst>(Latch->getTerminator()))
173       if (BI->isConditional())
174         return dyn_cast<ICmpInst>(BI->getCondition());
175 
176   return nullptr;
177 }
178 
179 /// Return the final value of the loop induction variable if found.
180 static Value *findFinalIVValue(const Loop &L, const PHINode &IndVar,
181                                const Instruction &StepInst) {
182   ICmpInst *LatchCmpInst = L.getLatchCmpInst();
183   if (!LatchCmpInst)
184     return nullptr;
185 
186   Value *Op0 = LatchCmpInst->getOperand(0);
187   Value *Op1 = LatchCmpInst->getOperand(1);
188   if (Op0 == &IndVar || Op0 == &StepInst)
189     return Op1;
190 
191   if (Op1 == &IndVar || Op1 == &StepInst)
192     return Op0;
193 
194   return nullptr;
195 }
196 
197 Optional<Loop::LoopBounds> Loop::LoopBounds::getBounds(const Loop &L,
198                                                        PHINode &IndVar,
199                                                        ScalarEvolution &SE) {
200   InductionDescriptor IndDesc;
201   if (!InductionDescriptor::isInductionPHI(&IndVar, &L, &SE, IndDesc))
202     return None;
203 
204   Value *InitialIVValue = IndDesc.getStartValue();
205   Instruction *StepInst = IndDesc.getInductionBinOp();
206   if (!InitialIVValue || !StepInst)
207     return None;
208 
209   const SCEV *Step = IndDesc.getStep();
210   Value *StepInstOp1 = StepInst->getOperand(1);
211   Value *StepInstOp0 = StepInst->getOperand(0);
212   Value *StepValue = nullptr;
213   if (SE.getSCEV(StepInstOp1) == Step)
214     StepValue = StepInstOp1;
215   else if (SE.getSCEV(StepInstOp0) == Step)
216     StepValue = StepInstOp0;
217 
218   Value *FinalIVValue = findFinalIVValue(L, IndVar, *StepInst);
219   if (!FinalIVValue)
220     return None;
221 
222   return LoopBounds(L, *InitialIVValue, *StepInst, StepValue, *FinalIVValue,
223                     SE);
224 }
225 
226 using Direction = Loop::LoopBounds::Direction;
227 
228 ICmpInst::Predicate Loop::LoopBounds::getCanonicalPredicate() const {
229   BasicBlock *Latch = L.getLoopLatch();
230   assert(Latch && "Expecting valid latch");
231 
232   BranchInst *BI = dyn_cast_or_null<BranchInst>(Latch->getTerminator());
233   assert(BI && BI->isConditional() && "Expecting conditional latch branch");
234 
235   ICmpInst *LatchCmpInst = dyn_cast<ICmpInst>(BI->getCondition());
236   assert(LatchCmpInst &&
237          "Expecting the latch compare instruction to be a CmpInst");
238 
239   // Need to inverse the predicate when first successor is not the loop
240   // header
241   ICmpInst::Predicate Pred = (BI->getSuccessor(0) == L.getHeader())
242                                  ? LatchCmpInst->getPredicate()
243                                  : LatchCmpInst->getInversePredicate();
244 
245   if (LatchCmpInst->getOperand(0) == &getFinalIVValue())
246     Pred = ICmpInst::getSwappedPredicate(Pred);
247 
248   // Need to flip strictness of the predicate when the latch compare instruction
249   // is not using StepInst
250   if (LatchCmpInst->getOperand(0) == &getStepInst() ||
251       LatchCmpInst->getOperand(1) == &getStepInst())
252     return Pred;
253 
254   // Cannot flip strictness of NE and EQ
255   if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
256     return ICmpInst::getFlippedStrictnessPredicate(Pred);
257 
258   Direction D = getDirection();
259   if (D == Direction::Increasing)
260     return ICmpInst::ICMP_SLT;
261 
262   if (D == Direction::Decreasing)
263     return ICmpInst::ICMP_SGT;
264 
265   // If cannot determine the direction, then unable to find the canonical
266   // predicate
267   return ICmpInst::BAD_ICMP_PREDICATE;
268 }
269 
270 Direction Loop::LoopBounds::getDirection() const {
271   if (const SCEVAddRecExpr *StepAddRecExpr =
272           dyn_cast<SCEVAddRecExpr>(SE.getSCEV(&getStepInst())))
273     if (const SCEV *StepRecur = StepAddRecExpr->getStepRecurrence(SE)) {
274       if (SE.isKnownPositive(StepRecur))
275         return Direction::Increasing;
276       if (SE.isKnownNegative(StepRecur))
277         return Direction::Decreasing;
278     }
279 
280   return Direction::Unknown;
281 }
282 
283 Optional<Loop::LoopBounds> Loop::getBounds(ScalarEvolution &SE) const {
284   if (PHINode *IndVar = getInductionVariable(SE))
285     return LoopBounds::getBounds(*this, *IndVar, SE);
286 
287   return None;
288 }
289 
290 PHINode *Loop::getInductionVariable(ScalarEvolution &SE) const {
291   if (!isLoopSimplifyForm())
292     return nullptr;
293 
294   BasicBlock *Header = getHeader();
295   assert(Header && "Expected a valid loop header");
296   ICmpInst *CmpInst = getLatchCmpInst();
297   if (!CmpInst)
298     return nullptr;
299 
300   Value *LatchCmpOp0 = CmpInst->getOperand(0);
301   Value *LatchCmpOp1 = CmpInst->getOperand(1);
302 
303   for (PHINode &IndVar : Header->phis()) {
304     InductionDescriptor IndDesc;
305     if (!InductionDescriptor::isInductionPHI(&IndVar, this, &SE, IndDesc))
306       continue;
307 
308     BasicBlock *Latch = getLoopLatch();
309     Value *StepInst = IndVar.getIncomingValueForBlock(Latch);
310 
311     // case 1:
312     // IndVar = phi[{InitialValue, preheader}, {StepInst, latch}]
313     // StepInst = IndVar + step
314     // cmp = StepInst < FinalValue
315     if (StepInst == LatchCmpOp0 || StepInst == LatchCmpOp1)
316       return &IndVar;
317 
318     // case 2:
319     // IndVar = phi[{InitialValue, preheader}, {StepInst, latch}]
320     // StepInst = IndVar + step
321     // cmp = IndVar < FinalValue
322     if (&IndVar == LatchCmpOp0 || &IndVar == LatchCmpOp1)
323       return &IndVar;
324   }
325 
326   return nullptr;
327 }
328 
329 bool Loop::getInductionDescriptor(ScalarEvolution &SE,
330                                   InductionDescriptor &IndDesc) const {
331   if (PHINode *IndVar = getInductionVariable(SE))
332     return InductionDescriptor::isInductionPHI(IndVar, this, &SE, IndDesc);
333 
334   return false;
335 }
336 
337 bool Loop::isAuxiliaryInductionVariable(PHINode &AuxIndVar,
338                                         ScalarEvolution &SE) const {
339   // Located in the loop header
340   BasicBlock *Header = getHeader();
341   if (AuxIndVar.getParent() != Header)
342     return false;
343 
344   // No uses outside of the loop
345   for (User *U : AuxIndVar.users())
346     if (const Instruction *I = dyn_cast<Instruction>(U))
347       if (!contains(I))
348         return false;
349 
350   InductionDescriptor IndDesc;
351   if (!InductionDescriptor::isInductionPHI(&AuxIndVar, this, &SE, IndDesc))
352     return false;
353 
354   // The step instruction opcode should be add or sub.
355   if (IndDesc.getInductionOpcode() != Instruction::Add &&
356       IndDesc.getInductionOpcode() != Instruction::Sub)
357     return false;
358 
359   // Incremented by a loop invariant step for each loop iteration
360   return SE.isLoopInvariant(IndDesc.getStep(), this);
361 }
362 
363 BranchInst *Loop::getLoopGuardBranch() const {
364   if (!isLoopSimplifyForm())
365     return nullptr;
366 
367   BasicBlock *Preheader = getLoopPreheader();
368   assert(Preheader && getLoopLatch() &&
369          "Expecting a loop with valid preheader and latch");
370 
371   // Loop should be in rotate form.
372   if (!isRotatedForm())
373     return nullptr;
374 
375   // Disallow loops with more than one unique exit block, as we do not verify
376   // that GuardOtherSucc post dominates all exit blocks.
377   BasicBlock *ExitFromLatch = getUniqueExitBlock();
378   if (!ExitFromLatch)
379     return nullptr;
380 
381   BasicBlock *GuardBB = Preheader->getUniquePredecessor();
382   if (!GuardBB)
383     return nullptr;
384 
385   assert(GuardBB->getTerminator() && "Expecting valid guard terminator");
386 
387   BranchInst *GuardBI = dyn_cast<BranchInst>(GuardBB->getTerminator());
388   if (!GuardBI || GuardBI->isUnconditional())
389     return nullptr;
390 
391   BasicBlock *GuardOtherSucc = (GuardBI->getSuccessor(0) == Preheader)
392                                    ? GuardBI->getSuccessor(1)
393                                    : GuardBI->getSuccessor(0);
394 
395   // Check if ExitFromLatch (or any BasicBlock which is an empty unique
396   // successor of ExitFromLatch) is equal to GuardOtherSucc. If
397   // skipEmptyBlockUntil returns GuardOtherSucc, then the guard branch for the
398   // loop is GuardBI (return GuardBI), otherwise return nullptr.
399   if (&LoopNest::skipEmptyBlockUntil(ExitFromLatch, GuardOtherSucc,
400                                      /*CheckUniquePred=*/true) ==
401       GuardOtherSucc)
402     return GuardBI;
403   else
404     return nullptr;
405 }
406 
407 bool Loop::isCanonical(ScalarEvolution &SE) const {
408   InductionDescriptor IndDesc;
409   if (!getInductionDescriptor(SE, IndDesc))
410     return false;
411 
412   ConstantInt *Init = dyn_cast_or_null<ConstantInt>(IndDesc.getStartValue());
413   if (!Init || !Init->isZero())
414     return false;
415 
416   if (IndDesc.getInductionOpcode() != Instruction::Add)
417     return false;
418 
419   ConstantInt *Step = IndDesc.getConstIntStepValue();
420   if (!Step || !Step->isOne())
421     return false;
422 
423   return true;
424 }
425 
426 // Check that 'BB' doesn't have any uses outside of the 'L'
427 static bool isBlockInLCSSAForm(const Loop &L, const BasicBlock &BB,
428                                const DominatorTree &DT, bool IgnoreTokens) {
429   for (const Instruction &I : BB) {
430     // Tokens can't be used in PHI nodes and live-out tokens prevent loop
431     // optimizations, so for the purposes of considered LCSSA form, we
432     // can ignore them.
433     if (IgnoreTokens && I.getType()->isTokenTy())
434       continue;
435 
436     for (const Use &U : I.uses()) {
437       const Instruction *UI = cast<Instruction>(U.getUser());
438       const BasicBlock *UserBB = UI->getParent();
439 
440       // For practical purposes, we consider that the use in a PHI
441       // occurs in the respective predecessor block. For more info,
442       // see the `phi` doc in LangRef and the LCSSA doc.
443       if (const PHINode *P = dyn_cast<PHINode>(UI))
444         UserBB = P->getIncomingBlock(U);
445 
446       // Check the current block, as a fast-path, before checking whether
447       // the use is anywhere in the loop.  Most values are used in the same
448       // block they are defined in.  Also, blocks not reachable from the
449       // entry are special; uses in them don't need to go through PHIs.
450       if (UserBB != &BB && !L.contains(UserBB) &&
451           DT.isReachableFromEntry(UserBB))
452         return false;
453     }
454   }
455   return true;
456 }
457 
458 bool Loop::isLCSSAForm(const DominatorTree &DT, bool IgnoreTokens) const {
459   // For each block we check that it doesn't have any uses outside of this loop.
460   return all_of(this->blocks(), [&](const BasicBlock *BB) {
461     return isBlockInLCSSAForm(*this, *BB, DT, IgnoreTokens);
462   });
463 }
464 
465 bool Loop::isRecursivelyLCSSAForm(const DominatorTree &DT, const LoopInfo &LI,
466                                   bool IgnoreTokens) const {
467   // For each block we check that it doesn't have any uses outside of its
468   // innermost loop. This process will transitively guarantee that the current
469   // loop and all of the nested loops are in LCSSA form.
470   return all_of(this->blocks(), [&](const BasicBlock *BB) {
471     return isBlockInLCSSAForm(*LI.getLoopFor(BB), *BB, DT, IgnoreTokens);
472   });
473 }
474 
475 bool Loop::isLoopSimplifyForm() const {
476   // Normal-form loops have a preheader, a single backedge, and all of their
477   // exits have all their predecessors inside the loop.
478   return getLoopPreheader() && getLoopLatch() && hasDedicatedExits();
479 }
480 
481 // Routines that reform the loop CFG and split edges often fail on indirectbr.
482 bool Loop::isSafeToClone() const {
483   // Return false if any loop blocks contain indirectbrs, or there are any calls
484   // to noduplicate functions.
485   for (BasicBlock *BB : this->blocks()) {
486     if (isa<IndirectBrInst>(BB->getTerminator()))
487       return false;
488 
489     for (Instruction &I : *BB)
490       if (auto *CB = dyn_cast<CallBase>(&I))
491         if (CB->cannotDuplicate())
492           return false;
493   }
494   return true;
495 }
496 
497 MDNode *Loop::getLoopID() const {
498   MDNode *LoopID = nullptr;
499 
500   // Go through the latch blocks and check the terminator for the metadata.
501   SmallVector<BasicBlock *, 4> LatchesBlocks;
502   getLoopLatches(LatchesBlocks);
503   for (BasicBlock *BB : LatchesBlocks) {
504     Instruction *TI = BB->getTerminator();
505     MDNode *MD = TI->getMetadata(LLVMContext::MD_loop);
506 
507     if (!MD)
508       return nullptr;
509 
510     if (!LoopID)
511       LoopID = MD;
512     else if (MD != LoopID)
513       return nullptr;
514   }
515   if (!LoopID || LoopID->getNumOperands() == 0 ||
516       LoopID->getOperand(0) != LoopID)
517     return nullptr;
518   return LoopID;
519 }
520 
521 void Loop::setLoopID(MDNode *LoopID) const {
522   assert((!LoopID || LoopID->getNumOperands() > 0) &&
523          "Loop ID needs at least one operand");
524   assert((!LoopID || LoopID->getOperand(0) == LoopID) &&
525          "Loop ID should refer to itself");
526 
527   SmallVector<BasicBlock *, 4> LoopLatches;
528   getLoopLatches(LoopLatches);
529   for (BasicBlock *BB : LoopLatches)
530     BB->getTerminator()->setMetadata(LLVMContext::MD_loop, LoopID);
531 }
532 
533 void Loop::setLoopAlreadyUnrolled() {
534   LLVMContext &Context = getHeader()->getContext();
535 
536   MDNode *DisableUnrollMD =
537       MDNode::get(Context, MDString::get(Context, "llvm.loop.unroll.disable"));
538   MDNode *LoopID = getLoopID();
539   MDNode *NewLoopID = makePostTransformationMetadata(
540       Context, LoopID, {"llvm.loop.unroll."}, {DisableUnrollMD});
541   setLoopID(NewLoopID);
542 }
543 
544 void Loop::setLoopMustProgress() {
545   LLVMContext &Context = getHeader()->getContext();
546 
547   MDNode *MustProgress = findOptionMDForLoop(this, "llvm.loop.mustprogress");
548 
549   if (MustProgress)
550     return;
551 
552   MDNode *MustProgressMD =
553       MDNode::get(Context, MDString::get(Context, "llvm.loop.mustprogress"));
554   MDNode *LoopID = getLoopID();
555   MDNode *NewLoopID =
556       makePostTransformationMetadata(Context, LoopID, {}, {MustProgressMD});
557   setLoopID(NewLoopID);
558 }
559 
560 bool Loop::isAnnotatedParallel() const {
561   MDNode *DesiredLoopIdMetadata = getLoopID();
562 
563   if (!DesiredLoopIdMetadata)
564     return false;
565 
566   MDNode *ParallelAccesses =
567       findOptionMDForLoop(this, "llvm.loop.parallel_accesses");
568   SmallPtrSet<MDNode *, 4>
569       ParallelAccessGroups; // For scalable 'contains' check.
570   if (ParallelAccesses) {
571     for (const MDOperand &MD : drop_begin(ParallelAccesses->operands())) {
572       MDNode *AccGroup = cast<MDNode>(MD.get());
573       assert(isValidAsAccessGroup(AccGroup) &&
574              "List item must be an access group");
575       ParallelAccessGroups.insert(AccGroup);
576     }
577   }
578 
579   // The loop branch contains the parallel loop metadata. In order to ensure
580   // that any parallel-loop-unaware optimization pass hasn't added loop-carried
581   // dependencies (thus converted the loop back to a sequential loop), check
582   // that all the memory instructions in the loop belong to an access group that
583   // is parallel to this loop.
584   for (BasicBlock *BB : this->blocks()) {
585     for (Instruction &I : *BB) {
586       if (!I.mayReadOrWriteMemory())
587         continue;
588 
589       if (MDNode *AccessGroup = I.getMetadata(LLVMContext::MD_access_group)) {
590         auto ContainsAccessGroup = [&ParallelAccessGroups](MDNode *AG) -> bool {
591           if (AG->getNumOperands() == 0) {
592             assert(isValidAsAccessGroup(AG) && "Item must be an access group");
593             return ParallelAccessGroups.count(AG);
594           }
595 
596           for (const MDOperand &AccessListItem : AG->operands()) {
597             MDNode *AccGroup = cast<MDNode>(AccessListItem.get());
598             assert(isValidAsAccessGroup(AccGroup) &&
599                    "List item must be an access group");
600             if (ParallelAccessGroups.count(AccGroup))
601               return true;
602           }
603           return false;
604         };
605 
606         if (ContainsAccessGroup(AccessGroup))
607           continue;
608       }
609 
610       // The memory instruction can refer to the loop identifier metadata
611       // directly or indirectly through another list metadata (in case of
612       // nested parallel loops). The loop identifier metadata refers to
613       // itself so we can check both cases with the same routine.
614       MDNode *LoopIdMD =
615           I.getMetadata(LLVMContext::MD_mem_parallel_loop_access);
616 
617       if (!LoopIdMD)
618         return false;
619 
620       if (!llvm::is_contained(LoopIdMD->operands(), DesiredLoopIdMetadata))
621         return false;
622     }
623   }
624   return true;
625 }
626 
627 DebugLoc Loop::getStartLoc() const { return getLocRange().getStart(); }
628 
629 Loop::LocRange Loop::getLocRange() const {
630   // If we have a debug location in the loop ID, then use it.
631   if (MDNode *LoopID = getLoopID()) {
632     DebugLoc Start;
633     // We use the first DebugLoc in the header as the start location of the loop
634     // and if there is a second DebugLoc in the header we use it as end location
635     // of the loop.
636     for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
637       if (DILocation *L = dyn_cast<DILocation>(LoopID->getOperand(i))) {
638         if (!Start)
639           Start = DebugLoc(L);
640         else
641           return LocRange(Start, DebugLoc(L));
642       }
643     }
644 
645     if (Start)
646       return LocRange(Start);
647   }
648 
649   // Try the pre-header first.
650   if (BasicBlock *PHeadBB = getLoopPreheader())
651     if (DebugLoc DL = PHeadBB->getTerminator()->getDebugLoc())
652       return LocRange(DL);
653 
654   // If we have no pre-header or there are no instructions with debug
655   // info in it, try the header.
656   if (BasicBlock *HeadBB = getHeader())
657     return LocRange(HeadBB->getTerminator()->getDebugLoc());
658 
659   return LocRange();
660 }
661 
662 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
663 LLVM_DUMP_METHOD void Loop::dump() const { print(dbgs()); }
664 
665 LLVM_DUMP_METHOD void Loop::dumpVerbose() const {
666   print(dbgs(), /*Verbose=*/true);
667 }
668 #endif
669 
670 //===----------------------------------------------------------------------===//
671 // UnloopUpdater implementation
672 //
673 
674 namespace {
675 /// Find the new parent loop for all blocks within the "unloop" whose last
676 /// backedges has just been removed.
677 class UnloopUpdater {
678   Loop &Unloop;
679   LoopInfo *LI;
680 
681   LoopBlocksDFS DFS;
682 
683   // Map unloop's immediate subloops to their nearest reachable parents. Nested
684   // loops within these subloops will not change parents. However, an immediate
685   // subloop's new parent will be the nearest loop reachable from either its own
686   // exits *or* any of its nested loop's exits.
687   DenseMap<Loop *, Loop *> SubloopParents;
688 
689   // Flag the presence of an irreducible backedge whose destination is a block
690   // directly contained by the original unloop.
691   bool FoundIB = false;
692 
693 public:
694   UnloopUpdater(Loop *UL, LoopInfo *LInfo) : Unloop(*UL), LI(LInfo), DFS(UL) {}
695 
696   void updateBlockParents();
697 
698   void removeBlocksFromAncestors();
699 
700   void updateSubloopParents();
701 
702 protected:
703   Loop *getNearestLoop(BasicBlock *BB, Loop *BBLoop);
704 };
705 } // end anonymous namespace
706 
707 /// Update the parent loop for all blocks that are directly contained within the
708 /// original "unloop".
709 void UnloopUpdater::updateBlockParents() {
710   if (Unloop.getNumBlocks()) {
711     // Perform a post order CFG traversal of all blocks within this loop,
712     // propagating the nearest loop from successors to predecessors.
713     LoopBlocksTraversal Traversal(DFS, LI);
714     for (BasicBlock *POI : Traversal) {
715 
716       Loop *L = LI->getLoopFor(POI);
717       Loop *NL = getNearestLoop(POI, L);
718 
719       if (NL != L) {
720         // For reducible loops, NL is now an ancestor of Unloop.
721         assert((NL != &Unloop && (!NL || NL->contains(&Unloop))) &&
722                "uninitialized successor");
723         LI->changeLoopFor(POI, NL);
724       } else {
725         // Or the current block is part of a subloop, in which case its parent
726         // is unchanged.
727         assert((FoundIB || Unloop.contains(L)) && "uninitialized successor");
728       }
729     }
730   }
731   // Each irreducible loop within the unloop induces a round of iteration using
732   // the DFS result cached by Traversal.
733   bool Changed = FoundIB;
734   for (unsigned NIters = 0; Changed; ++NIters) {
735     assert(NIters < Unloop.getNumBlocks() && "runaway iterative algorithm");
736     (void) NIters;
737 
738     // Iterate over the postorder list of blocks, propagating the nearest loop
739     // from successors to predecessors as before.
740     Changed = false;
741     for (LoopBlocksDFS::POIterator POI = DFS.beginPostorder(),
742                                    POE = DFS.endPostorder();
743          POI != POE; ++POI) {
744 
745       Loop *L = LI->getLoopFor(*POI);
746       Loop *NL = getNearestLoop(*POI, L);
747       if (NL != L) {
748         assert(NL != &Unloop && (!NL || NL->contains(&Unloop)) &&
749                "uninitialized successor");
750         LI->changeLoopFor(*POI, NL);
751         Changed = true;
752       }
753     }
754   }
755 }
756 
757 /// Remove unloop's blocks from all ancestors below their new parents.
758 void UnloopUpdater::removeBlocksFromAncestors() {
759   // Remove all unloop's blocks (including those in nested subloops) from
760   // ancestors below the new parent loop.
761   for (BasicBlock *BB : Unloop.blocks()) {
762     Loop *OuterParent = LI->getLoopFor(BB);
763     if (Unloop.contains(OuterParent)) {
764       while (OuterParent->getParentLoop() != &Unloop)
765         OuterParent = OuterParent->getParentLoop();
766       OuterParent = SubloopParents[OuterParent];
767     }
768     // Remove blocks from former Ancestors except Unloop itself which will be
769     // deleted.
770     for (Loop *OldParent = Unloop.getParentLoop(); OldParent != OuterParent;
771          OldParent = OldParent->getParentLoop()) {
772       assert(OldParent && "new loop is not an ancestor of the original");
773       OldParent->removeBlockFromLoop(BB);
774     }
775   }
776 }
777 
778 /// Update the parent loop for all subloops directly nested within unloop.
779 void UnloopUpdater::updateSubloopParents() {
780   while (!Unloop.isInnermost()) {
781     Loop *Subloop = *std::prev(Unloop.end());
782     Unloop.removeChildLoop(std::prev(Unloop.end()));
783 
784     assert(SubloopParents.count(Subloop) && "DFS failed to visit subloop");
785     if (Loop *Parent = SubloopParents[Subloop])
786       Parent->addChildLoop(Subloop);
787     else
788       LI->addTopLevelLoop(Subloop);
789   }
790 }
791 
792 /// Return the nearest parent loop among this block's successors. If a successor
793 /// is a subloop header, consider its parent to be the nearest parent of the
794 /// subloop's exits.
795 ///
796 /// For subloop blocks, simply update SubloopParents and return NULL.
797 Loop *UnloopUpdater::getNearestLoop(BasicBlock *BB, Loop *BBLoop) {
798 
799   // Initially for blocks directly contained by Unloop, NearLoop == Unloop and
800   // is considered uninitialized.
801   Loop *NearLoop = BBLoop;
802 
803   Loop *Subloop = nullptr;
804   if (NearLoop != &Unloop && Unloop.contains(NearLoop)) {
805     Subloop = NearLoop;
806     // Find the subloop ancestor that is directly contained within Unloop.
807     while (Subloop->getParentLoop() != &Unloop) {
808       Subloop = Subloop->getParentLoop();
809       assert(Subloop && "subloop is not an ancestor of the original loop");
810     }
811     // Get the current nearest parent of the Subloop exits, initially Unloop.
812     NearLoop = SubloopParents.insert({Subloop, &Unloop}).first->second;
813   }
814 
815   succ_iterator I = succ_begin(BB), E = succ_end(BB);
816   if (I == E) {
817     assert(!Subloop && "subloop blocks must have a successor");
818     NearLoop = nullptr; // unloop blocks may now exit the function.
819   }
820   for (; I != E; ++I) {
821     if (*I == BB)
822       continue; // self loops are uninteresting
823 
824     Loop *L = LI->getLoopFor(*I);
825     if (L == &Unloop) {
826       // This successor has not been processed. This path must lead to an
827       // irreducible backedge.
828       assert((FoundIB || !DFS.hasPostorder(*I)) && "should have seen IB");
829       FoundIB = true;
830     }
831     if (L != &Unloop && Unloop.contains(L)) {
832       // Successor is in a subloop.
833       if (Subloop)
834         continue; // Branching within subloops. Ignore it.
835 
836       // BB branches from the original into a subloop header.
837       assert(L->getParentLoop() == &Unloop && "cannot skip into nested loops");
838 
839       // Get the current nearest parent of the Subloop's exits.
840       L = SubloopParents[L];
841       // L could be Unloop if the only exit was an irreducible backedge.
842     }
843     if (L == &Unloop) {
844       continue;
845     }
846     // Handle critical edges from Unloop into a sibling loop.
847     if (L && !L->contains(&Unloop)) {
848       L = L->getParentLoop();
849     }
850     // Remember the nearest parent loop among successors or subloop exits.
851     if (NearLoop == &Unloop || !NearLoop || NearLoop->contains(L))
852       NearLoop = L;
853   }
854   if (Subloop) {
855     SubloopParents[Subloop] = NearLoop;
856     return BBLoop;
857   }
858   return NearLoop;
859 }
860 
861 LoopInfo::LoopInfo(const DomTreeBase<BasicBlock> &DomTree) { analyze(DomTree); }
862 
863 bool LoopInfo::invalidate(Function &F, const PreservedAnalyses &PA,
864                           FunctionAnalysisManager::Invalidator &) {
865   // Check whether the analysis, all analyses on functions, or the function's
866   // CFG have been preserved.
867   auto PAC = PA.getChecker<LoopAnalysis>();
868   return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>() ||
869            PAC.preservedSet<CFGAnalyses>());
870 }
871 
872 void LoopInfo::erase(Loop *Unloop) {
873   assert(!Unloop->isInvalid() && "Loop has already been erased!");
874 
875   auto InvalidateOnExit = make_scope_exit([&]() { destroy(Unloop); });
876 
877   // First handle the special case of no parent loop to simplify the algorithm.
878   if (Unloop->isOutermost()) {
879     // Since BBLoop had no parent, Unloop blocks are no longer in a loop.
880     for (BasicBlock *BB : Unloop->blocks()) {
881       // Don't reparent blocks in subloops.
882       if (getLoopFor(BB) != Unloop)
883         continue;
884 
885       // Blocks no longer have a parent but are still referenced by Unloop until
886       // the Unloop object is deleted.
887       changeLoopFor(BB, nullptr);
888     }
889 
890     // Remove the loop from the top-level LoopInfo object.
891     for (iterator I = begin();; ++I) {
892       assert(I != end() && "Couldn't find loop");
893       if (*I == Unloop) {
894         removeLoop(I);
895         break;
896       }
897     }
898 
899     // Move all of the subloops to the top-level.
900     while (!Unloop->isInnermost())
901       addTopLevelLoop(Unloop->removeChildLoop(std::prev(Unloop->end())));
902 
903     return;
904   }
905 
906   // Update the parent loop for all blocks within the loop. Blocks within
907   // subloops will not change parents.
908   UnloopUpdater Updater(Unloop, this);
909   Updater.updateBlockParents();
910 
911   // Remove blocks from former ancestor loops.
912   Updater.removeBlocksFromAncestors();
913 
914   // Add direct subloops as children in their new parent loop.
915   Updater.updateSubloopParents();
916 
917   // Remove unloop from its parent loop.
918   Loop *ParentLoop = Unloop->getParentLoop();
919   for (Loop::iterator I = ParentLoop->begin();; ++I) {
920     assert(I != ParentLoop->end() && "Couldn't find loop");
921     if (*I == Unloop) {
922       ParentLoop->removeChildLoop(I);
923       break;
924     }
925   }
926 }
927 
928 bool
929 LoopInfo::wouldBeOutOfLoopUseRequiringLCSSA(const Value *V,
930                                             const BasicBlock *ExitBB) const {
931   if (V->getType()->isTokenTy())
932     // We can't form PHIs of token type, so the definition of LCSSA excludes
933     // values of that type.
934     return false;
935 
936   const Instruction *I = dyn_cast<Instruction>(V);
937   if (!I)
938     return false;
939   const Loop *L = getLoopFor(I->getParent());
940   if (!L)
941     return false;
942   if (L->contains(ExitBB))
943     // Could be an exit bb of a subloop and contained in defining loop
944     return false;
945 
946   // We found a (new) out-of-loop use location, for a value defined in-loop.
947   // (Note that because of LCSSA, we don't have to account for values defined
948   // in sibling loops.  Such values will have LCSSA phis of their own in the
949   // common parent loop.)
950   return true;
951 }
952 
953 AnalysisKey LoopAnalysis::Key;
954 
955 LoopInfo LoopAnalysis::run(Function &F, FunctionAnalysisManager &AM) {
956   // FIXME: Currently we create a LoopInfo from scratch for every function.
957   // This may prove to be too wasteful due to deallocating and re-allocating
958   // memory each time for the underlying map and vector datastructures. At some
959   // point it may prove worthwhile to use a freelist and recycle LoopInfo
960   // objects. I don't want to add that kind of complexity until the scope of
961   // the problem is better understood.
962   LoopInfo LI;
963   LI.analyze(AM.getResult<DominatorTreeAnalysis>(F));
964   return LI;
965 }
966 
967 PreservedAnalyses LoopPrinterPass::run(Function &F,
968                                        FunctionAnalysisManager &AM) {
969   AM.getResult<LoopAnalysis>(F).print(OS);
970   return PreservedAnalyses::all();
971 }
972 
973 void llvm::printLoop(Loop &L, raw_ostream &OS, const std::string &Banner) {
974 
975   if (forcePrintModuleIR()) {
976     // handling -print-module-scope
977     OS << Banner << " (loop: ";
978     L.getHeader()->printAsOperand(OS, false);
979     OS << ")\n";
980 
981     // printing whole module
982     OS << *L.getHeader()->getModule();
983     return;
984   }
985 
986   OS << Banner;
987 
988   auto *PreHeader = L.getLoopPreheader();
989   if (PreHeader) {
990     OS << "\n; Preheader:";
991     PreHeader->print(OS);
992     OS << "\n; Loop:";
993   }
994 
995   for (auto *Block : L.blocks())
996     if (Block)
997       Block->print(OS);
998     else
999       OS << "Printing <null> block";
1000 
1001   SmallVector<BasicBlock *, 8> ExitBlocks;
1002   L.getExitBlocks(ExitBlocks);
1003   if (!ExitBlocks.empty()) {
1004     OS << "\n; Exit blocks";
1005     for (auto *Block : ExitBlocks)
1006       if (Block)
1007         Block->print(OS);
1008       else
1009         OS << "Printing <null> block";
1010   }
1011 }
1012 
1013 MDNode *llvm::findOptionMDForLoopID(MDNode *LoopID, StringRef Name) {
1014   // No loop metadata node, no loop properties.
1015   if (!LoopID)
1016     return nullptr;
1017 
1018   // First operand should refer to the metadata node itself, for legacy reasons.
1019   assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1020   assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1021 
1022   // Iterate over the metdata node operands and look for MDString metadata.
1023   for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
1024     MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
1025     if (!MD || MD->getNumOperands() < 1)
1026       continue;
1027     MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1028     if (!S)
1029       continue;
1030     // Return the operand node if MDString holds expected metadata.
1031     if (Name.equals(S->getString()))
1032       return MD;
1033   }
1034 
1035   // Loop property not found.
1036   return nullptr;
1037 }
1038 
1039 MDNode *llvm::findOptionMDForLoop(const Loop *TheLoop, StringRef Name) {
1040   return findOptionMDForLoopID(TheLoop->getLoopID(), Name);
1041 }
1042 
1043 /// Find string metadata for loop
1044 ///
1045 /// If it has a value (e.g. {"llvm.distribute", 1} return the value as an
1046 /// operand or null otherwise.  If the string metadata is not found return
1047 /// Optional's not-a-value.
1048 Optional<const MDOperand *> llvm::findStringMetadataForLoop(const Loop *TheLoop,
1049                                                             StringRef Name) {
1050   MDNode *MD = findOptionMDForLoop(TheLoop, Name);
1051   if (!MD)
1052     return None;
1053   switch (MD->getNumOperands()) {
1054   case 1:
1055     return nullptr;
1056   case 2:
1057     return &MD->getOperand(1);
1058   default:
1059     llvm_unreachable("loop metadata has 0 or 1 operand");
1060   }
1061 }
1062 
1063 Optional<bool> llvm::getOptionalBoolLoopAttribute(const Loop *TheLoop,
1064                                                   StringRef Name) {
1065   MDNode *MD = findOptionMDForLoop(TheLoop, Name);
1066   if (!MD)
1067     return None;
1068   switch (MD->getNumOperands()) {
1069   case 1:
1070     // When the value is absent it is interpreted as 'attribute set'.
1071     return true;
1072   case 2:
1073     if (ConstantInt *IntMD =
1074             mdconst::extract_or_null<ConstantInt>(MD->getOperand(1).get()))
1075       return IntMD->getZExtValue();
1076     return true;
1077   }
1078   llvm_unreachable("unexpected number of options");
1079 }
1080 
1081 bool llvm::getBooleanLoopAttribute(const Loop *TheLoop, StringRef Name) {
1082   return getOptionalBoolLoopAttribute(TheLoop, Name).value_or(false);
1083 }
1084 
1085 llvm::Optional<int> llvm::getOptionalIntLoopAttribute(const Loop *TheLoop,
1086                                                       StringRef Name) {
1087   const MDOperand *AttrMD =
1088       findStringMetadataForLoop(TheLoop, Name).value_or(nullptr);
1089   if (!AttrMD)
1090     return None;
1091 
1092   ConstantInt *IntMD = mdconst::extract_or_null<ConstantInt>(AttrMD->get());
1093   if (!IntMD)
1094     return None;
1095 
1096   return IntMD->getSExtValue();
1097 }
1098 
1099 int llvm::getIntLoopAttribute(const Loop *TheLoop, StringRef Name,
1100                               int Default) {
1101   return getOptionalIntLoopAttribute(TheLoop, Name).value_or(Default);
1102 }
1103 
1104 bool llvm::isFinite(const Loop *L) {
1105   return L->getHeader()->getParent()->willReturn();
1106 }
1107 
1108 static const char *LLVMLoopMustProgress = "llvm.loop.mustprogress";
1109 
1110 bool llvm::hasMustProgress(const Loop *L) {
1111   return getBooleanLoopAttribute(L, LLVMLoopMustProgress);
1112 }
1113 
1114 bool llvm::isMustProgress(const Loop *L) {
1115   return L->getHeader()->getParent()->mustProgress() || hasMustProgress(L);
1116 }
1117 
1118 bool llvm::isValidAsAccessGroup(MDNode *Node) {
1119   return Node->getNumOperands() == 0 && Node->isDistinct();
1120 }
1121 
1122 MDNode *llvm::makePostTransformationMetadata(LLVMContext &Context,
1123                                              MDNode *OrigLoopID,
1124                                              ArrayRef<StringRef> RemovePrefixes,
1125                                              ArrayRef<MDNode *> AddAttrs) {
1126   // First remove any existing loop metadata related to this transformation.
1127   SmallVector<Metadata *, 4> MDs;
1128 
1129   // Reserve first location for self reference to the LoopID metadata node.
1130   MDs.push_back(nullptr);
1131 
1132   // Remove metadata for the transformation that has been applied or that became
1133   // outdated.
1134   if (OrigLoopID) {
1135     for (unsigned i = 1, ie = OrigLoopID->getNumOperands(); i < ie; ++i) {
1136       bool IsVectorMetadata = false;
1137       Metadata *Op = OrigLoopID->getOperand(i);
1138       if (MDNode *MD = dyn_cast<MDNode>(Op)) {
1139         const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
1140         if (S)
1141           IsVectorMetadata =
1142               llvm::any_of(RemovePrefixes, [S](StringRef Prefix) -> bool {
1143                 return S->getString().startswith(Prefix);
1144               });
1145       }
1146       if (!IsVectorMetadata)
1147         MDs.push_back(Op);
1148     }
1149   }
1150 
1151   // Add metadata to avoid reapplying a transformation, such as
1152   // llvm.loop.unroll.disable and llvm.loop.isvectorized.
1153   MDs.append(AddAttrs.begin(), AddAttrs.end());
1154 
1155   MDNode *NewLoopID = MDNode::getDistinct(Context, MDs);
1156   // Replace the temporary node with a self-reference.
1157   NewLoopID->replaceOperandWith(0, NewLoopID);
1158   return NewLoopID;
1159 }
1160 
1161 //===----------------------------------------------------------------------===//
1162 // LoopInfo implementation
1163 //
1164 
1165 LoopInfoWrapperPass::LoopInfoWrapperPass() : FunctionPass(ID) {
1166   initializeLoopInfoWrapperPassPass(*PassRegistry::getPassRegistry());
1167 }
1168 
1169 char LoopInfoWrapperPass::ID = 0;
1170 INITIALIZE_PASS_BEGIN(LoopInfoWrapperPass, "loops", "Natural Loop Information",
1171                       true, true)
1172 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1173 INITIALIZE_PASS_END(LoopInfoWrapperPass, "loops", "Natural Loop Information",
1174                     true, true)
1175 
1176 bool LoopInfoWrapperPass::runOnFunction(Function &) {
1177   releaseMemory();
1178   LI.analyze(getAnalysis<DominatorTreeWrapperPass>().getDomTree());
1179   return false;
1180 }
1181 
1182 void LoopInfoWrapperPass::verifyAnalysis() const {
1183   // LoopInfoWrapperPass is a FunctionPass, but verifying every loop in the
1184   // function each time verifyAnalysis is called is very expensive. The
1185   // -verify-loop-info option can enable this. In order to perform some
1186   // checking by default, LoopPass has been taught to call verifyLoop manually
1187   // during loop pass sequences.
1188   if (VerifyLoopInfo) {
1189     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1190     LI.verify(DT);
1191   }
1192 }
1193 
1194 void LoopInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1195   AU.setPreservesAll();
1196   AU.addRequiredTransitive<DominatorTreeWrapperPass>();
1197 }
1198 
1199 void LoopInfoWrapperPass::print(raw_ostream &OS, const Module *) const {
1200   LI.print(OS);
1201 }
1202 
1203 PreservedAnalyses LoopVerifierPass::run(Function &F,
1204                                         FunctionAnalysisManager &AM) {
1205   LoopInfo &LI = AM.getResult<LoopAnalysis>(F);
1206   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1207   LI.verify(DT);
1208   return PreservedAnalyses::all();
1209 }
1210 
1211 //===----------------------------------------------------------------------===//
1212 // LoopBlocksDFS implementation
1213 //
1214 
1215 /// Traverse the loop blocks and store the DFS result.
1216 /// Useful for clients that just want the final DFS result and don't need to
1217 /// visit blocks during the initial traversal.
1218 void LoopBlocksDFS::perform(LoopInfo *LI) {
1219   LoopBlocksTraversal Traversal(*this, LI);
1220   for (LoopBlocksTraversal::POTIterator POI = Traversal.begin(),
1221                                         POE = Traversal.end();
1222        POI != POE; ++POI)
1223     ;
1224 }
1225