xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/GuardWidening.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===- GuardWidening.cpp - ---- Guard widening ----------------------------===//
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 implements the guard widening pass.  The semantics of the
10 // @llvm.experimental.guard intrinsic lets LLVM transform it so that it fails
11 // more often that it did before the transform.  This optimization is called
12 // "widening" and can be used hoist and common runtime checks in situations like
13 // these:
14 //
15 //    %cmp0 = 7 u< Length
16 //    call @llvm.experimental.guard(i1 %cmp0) [ "deopt"(...) ]
17 //    call @unknown_side_effects()
18 //    %cmp1 = 9 u< Length
19 //    call @llvm.experimental.guard(i1 %cmp1) [ "deopt"(...) ]
20 //    ...
21 //
22 // =>
23 //
24 //    %cmp0 = 9 u< Length
25 //    call @llvm.experimental.guard(i1 %cmp0) [ "deopt"(...) ]
26 //    call @unknown_side_effects()
27 //    ...
28 //
29 // If %cmp0 is false, @llvm.experimental.guard will "deoptimize" back to a
30 // generic implementation of the same function, which will have the correct
31 // semantics from that point onward.  It is always _legal_ to deoptimize (so
32 // replacing %cmp0 with false is "correct"), though it may not always be
33 // profitable to do so.
34 //
35 // NB! This pass is a work in progress.  It hasn't been tuned to be "production
36 // ready" yet.  It is known to have quadriatic running time and will not scale
37 // to large numbers of guards
38 //
39 //===----------------------------------------------------------------------===//
40 
41 #include "llvm/Transforms/Scalar/GuardWidening.h"
42 #include "llvm/ADT/DenseMap.h"
43 #include "llvm/ADT/DepthFirstIterator.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/Analysis/AssumptionCache.h"
46 #include "llvm/Analysis/GuardUtils.h"
47 #include "llvm/Analysis/LoopInfo.h"
48 #include "llvm/Analysis/MemorySSAUpdater.h"
49 #include "llvm/Analysis/PostDominators.h"
50 #include "llvm/Analysis/ValueTracking.h"
51 #include "llvm/IR/ConstantRange.h"
52 #include "llvm/IR/Dominators.h"
53 #include "llvm/IR/IRBuilder.h"
54 #include "llvm/IR/IntrinsicInst.h"
55 #include "llvm/IR/Module.h"
56 #include "llvm/IR/PatternMatch.h"
57 #include "llvm/Support/CommandLine.h"
58 #include "llvm/Support/Debug.h"
59 #include "llvm/Support/KnownBits.h"
60 #include "llvm/Transforms/Scalar.h"
61 #include "llvm/Transforms/Utils/GuardUtils.h"
62 #include "llvm/Transforms/Utils/LoopUtils.h"
63 #include <functional>
64 
65 using namespace llvm;
66 
67 #define DEBUG_TYPE "guard-widening"
68 
69 STATISTIC(GuardsEliminated, "Number of eliminated guards");
70 STATISTIC(CondBranchEliminated, "Number of eliminated conditional branches");
71 STATISTIC(FreezeAdded, "Number of freeze instruction introduced");
72 
73 static cl::opt<bool>
74     WidenBranchGuards("guard-widening-widen-branch-guards", cl::Hidden,
75                       cl::desc("Whether or not we should widen guards  "
76                                "expressed as branches by widenable conditions"),
77                       cl::init(true));
78 
79 namespace {
80 
81 // Get the condition of \p I. It can either be a guard or a conditional branch.
getCondition(Instruction * I)82 static Value *getCondition(Instruction *I) {
83   if (IntrinsicInst *GI = dyn_cast<IntrinsicInst>(I)) {
84     assert(GI->getIntrinsicID() == Intrinsic::experimental_guard &&
85            "Bad guard intrinsic?");
86     return GI->getArgOperand(0);
87   }
88   Value *Cond, *WC;
89   BasicBlock *IfTrueBB, *IfFalseBB;
90   if (parseWidenableBranch(I, Cond, WC, IfTrueBB, IfFalseBB))
91     return Cond;
92 
93   return cast<BranchInst>(I)->getCondition();
94 }
95 
96 // Set the condition for \p I to \p NewCond. \p I can either be a guard or a
97 // conditional branch.
setCondition(Instruction * I,Value * NewCond)98 static void setCondition(Instruction *I, Value *NewCond) {
99   if (IntrinsicInst *GI = dyn_cast<IntrinsicInst>(I)) {
100     assert(GI->getIntrinsicID() == Intrinsic::experimental_guard &&
101            "Bad guard intrinsic?");
102     GI->setArgOperand(0, NewCond);
103     return;
104   }
105   cast<BranchInst>(I)->setCondition(NewCond);
106 }
107 
108 // Eliminates the guard instruction properly.
eliminateGuard(Instruction * GuardInst,MemorySSAUpdater * MSSAU)109 static void eliminateGuard(Instruction *GuardInst, MemorySSAUpdater *MSSAU) {
110   GuardInst->eraseFromParent();
111   if (MSSAU)
112     MSSAU->removeMemoryAccess(GuardInst);
113   ++GuardsEliminated;
114 }
115 
116 /// Find a point at which the widened condition of \p Guard should be inserted.
117 /// When it is represented as intrinsic call, we can do it right before the call
118 /// instruction. However, when we are dealing with widenable branch, we must
119 /// account for the following situation: widening should not turn a
120 /// loop-invariant condition into a loop-variant. It means that if
121 /// widenable.condition() call is invariant (w.r.t. any loop), the new wide
122 /// condition should stay invariant. Otherwise there can be a miscompile, like
123 /// the one described at https://github.com/llvm/llvm-project/issues/60234. The
124 /// safest way to do it is to expand the new condition at WC's block.
125 static std::optional<BasicBlock::iterator>
findInsertionPointForWideCondition(Instruction * WCOrGuard)126 findInsertionPointForWideCondition(Instruction *WCOrGuard) {
127   if (isGuard(WCOrGuard))
128     return WCOrGuard->getIterator();
129   if (auto WC = extractWidenableCondition(WCOrGuard))
130     return cast<Instruction>(WC)->getIterator();
131   return std::nullopt;
132 }
133 
134 class GuardWideningImpl {
135   DominatorTree &DT;
136   PostDominatorTree *PDT;
137   LoopInfo &LI;
138   AssumptionCache &AC;
139   MemorySSAUpdater *MSSAU;
140 
141   /// Together, these describe the region of interest.  This might be all of
142   /// the blocks within a function, or only a given loop's blocks and preheader.
143   DomTreeNode *Root;
144   std::function<bool(BasicBlock*)> BlockFilter;
145 
146   /// The set of guards and conditional branches whose conditions have been
147   /// widened into dominating guards.
148   SmallVector<Instruction *, 16> EliminatedGuardsAndBranches;
149 
150   /// The set of guards which have been widened to include conditions to other
151   /// guards.
152   DenseSet<Instruction *> WidenedGuards;
153 
154   /// Try to eliminate instruction \p Instr by widening it into an earlier
155   /// dominating guard.  \p DFSI is the DFS iterator on the dominator tree that
156   /// is currently visiting the block containing \p Guard, and \p GuardsPerBlock
157   /// maps BasicBlocks to the set of guards seen in that block.
158   bool eliminateInstrViaWidening(
159       Instruction *Instr, const df_iterator<DomTreeNode *> &DFSI,
160       const DenseMap<BasicBlock *, SmallVector<Instruction *, 8>>
161           &GuardsPerBlock);
162 
163   /// Used to keep track of which widening potential is more effective.
164   enum WideningScore {
165     /// Don't widen.
166     WS_IllegalOrNegative,
167 
168     /// Widening is performance neutral as far as the cycles spent in check
169     /// conditions goes (but can still help, e.g., code layout, having less
170     /// deopt state).
171     WS_Neutral,
172 
173     /// Widening is profitable.
174     WS_Positive,
175 
176     /// Widening is very profitable.  Not significantly different from \c
177     /// WS_Positive, except by the order.
178     WS_VeryPositive
179   };
180 
181   static StringRef scoreTypeToString(WideningScore WS);
182 
183   /// Compute the score for widening the condition in \p DominatedInstr
184   /// into \p WideningPoint.
185   WideningScore computeWideningScore(Instruction *DominatedInstr,
186                                      Instruction *ToWiden,
187                                      BasicBlock::iterator WideningPoint,
188                                      SmallVectorImpl<Value *> &ChecksToHoist,
189                                      SmallVectorImpl<Value *> &ChecksToWiden);
190 
191   /// Helper to check if \p V can be hoisted to \p InsertPos.
canBeHoistedTo(const Value * V,BasicBlock::iterator InsertPos) const192   bool canBeHoistedTo(const Value *V, BasicBlock::iterator InsertPos) const {
193     SmallPtrSet<const Instruction *, 8> Visited;
194     return canBeHoistedTo(V, InsertPos, Visited);
195   }
196 
197   bool canBeHoistedTo(const Value *V, BasicBlock::iterator InsertPos,
198                       SmallPtrSetImpl<const Instruction *> &Visited) const;
199 
canBeHoistedTo(const SmallVectorImpl<Value * > & Checks,BasicBlock::iterator InsertPos) const200   bool canBeHoistedTo(const SmallVectorImpl<Value *> &Checks,
201                       BasicBlock::iterator InsertPos) const {
202     return all_of(Checks,
203                   [&](const Value *V) { return canBeHoistedTo(V, InsertPos); });
204   }
205   /// Helper to hoist \p V to \p InsertPos.  Guaranteed to succeed if \c
206   /// canBeHoistedTo returned true.
207   void makeAvailableAt(Value *V, BasicBlock::iterator InsertPos) const;
208 
makeAvailableAt(const SmallVectorImpl<Value * > & Checks,BasicBlock::iterator InsertPos) const209   void makeAvailableAt(const SmallVectorImpl<Value *> &Checks,
210                        BasicBlock::iterator InsertPos) const {
211     for (Value *V : Checks)
212       makeAvailableAt(V, InsertPos);
213   }
214 
215   /// Common helper used by \c widenGuard and \c isWideningCondProfitable.  Try
216   /// to generate an expression computing the logical AND of \p ChecksToHoist
217   /// and \p ChecksToWiden. Return true if the expression computing the AND is
218   /// only as expensive as computing one of the set of expressions. If \p
219   /// InsertPt is true then actually generate the resulting expression, make it
220   /// available at \p InsertPt and return it in \p Result (else no change to the
221   /// IR is made).
222   std::optional<Value *>
223   mergeChecks(SmallVectorImpl<Value *> &ChecksToHoist,
224               SmallVectorImpl<Value *> &ChecksToWiden,
225               std::optional<BasicBlock::iterator> InsertPt);
226 
227   /// Generate the logical AND of \p ChecksToHoist and \p OldCondition and make
228   /// it available at InsertPt
229   Value *hoistChecks(SmallVectorImpl<Value *> &ChecksToHoist,
230                      Value *OldCondition, BasicBlock::iterator InsertPt);
231 
232   /// Adds freeze to Orig and push it as far as possible very aggressively.
233   /// Also replaces all uses of frozen instruction with frozen version.
234   Value *freezeAndPush(Value *Orig, BasicBlock::iterator InsertPt);
235 
236   /// Represents a range check of the form \c Base + \c Offset u< \c Length,
237   /// with the constraint that \c Length is not negative.  \c CheckInst is the
238   /// pre-existing instruction in the IR that computes the result of this range
239   /// check.
240   class RangeCheck {
241     const Value *Base;
242     const ConstantInt *Offset;
243     const Value *Length;
244     ICmpInst *CheckInst;
245 
246   public:
RangeCheck(const Value * Base,const ConstantInt * Offset,const Value * Length,ICmpInst * CheckInst)247     explicit RangeCheck(const Value *Base, const ConstantInt *Offset,
248                         const Value *Length, ICmpInst *CheckInst)
249         : Base(Base), Offset(Offset), Length(Length), CheckInst(CheckInst) {}
250 
setBase(const Value * NewBase)251     void setBase(const Value *NewBase) { Base = NewBase; }
setOffset(const ConstantInt * NewOffset)252     void setOffset(const ConstantInt *NewOffset) { Offset = NewOffset; }
253 
getBase() const254     const Value *getBase() const { return Base; }
getOffset() const255     const ConstantInt *getOffset() const { return Offset; }
getOffsetValue() const256     const APInt &getOffsetValue() const { return getOffset()->getValue(); }
getLength() const257     const Value *getLength() const { return Length; };
getCheckInst() const258     ICmpInst *getCheckInst() const { return CheckInst; }
259 
print(raw_ostream & OS,bool PrintTypes=false)260     void print(raw_ostream &OS, bool PrintTypes = false) {
261       OS << "Base: ";
262       Base->printAsOperand(OS, PrintTypes);
263       OS << " Offset: ";
264       Offset->printAsOperand(OS, PrintTypes);
265       OS << " Length: ";
266       Length->printAsOperand(OS, PrintTypes);
267     }
268 
dump()269     LLVM_DUMP_METHOD void dump() {
270       print(dbgs());
271       dbgs() << "\n";
272     }
273   };
274 
275   /// Parse \p ToParse into a conjunction (logical-and) of range checks; and
276   /// append them to \p Checks.  Returns true on success, may clobber \c Checks
277   /// on failure.
parseRangeChecks(SmallVectorImpl<Value * > & ToParse,SmallVectorImpl<RangeCheck> & Checks)278   bool parseRangeChecks(SmallVectorImpl<Value *> &ToParse,
279                         SmallVectorImpl<RangeCheck> &Checks) {
280     for (auto CheckCond : ToParse) {
281       if (!parseRangeChecks(CheckCond, Checks))
282         return false;
283     }
284     return true;
285   }
286 
287   bool parseRangeChecks(Value *CheckCond, SmallVectorImpl<RangeCheck> &Checks);
288 
289   /// Combine the checks in \p Checks into a smaller set of checks and append
290   /// them into \p CombinedChecks.  Return true on success (i.e. all of checks
291   /// in \p Checks were combined into \p CombinedChecks).  Clobbers \p Checks
292   /// and \p CombinedChecks on success and on failure.
293   bool combineRangeChecks(SmallVectorImpl<RangeCheck> &Checks,
294                           SmallVectorImpl<RangeCheck> &CombinedChecks) const;
295 
296   /// Can we compute the logical AND of \p ChecksToHoist and \p ChecksToWiden
297   /// for the price of computing only one of the set of expressions?
isWideningCondProfitable(SmallVectorImpl<Value * > & ChecksToHoist,SmallVectorImpl<Value * > & ChecksToWiden)298   bool isWideningCondProfitable(SmallVectorImpl<Value *> &ChecksToHoist,
299                                 SmallVectorImpl<Value *> &ChecksToWiden) {
300     return mergeChecks(ChecksToHoist, ChecksToWiden, /*InsertPt=*/std::nullopt)
301         .has_value();
302   }
303 
304   /// Widen \p ChecksToWiden to fail if any of \p ChecksToHoist is false
widenGuard(SmallVectorImpl<Value * > & ChecksToHoist,SmallVectorImpl<Value * > & ChecksToWiden,Instruction * ToWiden)305   void widenGuard(SmallVectorImpl<Value *> &ChecksToHoist,
306                   SmallVectorImpl<Value *> &ChecksToWiden,
307                   Instruction *ToWiden) {
308     auto InsertPt = findInsertionPointForWideCondition(ToWiden);
309     auto MergedCheck = mergeChecks(ChecksToHoist, ChecksToWiden, InsertPt);
310     Value *Result = MergedCheck ? *MergedCheck
311                                 : hoistChecks(ChecksToHoist,
312                                               getCondition(ToWiden), *InsertPt);
313 
314     if (isGuardAsWidenableBranch(ToWiden)) {
315       setWidenableBranchCond(cast<BranchInst>(ToWiden), Result);
316       return;
317     }
318     setCondition(ToWiden, Result);
319   }
320 
321 public:
GuardWideningImpl(DominatorTree & DT,PostDominatorTree * PDT,LoopInfo & LI,AssumptionCache & AC,MemorySSAUpdater * MSSAU,DomTreeNode * Root,std::function<bool (BasicBlock *)> BlockFilter)322   explicit GuardWideningImpl(DominatorTree &DT, PostDominatorTree *PDT,
323                              LoopInfo &LI, AssumptionCache &AC,
324                              MemorySSAUpdater *MSSAU, DomTreeNode *Root,
325                              std::function<bool(BasicBlock *)> BlockFilter)
326       : DT(DT), PDT(PDT), LI(LI), AC(AC), MSSAU(MSSAU), Root(Root),
327         BlockFilter(BlockFilter) {}
328 
329   /// The entry point for this pass.
330   bool run();
331 };
332 }
333 
isSupportedGuardInstruction(const Instruction * Insn)334 static bool isSupportedGuardInstruction(const Instruction *Insn) {
335   if (isGuard(Insn))
336     return true;
337   if (WidenBranchGuards && isGuardAsWidenableBranch(Insn))
338     return true;
339   return false;
340 }
341 
run()342 bool GuardWideningImpl::run() {
343   DenseMap<BasicBlock *, SmallVector<Instruction *, 8>> GuardsInBlock;
344   bool Changed = false;
345   for (auto DFI = df_begin(Root), DFE = df_end(Root);
346        DFI != DFE; ++DFI) {
347     auto *BB = (*DFI)->getBlock();
348     if (!BlockFilter(BB))
349       continue;
350 
351     auto &CurrentList = GuardsInBlock[BB];
352 
353     for (auto &I : *BB)
354       if (isSupportedGuardInstruction(&I))
355         CurrentList.push_back(cast<Instruction>(&I));
356 
357     for (auto *II : CurrentList)
358       Changed |= eliminateInstrViaWidening(II, DFI, GuardsInBlock);
359   }
360 
361   assert(EliminatedGuardsAndBranches.empty() || Changed);
362   for (auto *I : EliminatedGuardsAndBranches)
363     if (!WidenedGuards.count(I)) {
364       assert(isa<ConstantInt>(getCondition(I)) && "Should be!");
365       if (isSupportedGuardInstruction(I))
366         eliminateGuard(I, MSSAU);
367       else {
368         assert(isa<BranchInst>(I) &&
369                "Eliminated something other than guard or branch?");
370         ++CondBranchEliminated;
371       }
372     }
373 
374   return Changed;
375 }
376 
eliminateInstrViaWidening(Instruction * Instr,const df_iterator<DomTreeNode * > & DFSI,const DenseMap<BasicBlock *,SmallVector<Instruction *,8>> & GuardsInBlock)377 bool GuardWideningImpl::eliminateInstrViaWidening(
378     Instruction *Instr, const df_iterator<DomTreeNode *> &DFSI,
379     const DenseMap<BasicBlock *, SmallVector<Instruction *, 8>>
380         &GuardsInBlock) {
381   SmallVector<Value *> ChecksToHoist;
382   parseWidenableGuard(Instr, ChecksToHoist);
383   // Ignore trivial true or false conditions. These instructions will be
384   // trivially eliminated by any cleanup pass. Do not erase them because other
385   // guards can possibly be widened into them.
386   if (ChecksToHoist.empty() ||
387       (ChecksToHoist.size() == 1 && isa<ConstantInt>(ChecksToHoist.front())))
388     return false;
389 
390   Instruction *BestSoFar = nullptr;
391   auto BestScoreSoFar = WS_IllegalOrNegative;
392 
393   // In the set of dominating guards, find the one we can merge GuardInst with
394   // for the most profit.
395   for (unsigned i = 0, e = DFSI.getPathLength(); i != e; ++i) {
396     auto *CurBB = DFSI.getPath(i)->getBlock();
397     if (!BlockFilter(CurBB))
398       break;
399     assert(GuardsInBlock.count(CurBB) && "Must have been populated by now!");
400     const auto &GuardsInCurBB = GuardsInBlock.find(CurBB)->second;
401 
402     auto I = GuardsInCurBB.begin();
403     auto E = Instr->getParent() == CurBB ? find(GuardsInCurBB, Instr)
404                                          : GuardsInCurBB.end();
405 
406 #ifndef NDEBUG
407     {
408       unsigned Index = 0;
409       for (auto &I : *CurBB) {
410         if (Index == GuardsInCurBB.size())
411           break;
412         if (GuardsInCurBB[Index] == &I)
413           Index++;
414       }
415       assert(Index == GuardsInCurBB.size() &&
416              "Guards expected to be in order!");
417     }
418 #endif
419 
420     assert((i == (e - 1)) == (Instr->getParent() == CurBB) && "Bad DFS?");
421 
422     for (auto *Candidate : make_range(I, E)) {
423       auto WideningPoint = findInsertionPointForWideCondition(Candidate);
424       if (!WideningPoint)
425         continue;
426       SmallVector<Value *> CandidateChecks;
427       parseWidenableGuard(Candidate, CandidateChecks);
428       auto Score = computeWideningScore(Instr, Candidate, *WideningPoint,
429                                         ChecksToHoist, CandidateChecks);
430       LLVM_DEBUG(dbgs() << "Score between " << *Instr << " and " << *Candidate
431                         << " is " << scoreTypeToString(Score) << "\n");
432       if (Score > BestScoreSoFar) {
433         BestScoreSoFar = Score;
434         BestSoFar = Candidate;
435       }
436     }
437   }
438 
439   if (BestScoreSoFar == WS_IllegalOrNegative) {
440     LLVM_DEBUG(dbgs() << "Did not eliminate guard " << *Instr << "\n");
441     return false;
442   }
443 
444   assert(BestSoFar != Instr && "Should have never visited same guard!");
445   assert(DT.dominates(BestSoFar, Instr) && "Should be!");
446 
447   LLVM_DEBUG(dbgs() << "Widening " << *Instr << " into " << *BestSoFar
448                     << " with score " << scoreTypeToString(BestScoreSoFar)
449                     << "\n");
450   SmallVector<Value *> ChecksToWiden;
451   parseWidenableGuard(BestSoFar, ChecksToWiden);
452   widenGuard(ChecksToHoist, ChecksToWiden, BestSoFar);
453   auto NewGuardCondition = ConstantInt::getTrue(Instr->getContext());
454   setCondition(Instr, NewGuardCondition);
455   EliminatedGuardsAndBranches.push_back(Instr);
456   WidenedGuards.insert(BestSoFar);
457   return true;
458 }
459 
computeWideningScore(Instruction * DominatedInstr,Instruction * ToWiden,BasicBlock::iterator WideningPoint,SmallVectorImpl<Value * > & ChecksToHoist,SmallVectorImpl<Value * > & ChecksToWiden)460 GuardWideningImpl::WideningScore GuardWideningImpl::computeWideningScore(
461     Instruction *DominatedInstr, Instruction *ToWiden,
462     BasicBlock::iterator WideningPoint, SmallVectorImpl<Value *> &ChecksToHoist,
463     SmallVectorImpl<Value *> &ChecksToWiden) {
464   Loop *DominatedInstrLoop = LI.getLoopFor(DominatedInstr->getParent());
465   Loop *DominatingGuardLoop = LI.getLoopFor(WideningPoint->getParent());
466   bool HoistingOutOfLoop = false;
467 
468   if (DominatingGuardLoop != DominatedInstrLoop) {
469     // Be conservative and don't widen into a sibling loop.  TODO: If the
470     // sibling is colder, we should consider allowing this.
471     if (DominatingGuardLoop &&
472         !DominatingGuardLoop->contains(DominatedInstrLoop))
473       return WS_IllegalOrNegative;
474 
475     HoistingOutOfLoop = true;
476   }
477 
478   if (!canBeHoistedTo(ChecksToHoist, WideningPoint))
479     return WS_IllegalOrNegative;
480   // Further in the GuardWideningImpl::hoistChecks the entire condition might be
481   // widened, not the parsed list of checks. So we need to check the possibility
482   // of that condition hoisting.
483   if (!canBeHoistedTo(getCondition(ToWiden), WideningPoint))
484     return WS_IllegalOrNegative;
485 
486   // If the guard was conditional executed, it may never be reached
487   // dynamically.  There are two potential downsides to hoisting it out of the
488   // conditionally executed region: 1) we may spuriously deopt without need and
489   // 2) we have the extra cost of computing the guard condition in the common
490   // case.  At the moment, we really only consider the second in our heuristic
491   // here.  TODO: evaluate cost model for spurious deopt
492   // NOTE: As written, this also lets us hoist right over another guard which
493   // is essentially just another spelling for control flow.
494   if (isWideningCondProfitable(ChecksToHoist, ChecksToWiden))
495     return HoistingOutOfLoop ? WS_VeryPositive : WS_Positive;
496 
497   if (HoistingOutOfLoop)
498     return WS_Positive;
499 
500   // For a given basic block \p BB, return its successor which is guaranteed or
501   // highly likely will be taken as its successor.
502   auto GetLikelySuccessor = [](const BasicBlock * BB)->const BasicBlock * {
503     if (auto *UniqueSucc = BB->getUniqueSuccessor())
504       return UniqueSucc;
505     auto *Term = BB->getTerminator();
506     Value *Cond = nullptr;
507     const BasicBlock *IfTrue = nullptr, *IfFalse = nullptr;
508     using namespace PatternMatch;
509     if (!match(Term, m_Br(m_Value(Cond), m_BasicBlock(IfTrue),
510                           m_BasicBlock(IfFalse))))
511       return nullptr;
512     // For constant conditions, only one dynamical successor is possible
513     if (auto *ConstCond = dyn_cast<ConstantInt>(Cond))
514       return ConstCond->isAllOnesValue() ? IfTrue : IfFalse;
515     // If one of successors ends with deopt, another one is likely.
516     if (IfFalse->getPostdominatingDeoptimizeCall())
517       return IfTrue;
518     if (IfTrue->getPostdominatingDeoptimizeCall())
519       return IfFalse;
520     // TODO: Use branch frequency metatada to allow hoisting through non-deopt
521     // branches?
522     return nullptr;
523   };
524 
525   // Returns true if we might be hoisting above explicit control flow into a
526   // considerably hotter block.  Note that this completely ignores implicit
527   // control flow (guards, calls which throw, etc...).  That choice appears
528   // arbitrary (we assume that implicit control flow exits are all rare).
529   auto MaybeHoistingToHotterBlock = [&]() {
530     const auto *DominatingBlock = WideningPoint->getParent();
531     const auto *DominatedBlock = DominatedInstr->getParent();
532 
533     // Descend as low as we can, always taking the likely successor.
534     assert(DT.isReachableFromEntry(DominatingBlock) && "Unreached code");
535     assert(DT.isReachableFromEntry(DominatedBlock) && "Unreached code");
536     assert(DT.dominates(DominatingBlock, DominatedBlock) && "No dominance");
537     while (DominatedBlock != DominatingBlock) {
538       auto *LikelySucc = GetLikelySuccessor(DominatingBlock);
539       // No likely successor?
540       if (!LikelySucc)
541         break;
542       // Only go down the dominator tree.
543       if (!DT.properlyDominates(DominatingBlock, LikelySucc))
544         break;
545       DominatingBlock = LikelySucc;
546     }
547 
548     // Found?
549     if (DominatedBlock == DominatingBlock)
550       return false;
551     // We followed the likely successor chain and went past the dominated
552     // block. It means that the dominated guard is in dead/very cold code.
553     if (!DT.dominates(DominatingBlock, DominatedBlock))
554       return true;
555     // TODO: diamond, triangle cases
556     if (!PDT)
557       return true;
558     return !PDT->dominates(DominatedBlock, DominatingBlock);
559   };
560 
561   return MaybeHoistingToHotterBlock() ? WS_IllegalOrNegative : WS_Neutral;
562 }
563 
canBeHoistedTo(const Value * V,BasicBlock::iterator Loc,SmallPtrSetImpl<const Instruction * > & Visited) const564 bool GuardWideningImpl::canBeHoistedTo(
565     const Value *V, BasicBlock::iterator Loc,
566     SmallPtrSetImpl<const Instruction *> &Visited) const {
567   auto *Inst = dyn_cast<Instruction>(V);
568   if (!Inst || DT.dominates(Inst, Loc) || Visited.count(Inst))
569     return true;
570 
571   if (!isSafeToSpeculativelyExecute(Inst, Loc, &AC, &DT) ||
572       Inst->mayReadFromMemory())
573     return false;
574 
575   Visited.insert(Inst);
576 
577   // We only want to go _up_ the dominance chain when recursing.
578   assert(!isa<PHINode>(Loc) &&
579          "PHIs should return false for isSafeToSpeculativelyExecute");
580   assert(DT.isReachableFromEntry(Inst->getParent()) &&
581          "We did a DFS from the block entry!");
582   return all_of(Inst->operands(),
583                 [&](Value *Op) { return canBeHoistedTo(Op, Loc, Visited); });
584 }
585 
makeAvailableAt(Value * V,BasicBlock::iterator Loc) const586 void GuardWideningImpl::makeAvailableAt(Value *V,
587                                         BasicBlock::iterator Loc) const {
588   auto *Inst = dyn_cast<Instruction>(V);
589   if (!Inst || DT.dominates(Inst, Loc))
590     return;
591 
592   assert(isSafeToSpeculativelyExecute(Inst, Loc, &AC, &DT) &&
593          !Inst->mayReadFromMemory() &&
594          "Should've checked with canBeHoistedTo!");
595 
596   for (Value *Op : Inst->operands())
597     makeAvailableAt(Op, Loc);
598 
599   Inst->moveBefore(*Loc->getParent(), Loc);
600 }
601 
602 // Return Instruction before which we can insert freeze for the value V as close
603 // to def as possible. If there is no place to add freeze, return empty.
604 static std::optional<BasicBlock::iterator>
getFreezeInsertPt(Value * V,const DominatorTree & DT)605 getFreezeInsertPt(Value *V, const DominatorTree &DT) {
606   auto *I = dyn_cast<Instruction>(V);
607   if (!I)
608     return DT.getRoot()->getFirstNonPHIOrDbgOrAlloca()->getIterator();
609 
610   std::optional<BasicBlock::iterator> Res = I->getInsertionPointAfterDef();
611   // If there is no place to add freeze - return nullptr.
612   if (!Res || !DT.dominates(I, &**Res))
613     return std::nullopt;
614 
615   Instruction *ResInst = &**Res;
616 
617   // If there is a User dominated by original I, then it should be dominated
618   // by Freeze instruction as well.
619   if (any_of(I->users(), [&](User *U) {
620         Instruction *User = cast<Instruction>(U);
621         return ResInst != User && DT.dominates(I, User) &&
622                !DT.dominates(ResInst, User);
623       }))
624     return std::nullopt;
625   return Res;
626 }
627 
freezeAndPush(Value * Orig,BasicBlock::iterator InsertPt)628 Value *GuardWideningImpl::freezeAndPush(Value *Orig,
629                                         BasicBlock::iterator InsertPt) {
630   if (isGuaranteedNotToBePoison(Orig, nullptr, InsertPt, &DT))
631     return Orig;
632   std::optional<BasicBlock::iterator> InsertPtAtDef =
633       getFreezeInsertPt(Orig, DT);
634   if (!InsertPtAtDef) {
635     FreezeInst *FI = new FreezeInst(Orig, "gw.freeze");
636     FI->insertBefore(*InsertPt->getParent(), InsertPt);
637     return FI;
638   }
639   if (isa<Constant>(Orig) || isa<GlobalValue>(Orig)) {
640     BasicBlock::iterator InsertPt = *InsertPtAtDef;
641     FreezeInst *FI = new FreezeInst(Orig, "gw.freeze");
642     FI->insertBefore(*InsertPt->getParent(), InsertPt);
643     return FI;
644   }
645 
646   SmallSet<Value *, 16> Visited;
647   SmallVector<Value *, 16> Worklist;
648   SmallSet<Instruction *, 16> DropPoisonFlags;
649   SmallVector<Value *, 16> NeedFreeze;
650   DenseMap<Value *, FreezeInst *> CacheOfFreezes;
651 
652   // A bit overloaded data structures. Visited contains constant/GV
653   // if we already met it. In this case CacheOfFreezes has a freeze if it is
654   // required.
655   auto handleConstantOrGlobal = [&](Use &U) {
656     Value *Def = U.get();
657     if (!isa<Constant>(Def) && !isa<GlobalValue>(Def))
658       return false;
659 
660     if (Visited.insert(Def).second) {
661       if (isGuaranteedNotToBePoison(Def, nullptr, InsertPt, &DT))
662         return true;
663       BasicBlock::iterator InsertPt = *getFreezeInsertPt(Def, DT);
664       FreezeInst *FI = new FreezeInst(Def, Def->getName() + ".gw.fr");
665       FI->insertBefore(*InsertPt->getParent(), InsertPt);
666       CacheOfFreezes[Def] = FI;
667     }
668 
669     if (CacheOfFreezes.count(Def))
670       U.set(CacheOfFreezes[Def]);
671     return true;
672   };
673 
674   Worklist.push_back(Orig);
675   while (!Worklist.empty()) {
676     Value *V = Worklist.pop_back_val();
677     if (!Visited.insert(V).second)
678       continue;
679 
680     if (isGuaranteedNotToBePoison(V, nullptr, InsertPt, &DT))
681       continue;
682 
683     Instruction *I = dyn_cast<Instruction>(V);
684     if (!I || canCreateUndefOrPoison(cast<Operator>(I),
685                                      /*ConsiderFlagsAndMetadata*/ false)) {
686       NeedFreeze.push_back(V);
687       continue;
688     }
689     // Check all operands. If for any of them we cannot insert Freeze,
690     // stop here. Otherwise, iterate.
691     if (any_of(I->operands(), [&](Value *Op) {
692           return isa<Instruction>(Op) && !getFreezeInsertPt(Op, DT);
693         })) {
694       NeedFreeze.push_back(I);
695       continue;
696     }
697     DropPoisonFlags.insert(I);
698     for (Use &U : I->operands())
699       if (!handleConstantOrGlobal(U))
700         Worklist.push_back(U.get());
701   }
702   for (Instruction *I : DropPoisonFlags)
703     I->dropPoisonGeneratingAnnotations();
704 
705   Value *Result = Orig;
706   for (Value *V : NeedFreeze) {
707     BasicBlock::iterator FreezeInsertPt = *getFreezeInsertPt(V, DT);
708     FreezeInst *FI = new FreezeInst(V, V->getName() + ".gw.fr");
709     FI->insertBefore(*FreezeInsertPt->getParent(), FreezeInsertPt);
710     ++FreezeAdded;
711     if (V == Orig)
712       Result = FI;
713     V->replaceUsesWithIf(
714         FI, [&](const Use & U)->bool { return U.getUser() != FI; });
715   }
716 
717   return Result;
718 }
719 
720 std::optional<Value *>
mergeChecks(SmallVectorImpl<Value * > & ChecksToHoist,SmallVectorImpl<Value * > & ChecksToWiden,std::optional<BasicBlock::iterator> InsertPt)721 GuardWideningImpl::mergeChecks(SmallVectorImpl<Value *> &ChecksToHoist,
722                                SmallVectorImpl<Value *> &ChecksToWiden,
723                                std::optional<BasicBlock::iterator> InsertPt) {
724   using namespace llvm::PatternMatch;
725 
726   Value *Result = nullptr;
727   {
728     // L >u C0 && L >u C1  ->  L >u max(C0, C1)
729     ConstantInt *RHS0, *RHS1;
730     Value *LHS;
731     ICmpInst::Predicate Pred0, Pred1;
732     // TODO: Support searching for pairs to merge from both whole lists of
733     // ChecksToHoist and ChecksToWiden.
734     if (ChecksToWiden.size() == 1 && ChecksToHoist.size() == 1 &&
735         match(ChecksToWiden.front(),
736               m_ICmp(Pred0, m_Value(LHS), m_ConstantInt(RHS0))) &&
737         match(ChecksToHoist.front(),
738               m_ICmp(Pred1, m_Specific(LHS), m_ConstantInt(RHS1)))) {
739 
740       ConstantRange CR0 =
741           ConstantRange::makeExactICmpRegion(Pred0, RHS0->getValue());
742       ConstantRange CR1 =
743           ConstantRange::makeExactICmpRegion(Pred1, RHS1->getValue());
744 
745       // Given what we're doing here and the semantics of guards, it would
746       // be correct to use a subset intersection, but that may be too
747       // aggressive in cases we care about.
748       if (std::optional<ConstantRange> Intersect =
749               CR0.exactIntersectWith(CR1)) {
750         APInt NewRHSAP;
751         CmpInst::Predicate Pred;
752         if (Intersect->getEquivalentICmp(Pred, NewRHSAP)) {
753           if (InsertPt) {
754             ConstantInt *NewRHS =
755                 ConstantInt::get((*InsertPt)->getContext(), NewRHSAP);
756             assert(canBeHoistedTo(LHS, *InsertPt) && "must be");
757             makeAvailableAt(LHS, *InsertPt);
758             Result = new ICmpInst(*InsertPt, Pred, LHS, NewRHS, "wide.chk");
759           }
760           return Result;
761         }
762       }
763     }
764   }
765 
766   {
767     SmallVector<GuardWideningImpl::RangeCheck, 4> Checks, CombinedChecks;
768     if (parseRangeChecks(ChecksToWiden, Checks) &&
769         parseRangeChecks(ChecksToHoist, Checks) &&
770         combineRangeChecks(Checks, CombinedChecks)) {
771       if (InsertPt) {
772         for (auto &RC : CombinedChecks) {
773           makeAvailableAt(RC.getCheckInst(), *InsertPt);
774           if (Result)
775             Result = BinaryOperator::CreateAnd(RC.getCheckInst(), Result, "",
776                                                *InsertPt);
777           else
778             Result = RC.getCheckInst();
779         }
780         assert(Result && "Failed to find result value");
781         Result->setName("wide.chk");
782         Result = freezeAndPush(Result, *InsertPt);
783       }
784       return Result;
785     }
786   }
787   // We were not able to compute ChecksToHoist AND ChecksToWiden for the price
788   // of one.
789   return std::nullopt;
790 }
791 
hoistChecks(SmallVectorImpl<Value * > & ChecksToHoist,Value * OldCondition,BasicBlock::iterator InsertPt)792 Value *GuardWideningImpl::hoistChecks(SmallVectorImpl<Value *> &ChecksToHoist,
793                                       Value *OldCondition,
794                                       BasicBlock::iterator InsertPt) {
795   assert(!ChecksToHoist.empty());
796   IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
797   makeAvailableAt(ChecksToHoist, InsertPt);
798   makeAvailableAt(OldCondition, InsertPt);
799   Value *Result = Builder.CreateAnd(ChecksToHoist);
800   Result = freezeAndPush(Result, InsertPt);
801   Result = Builder.CreateAnd(OldCondition, Result);
802   Result->setName("wide.chk");
803   return Result;
804 }
805 
parseRangeChecks(Value * CheckCond,SmallVectorImpl<GuardWideningImpl::RangeCheck> & Checks)806 bool GuardWideningImpl::parseRangeChecks(
807     Value *CheckCond, SmallVectorImpl<GuardWideningImpl::RangeCheck> &Checks) {
808   using namespace llvm::PatternMatch;
809 
810   auto *IC = dyn_cast<ICmpInst>(CheckCond);
811   if (!IC || !IC->getOperand(0)->getType()->isIntegerTy() ||
812       (IC->getPredicate() != ICmpInst::ICMP_ULT &&
813        IC->getPredicate() != ICmpInst::ICMP_UGT))
814     return false;
815 
816   const Value *CmpLHS = IC->getOperand(0), *CmpRHS = IC->getOperand(1);
817   if (IC->getPredicate() == ICmpInst::ICMP_UGT)
818     std::swap(CmpLHS, CmpRHS);
819 
820   auto &DL = IC->getDataLayout();
821 
822   GuardWideningImpl::RangeCheck Check(
823       CmpLHS, cast<ConstantInt>(ConstantInt::getNullValue(CmpRHS->getType())),
824       CmpRHS, IC);
825 
826   if (!isKnownNonNegative(Check.getLength(), DL))
827     return false;
828 
829   // What we have in \c Check now is a correct interpretation of \p CheckCond.
830   // Try to see if we can move some constant offsets into the \c Offset field.
831 
832   bool Changed;
833   auto &Ctx = CheckCond->getContext();
834 
835   do {
836     Value *OpLHS;
837     ConstantInt *OpRHS;
838     Changed = false;
839 
840 #ifndef NDEBUG
841     auto *BaseInst = dyn_cast<Instruction>(Check.getBase());
842     assert((!BaseInst || DT.isReachableFromEntry(BaseInst->getParent())) &&
843            "Unreachable instruction?");
844 #endif
845 
846     if (match(Check.getBase(), m_Add(m_Value(OpLHS), m_ConstantInt(OpRHS)))) {
847       Check.setBase(OpLHS);
848       APInt NewOffset = Check.getOffsetValue() + OpRHS->getValue();
849       Check.setOffset(ConstantInt::get(Ctx, NewOffset));
850       Changed = true;
851     } else if (match(Check.getBase(),
852                      m_Or(m_Value(OpLHS), m_ConstantInt(OpRHS)))) {
853       KnownBits Known = computeKnownBits(OpLHS, DL);
854       if ((OpRHS->getValue() & Known.Zero) == OpRHS->getValue()) {
855         Check.setBase(OpLHS);
856         APInt NewOffset = Check.getOffsetValue() + OpRHS->getValue();
857         Check.setOffset(ConstantInt::get(Ctx, NewOffset));
858         Changed = true;
859       }
860     }
861   } while (Changed);
862 
863   Checks.push_back(Check);
864   return true;
865 }
866 
combineRangeChecks(SmallVectorImpl<GuardWideningImpl::RangeCheck> & Checks,SmallVectorImpl<GuardWideningImpl::RangeCheck> & RangeChecksOut) const867 bool GuardWideningImpl::combineRangeChecks(
868     SmallVectorImpl<GuardWideningImpl::RangeCheck> &Checks,
869     SmallVectorImpl<GuardWideningImpl::RangeCheck> &RangeChecksOut) const {
870   unsigned OldCount = Checks.size();
871   while (!Checks.empty()) {
872     // Pick all of the range checks with a specific base and length, and try to
873     // merge them.
874     const Value *CurrentBase = Checks.front().getBase();
875     const Value *CurrentLength = Checks.front().getLength();
876 
877     SmallVector<GuardWideningImpl::RangeCheck, 3> CurrentChecks;
878 
879     auto IsCurrentCheck = [&](GuardWideningImpl::RangeCheck &RC) {
880       return RC.getBase() == CurrentBase && RC.getLength() == CurrentLength;
881     };
882 
883     copy_if(Checks, std::back_inserter(CurrentChecks), IsCurrentCheck);
884     erase_if(Checks, IsCurrentCheck);
885 
886     assert(CurrentChecks.size() != 0 && "We know we have at least one!");
887 
888     if (CurrentChecks.size() < 3) {
889       llvm::append_range(RangeChecksOut, CurrentChecks);
890       continue;
891     }
892 
893     // CurrentChecks.size() will typically be 3 here, but so far there has been
894     // no need to hard-code that fact.
895 
896     llvm::sort(CurrentChecks, [&](const GuardWideningImpl::RangeCheck &LHS,
897                                   const GuardWideningImpl::RangeCheck &RHS) {
898       return LHS.getOffsetValue().slt(RHS.getOffsetValue());
899     });
900 
901     // Note: std::sort should not invalidate the ChecksStart iterator.
902 
903     const ConstantInt *MinOffset = CurrentChecks.front().getOffset();
904     const ConstantInt *MaxOffset = CurrentChecks.back().getOffset();
905 
906     unsigned BitWidth = MaxOffset->getValue().getBitWidth();
907     if ((MaxOffset->getValue() - MinOffset->getValue())
908             .ugt(APInt::getSignedMinValue(BitWidth)))
909       return false;
910 
911     APInt MaxDiff = MaxOffset->getValue() - MinOffset->getValue();
912     const APInt &HighOffset = MaxOffset->getValue();
913     auto OffsetOK = [&](const GuardWideningImpl::RangeCheck &RC) {
914       return (HighOffset - RC.getOffsetValue()).ult(MaxDiff);
915     };
916 
917     if (MaxDiff.isMinValue() || !all_of(drop_begin(CurrentChecks), OffsetOK))
918       return false;
919 
920     // We have a series of f+1 checks as:
921     //
922     //   I+k_0 u< L   ... Chk_0
923     //   I+k_1 u< L   ... Chk_1
924     //   ...
925     //   I+k_f u< L   ... Chk_f
926     //
927     //     with forall i in [0,f]: k_f-k_i u< k_f-k_0  ... Precond_0
928     //          k_f-k_0 u< INT_MIN+k_f                 ... Precond_1
929     //          k_f != k_0                             ... Precond_2
930     //
931     // Claim:
932     //   Chk_0 AND Chk_f  implies all the other checks
933     //
934     // Informal proof sketch:
935     //
936     // We will show that the integer range [I+k_0,I+k_f] does not unsigned-wrap
937     // (i.e. going from I+k_0 to I+k_f does not cross the -1,0 boundary) and
938     // thus I+k_f is the greatest unsigned value in that range.
939     //
940     // This combined with Ckh_(f+1) shows that everything in that range is u< L.
941     // Via Precond_0 we know that all of the indices in Chk_0 through Chk_(f+1)
942     // lie in [I+k_0,I+k_f], this proving our claim.
943     //
944     // To see that [I+k_0,I+k_f] is not a wrapping range, note that there are
945     // two possibilities: I+k_0 u< I+k_f or I+k_0 >u I+k_f (they can't be equal
946     // since k_0 != k_f).  In the former case, [I+k_0,I+k_f] is not a wrapping
947     // range by definition, and the latter case is impossible:
948     //
949     //   0-----I+k_f---I+k_0----L---INT_MAX,INT_MIN------------------(-1)
950     //   xxxxxx             xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
951     //
952     // For Chk_0 to succeed, we'd have to have k_f-k_0 (the range highlighted
953     // with 'x' above) to be at least >u INT_MIN.
954 
955     RangeChecksOut.emplace_back(CurrentChecks.front());
956     RangeChecksOut.emplace_back(CurrentChecks.back());
957   }
958 
959   assert(RangeChecksOut.size() <= OldCount && "We pessimized!");
960   return RangeChecksOut.size() != OldCount;
961 }
962 
963 #ifndef NDEBUG
scoreTypeToString(WideningScore WS)964 StringRef GuardWideningImpl::scoreTypeToString(WideningScore WS) {
965   switch (WS) {
966   case WS_IllegalOrNegative:
967     return "IllegalOrNegative";
968   case WS_Neutral:
969     return "Neutral";
970   case WS_Positive:
971     return "Positive";
972   case WS_VeryPositive:
973     return "VeryPositive";
974   }
975 
976   llvm_unreachable("Fully covered switch above!");
977 }
978 #endif
979 
run(Function & F,FunctionAnalysisManager & AM)980 PreservedAnalyses GuardWideningPass::run(Function &F,
981                                          FunctionAnalysisManager &AM) {
982   // Avoid requesting analyses if there are no guards or widenable conditions.
983   auto *GuardDecl = F.getParent()->getFunction(
984       Intrinsic::getName(Intrinsic::experimental_guard));
985   bool HasIntrinsicGuards = GuardDecl && !GuardDecl->use_empty();
986   auto *WCDecl = F.getParent()->getFunction(
987       Intrinsic::getName(Intrinsic::experimental_widenable_condition));
988   bool HasWidenableConditions = WCDecl && !WCDecl->use_empty();
989   if (!HasIntrinsicGuards && !HasWidenableConditions)
990     return PreservedAnalyses::all();
991   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
992   auto &LI = AM.getResult<LoopAnalysis>(F);
993   auto &PDT = AM.getResult<PostDominatorTreeAnalysis>(F);
994   auto &AC = AM.getResult<AssumptionAnalysis>(F);
995   auto *MSSAA = AM.getCachedResult<MemorySSAAnalysis>(F);
996   std::unique_ptr<MemorySSAUpdater> MSSAU;
997   if (MSSAA)
998     MSSAU = std::make_unique<MemorySSAUpdater>(&MSSAA->getMSSA());
999   if (!GuardWideningImpl(DT, &PDT, LI, AC, MSSAU ? MSSAU.get() : nullptr,
1000                          DT.getRootNode(), [](BasicBlock *) { return true; })
1001            .run())
1002     return PreservedAnalyses::all();
1003 
1004   PreservedAnalyses PA;
1005   PA.preserveSet<CFGAnalyses>();
1006   PA.preserve<MemorySSAAnalysis>();
1007   return PA;
1008 }
1009 
run(Loop & L,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater & U)1010 PreservedAnalyses GuardWideningPass::run(Loop &L, LoopAnalysisManager &AM,
1011                                          LoopStandardAnalysisResults &AR,
1012                                          LPMUpdater &U) {
1013   BasicBlock *RootBB = L.getLoopPredecessor();
1014   if (!RootBB)
1015     RootBB = L.getHeader();
1016   auto BlockFilter = [&](BasicBlock *BB) {
1017     return BB == RootBB || L.contains(BB);
1018   };
1019   std::unique_ptr<MemorySSAUpdater> MSSAU;
1020   if (AR.MSSA)
1021     MSSAU = std::make_unique<MemorySSAUpdater>(AR.MSSA);
1022   if (!GuardWideningImpl(AR.DT, nullptr, AR.LI, AR.AC,
1023                          MSSAU ? MSSAU.get() : nullptr, AR.DT.getNode(RootBB),
1024                          BlockFilter)
1025            .run())
1026     return PreservedAnalyses::all();
1027 
1028   auto PA = getLoopPassPreservedAnalyses();
1029   if (AR.MSSA)
1030     PA.preserve<MemorySSAAnalysis>();
1031   return PA;
1032 }
1033