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/BranchProbabilityInfo.h"
46 #include "llvm/Analysis/GuardUtils.h"
47 #include "llvm/Analysis/LoopInfo.h"
48 #include "llvm/Analysis/LoopPass.h"
49 #include "llvm/Analysis/MemorySSAUpdater.h"
50 #include "llvm/Analysis/PostDominators.h"
51 #include "llvm/Analysis/ValueTracking.h"
52 #include "llvm/IR/ConstantRange.h"
53 #include "llvm/IR/Dominators.h"
54 #include "llvm/IR/IntrinsicInst.h"
55 #include "llvm/IR/PatternMatch.h"
56 #include "llvm/InitializePasses.h"
57 #include "llvm/Pass.h"
58 #include "llvm/Support/CommandLine.h"
59 #include "llvm/Support/Debug.h"
60 #include "llvm/Support/KnownBits.h"
61 #include "llvm/Transforms/Scalar.h"
62 #include "llvm/Transforms/Utils/GuardUtils.h"
63 #include "llvm/Transforms/Utils/LoopUtils.h"
64 #include <functional>
65 
66 using namespace llvm;
67 
68 #define DEBUG_TYPE "guard-widening"
69 
70 STATISTIC(GuardsEliminated, "Number of eliminated guards");
71 STATISTIC(CondBranchEliminated, "Number of eliminated conditional branches");
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.
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.
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.
109 static void eliminateGuard(Instruction *GuardInst, MemorySSAUpdater *MSSAU) {
110   GuardInst->eraseFromParent();
111   if (MSSAU)
112     MSSAU->removeMemoryAccess(GuardInst);
113   ++GuardsEliminated;
114 }
115 
116 class GuardWideningImpl {
117   DominatorTree &DT;
118   PostDominatorTree *PDT;
119   LoopInfo &LI;
120   MemorySSAUpdater *MSSAU;
121 
122   /// Together, these describe the region of interest.  This might be all of
123   /// the blocks within a function, or only a given loop's blocks and preheader.
124   DomTreeNode *Root;
125   std::function<bool(BasicBlock*)> BlockFilter;
126 
127   /// The set of guards and conditional branches whose conditions have been
128   /// widened into dominating guards.
129   SmallVector<Instruction *, 16> EliminatedGuardsAndBranches;
130 
131   /// The set of guards which have been widened to include conditions to other
132   /// guards.
133   DenseSet<Instruction *> WidenedGuards;
134 
135   /// Try to eliminate instruction \p Instr by widening it into an earlier
136   /// dominating guard.  \p DFSI is the DFS iterator on the dominator tree that
137   /// is currently visiting the block containing \p Guard, and \p GuardsPerBlock
138   /// maps BasicBlocks to the set of guards seen in that block.
139   bool eliminateInstrViaWidening(
140       Instruction *Instr, const df_iterator<DomTreeNode *> &DFSI,
141       const DenseMap<BasicBlock *, SmallVector<Instruction *, 8>> &
142           GuardsPerBlock, bool InvertCondition = false);
143 
144   /// Used to keep track of which widening potential is more effective.
145   enum WideningScore {
146     /// Don't widen.
147     WS_IllegalOrNegative,
148 
149     /// Widening is performance neutral as far as the cycles spent in check
150     /// conditions goes (but can still help, e.g., code layout, having less
151     /// deopt state).
152     WS_Neutral,
153 
154     /// Widening is profitable.
155     WS_Positive,
156 
157     /// Widening is very profitable.  Not significantly different from \c
158     /// WS_Positive, except by the order.
159     WS_VeryPositive
160   };
161 
162   static StringRef scoreTypeToString(WideningScore WS);
163 
164   /// Compute the score for widening the condition in \p DominatedInstr
165   /// into \p DominatingGuard. If \p InvertCond is set, then we widen the
166   /// inverted condition of the dominating guard.
167   WideningScore computeWideningScore(Instruction *DominatedInstr,
168                                      Instruction *DominatingGuard,
169                                      bool InvertCond);
170 
171   /// Helper to check if \p V can be hoisted to \p InsertPos.
172   bool isAvailableAt(const Value *V, const Instruction *InsertPos) const {
173     SmallPtrSet<const Instruction *, 8> Visited;
174     return isAvailableAt(V, InsertPos, Visited);
175   }
176 
177   bool isAvailableAt(const Value *V, const Instruction *InsertPos,
178                      SmallPtrSetImpl<const Instruction *> &Visited) const;
179 
180   /// Helper to hoist \p V to \p InsertPos.  Guaranteed to succeed if \c
181   /// isAvailableAt returned true.
182   void makeAvailableAt(Value *V, Instruction *InsertPos) const;
183 
184   /// Common helper used by \c widenGuard and \c isWideningCondProfitable.  Try
185   /// to generate an expression computing the logical AND of \p Cond0 and (\p
186   /// Cond1 XOR \p InvertCondition).
187   /// Return true if the expression computing the AND is only as
188   /// expensive as computing one of the two. If \p InsertPt is true then
189   /// actually generate the resulting expression, make it available at \p
190   /// InsertPt and return it in \p Result (else no change to the IR is made).
191   bool widenCondCommon(Value *Cond0, Value *Cond1, Instruction *InsertPt,
192                        Value *&Result, bool InvertCondition);
193 
194   /// Represents a range check of the form \c Base + \c Offset u< \c Length,
195   /// with the constraint that \c Length is not negative.  \c CheckInst is the
196   /// pre-existing instruction in the IR that computes the result of this range
197   /// check.
198   class RangeCheck {
199     const Value *Base;
200     const ConstantInt *Offset;
201     const Value *Length;
202     ICmpInst *CheckInst;
203 
204   public:
205     explicit RangeCheck(const Value *Base, const ConstantInt *Offset,
206                         const Value *Length, ICmpInst *CheckInst)
207         : Base(Base), Offset(Offset), Length(Length), CheckInst(CheckInst) {}
208 
209     void setBase(const Value *NewBase) { Base = NewBase; }
210     void setOffset(const ConstantInt *NewOffset) { Offset = NewOffset; }
211 
212     const Value *getBase() const { return Base; }
213     const ConstantInt *getOffset() const { return Offset; }
214     const APInt &getOffsetValue() const { return getOffset()->getValue(); }
215     const Value *getLength() const { return Length; };
216     ICmpInst *getCheckInst() const { return CheckInst; }
217 
218     void print(raw_ostream &OS, bool PrintTypes = false) {
219       OS << "Base: ";
220       Base->printAsOperand(OS, PrintTypes);
221       OS << " Offset: ";
222       Offset->printAsOperand(OS, PrintTypes);
223       OS << " Length: ";
224       Length->printAsOperand(OS, PrintTypes);
225     }
226 
227     LLVM_DUMP_METHOD void dump() {
228       print(dbgs());
229       dbgs() << "\n";
230     }
231   };
232 
233   /// Parse \p CheckCond into a conjunction (logical-and) of range checks; and
234   /// append them to \p Checks.  Returns true on success, may clobber \c Checks
235   /// on failure.
236   bool parseRangeChecks(Value *CheckCond, SmallVectorImpl<RangeCheck> &Checks) {
237     SmallPtrSet<const Value *, 8> Visited;
238     return parseRangeChecks(CheckCond, Checks, Visited);
239   }
240 
241   bool parseRangeChecks(Value *CheckCond, SmallVectorImpl<RangeCheck> &Checks,
242                         SmallPtrSetImpl<const Value *> &Visited);
243 
244   /// Combine the checks in \p Checks into a smaller set of checks and append
245   /// them into \p CombinedChecks.  Return true on success (i.e. all of checks
246   /// in \p Checks were combined into \p CombinedChecks).  Clobbers \p Checks
247   /// and \p CombinedChecks on success and on failure.
248   bool combineRangeChecks(SmallVectorImpl<RangeCheck> &Checks,
249                           SmallVectorImpl<RangeCheck> &CombinedChecks) const;
250 
251   /// Can we compute the logical AND of \p Cond0 and \p Cond1 for the price of
252   /// computing only one of the two expressions?
253   bool isWideningCondProfitable(Value *Cond0, Value *Cond1, bool InvertCond) {
254     Value *ResultUnused;
255     return widenCondCommon(Cond0, Cond1, /*InsertPt=*/nullptr, ResultUnused,
256                            InvertCond);
257   }
258 
259   /// If \p InvertCondition is false, Widen \p ToWiden to fail if
260   /// \p NewCondition is false, otherwise make it fail if \p NewCondition is
261   /// true (in addition to whatever it is already checking).
262   void widenGuard(Instruction *ToWiden, Value *NewCondition,
263                   bool InvertCondition) {
264     Value *Result;
265 
266     widenCondCommon(getCondition(ToWiden), NewCondition, ToWiden, Result,
267                     InvertCondition);
268     if (isGuardAsWidenableBranch(ToWiden)) {
269       setWidenableBranchCond(cast<BranchInst>(ToWiden), Result);
270       return;
271     }
272     setCondition(ToWiden, Result);
273   }
274 
275 public:
276   explicit GuardWideningImpl(DominatorTree &DT, PostDominatorTree *PDT,
277                              LoopInfo &LI, MemorySSAUpdater *MSSAU,
278                              DomTreeNode *Root,
279                              std::function<bool(BasicBlock*)> BlockFilter)
280       : DT(DT), PDT(PDT), LI(LI), MSSAU(MSSAU), Root(Root),
281         BlockFilter(BlockFilter) {}
282 
283   /// The entry point for this pass.
284   bool run();
285 };
286 }
287 
288 static bool isSupportedGuardInstruction(const Instruction *Insn) {
289   if (isGuard(Insn))
290     return true;
291   if (WidenBranchGuards && isGuardAsWidenableBranch(Insn))
292     return true;
293   return false;
294 }
295 
296 bool GuardWideningImpl::run() {
297   DenseMap<BasicBlock *, SmallVector<Instruction *, 8>> GuardsInBlock;
298   bool Changed = false;
299   for (auto DFI = df_begin(Root), DFE = df_end(Root);
300        DFI != DFE; ++DFI) {
301     auto *BB = (*DFI)->getBlock();
302     if (!BlockFilter(BB))
303       continue;
304 
305     auto &CurrentList = GuardsInBlock[BB];
306 
307     for (auto &I : *BB)
308       if (isSupportedGuardInstruction(&I))
309         CurrentList.push_back(cast<Instruction>(&I));
310 
311     for (auto *II : CurrentList)
312       Changed |= eliminateInstrViaWidening(II, DFI, GuardsInBlock);
313   }
314 
315   assert(EliminatedGuardsAndBranches.empty() || Changed);
316   for (auto *I : EliminatedGuardsAndBranches)
317     if (!WidenedGuards.count(I)) {
318       assert(isa<ConstantInt>(getCondition(I)) && "Should be!");
319       if (isSupportedGuardInstruction(I))
320         eliminateGuard(I, MSSAU);
321       else {
322         assert(isa<BranchInst>(I) &&
323                "Eliminated something other than guard or branch?");
324         ++CondBranchEliminated;
325       }
326     }
327 
328   return Changed;
329 }
330 
331 bool GuardWideningImpl::eliminateInstrViaWidening(
332     Instruction *Instr, const df_iterator<DomTreeNode *> &DFSI,
333     const DenseMap<BasicBlock *, SmallVector<Instruction *, 8>> &
334         GuardsInBlock, bool InvertCondition) {
335   // Ignore trivial true or false conditions. These instructions will be
336   // trivially eliminated by any cleanup pass. Do not erase them because other
337   // guards can possibly be widened into them.
338   if (isa<ConstantInt>(getCondition(Instr)))
339     return false;
340 
341   Instruction *BestSoFar = nullptr;
342   auto BestScoreSoFar = WS_IllegalOrNegative;
343 
344   // In the set of dominating guards, find the one we can merge GuardInst with
345   // for the most profit.
346   for (unsigned i = 0, e = DFSI.getPathLength(); i != e; ++i) {
347     auto *CurBB = DFSI.getPath(i)->getBlock();
348     if (!BlockFilter(CurBB))
349       break;
350     assert(GuardsInBlock.count(CurBB) && "Must have been populated by now!");
351     const auto &GuardsInCurBB = GuardsInBlock.find(CurBB)->second;
352 
353     auto I = GuardsInCurBB.begin();
354     auto E = Instr->getParent() == CurBB ? find(GuardsInCurBB, Instr)
355                                          : GuardsInCurBB.end();
356 
357 #ifndef NDEBUG
358     {
359       unsigned Index = 0;
360       for (auto &I : *CurBB) {
361         if (Index == GuardsInCurBB.size())
362           break;
363         if (GuardsInCurBB[Index] == &I)
364           Index++;
365       }
366       assert(Index == GuardsInCurBB.size() &&
367              "Guards expected to be in order!");
368     }
369 #endif
370 
371     assert((i == (e - 1)) == (Instr->getParent() == CurBB) && "Bad DFS?");
372 
373     for (auto *Candidate : make_range(I, E)) {
374       auto Score = computeWideningScore(Instr, Candidate, InvertCondition);
375       LLVM_DEBUG(dbgs() << "Score between " << *getCondition(Instr)
376                         << " and " << *getCondition(Candidate) << " is "
377                         << scoreTypeToString(Score) << "\n");
378       if (Score > BestScoreSoFar) {
379         BestScoreSoFar = Score;
380         BestSoFar = Candidate;
381       }
382     }
383   }
384 
385   if (BestScoreSoFar == WS_IllegalOrNegative) {
386     LLVM_DEBUG(dbgs() << "Did not eliminate guard " << *Instr << "\n");
387     return false;
388   }
389 
390   assert(BestSoFar != Instr && "Should have never visited same guard!");
391   assert(DT.dominates(BestSoFar, Instr) && "Should be!");
392 
393   LLVM_DEBUG(dbgs() << "Widening " << *Instr << " into " << *BestSoFar
394                     << " with score " << scoreTypeToString(BestScoreSoFar)
395                     << "\n");
396   widenGuard(BestSoFar, getCondition(Instr), InvertCondition);
397   auto NewGuardCondition = InvertCondition
398                                ? ConstantInt::getFalse(Instr->getContext())
399                                : ConstantInt::getTrue(Instr->getContext());
400   setCondition(Instr, NewGuardCondition);
401   EliminatedGuardsAndBranches.push_back(Instr);
402   WidenedGuards.insert(BestSoFar);
403   return true;
404 }
405 
406 GuardWideningImpl::WideningScore
407 GuardWideningImpl::computeWideningScore(Instruction *DominatedInstr,
408                                         Instruction *DominatingGuard,
409                                         bool InvertCond) {
410   Loop *DominatedInstrLoop = LI.getLoopFor(DominatedInstr->getParent());
411   Loop *DominatingGuardLoop = LI.getLoopFor(DominatingGuard->getParent());
412   bool HoistingOutOfLoop = false;
413 
414   if (DominatingGuardLoop != DominatedInstrLoop) {
415     // Be conservative and don't widen into a sibling loop.  TODO: If the
416     // sibling is colder, we should consider allowing this.
417     if (DominatingGuardLoop &&
418         !DominatingGuardLoop->contains(DominatedInstrLoop))
419       return WS_IllegalOrNegative;
420 
421     HoistingOutOfLoop = true;
422   }
423 
424   if (!isAvailableAt(getCondition(DominatedInstr), DominatingGuard))
425     return WS_IllegalOrNegative;
426 
427   // If the guard was conditional executed, it may never be reached
428   // dynamically.  There are two potential downsides to hoisting it out of the
429   // conditionally executed region: 1) we may spuriously deopt without need and
430   // 2) we have the extra cost of computing the guard condition in the common
431   // case.  At the moment, we really only consider the second in our heuristic
432   // here.  TODO: evaluate cost model for spurious deopt
433   // NOTE: As written, this also lets us hoist right over another guard which
434   // is essentially just another spelling for control flow.
435   if (isWideningCondProfitable(getCondition(DominatedInstr),
436                                getCondition(DominatingGuard), InvertCond))
437     return HoistingOutOfLoop ? WS_VeryPositive : WS_Positive;
438 
439   if (HoistingOutOfLoop)
440     return WS_Positive;
441 
442   // Returns true if we might be hoisting above explicit control flow.  Note
443   // that this completely ignores implicit control flow (guards, calls which
444   // throw, etc...).  That choice appears arbitrary.
445   auto MaybeHoistingOutOfIf = [&]() {
446     auto *DominatingBlock = DominatingGuard->getParent();
447     auto *DominatedBlock = DominatedInstr->getParent();
448     if (isGuardAsWidenableBranch(DominatingGuard))
449       DominatingBlock = cast<BranchInst>(DominatingGuard)->getSuccessor(0);
450 
451     // Same Block?
452     if (DominatedBlock == DominatingBlock)
453       return false;
454     // Obvious successor (common loop header/preheader case)
455     if (DominatedBlock == DominatingBlock->getUniqueSuccessor())
456       return false;
457     // TODO: diamond, triangle cases
458     if (!PDT) return true;
459     return !PDT->dominates(DominatedBlock, DominatingBlock);
460   };
461 
462   return MaybeHoistingOutOfIf() ? WS_IllegalOrNegative : WS_Neutral;
463 }
464 
465 bool GuardWideningImpl::isAvailableAt(
466     const Value *V, const Instruction *Loc,
467     SmallPtrSetImpl<const Instruction *> &Visited) const {
468   auto *Inst = dyn_cast<Instruction>(V);
469   if (!Inst || DT.dominates(Inst, Loc) || Visited.count(Inst))
470     return true;
471 
472   if (!isSafeToSpeculativelyExecute(Inst, Loc, &DT) ||
473       Inst->mayReadFromMemory())
474     return false;
475 
476   Visited.insert(Inst);
477 
478   // We only want to go _up_ the dominance chain when recursing.
479   assert(!isa<PHINode>(Loc) &&
480          "PHIs should return false for isSafeToSpeculativelyExecute");
481   assert(DT.isReachableFromEntry(Inst->getParent()) &&
482          "We did a DFS from the block entry!");
483   return all_of(Inst->operands(),
484                 [&](Value *Op) { return isAvailableAt(Op, Loc, Visited); });
485 }
486 
487 void GuardWideningImpl::makeAvailableAt(Value *V, Instruction *Loc) const {
488   auto *Inst = dyn_cast<Instruction>(V);
489   if (!Inst || DT.dominates(Inst, Loc))
490     return;
491 
492   assert(isSafeToSpeculativelyExecute(Inst, Loc, &DT) &&
493          !Inst->mayReadFromMemory() && "Should've checked with isAvailableAt!");
494 
495   for (Value *Op : Inst->operands())
496     makeAvailableAt(Op, Loc);
497 
498   Inst->moveBefore(Loc);
499 }
500 
501 bool GuardWideningImpl::widenCondCommon(Value *Cond0, Value *Cond1,
502                                         Instruction *InsertPt, Value *&Result,
503                                         bool InvertCondition) {
504   using namespace llvm::PatternMatch;
505 
506   {
507     // L >u C0 && L >u C1  ->  L >u max(C0, C1)
508     ConstantInt *RHS0, *RHS1;
509     Value *LHS;
510     ICmpInst::Predicate Pred0, Pred1;
511     if (match(Cond0, m_ICmp(Pred0, m_Value(LHS), m_ConstantInt(RHS0))) &&
512         match(Cond1, m_ICmp(Pred1, m_Specific(LHS), m_ConstantInt(RHS1)))) {
513       if (InvertCondition)
514         Pred1 = ICmpInst::getInversePredicate(Pred1);
515 
516       ConstantRange CR0 =
517           ConstantRange::makeExactICmpRegion(Pred0, RHS0->getValue());
518       ConstantRange CR1 =
519           ConstantRange::makeExactICmpRegion(Pred1, RHS1->getValue());
520 
521       // Given what we're doing here and the semantics of guards, it would
522       // be correct to use a subset intersection, but that may be too
523       // aggressive in cases we care about.
524       if (Optional<ConstantRange> Intersect = CR0.exactIntersectWith(CR1)) {
525         APInt NewRHSAP;
526         CmpInst::Predicate Pred;
527         if (Intersect->getEquivalentICmp(Pred, NewRHSAP)) {
528           if (InsertPt) {
529             ConstantInt *NewRHS =
530                 ConstantInt::get(Cond0->getContext(), NewRHSAP);
531             Result = new ICmpInst(InsertPt, Pred, LHS, NewRHS, "wide.chk");
532           }
533           return true;
534         }
535       }
536     }
537   }
538 
539   {
540     SmallVector<GuardWideningImpl::RangeCheck, 4> Checks, CombinedChecks;
541     // TODO: Support InvertCondition case?
542     if (!InvertCondition &&
543         parseRangeChecks(Cond0, Checks) && parseRangeChecks(Cond1, Checks) &&
544         combineRangeChecks(Checks, CombinedChecks)) {
545       if (InsertPt) {
546         Result = nullptr;
547         for (auto &RC : CombinedChecks) {
548           makeAvailableAt(RC.getCheckInst(), InsertPt);
549           if (Result)
550             Result = BinaryOperator::CreateAnd(RC.getCheckInst(), Result, "",
551                                                InsertPt);
552           else
553             Result = RC.getCheckInst();
554         }
555         assert(Result && "Failed to find result value");
556         Result->setName("wide.chk");
557       }
558       return true;
559     }
560   }
561 
562   // Base case -- just logical-and the two conditions together.
563 
564   if (InsertPt) {
565     makeAvailableAt(Cond0, InsertPt);
566     makeAvailableAt(Cond1, InsertPt);
567     if (InvertCondition)
568       Cond1 = BinaryOperator::CreateNot(Cond1, "inverted", InsertPt);
569     Result = BinaryOperator::CreateAnd(Cond0, Cond1, "wide.chk", InsertPt);
570   }
571 
572   // We were not able to compute Cond0 AND Cond1 for the price of one.
573   return false;
574 }
575 
576 bool GuardWideningImpl::parseRangeChecks(
577     Value *CheckCond, SmallVectorImpl<GuardWideningImpl::RangeCheck> &Checks,
578     SmallPtrSetImpl<const Value *> &Visited) {
579   if (!Visited.insert(CheckCond).second)
580     return true;
581 
582   using namespace llvm::PatternMatch;
583 
584   {
585     Value *AndLHS, *AndRHS;
586     if (match(CheckCond, m_And(m_Value(AndLHS), m_Value(AndRHS))))
587       return parseRangeChecks(AndLHS, Checks) &&
588              parseRangeChecks(AndRHS, Checks);
589   }
590 
591   auto *IC = dyn_cast<ICmpInst>(CheckCond);
592   if (!IC || !IC->getOperand(0)->getType()->isIntegerTy() ||
593       (IC->getPredicate() != ICmpInst::ICMP_ULT &&
594        IC->getPredicate() != ICmpInst::ICMP_UGT))
595     return false;
596 
597   const Value *CmpLHS = IC->getOperand(0), *CmpRHS = IC->getOperand(1);
598   if (IC->getPredicate() == ICmpInst::ICMP_UGT)
599     std::swap(CmpLHS, CmpRHS);
600 
601   auto &DL = IC->getModule()->getDataLayout();
602 
603   GuardWideningImpl::RangeCheck Check(
604       CmpLHS, cast<ConstantInt>(ConstantInt::getNullValue(CmpRHS->getType())),
605       CmpRHS, IC);
606 
607   if (!isKnownNonNegative(Check.getLength(), DL))
608     return false;
609 
610   // What we have in \c Check now is a correct interpretation of \p CheckCond.
611   // Try to see if we can move some constant offsets into the \c Offset field.
612 
613   bool Changed;
614   auto &Ctx = CheckCond->getContext();
615 
616   do {
617     Value *OpLHS;
618     ConstantInt *OpRHS;
619     Changed = false;
620 
621 #ifndef NDEBUG
622     auto *BaseInst = dyn_cast<Instruction>(Check.getBase());
623     assert((!BaseInst || DT.isReachableFromEntry(BaseInst->getParent())) &&
624            "Unreachable instruction?");
625 #endif
626 
627     if (match(Check.getBase(), m_Add(m_Value(OpLHS), m_ConstantInt(OpRHS)))) {
628       Check.setBase(OpLHS);
629       APInt NewOffset = Check.getOffsetValue() + OpRHS->getValue();
630       Check.setOffset(ConstantInt::get(Ctx, NewOffset));
631       Changed = true;
632     } else if (match(Check.getBase(),
633                      m_Or(m_Value(OpLHS), m_ConstantInt(OpRHS)))) {
634       KnownBits Known = computeKnownBits(OpLHS, DL);
635       if ((OpRHS->getValue() & Known.Zero) == OpRHS->getValue()) {
636         Check.setBase(OpLHS);
637         APInt NewOffset = Check.getOffsetValue() + OpRHS->getValue();
638         Check.setOffset(ConstantInt::get(Ctx, NewOffset));
639         Changed = true;
640       }
641     }
642   } while (Changed);
643 
644   Checks.push_back(Check);
645   return true;
646 }
647 
648 bool GuardWideningImpl::combineRangeChecks(
649     SmallVectorImpl<GuardWideningImpl::RangeCheck> &Checks,
650     SmallVectorImpl<GuardWideningImpl::RangeCheck> &RangeChecksOut) const {
651   unsigned OldCount = Checks.size();
652   while (!Checks.empty()) {
653     // Pick all of the range checks with a specific base and length, and try to
654     // merge them.
655     const Value *CurrentBase = Checks.front().getBase();
656     const Value *CurrentLength = Checks.front().getLength();
657 
658     SmallVector<GuardWideningImpl::RangeCheck, 3> CurrentChecks;
659 
660     auto IsCurrentCheck = [&](GuardWideningImpl::RangeCheck &RC) {
661       return RC.getBase() == CurrentBase && RC.getLength() == CurrentLength;
662     };
663 
664     copy_if(Checks, std::back_inserter(CurrentChecks), IsCurrentCheck);
665     erase_if(Checks, IsCurrentCheck);
666 
667     assert(CurrentChecks.size() != 0 && "We know we have at least one!");
668 
669     if (CurrentChecks.size() < 3) {
670       llvm::append_range(RangeChecksOut, CurrentChecks);
671       continue;
672     }
673 
674     // CurrentChecks.size() will typically be 3 here, but so far there has been
675     // no need to hard-code that fact.
676 
677     llvm::sort(CurrentChecks, [&](const GuardWideningImpl::RangeCheck &LHS,
678                                   const GuardWideningImpl::RangeCheck &RHS) {
679       return LHS.getOffsetValue().slt(RHS.getOffsetValue());
680     });
681 
682     // Note: std::sort should not invalidate the ChecksStart iterator.
683 
684     const ConstantInt *MinOffset = CurrentChecks.front().getOffset();
685     const ConstantInt *MaxOffset = CurrentChecks.back().getOffset();
686 
687     unsigned BitWidth = MaxOffset->getValue().getBitWidth();
688     if ((MaxOffset->getValue() - MinOffset->getValue())
689             .ugt(APInt::getSignedMinValue(BitWidth)))
690       return false;
691 
692     APInt MaxDiff = MaxOffset->getValue() - MinOffset->getValue();
693     const APInt &HighOffset = MaxOffset->getValue();
694     auto OffsetOK = [&](const GuardWideningImpl::RangeCheck &RC) {
695       return (HighOffset - RC.getOffsetValue()).ult(MaxDiff);
696     };
697 
698     if (MaxDiff.isMinValue() || !all_of(drop_begin(CurrentChecks), OffsetOK))
699       return false;
700 
701     // We have a series of f+1 checks as:
702     //
703     //   I+k_0 u< L   ... Chk_0
704     //   I+k_1 u< L   ... Chk_1
705     //   ...
706     //   I+k_f u< L   ... Chk_f
707     //
708     //     with forall i in [0,f]: k_f-k_i u< k_f-k_0  ... Precond_0
709     //          k_f-k_0 u< INT_MIN+k_f                 ... Precond_1
710     //          k_f != k_0                             ... Precond_2
711     //
712     // Claim:
713     //   Chk_0 AND Chk_f  implies all the other checks
714     //
715     // Informal proof sketch:
716     //
717     // We will show that the integer range [I+k_0,I+k_f] does not unsigned-wrap
718     // (i.e. going from I+k_0 to I+k_f does not cross the -1,0 boundary) and
719     // thus I+k_f is the greatest unsigned value in that range.
720     //
721     // This combined with Ckh_(f+1) shows that everything in that range is u< L.
722     // Via Precond_0 we know that all of the indices in Chk_0 through Chk_(f+1)
723     // lie in [I+k_0,I+k_f], this proving our claim.
724     //
725     // To see that [I+k_0,I+k_f] is not a wrapping range, note that there are
726     // two possibilities: I+k_0 u< I+k_f or I+k_0 >u I+k_f (they can't be equal
727     // since k_0 != k_f).  In the former case, [I+k_0,I+k_f] is not a wrapping
728     // range by definition, and the latter case is impossible:
729     //
730     //   0-----I+k_f---I+k_0----L---INT_MAX,INT_MIN------------------(-1)
731     //   xxxxxx             xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
732     //
733     // For Chk_0 to succeed, we'd have to have k_f-k_0 (the range highlighted
734     // with 'x' above) to be at least >u INT_MIN.
735 
736     RangeChecksOut.emplace_back(CurrentChecks.front());
737     RangeChecksOut.emplace_back(CurrentChecks.back());
738   }
739 
740   assert(RangeChecksOut.size() <= OldCount && "We pessimized!");
741   return RangeChecksOut.size() != OldCount;
742 }
743 
744 #ifndef NDEBUG
745 StringRef GuardWideningImpl::scoreTypeToString(WideningScore WS) {
746   switch (WS) {
747   case WS_IllegalOrNegative:
748     return "IllegalOrNegative";
749   case WS_Neutral:
750     return "Neutral";
751   case WS_Positive:
752     return "Positive";
753   case WS_VeryPositive:
754     return "VeryPositive";
755   }
756 
757   llvm_unreachable("Fully covered switch above!");
758 }
759 #endif
760 
761 PreservedAnalyses GuardWideningPass::run(Function &F,
762                                          FunctionAnalysisManager &AM) {
763   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
764   auto &LI = AM.getResult<LoopAnalysis>(F);
765   auto &PDT = AM.getResult<PostDominatorTreeAnalysis>(F);
766   auto *MSSAA = AM.getCachedResult<MemorySSAAnalysis>(F);
767   std::unique_ptr<MemorySSAUpdater> MSSAU;
768   if (MSSAA)
769     MSSAU = std::make_unique<MemorySSAUpdater>(&MSSAA->getMSSA());
770   if (!GuardWideningImpl(DT, &PDT, LI, MSSAU ? MSSAU.get() : nullptr,
771                          DT.getRootNode(), [](BasicBlock *) { return true; })
772            .run())
773     return PreservedAnalyses::all();
774 
775   PreservedAnalyses PA;
776   PA.preserveSet<CFGAnalyses>();
777   PA.preserve<MemorySSAAnalysis>();
778   return PA;
779 }
780 
781 PreservedAnalyses GuardWideningPass::run(Loop &L, LoopAnalysisManager &AM,
782                                          LoopStandardAnalysisResults &AR,
783                                          LPMUpdater &U) {
784   BasicBlock *RootBB = L.getLoopPredecessor();
785   if (!RootBB)
786     RootBB = L.getHeader();
787   auto BlockFilter = [&](BasicBlock *BB) {
788     return BB == RootBB || L.contains(BB);
789   };
790   std::unique_ptr<MemorySSAUpdater> MSSAU;
791   if (AR.MSSA)
792     MSSAU = std::make_unique<MemorySSAUpdater>(AR.MSSA);
793   if (!GuardWideningImpl(AR.DT, nullptr, AR.LI, MSSAU ? MSSAU.get() : nullptr,
794                          AR.DT.getNode(RootBB), BlockFilter).run())
795     return PreservedAnalyses::all();
796 
797   auto PA = getLoopPassPreservedAnalyses();
798   if (AR.MSSA)
799     PA.preserve<MemorySSAAnalysis>();
800   return PA;
801 }
802 
803 namespace {
804 struct GuardWideningLegacyPass : public FunctionPass {
805   static char ID;
806 
807   GuardWideningLegacyPass() : FunctionPass(ID) {
808     initializeGuardWideningLegacyPassPass(*PassRegistry::getPassRegistry());
809   }
810 
811   bool runOnFunction(Function &F) override {
812     if (skipFunction(F))
813       return false;
814     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
815     auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
816     auto &PDT = getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree();
817     auto *MSSAWP = getAnalysisIfAvailable<MemorySSAWrapperPass>();
818     std::unique_ptr<MemorySSAUpdater> MSSAU;
819     if (MSSAWP)
820       MSSAU = std::make_unique<MemorySSAUpdater>(&MSSAWP->getMSSA());
821     return GuardWideningImpl(DT, &PDT, LI, MSSAU ? MSSAU.get() : nullptr,
822                              DT.getRootNode(),
823                              [](BasicBlock *) { return true; })
824         .run();
825   }
826 
827   void getAnalysisUsage(AnalysisUsage &AU) const override {
828     AU.setPreservesCFG();
829     AU.addRequired<DominatorTreeWrapperPass>();
830     AU.addRequired<PostDominatorTreeWrapperPass>();
831     AU.addRequired<LoopInfoWrapperPass>();
832     AU.addPreserved<MemorySSAWrapperPass>();
833   }
834 };
835 
836 /// Same as above, but restricted to a single loop at a time.  Can be
837 /// scheduled with other loop passes w/o breaking out of LPM
838 struct LoopGuardWideningLegacyPass : public LoopPass {
839   static char ID;
840 
841   LoopGuardWideningLegacyPass() : LoopPass(ID) {
842     initializeLoopGuardWideningLegacyPassPass(*PassRegistry::getPassRegistry());
843   }
844 
845   bool runOnLoop(Loop *L, LPPassManager &LPM) override {
846     if (skipLoop(L))
847       return false;
848     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
849     auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
850     auto *PDTWP = getAnalysisIfAvailable<PostDominatorTreeWrapperPass>();
851     auto *PDT = PDTWP ? &PDTWP->getPostDomTree() : nullptr;
852     auto *MSSAWP = getAnalysisIfAvailable<MemorySSAWrapperPass>();
853     std::unique_ptr<MemorySSAUpdater> MSSAU;
854     if (MSSAWP)
855       MSSAU = std::make_unique<MemorySSAUpdater>(&MSSAWP->getMSSA());
856 
857     BasicBlock *RootBB = L->getLoopPredecessor();
858     if (!RootBB)
859       RootBB = L->getHeader();
860     auto BlockFilter = [&](BasicBlock *BB) {
861       return BB == RootBB || L->contains(BB);
862     };
863     return GuardWideningImpl(DT, PDT, LI, MSSAU ? MSSAU.get() : nullptr,
864                              DT.getNode(RootBB), BlockFilter).run();
865   }
866 
867   void getAnalysisUsage(AnalysisUsage &AU) const override {
868     AU.setPreservesCFG();
869     getLoopAnalysisUsage(AU);
870     AU.addPreserved<PostDominatorTreeWrapperPass>();
871     AU.addPreserved<MemorySSAWrapperPass>();
872   }
873 };
874 }
875 
876 char GuardWideningLegacyPass::ID = 0;
877 char LoopGuardWideningLegacyPass::ID = 0;
878 
879 INITIALIZE_PASS_BEGIN(GuardWideningLegacyPass, "guard-widening", "Widen guards",
880                       false, false)
881 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
882 INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)
883 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
884 INITIALIZE_PASS_END(GuardWideningLegacyPass, "guard-widening", "Widen guards",
885                     false, false)
886 
887 INITIALIZE_PASS_BEGIN(LoopGuardWideningLegacyPass, "loop-guard-widening",
888                       "Widen guards (within a single loop, as a loop pass)",
889                       false, false)
890 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
891 INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)
892 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
893 INITIALIZE_PASS_END(LoopGuardWideningLegacyPass, "loop-guard-widening",
894                     "Widen guards (within a single loop, as a loop pass)",
895                     false, false)
896 
897 FunctionPass *llvm::createGuardWideningPass() {
898   return new GuardWideningLegacyPass();
899 }
900 
901 Pass *llvm::createLoopGuardWideningPass() {
902   return new LoopGuardWideningLegacyPass();
903 }
904