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/GuardUtils.h"
46 #include "llvm/Analysis/LoopInfo.h"
47 #include "llvm/Analysis/LoopPass.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/IntrinsicInst.h"
54 #include "llvm/IR/PatternMatch.h"
55 #include "llvm/InitializePasses.h"
56 #include "llvm/Pass.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 
72 static cl::opt<bool>
73     WidenBranchGuards("guard-widening-widen-branch-guards", cl::Hidden,
74                       cl::desc("Whether or not we should widen guards  "
75                                "expressed as branches by widenable conditions"),
76                       cl::init(true));
77 
78 namespace {
79 
80 // Get the condition of \p I. It can either be a guard or a conditional branch.
81 static Value *getCondition(Instruction *I) {
82   if (IntrinsicInst *GI = dyn_cast<IntrinsicInst>(I)) {
83     assert(GI->getIntrinsicID() == Intrinsic::experimental_guard &&
84            "Bad guard intrinsic?");
85     return GI->getArgOperand(0);
86   }
87   Value *Cond, *WC;
88   BasicBlock *IfTrueBB, *IfFalseBB;
89   if (parseWidenableBranch(I, Cond, WC, IfTrueBB, IfFalseBB))
90     return Cond;
91 
92   return cast<BranchInst>(I)->getCondition();
93 }
94 
95 // Set the condition for \p I to \p NewCond. \p I can either be a guard or a
96 // conditional branch.
97 static void setCondition(Instruction *I, Value *NewCond) {
98   if (IntrinsicInst *GI = dyn_cast<IntrinsicInst>(I)) {
99     assert(GI->getIntrinsicID() == Intrinsic::experimental_guard &&
100            "Bad guard intrinsic?");
101     GI->setArgOperand(0, NewCond);
102     return;
103   }
104   cast<BranchInst>(I)->setCondition(NewCond);
105 }
106 
107 // Eliminates the guard instruction properly.
108 static void eliminateGuard(Instruction *GuardInst, MemorySSAUpdater *MSSAU) {
109   GuardInst->eraseFromParent();
110   if (MSSAU)
111     MSSAU->removeMemoryAccess(GuardInst);
112   ++GuardsEliminated;
113 }
114 
115 class GuardWideningImpl {
116   DominatorTree &DT;
117   PostDominatorTree *PDT;
118   LoopInfo &LI;
119   MemorySSAUpdater *MSSAU;
120 
121   /// Together, these describe the region of interest.  This might be all of
122   /// the blocks within a function, or only a given loop's blocks and preheader.
123   DomTreeNode *Root;
124   std::function<bool(BasicBlock*)> BlockFilter;
125 
126   /// The set of guards and conditional branches whose conditions have been
127   /// widened into dominating guards.
128   SmallVector<Instruction *, 16> EliminatedGuardsAndBranches;
129 
130   /// The set of guards which have been widened to include conditions to other
131   /// guards.
132   DenseSet<Instruction *> WidenedGuards;
133 
134   /// Try to eliminate instruction \p Instr by widening it into an earlier
135   /// dominating guard.  \p DFSI is the DFS iterator on the dominator tree that
136   /// is currently visiting the block containing \p Guard, and \p GuardsPerBlock
137   /// maps BasicBlocks to the set of guards seen in that block.
138   bool eliminateInstrViaWidening(
139       Instruction *Instr, const df_iterator<DomTreeNode *> &DFSI,
140       const DenseMap<BasicBlock *, SmallVector<Instruction *, 8>> &
141           GuardsPerBlock, bool InvertCondition = false);
142 
143   /// Used to keep track of which widening potential is more effective.
144   enum WideningScore {
145     /// Don't widen.
146     WS_IllegalOrNegative,
147 
148     /// Widening is performance neutral as far as the cycles spent in check
149     /// conditions goes (but can still help, e.g., code layout, having less
150     /// deopt state).
151     WS_Neutral,
152 
153     /// Widening is profitable.
154     WS_Positive,
155 
156     /// Widening is very profitable.  Not significantly different from \c
157     /// WS_Positive, except by the order.
158     WS_VeryPositive
159   };
160 
161   static StringRef scoreTypeToString(WideningScore WS);
162 
163   /// Compute the score for widening the condition in \p DominatedInstr
164   /// into \p DominatingGuard. If \p InvertCond is set, then we widen the
165   /// inverted condition of the dominating guard.
166   WideningScore computeWideningScore(Instruction *DominatedInstr,
167                                      Instruction *DominatingGuard,
168                                      bool InvertCond);
169 
170   /// Helper to check if \p V can be hoisted to \p InsertPos.
171   bool isAvailableAt(const Value *V, const Instruction *InsertPos) const {
172     SmallPtrSet<const Instruction *, 8> Visited;
173     return isAvailableAt(V, InsertPos, Visited);
174   }
175 
176   bool isAvailableAt(const Value *V, const Instruction *InsertPos,
177                      SmallPtrSetImpl<const Instruction *> &Visited) const;
178 
179   /// Helper to hoist \p V to \p InsertPos.  Guaranteed to succeed if \c
180   /// isAvailableAt returned true.
181   void makeAvailableAt(Value *V, Instruction *InsertPos) const;
182 
183   /// Common helper used by \c widenGuard and \c isWideningCondProfitable.  Try
184   /// to generate an expression computing the logical AND of \p Cond0 and (\p
185   /// Cond1 XOR \p InvertCondition).
186   /// Return true if the expression computing the AND is only as
187   /// expensive as computing one of the two. If \p InsertPt is true then
188   /// actually generate the resulting expression, make it available at \p
189   /// InsertPt and return it in \p Result (else no change to the IR is made).
190   bool widenCondCommon(Value *Cond0, Value *Cond1, Instruction *InsertPt,
191                        Value *&Result, bool InvertCondition);
192 
193   /// Represents a range check of the form \c Base + \c Offset u< \c Length,
194   /// with the constraint that \c Length is not negative.  \c CheckInst is the
195   /// pre-existing instruction in the IR that computes the result of this range
196   /// check.
197   class RangeCheck {
198     const Value *Base;
199     const ConstantInt *Offset;
200     const Value *Length;
201     ICmpInst *CheckInst;
202 
203   public:
204     explicit RangeCheck(const Value *Base, const ConstantInt *Offset,
205                         const Value *Length, ICmpInst *CheckInst)
206         : Base(Base), Offset(Offset), Length(Length), CheckInst(CheckInst) {}
207 
208     void setBase(const Value *NewBase) { Base = NewBase; }
209     void setOffset(const ConstantInt *NewOffset) { Offset = NewOffset; }
210 
211     const Value *getBase() const { return Base; }
212     const ConstantInt *getOffset() const { return Offset; }
213     const APInt &getOffsetValue() const { return getOffset()->getValue(); }
214     const Value *getLength() const { return Length; };
215     ICmpInst *getCheckInst() const { return CheckInst; }
216 
217     void print(raw_ostream &OS, bool PrintTypes = false) {
218       OS << "Base: ";
219       Base->printAsOperand(OS, PrintTypes);
220       OS << " Offset: ";
221       Offset->printAsOperand(OS, PrintTypes);
222       OS << " Length: ";
223       Length->printAsOperand(OS, PrintTypes);
224     }
225 
226     LLVM_DUMP_METHOD void dump() {
227       print(dbgs());
228       dbgs() << "\n";
229     }
230   };
231 
232   /// Parse \p CheckCond into a conjunction (logical-and) of range checks; and
233   /// append them to \p Checks.  Returns true on success, may clobber \c Checks
234   /// on failure.
235   bool parseRangeChecks(Value *CheckCond, SmallVectorImpl<RangeCheck> &Checks) {
236     SmallPtrSet<const Value *, 8> Visited;
237     return parseRangeChecks(CheckCond, Checks, Visited);
238   }
239 
240   bool parseRangeChecks(Value *CheckCond, SmallVectorImpl<RangeCheck> &Checks,
241                         SmallPtrSetImpl<const Value *> &Visited);
242 
243   /// Combine the checks in \p Checks into a smaller set of checks and append
244   /// them into \p CombinedChecks.  Return true on success (i.e. all of checks
245   /// in \p Checks were combined into \p CombinedChecks).  Clobbers \p Checks
246   /// and \p CombinedChecks on success and on failure.
247   bool combineRangeChecks(SmallVectorImpl<RangeCheck> &Checks,
248                           SmallVectorImpl<RangeCheck> &CombinedChecks) const;
249 
250   /// Can we compute the logical AND of \p Cond0 and \p Cond1 for the price of
251   /// computing only one of the two expressions?
252   bool isWideningCondProfitable(Value *Cond0, Value *Cond1, bool InvertCond) {
253     Value *ResultUnused;
254     return widenCondCommon(Cond0, Cond1, /*InsertPt=*/nullptr, ResultUnused,
255                            InvertCond);
256   }
257 
258   /// If \p InvertCondition is false, Widen \p ToWiden to fail if
259   /// \p NewCondition is false, otherwise make it fail if \p NewCondition is
260   /// true (in addition to whatever it is already checking).
261   void widenGuard(Instruction *ToWiden, Value *NewCondition,
262                   bool InvertCondition) {
263     Value *Result;
264 
265     widenCondCommon(getCondition(ToWiden), NewCondition, ToWiden, Result,
266                     InvertCondition);
267     if (isGuardAsWidenableBranch(ToWiden)) {
268       setWidenableBranchCond(cast<BranchInst>(ToWiden), Result);
269       return;
270     }
271     setCondition(ToWiden, Result);
272   }
273 
274 public:
275   explicit GuardWideningImpl(DominatorTree &DT, PostDominatorTree *PDT,
276                              LoopInfo &LI, MemorySSAUpdater *MSSAU,
277                              DomTreeNode *Root,
278                              std::function<bool(BasicBlock*)> BlockFilter)
279       : DT(DT), PDT(PDT), LI(LI), MSSAU(MSSAU), Root(Root),
280         BlockFilter(BlockFilter) {}
281 
282   /// The entry point for this pass.
283   bool run();
284 };
285 }
286 
287 static bool isSupportedGuardInstruction(const Instruction *Insn) {
288   if (isGuard(Insn))
289     return true;
290   if (WidenBranchGuards && isGuardAsWidenableBranch(Insn))
291     return true;
292   return false;
293 }
294 
295 bool GuardWideningImpl::run() {
296   DenseMap<BasicBlock *, SmallVector<Instruction *, 8>> GuardsInBlock;
297   bool Changed = false;
298   for (auto DFI = df_begin(Root), DFE = df_end(Root);
299        DFI != DFE; ++DFI) {
300     auto *BB = (*DFI)->getBlock();
301     if (!BlockFilter(BB))
302       continue;
303 
304     auto &CurrentList = GuardsInBlock[BB];
305 
306     for (auto &I : *BB)
307       if (isSupportedGuardInstruction(&I))
308         CurrentList.push_back(cast<Instruction>(&I));
309 
310     for (auto *II : CurrentList)
311       Changed |= eliminateInstrViaWidening(II, DFI, GuardsInBlock);
312   }
313 
314   assert(EliminatedGuardsAndBranches.empty() || Changed);
315   for (auto *I : EliminatedGuardsAndBranches)
316     if (!WidenedGuards.count(I)) {
317       assert(isa<ConstantInt>(getCondition(I)) && "Should be!");
318       if (isSupportedGuardInstruction(I))
319         eliminateGuard(I, MSSAU);
320       else {
321         assert(isa<BranchInst>(I) &&
322                "Eliminated something other than guard or branch?");
323         ++CondBranchEliminated;
324       }
325     }
326 
327   return Changed;
328 }
329 
330 bool GuardWideningImpl::eliminateInstrViaWidening(
331     Instruction *Instr, const df_iterator<DomTreeNode *> &DFSI,
332     const DenseMap<BasicBlock *, SmallVector<Instruction *, 8>> &
333         GuardsInBlock, bool InvertCondition) {
334   // Ignore trivial true or false conditions. These instructions will be
335   // trivially eliminated by any cleanup pass. Do not erase them because other
336   // guards can possibly be widened into them.
337   if (isa<ConstantInt>(getCondition(Instr)))
338     return false;
339 
340   Instruction *BestSoFar = nullptr;
341   auto BestScoreSoFar = WS_IllegalOrNegative;
342 
343   // In the set of dominating guards, find the one we can merge GuardInst with
344   // for the most profit.
345   for (unsigned i = 0, e = DFSI.getPathLength(); i != e; ++i) {
346     auto *CurBB = DFSI.getPath(i)->getBlock();
347     if (!BlockFilter(CurBB))
348       break;
349     assert(GuardsInBlock.count(CurBB) && "Must have been populated by now!");
350     const auto &GuardsInCurBB = GuardsInBlock.find(CurBB)->second;
351 
352     auto I = GuardsInCurBB.begin();
353     auto E = Instr->getParent() == CurBB ? find(GuardsInCurBB, Instr)
354                                          : GuardsInCurBB.end();
355 
356 #ifndef NDEBUG
357     {
358       unsigned Index = 0;
359       for (auto &I : *CurBB) {
360         if (Index == GuardsInCurBB.size())
361           break;
362         if (GuardsInCurBB[Index] == &I)
363           Index++;
364       }
365       assert(Index == GuardsInCurBB.size() &&
366              "Guards expected to be in order!");
367     }
368 #endif
369 
370     assert((i == (e - 1)) == (Instr->getParent() == CurBB) && "Bad DFS?");
371 
372     for (auto *Candidate : make_range(I, E)) {
373       auto Score = computeWideningScore(Instr, Candidate, InvertCondition);
374       LLVM_DEBUG(dbgs() << "Score between " << *getCondition(Instr)
375                         << " and " << *getCondition(Candidate) << " is "
376                         << scoreTypeToString(Score) << "\n");
377       if (Score > BestScoreSoFar) {
378         BestScoreSoFar = Score;
379         BestSoFar = Candidate;
380       }
381     }
382   }
383 
384   if (BestScoreSoFar == WS_IllegalOrNegative) {
385     LLVM_DEBUG(dbgs() << "Did not eliminate guard " << *Instr << "\n");
386     return false;
387   }
388 
389   assert(BestSoFar != Instr && "Should have never visited same guard!");
390   assert(DT.dominates(BestSoFar, Instr) && "Should be!");
391 
392   LLVM_DEBUG(dbgs() << "Widening " << *Instr << " into " << *BestSoFar
393                     << " with score " << scoreTypeToString(BestScoreSoFar)
394                     << "\n");
395   widenGuard(BestSoFar, getCondition(Instr), InvertCondition);
396   auto NewGuardCondition = InvertCondition
397                                ? ConstantInt::getFalse(Instr->getContext())
398                                : ConstantInt::getTrue(Instr->getContext());
399   setCondition(Instr, NewGuardCondition);
400   EliminatedGuardsAndBranches.push_back(Instr);
401   WidenedGuards.insert(BestSoFar);
402   return true;
403 }
404 
405 GuardWideningImpl::WideningScore
406 GuardWideningImpl::computeWideningScore(Instruction *DominatedInstr,
407                                         Instruction *DominatingGuard,
408                                         bool InvertCond) {
409   Loop *DominatedInstrLoop = LI.getLoopFor(DominatedInstr->getParent());
410   Loop *DominatingGuardLoop = LI.getLoopFor(DominatingGuard->getParent());
411   bool HoistingOutOfLoop = false;
412 
413   if (DominatingGuardLoop != DominatedInstrLoop) {
414     // Be conservative and don't widen into a sibling loop.  TODO: If the
415     // sibling is colder, we should consider allowing this.
416     if (DominatingGuardLoop &&
417         !DominatingGuardLoop->contains(DominatedInstrLoop))
418       return WS_IllegalOrNegative;
419 
420     HoistingOutOfLoop = true;
421   }
422 
423   if (!isAvailableAt(getCondition(DominatedInstr), DominatingGuard))
424     return WS_IllegalOrNegative;
425 
426   // If the guard was conditional executed, it may never be reached
427   // dynamically.  There are two potential downsides to hoisting it out of the
428   // conditionally executed region: 1) we may spuriously deopt without need and
429   // 2) we have the extra cost of computing the guard condition in the common
430   // case.  At the moment, we really only consider the second in our heuristic
431   // here.  TODO: evaluate cost model for spurious deopt
432   // NOTE: As written, this also lets us hoist right over another guard which
433   // is essentially just another spelling for control flow.
434   if (isWideningCondProfitable(getCondition(DominatedInstr),
435                                getCondition(DominatingGuard), InvertCond))
436     return HoistingOutOfLoop ? WS_VeryPositive : WS_Positive;
437 
438   if (HoistingOutOfLoop)
439     return WS_Positive;
440 
441   // Returns true if we might be hoisting above explicit control flow.  Note
442   // that this completely ignores implicit control flow (guards, calls which
443   // throw, etc...).  That choice appears arbitrary.
444   auto MaybeHoistingOutOfIf = [&]() {
445     auto *DominatingBlock = DominatingGuard->getParent();
446     auto *DominatedBlock = DominatedInstr->getParent();
447     if (isGuardAsWidenableBranch(DominatingGuard))
448       DominatingBlock = cast<BranchInst>(DominatingGuard)->getSuccessor(0);
449 
450     // Same Block?
451     if (DominatedBlock == DominatingBlock)
452       return false;
453     // Obvious successor (common loop header/preheader case)
454     if (DominatedBlock == DominatingBlock->getUniqueSuccessor())
455       return false;
456     // TODO: diamond, triangle cases
457     if (!PDT) return true;
458     return !PDT->dominates(DominatedBlock, DominatingBlock);
459   };
460 
461   return MaybeHoistingOutOfIf() ? WS_IllegalOrNegative : WS_Neutral;
462 }
463 
464 bool GuardWideningImpl::isAvailableAt(
465     const Value *V, const Instruction *Loc,
466     SmallPtrSetImpl<const Instruction *> &Visited) const {
467   auto *Inst = dyn_cast<Instruction>(V);
468   if (!Inst || DT.dominates(Inst, Loc) || Visited.count(Inst))
469     return true;
470 
471   if (!isSafeToSpeculativelyExecute(Inst, Loc, &DT) ||
472       Inst->mayReadFromMemory())
473     return false;
474 
475   Visited.insert(Inst);
476 
477   // We only want to go _up_ the dominance chain when recursing.
478   assert(!isa<PHINode>(Loc) &&
479          "PHIs should return false for isSafeToSpeculativelyExecute");
480   assert(DT.isReachableFromEntry(Inst->getParent()) &&
481          "We did a DFS from the block entry!");
482   return all_of(Inst->operands(),
483                 [&](Value *Op) { return isAvailableAt(Op, Loc, Visited); });
484 }
485 
486 void GuardWideningImpl::makeAvailableAt(Value *V, Instruction *Loc) const {
487   auto *Inst = dyn_cast<Instruction>(V);
488   if (!Inst || DT.dominates(Inst, Loc))
489     return;
490 
491   assert(isSafeToSpeculativelyExecute(Inst, Loc, &DT) &&
492          !Inst->mayReadFromMemory() && "Should've checked with isAvailableAt!");
493 
494   for (Value *Op : Inst->operands())
495     makeAvailableAt(Op, Loc);
496 
497   Inst->moveBefore(Loc);
498   // If we moved instruction before guard we must clean poison generating flags.
499   Inst->dropPoisonGeneratingFlags();
500 }
501 
502 bool GuardWideningImpl::widenCondCommon(Value *Cond0, Value *Cond1,
503                                         Instruction *InsertPt, Value *&Result,
504                                         bool InvertCondition) {
505   using namespace llvm::PatternMatch;
506 
507   {
508     // L >u C0 && L >u C1  ->  L >u max(C0, C1)
509     ConstantInt *RHS0, *RHS1;
510     Value *LHS;
511     ICmpInst::Predicate Pred0, Pred1;
512     if (match(Cond0, m_ICmp(Pred0, m_Value(LHS), m_ConstantInt(RHS0))) &&
513         match(Cond1, m_ICmp(Pred1, m_Specific(LHS), m_ConstantInt(RHS1)))) {
514       if (InvertCondition)
515         Pred1 = ICmpInst::getInversePredicate(Pred1);
516 
517       ConstantRange CR0 =
518           ConstantRange::makeExactICmpRegion(Pred0, RHS0->getValue());
519       ConstantRange CR1 =
520           ConstantRange::makeExactICmpRegion(Pred1, RHS1->getValue());
521 
522       // Given what we're doing here and the semantics of guards, it would
523       // be correct to use a subset intersection, but that may be too
524       // aggressive in cases we care about.
525       if (Optional<ConstantRange> Intersect = CR0.exactIntersectWith(CR1)) {
526         APInt NewRHSAP;
527         CmpInst::Predicate Pred;
528         if (Intersect->getEquivalentICmp(Pred, NewRHSAP)) {
529           if (InsertPt) {
530             ConstantInt *NewRHS =
531                 ConstantInt::get(Cond0->getContext(), NewRHSAP);
532             Result = new ICmpInst(InsertPt, Pred, LHS, NewRHS, "wide.chk");
533           }
534           return true;
535         }
536       }
537     }
538   }
539 
540   {
541     SmallVector<GuardWideningImpl::RangeCheck, 4> Checks, CombinedChecks;
542     // TODO: Support InvertCondition case?
543     if (!InvertCondition &&
544         parseRangeChecks(Cond0, Checks) && parseRangeChecks(Cond1, Checks) &&
545         combineRangeChecks(Checks, CombinedChecks)) {
546       if (InsertPt) {
547         Result = nullptr;
548         for (auto &RC : CombinedChecks) {
549           makeAvailableAt(RC.getCheckInst(), InsertPt);
550           if (Result)
551             Result = BinaryOperator::CreateAnd(RC.getCheckInst(), Result, "",
552                                                InsertPt);
553           else
554             Result = RC.getCheckInst();
555         }
556         assert(Result && "Failed to find result value");
557         Result->setName("wide.chk");
558       }
559       return true;
560     }
561   }
562 
563   // Base case -- just logical-and the two conditions together.
564 
565   if (InsertPt) {
566     makeAvailableAt(Cond0, InsertPt);
567     makeAvailableAt(Cond1, InsertPt);
568     if (InvertCondition)
569       Cond1 = BinaryOperator::CreateNot(Cond1, "inverted", InsertPt);
570     Result = BinaryOperator::CreateAnd(Cond0, Cond1, "wide.chk", InsertPt);
571   }
572 
573   // We were not able to compute Cond0 AND Cond1 for the price of one.
574   return false;
575 }
576 
577 bool GuardWideningImpl::parseRangeChecks(
578     Value *CheckCond, SmallVectorImpl<GuardWideningImpl::RangeCheck> &Checks,
579     SmallPtrSetImpl<const Value *> &Visited) {
580   if (!Visited.insert(CheckCond).second)
581     return true;
582 
583   using namespace llvm::PatternMatch;
584 
585   {
586     Value *AndLHS, *AndRHS;
587     if (match(CheckCond, m_And(m_Value(AndLHS), m_Value(AndRHS))))
588       return parseRangeChecks(AndLHS, Checks) &&
589              parseRangeChecks(AndRHS, Checks);
590   }
591 
592   auto *IC = dyn_cast<ICmpInst>(CheckCond);
593   if (!IC || !IC->getOperand(0)->getType()->isIntegerTy() ||
594       (IC->getPredicate() != ICmpInst::ICMP_ULT &&
595        IC->getPredicate() != ICmpInst::ICMP_UGT))
596     return false;
597 
598   const Value *CmpLHS = IC->getOperand(0), *CmpRHS = IC->getOperand(1);
599   if (IC->getPredicate() == ICmpInst::ICMP_UGT)
600     std::swap(CmpLHS, CmpRHS);
601 
602   auto &DL = IC->getModule()->getDataLayout();
603 
604   GuardWideningImpl::RangeCheck Check(
605       CmpLHS, cast<ConstantInt>(ConstantInt::getNullValue(CmpRHS->getType())),
606       CmpRHS, IC);
607 
608   if (!isKnownNonNegative(Check.getLength(), DL))
609     return false;
610 
611   // What we have in \c Check now is a correct interpretation of \p CheckCond.
612   // Try to see if we can move some constant offsets into the \c Offset field.
613 
614   bool Changed;
615   auto &Ctx = CheckCond->getContext();
616 
617   do {
618     Value *OpLHS;
619     ConstantInt *OpRHS;
620     Changed = false;
621 
622 #ifndef NDEBUG
623     auto *BaseInst = dyn_cast<Instruction>(Check.getBase());
624     assert((!BaseInst || DT.isReachableFromEntry(BaseInst->getParent())) &&
625            "Unreachable instruction?");
626 #endif
627 
628     if (match(Check.getBase(), m_Add(m_Value(OpLHS), m_ConstantInt(OpRHS)))) {
629       Check.setBase(OpLHS);
630       APInt NewOffset = Check.getOffsetValue() + OpRHS->getValue();
631       Check.setOffset(ConstantInt::get(Ctx, NewOffset));
632       Changed = true;
633     } else if (match(Check.getBase(),
634                      m_Or(m_Value(OpLHS), m_ConstantInt(OpRHS)))) {
635       KnownBits Known = computeKnownBits(OpLHS, DL);
636       if ((OpRHS->getValue() & Known.Zero) == OpRHS->getValue()) {
637         Check.setBase(OpLHS);
638         APInt NewOffset = Check.getOffsetValue() + OpRHS->getValue();
639         Check.setOffset(ConstantInt::get(Ctx, NewOffset));
640         Changed = true;
641       }
642     }
643   } while (Changed);
644 
645   Checks.push_back(Check);
646   return true;
647 }
648 
649 bool GuardWideningImpl::combineRangeChecks(
650     SmallVectorImpl<GuardWideningImpl::RangeCheck> &Checks,
651     SmallVectorImpl<GuardWideningImpl::RangeCheck> &RangeChecksOut) const {
652   unsigned OldCount = Checks.size();
653   while (!Checks.empty()) {
654     // Pick all of the range checks with a specific base and length, and try to
655     // merge them.
656     const Value *CurrentBase = Checks.front().getBase();
657     const Value *CurrentLength = Checks.front().getLength();
658 
659     SmallVector<GuardWideningImpl::RangeCheck, 3> CurrentChecks;
660 
661     auto IsCurrentCheck = [&](GuardWideningImpl::RangeCheck &RC) {
662       return RC.getBase() == CurrentBase && RC.getLength() == CurrentLength;
663     };
664 
665     copy_if(Checks, std::back_inserter(CurrentChecks), IsCurrentCheck);
666     erase_if(Checks, IsCurrentCheck);
667 
668     assert(CurrentChecks.size() != 0 && "We know we have at least one!");
669 
670     if (CurrentChecks.size() < 3) {
671       llvm::append_range(RangeChecksOut, CurrentChecks);
672       continue;
673     }
674 
675     // CurrentChecks.size() will typically be 3 here, but so far there has been
676     // no need to hard-code that fact.
677 
678     llvm::sort(CurrentChecks, [&](const GuardWideningImpl::RangeCheck &LHS,
679                                   const GuardWideningImpl::RangeCheck &RHS) {
680       return LHS.getOffsetValue().slt(RHS.getOffsetValue());
681     });
682 
683     // Note: std::sort should not invalidate the ChecksStart iterator.
684 
685     const ConstantInt *MinOffset = CurrentChecks.front().getOffset();
686     const ConstantInt *MaxOffset = CurrentChecks.back().getOffset();
687 
688     unsigned BitWidth = MaxOffset->getValue().getBitWidth();
689     if ((MaxOffset->getValue() - MinOffset->getValue())
690             .ugt(APInt::getSignedMinValue(BitWidth)))
691       return false;
692 
693     APInt MaxDiff = MaxOffset->getValue() - MinOffset->getValue();
694     const APInt &HighOffset = MaxOffset->getValue();
695     auto OffsetOK = [&](const GuardWideningImpl::RangeCheck &RC) {
696       return (HighOffset - RC.getOffsetValue()).ult(MaxDiff);
697     };
698 
699     if (MaxDiff.isMinValue() || !all_of(drop_begin(CurrentChecks), OffsetOK))
700       return false;
701 
702     // We have a series of f+1 checks as:
703     //
704     //   I+k_0 u< L   ... Chk_0
705     //   I+k_1 u< L   ... Chk_1
706     //   ...
707     //   I+k_f u< L   ... Chk_f
708     //
709     //     with forall i in [0,f]: k_f-k_i u< k_f-k_0  ... Precond_0
710     //          k_f-k_0 u< INT_MIN+k_f                 ... Precond_1
711     //          k_f != k_0                             ... Precond_2
712     //
713     // Claim:
714     //   Chk_0 AND Chk_f  implies all the other checks
715     //
716     // Informal proof sketch:
717     //
718     // We will show that the integer range [I+k_0,I+k_f] does not unsigned-wrap
719     // (i.e. going from I+k_0 to I+k_f does not cross the -1,0 boundary) and
720     // thus I+k_f is the greatest unsigned value in that range.
721     //
722     // This combined with Ckh_(f+1) shows that everything in that range is u< L.
723     // Via Precond_0 we know that all of the indices in Chk_0 through Chk_(f+1)
724     // lie in [I+k_0,I+k_f], this proving our claim.
725     //
726     // To see that [I+k_0,I+k_f] is not a wrapping range, note that there are
727     // two possibilities: I+k_0 u< I+k_f or I+k_0 >u I+k_f (they can't be equal
728     // since k_0 != k_f).  In the former case, [I+k_0,I+k_f] is not a wrapping
729     // range by definition, and the latter case is impossible:
730     //
731     //   0-----I+k_f---I+k_0----L---INT_MAX,INT_MIN------------------(-1)
732     //   xxxxxx             xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
733     //
734     // For Chk_0 to succeed, we'd have to have k_f-k_0 (the range highlighted
735     // with 'x' above) to be at least >u INT_MIN.
736 
737     RangeChecksOut.emplace_back(CurrentChecks.front());
738     RangeChecksOut.emplace_back(CurrentChecks.back());
739   }
740 
741   assert(RangeChecksOut.size() <= OldCount && "We pessimized!");
742   return RangeChecksOut.size() != OldCount;
743 }
744 
745 #ifndef NDEBUG
746 StringRef GuardWideningImpl::scoreTypeToString(WideningScore WS) {
747   switch (WS) {
748   case WS_IllegalOrNegative:
749     return "IllegalOrNegative";
750   case WS_Neutral:
751     return "Neutral";
752   case WS_Positive:
753     return "Positive";
754   case WS_VeryPositive:
755     return "VeryPositive";
756   }
757 
758   llvm_unreachable("Fully covered switch above!");
759 }
760 #endif
761 
762 PreservedAnalyses GuardWideningPass::run(Function &F,
763                                          FunctionAnalysisManager &AM) {
764   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
765   auto &LI = AM.getResult<LoopAnalysis>(F);
766   auto &PDT = AM.getResult<PostDominatorTreeAnalysis>(F);
767   auto *MSSAA = AM.getCachedResult<MemorySSAAnalysis>(F);
768   std::unique_ptr<MemorySSAUpdater> MSSAU;
769   if (MSSAA)
770     MSSAU = std::make_unique<MemorySSAUpdater>(&MSSAA->getMSSA());
771   if (!GuardWideningImpl(DT, &PDT, LI, MSSAU ? MSSAU.get() : nullptr,
772                          DT.getRootNode(), [](BasicBlock *) { return true; })
773            .run())
774     return PreservedAnalyses::all();
775 
776   PreservedAnalyses PA;
777   PA.preserveSet<CFGAnalyses>();
778   PA.preserve<MemorySSAAnalysis>();
779   return PA;
780 }
781 
782 PreservedAnalyses GuardWideningPass::run(Loop &L, LoopAnalysisManager &AM,
783                                          LoopStandardAnalysisResults &AR,
784                                          LPMUpdater &U) {
785   BasicBlock *RootBB = L.getLoopPredecessor();
786   if (!RootBB)
787     RootBB = L.getHeader();
788   auto BlockFilter = [&](BasicBlock *BB) {
789     return BB == RootBB || L.contains(BB);
790   };
791   std::unique_ptr<MemorySSAUpdater> MSSAU;
792   if (AR.MSSA)
793     MSSAU = std::make_unique<MemorySSAUpdater>(AR.MSSA);
794   if (!GuardWideningImpl(AR.DT, nullptr, AR.LI, MSSAU ? MSSAU.get() : nullptr,
795                          AR.DT.getNode(RootBB), BlockFilter).run())
796     return PreservedAnalyses::all();
797 
798   auto PA = getLoopPassPreservedAnalyses();
799   if (AR.MSSA)
800     PA.preserve<MemorySSAAnalysis>();
801   return PA;
802 }
803 
804 namespace {
805 struct GuardWideningLegacyPass : public FunctionPass {
806   static char ID;
807 
808   GuardWideningLegacyPass() : FunctionPass(ID) {
809     initializeGuardWideningLegacyPassPass(*PassRegistry::getPassRegistry());
810   }
811 
812   bool runOnFunction(Function &F) override {
813     if (skipFunction(F))
814       return false;
815     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
816     auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
817     auto &PDT = getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree();
818     auto *MSSAWP = getAnalysisIfAvailable<MemorySSAWrapperPass>();
819     std::unique_ptr<MemorySSAUpdater> MSSAU;
820     if (MSSAWP)
821       MSSAU = std::make_unique<MemorySSAUpdater>(&MSSAWP->getMSSA());
822     return GuardWideningImpl(DT, &PDT, LI, MSSAU ? MSSAU.get() : nullptr,
823                              DT.getRootNode(),
824                              [](BasicBlock *) { return true; })
825         .run();
826   }
827 
828   void getAnalysisUsage(AnalysisUsage &AU) const override {
829     AU.setPreservesCFG();
830     AU.addRequired<DominatorTreeWrapperPass>();
831     AU.addRequired<PostDominatorTreeWrapperPass>();
832     AU.addRequired<LoopInfoWrapperPass>();
833     AU.addPreserved<MemorySSAWrapperPass>();
834   }
835 };
836 
837 /// Same as above, but restricted to a single loop at a time.  Can be
838 /// scheduled with other loop passes w/o breaking out of LPM
839 struct LoopGuardWideningLegacyPass : public LoopPass {
840   static char ID;
841 
842   LoopGuardWideningLegacyPass() : LoopPass(ID) {
843     initializeLoopGuardWideningLegacyPassPass(*PassRegistry::getPassRegistry());
844   }
845 
846   bool runOnLoop(Loop *L, LPPassManager &LPM) override {
847     if (skipLoop(L))
848       return false;
849     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
850     auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
851     auto *PDTWP = getAnalysisIfAvailable<PostDominatorTreeWrapperPass>();
852     auto *PDT = PDTWP ? &PDTWP->getPostDomTree() : nullptr;
853     auto *MSSAWP = getAnalysisIfAvailable<MemorySSAWrapperPass>();
854     std::unique_ptr<MemorySSAUpdater> MSSAU;
855     if (MSSAWP)
856       MSSAU = std::make_unique<MemorySSAUpdater>(&MSSAWP->getMSSA());
857 
858     BasicBlock *RootBB = L->getLoopPredecessor();
859     if (!RootBB)
860       RootBB = L->getHeader();
861     auto BlockFilter = [&](BasicBlock *BB) {
862       return BB == RootBB || L->contains(BB);
863     };
864     return GuardWideningImpl(DT, PDT, LI, MSSAU ? MSSAU.get() : nullptr,
865                              DT.getNode(RootBB), BlockFilter).run();
866   }
867 
868   void getAnalysisUsage(AnalysisUsage &AU) const override {
869     AU.setPreservesCFG();
870     getLoopAnalysisUsage(AU);
871     AU.addPreserved<PostDominatorTreeWrapperPass>();
872     AU.addPreserved<MemorySSAWrapperPass>();
873   }
874 };
875 }
876 
877 char GuardWideningLegacyPass::ID = 0;
878 char LoopGuardWideningLegacyPass::ID = 0;
879 
880 INITIALIZE_PASS_BEGIN(GuardWideningLegacyPass, "guard-widening", "Widen guards",
881                       false, false)
882 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
883 INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)
884 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
885 INITIALIZE_PASS_END(GuardWideningLegacyPass, "guard-widening", "Widen guards",
886                     false, false)
887 
888 INITIALIZE_PASS_BEGIN(LoopGuardWideningLegacyPass, "loop-guard-widening",
889                       "Widen guards (within a single loop, as a loop pass)",
890                       false, false)
891 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
892 INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)
893 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
894 INITIALIZE_PASS_END(LoopGuardWideningLegacyPass, "loop-guard-widening",
895                     "Widen guards (within a single loop, as a loop pass)",
896                     false, false)
897 
898 FunctionPass *llvm::createGuardWideningPass() {
899   return new GuardWideningLegacyPass();
900 }
901 
902 Pass *llvm::createLoopGuardWideningPass() {
903   return new LoopGuardWideningLegacyPass();
904 }
905