[LoopUnroll] Keep the loop test only on the first iteration of max-or-zero loops

[john.brawn]

Summary

When we have a loop with a known upper bound on the number of iterations, and furthermore know that either the number of iterations will be either exactly that upper bound or zero, then we can fully unroll up to that upper bound keeping only the first loop test to check for the zero iteration case.

Most of the work here is in plumbing this 'max-or-zero' information from the part of scalar evolution where it's detected through to loop unrolling. I've also gone for the safe default of 'false' everywhere but howManyLessThans which could probably be improved.

[jmolloy]

The LoopUnroll changes look very reasonable to me, but I'd like Silviu's opinion on the SCEV plumbing.

[haicheng]

The loop unrolling part looks good to me, too. I am not an expert of SCEV.

include/llvm/Transforms/Utils/UnrollLoop.h

35

Thank you for catching my mistake.

[sbaranga]

include/llvm/Analysis/ScalarEvolution.h

1395

I've only taken a brief look at this, but it might be a nicer interface to add a getMaxOrZeroBackedgeTakenCount. That way we don't have to add the MaxOrZero everywhere (same in other places) and would look more like the existing code.

lib/Analysis/ScalarEvolution.cpp

5756–5757

Does this work with more than one loop exit and with predicated loop exits?

I'm not certain this would hold if we have at least one loop exit that doesn't dominate the latch, given the very strict definition (the backedge gets taken exactly 0 or MaxNotTaken times).

[john.brawn]

lib/Analysis/ScalarEvolution.cpp

5756–5757

That's handled down on line 5759 - MaxOrZero is set to false on loops with more than one exit.

[sbaranga]

lib/Analysis/ScalarEvolution.cpp

5756–5757

Good point, I've missed that. It should be fine then.

[john.brawn]

Pass the MaxOrZero information through to loop unrolling using a separate function instead of via a pointer argument.

[sbaranga]

LGTM for the SCEV part.

include/llvm/Analysis/ScalarEvolution.h

private:
   /// predicate by splitting it into a set of independent predicates.
   bool ProvingSplitPredicate;
 
   /// Information about the number of loop iterations for which a loop exit's
   /// branch condition evaluates to the not-taken path.  This is a temporary
   /// pair of exact and max expressions that are eventually summarized in
   /// ExitNotTakenInfo and BackedgeTakenInfo.
   struct ExitLimit {
-    const SCEV *ExactNotTaken;
-    const SCEV *MaxNotTaken;
+    const SCEV *ExactNotTaken; //< The exit is not taken exactly this many times
+    const SCEV *MaxNotTaken; //< The exit is not taken at most this many times
+    bool MaxOrZero; //< Not taken either exactly MaxNotTaken or zero times
 
     /// A set of predicate guards for this ExitLimit. The result is only valid
     /// if all of the predicates in \c Predicates evaluate to 'true' at
     /// run-time.
     SmallPtrSet<const SCEVPredicate *, 4> Predicates;
 
     void addPredicate(const SCEVPredicate *P) {
       assert(!isa<SCEVUnionPredicate>(P) && "Only add leaf predicates here!");
       Predicates.insert(P);
     }
 
-    /*implicit*/ ExitLimit(const SCEV *E) : ExactNotTaken(E), MaxNotTaken(E) {}
+    /*implicit*/ ExitLimit(const SCEV *E)
+        : ExactNotTaken(E), MaxNotTaken(E), MaxOrZero(false) {}
 
     ExitLimit(
-        const SCEV *E, const SCEV *M,
+        const SCEV *E, const SCEV *M, bool MaxOrZero,
         ArrayRef<const SmallPtrSetImpl<const SCEVPredicate *> *> PredSetList)
-        : ExactNotTaken(E), MaxNotTaken(M) {
+        : ExactNotTaken(E), MaxNotTaken(M), MaxOrZero(MaxOrZero) {
       assert((isa<SCEVCouldNotCompute>(ExactNotTaken) ||
               !isa<SCEVCouldNotCompute>(MaxNotTaken)) &&
              "Exact is not allowed to be less precise than Max");
       for (auto *PredSet : PredSetList)
         for (auto *P : *PredSet)
           addPredicate(P);
     }
 
-    ExitLimit(const SCEV *E, const SCEV *M,
+    ExitLimit(const SCEV *E, const SCEV *M, bool MaxOrZero,
               const SmallPtrSetImpl<const SCEVPredicate *> &PredSet)
-        : ExitLimit(E, M, {&PredSet}) {}
+        : ExitLimit(E, M, MaxOrZero, {&PredSet}) {}
 
-    ExitLimit(const SCEV *E, const SCEV *M) : ExitLimit(E, M, None) {}
+    ExitLimit(const SCEV *E, const SCEV *M, bool MaxOrZero)
+        : ExitLimit(E, M, MaxOrZero, None) {}
 
     /// Test whether this ExitLimit contains any computed information, or
     /// whether it's all SCEVCouldNotCompute values.
     bool hasAnyInfo() const {
       return !isa<SCEVCouldNotCompute>(ExactNotTaken) ||
              !isa<SCEVCouldNotCompute>(MaxNotTaken);
     }
 
class BackedgeTakenInfo {
     /// least maximum backedge-taken count of the loop that is known, or a
     /// SCEVCouldNotCompute. This expression is only valid if the predicates
     /// associated with all loop exits are true.
     ///
     /// The integer part of \c MaxAndComplete is a boolean indicating if \c
     /// ExitNotTaken has an element for every exiting block in the loop.
     PointerIntPair<const SCEV *, 1> MaxAndComplete;
 
+    /// True iff the backedge is taken either exactly Max or zero times.
+    bool MaxOrZero;
+
     /// \name Helper projection functions on \c MaxAndComplete.
     /// @{
     bool isComplete() const { return MaxAndComplete.getInt(); }
     const SCEV *getMax() const { return MaxAndComplete.getPointer(); }
     /// @}
 
   public:
     BackedgeTakenInfo() : MaxAndComplete(nullptr, 0) {}
 
     BackedgeTakenInfo(const BackedgeTakenInfo &) = delete;
 
     BackedgeTakenInfo(BackedgeTakenInfo &&Other) {
       ExitNotTaken = std::move(Other.ExitNotTaken);
       MaxAndComplete = std::move(Other.MaxAndComplete);
+      MaxOrZero = std::move(Other.MaxOrZero);
     }
 
     BackedgeTakenInfo &operator=(BackedgeTakenInfo &&Other) {
       ExitNotTaken = std::move(Other.ExitNotTaken);
       MaxAndComplete = std::move(Other.MaxAndComplete);
+      MaxOrZero = std::move(Other.MaxOrZero);
       return *this;
     }
 
     typedef std::pair<BasicBlock *, ExitLimit> EdgeExitInfo;
 
     /// Initialize BackedgeTakenInfo from a list of exact exit counts.
     BackedgeTakenInfo(SmallVectorImpl<EdgeExitInfo> &&ExitCounts, bool Complete,
-                      const SCEV *MaxCount);
+                      const SCEV *MaxCount, bool MaxOrZero);
 
     /// Test whether this BackedgeTakenInfo contains any computed information,
     /// or whether it's all SCEVCouldNotCompute values.
     bool hasAnyInfo() const {
       return !ExitNotTaken.empty() || !isa<SCEVCouldNotCompute>(getMax());
     }
 
     /// Test whether this BackedgeTakenInfo contains complete information.
public:
     /// edge, or SCEVCouldNotCompute. The loop is guaranteed not to exit via
     /// this block before this number of iterations, but may exit via another
     /// block.
     const SCEV *getExact(BasicBlock *ExitingBlock, ScalarEvolution *SE) const;
 
     /// Get the max backedge taken count for the loop.
     const SCEV *getMax(ScalarEvolution *SE) const;
 
+    /// Return true if the number of times this backedge is taken is either the
+    /// value returned by getMax or zero.
+    bool isMaxOrZero(ScalarEvolution *SE) const;
+
     /// Return true if any backedge taken count expressions refer to the given
     /// subexpression.
     bool hasOperand(const SCEV *S, ScalarEvolution *SE) const;
 
     /// Invalidate this result and free associated memory.
     void clear();
   };
 
public:
   /// checks and can be used to perform loop versioning.
   const SCEV *getPredicatedBackedgeTakenCount(const Loop *L,
                                               SCEVUnionPredicate &Predicates);
 
   /// Similar to getBackedgeTakenCount, except return the least SCEV value
   /// that is known never to be less than the actual backedge taken count.
   const SCEV *getMaxBackedgeTakenCount(const Loop *L);
 
+  /// Return true if the backedge taken count is either the value returned by
+  /// getMaxBackedgeTakenCount or zero.
+  bool isBackedgeTakenCountMaxOrZero(const Loop *L);
+
   /// Return true if the specified loop has an analyzable loop-invariant
   /// backedge-taken count.
   bool hasLoopInvariantBackedgeTakenCount(const Loop *L);
 
   /// This method should be called by the client when it has changed a loop in
   /// a way that may effect ScalarEvolution's ability to compute a trip count,
   /// or if the loop is deleted.  This call is potentially expensive for large
   /// loop bodies.

include/llvm/Transforms/Utils/UnrollLoop.h

 class LPPassManager;
 class MDNode;
 class Pass;
 class OptimizationRemarkEmitter;
 class ScalarEvolution;
 
 bool UnrollLoop(Loop *L, unsigned Count, unsigned TripCount, bool Force,
                 bool AllowRuntime, bool AllowExpensiveTripCount,
-                bool UseUpperBound, unsigned TripMultiple, LoopInfo *LI,
-                ScalarEvolution *SE, DominatorTree *DT, AssumptionCache *AC,
+                bool PreserveCondBr, bool PreserveOnlyFirst,
+                unsigned TripMultiple, LoopInfo *LI, ScalarEvolution *SE,
+                DominatorTree *DT, AssumptionCache *AC,
                 OptimizationRemarkEmitter *ORE, bool PreserveLCSSA);
 
 bool UnrollRuntimeLoopRemainder(Loop *L, unsigned Count,
                                 bool AllowExpensiveTripCount,
                                 bool UseEpilogRemainder, LoopInfo *LI,
                                 ScalarEvolution *SE, DominatorTree *DT,
                                 bool PreserveLCSSA);
 
 MDNode *GetUnrollMetadata(MDNode *LoopID, StringRef Name);
 }
 
 #endif

lib/Analysis/ScalarEvolution.cpp

 }
 
 /// Similar to getBackedgeTakenCount, except return the least SCEV value that is
 /// known never to be less than the actual backedge taken count.
 const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) {
   return getBackedgeTakenInfo(L).getMax(this);
 }
 
+bool ScalarEvolution::isBackedgeTakenCountMaxOrZero(const Loop *L) {
+  return getBackedgeTakenInfo(L).isMaxOrZero(this);
+}
+
 /// Push PHI nodes in the header of the given loop onto the given Worklist.
 static void
 PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) {
   BasicBlock *Header = L->getHeader();
 
   // Push all Loop-header PHIs onto the Worklist stack.
   for (BasicBlock::iterator I = Header->begin();
        PHINode *PN = dyn_cast<PHINode>(I); ++I)
ScalarEvolution::BackedgeTakenInfo::getMax(ScalarEvolution *SE) const {
   };
 
   if (any_of(ExitNotTaken, PredicateNotAlwaysTrue) || !getMax())
     return SE->getCouldNotCompute();
 
   return getMax();
 }
 
+bool ScalarEvolution::BackedgeTakenInfo::isMaxOrZero(ScalarEvolution *SE) const {
+  auto PredicateNotAlwaysTrue = [](const ExitNotTakenInfo &ENT) {
+    return !ENT.hasAlwaysTruePredicate();
+  };
+  return MaxOrZero && !any_of(ExitNotTaken, PredicateNotAlwaysTrue);
+}
+
 bool ScalarEvolution::BackedgeTakenInfo::hasOperand(const SCEV *S,
                                                     ScalarEvolution *SE) const {
   if (getMax() && getMax() != SE->getCouldNotCompute() &&
       SE->hasOperand(getMax(), S))
     return true;
 
   for (auto &ENT : ExitNotTaken)
     if (ENT.ExactNotTaken != SE->getCouldNotCompute() &&
         SE->hasOperand(ENT.ExactNotTaken, S))
       return true;
 
   return false;
 }
 
 /// Allocate memory for BackedgeTakenInfo and copy the not-taken count of each
 /// computable exit into a persistent ExitNotTakenInfo array.
 ScalarEvolution::BackedgeTakenInfo::BackedgeTakenInfo(
     SmallVectorImpl<ScalarEvolution::BackedgeTakenInfo::EdgeExitInfo>
         &&ExitCounts,
-    bool Complete, const SCEV *MaxCount)
-    : MaxAndComplete(MaxCount, Complete) {
+    bool Complete, const SCEV *MaxCount, bool MaxOrZero)
+    : MaxAndComplete(MaxCount, Complete), MaxOrZero(MaxOrZero) {
   typedef ScalarEvolution::BackedgeTakenInfo::EdgeExitInfo EdgeExitInfo;
   ExitNotTaken.reserve(ExitCounts.size());
   std::transform(
       ExitCounts.begin(), ExitCounts.end(), std::back_inserter(ExitNotTaken),
       [&](const EdgeExitInfo &EEI) {
         BasicBlock *ExitBB = EEI.first;
         const ExitLimit &EL = EEI.second;
         if (EL.Predicates.empty())
ScalarEvolution::computeBackedgeTakenCount(const Loop *L,
 
   typedef ScalarEvolution::BackedgeTakenInfo::EdgeExitInfo EdgeExitInfo;
 
   SmallVector<EdgeExitInfo, 4> ExitCounts;
   bool CouldComputeBECount = true;
   BasicBlock *Latch = L->getLoopLatch(); // may be NULL.
   const SCEV *MustExitMaxBECount = nullptr;
   const SCEV *MayExitMaxBECount = nullptr;
+  bool MustExitMaxOrZero = false;
 
   // Compute the ExitLimit for each loop exit. Use this to populate ExitCounts
   // and compute maxBECount.
   // Do a union of all the predicates here.
   for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
     BasicBlock *ExitBB = ExitingBlocks[i];
     ExitLimit EL = computeExitLimit(L, ExitBB, AllowPredicates);
 
for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
     // If the exit dominates the loop latch, it is a LoopMustExit otherwise it
     // is a LoopMayExit.  If any computable LoopMustExit is found, then
     // MaxBECount is the minimum EL.MaxNotTaken of computable
     // LoopMustExits. Otherwise, MaxBECount is conservatively the maximum
     // EL.MaxNotTaken, where CouldNotCompute is considered greater than any
     // computable EL.MaxNotTaken.
     if (EL.MaxNotTaken != getCouldNotCompute() && Latch &&
         DT.dominates(ExitBB, Latch)) {
-      if (!MustExitMaxBECount)
+      if (!MustExitMaxBECount) {
         MustExitMaxBECount = EL.MaxNotTaken;
-      else {
+        MustExitMaxOrZero = EL.MaxOrZero;
+      } else {
         MustExitMaxBECount =
             getUMinFromMismatchedTypes(MustExitMaxBECount, EL.MaxNotTaken);
       }
     } else if (MayExitMaxBECount != getCouldNotCompute()) {
       if (!MayExitMaxBECount || EL.MaxNotTaken == getCouldNotCompute())
         MayExitMaxBECount = EL.MaxNotTaken;
       else {
         MayExitMaxBECount =
             getUMaxFromMismatchedTypes(MayExitMaxBECount, EL.MaxNotTaken);
       }
     }
   }
   const SCEV *MaxBECount = MustExitMaxBECount ? MustExitMaxBECount :
     (MayExitMaxBECount ? MayExitMaxBECount : getCouldNotCompute());
+  // The loop backedge will be taken the maximum or zero times if there's
+  // a single exit that must be taken the maximum or zero times.
+  bool MaxOrZero = (MustExitMaxOrZero && ExitingBlocks.size() == 1);
   return BackedgeTakenInfo(std::move(ExitCounts), CouldComputeBECount,
-                           MaxBECount);
+                           MaxBECount, MaxOrZero);
 }
 
 ScalarEvolution::ExitLimit
 ScalarEvolution::computeExitLimit(const Loop *L, BasicBlock *ExitingBlock,
                                   bool AllowPredicates) {
 
   // Okay, we've chosen an exiting block.  See what condition causes us to exit
   // at this block and remember the exit block and whether all other targets
if (BO->getOpcode() == Instruction::And) {
       // to be more aggressive when computing BECount than when computing
       // MaxBECount.  In these cases it is possible for EL0.ExactNotTaken and
       // EL1.ExactNotTaken to match, but for EL0.MaxNotTaken and EL1.MaxNotTaken
       // to not.
       if (isa<SCEVCouldNotCompute>(MaxBECount) &&
           !isa<SCEVCouldNotCompute>(BECount))
         MaxBECount = BECount;
 
-      return ExitLimit(BECount, MaxBECount, {&EL0.Predicates, &EL1.Predicates});
+      return ExitLimit(BECount, MaxBECount, false,
+                       {&EL0.Predicates, &EL1.Predicates});
     }
     if (BO->getOpcode() == Instruction::Or) {
       // Recurse on the operands of the or.
       bool EitherMayExit = L->contains(FBB);
       ExitLimit EL0 = computeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB,
                                                ControlsExit && !EitherMayExit,
                                                AllowPredicates);
       ExitLimit EL1 = computeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB,
if (BO->getOpcode() == Instruction::Or) {
         // For now, be conservative.
         assert(L->contains(TBB) && "Loop block has no successor in loop!");
         if (EL0.MaxNotTaken == EL1.MaxNotTaken)
           MaxBECount = EL0.MaxNotTaken;
         if (EL0.ExactNotTaken == EL1.ExactNotTaken)
           BECount = EL0.ExactNotTaken;
       }
 
-      return ExitLimit(BECount, MaxBECount, {&EL0.Predicates, &EL1.Predicates});
+      return ExitLimit(BECount, MaxBECount, false,
+                       {&EL0.Predicates, &EL1.Predicates});
     }
   }
 
   // With an icmp, it may be feasible to compute an exact backedge-taken count.
   // Proceed to the next level to examine the icmp.
   if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond)) {
     ExitLimit EL =
         computeExitLimitFromICmp(L, ExitCondICmp, TBB, FBB, ControlsExit);
auto *Result =
       ConstantFoldCompareInstOperands(Pred, StableValue, RHS, DL, &TLI);
   assert(Result->getType()->isIntegerTy(1) &&
          "Otherwise cannot be an operand to a branch instruction");
 
   if (Result->isZeroValue()) {
     unsigned BitWidth = getTypeSizeInBits(RHS->getType());
     const SCEV *UpperBound =
         getConstant(getEffectiveSCEVType(RHS->getType()), BitWidth);
-    return ExitLimit(getCouldNotCompute(), UpperBound);
+    return ExitLimit(getCouldNotCompute(), UpperBound, false);
   }
 
   return getCouldNotCompute();
 }
 
 /// Return true if we can constant fold an instruction of the specified type,
 /// assuming that all operands were constants.
 static bool CanConstantFold(const Instruction *I) {
if (auto Roots = SolveQuadraticEquation(AddRec, *this)) {
         if (!CB->getZExtValue())
           std::swap(R1, R2); // R1 is the minimum root now.
 
         // We can only use this value if the chrec ends up with an exact zero
         // value at this index.  When solving for "X*X != 5", for example, we
         // should not accept a root of 2.
         const SCEV *Val = AddRec->evaluateAtIteration(R1, *this);
         if (Val->isZero())
-          return ExitLimit(R1, R1, Predicates); // We found a quadratic root!
+          // We found a quadratic root!
+          return ExitLimit(R1, R1, false, Predicates);
       }
     }
     return getCouldNotCompute();
   }
 
   // Otherwise we can only handle this if it is affine.
   if (!AddRec->isAffine())
     return getCouldNotCompute();
if (StepC->getValue()->equalsInt(1) || StepC->getValue()->isAllOnesValue()) {
     if (!CountDown && CR.getUnsignedMin().isMinValue())
       // When counting up, the worst starting value is 1, not 0.
       MaxBECount = CR.getUnsignedMax().isMinValue()
         ? getConstant(APInt::getMinValue(CR.getBitWidth()))
         : getConstant(APInt::getMaxValue(CR.getBitWidth()));
     else
       MaxBECount = getConstant(CountDown ? CR.getUnsignedMax()
                                          : -CR.getUnsignedMin());
-    return ExitLimit(Distance, MaxBECount, Predicates);
+    return ExitLimit(Distance, MaxBECount, false, Predicates);
   }
 
   // As a special case, handle the instance where Step is a positive power of
   // two. In this case, determining whether Step divides Distance evenly can be
   // done by counting and comparing the number of trailing zeros of Step and
   // Distance.
   if (!CountDown) {
     const APInt &StepV = StepC->getAPInt();
if (StepV.isPowerOf2() &&
       // and a zero extend.
 
       unsigned NarrowWidth = StepV.getBitWidth() - StepV.countTrailingZeros();
       auto *NarrowTy = IntegerType::get(getContext(), NarrowWidth);
       auto *WideTy = Distance->getType();
 
       const SCEV *Limit =
           getZeroExtendExpr(getTruncateExpr(ModuloResult, NarrowTy), WideTy);
-      return ExitLimit(Limit, Limit, Predicates);
+      return ExitLimit(Limit, Limit, false, Predicates);
     }
   }
 
   // If the condition controls loop exit (the loop exits only if the expression
   // is true) and the addition is no-wrap we can use unsigned divide to
   // compute the backedge count.  In this case, the step may not divide the
   // distance, but we don't care because if the condition is "missed" the loop
   // will have undefined behavior due to wrapping.
   if (ControlsExit && AddRec->hasNoSelfWrap() &&
       loopHasNoAbnormalExits(AddRec->getLoop())) {
     const SCEV *Exact =
         getUDivExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step);
-    return ExitLimit(Exact, Exact, Predicates);
+    return ExitLimit(Exact, Exact, false, Predicates);
   }
 
   // Then, try to solve the above equation provided that Start is constant.
   if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start)) {
     const SCEV *E = SolveLinEquationWithOverflow(
         StepC->getValue()->getValue(), -StartC->getValue()->getValue(), *this);
-    return ExitLimit(E, E, Predicates);
+    return ExitLimit(E, E, false, Predicates);
   }
   return getCouldNotCompute();
 }
 
 ScalarEvolution::ExitLimit
 ScalarEvolution::howFarToNonZero(const SCEV *V, const Loop *L) {
   // Loops that look like: while (X == 0) are very strange indeed.  We don't
   // handle them yet except for the trivial case.  This could be expanded in the
ScalarEvolution::howManyLessThans(const SCEV *LHS, const SCEV *RHS,
   if (isLoopEntryGuardedByCond(L, Cond, getMinusSCEV(Start, Stride), RHS))
     BECount = BECountIfBackedgeTaken;
   else {
     End = IsSigned ? getSMaxExpr(RHS, Start) : getUMaxExpr(RHS, Start);
     BECount = computeBECount(getMinusSCEV(End, Start), Stride, false);
   }
 
   const SCEV *MaxBECount;
+  bool MaxOrZero = false;
   if (isa<SCEVConstant>(BECount))
     MaxBECount = BECount;
-  else if (isa<SCEVConstant>(BECountIfBackedgeTaken))
+  else if (isa<SCEVConstant>(BECountIfBackedgeTaken)) {
     // If we know exactly how many times the backedge will be taken if it's
     // taken at least once, then the backedge count will either be that or
     // zero.
     MaxBECount = BECountIfBackedgeTaken;
-  else {
+    MaxOrZero = true;
+  } else {
     // Calculate the maximum backedge count based on the range of values
     // permitted by Start, End, and Stride.
     APInt MinStart = IsSigned ? getSignedRange(Start).getSignedMin()
                               : getUnsignedRange(Start).getUnsignedMin();
 
     unsigned BitWidth = getTypeSizeInBits(LHS->getType());
 
     APInt StrideForMaxBECount;
else if (isa<SCEVConstant>(BECountIfBackedgeTaken)) {
 
     MaxBECount = computeBECount(getConstant(MaxEnd - MinStart),
                                 getConstant(StrideForMaxBECount), false);
   }
 
   if (isa<SCEVCouldNotCompute>(MaxBECount))
     MaxBECount = BECount;
 
-  return ExitLimit(BECount, MaxBECount, Predicates);
+  return ExitLimit(BECount, MaxBECount, MaxOrZero, Predicates);
 }
 
 ScalarEvolution::ExitLimit
 ScalarEvolution::howManyGreaterThans(const SCEV *LHS, const SCEV *RHS,
                                      const Loop *L, bool IsSigned,
                                      bool ControlsExit, bool AllowPredicates) {
   SmallPtrSet<const SCEVPredicate *, 4> Predicates;
   // We handle only IV > Invariant
if (isa<SCEVConstant>(BECount))
     MaxBECount = BECount;
   else
     MaxBECount = computeBECount(getConstant(MaxStart - MinEnd),
                                 getConstant(MinStride), false);
 
   if (isa<SCEVCouldNotCompute>(MaxBECount))
     MaxBECount = BECount;
 
-  return ExitLimit(BECount, MaxBECount, Predicates);
+  return ExitLimit(BECount, MaxBECount, false, Predicates);
 }
 
 const SCEV *SCEVAddRecExpr::getNumIterationsInRange(const ConstantRange &Range,
                                                     ScalarEvolution &SE) const {
   if (Range.isFullSet())  // Infinite loop.
     return SE.getCouldNotCompute();
 
   // If the start is a non-zero constant, shift the range to simplify things.
static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE,
 
   OS << "\n"
         "Loop ";
   L->getHeader()->printAsOperand(OS, /*PrintType=*/false);
   OS << ": ";
 
   if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
     OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L);
+    if (SE->isBackedgeTakenCountMaxOrZero(L))
+      OS << ", actual taken count either this or zero.";
   } else {
     OS << "Unpredictable max backedge-taken count. ";
   }
 
   OS << "\n"
         "Loop ";
   L->getHeader()->printAsOperand(OS, /*PrintType=*/false);
   OS << ": ";

lib/Transforms/Scalar/LoopUnrollPass.cpp

static bool tryToUnrollLoop(Loop *L, DominatorTree &DT, LoopInfo *LI,
   // case, the program would be ill-formed (on most architectures)
   // unless n were the same on all threads in a thread group.
   // Assuming n is the same on all threads, any kind of unrolling is
   // safe.  But currently llvm's notion of convergence isn't powerful
   // enough to express this.
   if (Convergent)
     UP.AllowRemainder = false;
 
-  // Try to find the trip count upper bound if it is allowed and we cannot find
-  // exact trip count.
-  if (UP.UpperBound) {
-    if (!TripCount) {
-      MaxTripCount = SE->getSmallConstantMaxTripCount(L);
-      // Only unroll with small upper bound.
-      if (MaxTripCount > UnrollMaxUpperBound)
-        MaxTripCount = 0;
+  // Try to find the trip count upper bound if we cannot find the exact trip
+  // count.
+  bool MaxOrZero = false;
+    MaxOrZero = SE->isBackedgeTakenCountMaxOrZero(L);
+    // We can unroll by the upper bound amount if it's generally allowed or if
+    // we know that the loop is executed either the upper bound or zero times.
+    // (MaxOrZero unrolling keeps only the first loop test, so the number of
+    // loop tests remains the same compared to the non-unrolled version, whereas
+    // the generic upper bound unrolling keeps all but the last loop test so the
+    // number of loop tests goes up which may end up being worse on targets with
+    // constriained branch predictor resources so is controlled by an option.)
+    // In addition we only unroll small upper bounds.
+    if (!(UP.UpperBound || MaxOrZero) || MaxTripCount > UnrollMaxUpperBound) {
     }
   }
 
   // computeUnrollCount() decides whether it is beneficial to use upper bound to
   // fully unroll the loop.
   bool UseUpperBound = false;
   bool IsCountSetExplicitly =
       computeUnrollCount(L, TTI, DT, LI, SE, &ORE, TripCount, MaxTripCount,
                          TripMultiple, LoopSize, UP, UseUpperBound);
   if (!UP.Count)
     return false;
   // Unroll factor (Count) must be less or equal to TripCount.
   if (TripCount && UP.Count > TripCount)
     UP.Count = TripCount;
 
   // Unroll the loop.
   if (!UnrollLoop(L, UP.Count, TripCount, UP.Force, UP.Runtime,
-                  UP.AllowExpensiveTripCount, UseUpperBound, TripMultiple, LI,
-                  SE, &DT, &AC, &ORE, PreserveLCSSA))
+                  UP.AllowExpensiveTripCount, UseUpperBound, MaxOrZero,
+                  TripMultiple, LI, SE, &DT, &AC, &ORE, PreserveLCSSA))
     return false;
 
   // If loop has an unroll count pragma or unrolled by explicitly set count
   // mark loop as unrolled to prevent unrolling beyond that requested.
   if (IsCountSetExplicitly)
     SetLoopAlreadyUnrolled(L);
   return true;
 }

lib/Transforms/Utils/LoopUnroll.cpp

 /// exits were taken. Note that UnrollLoop assumes that the loop counter test
 /// terminates LatchBlock in order to remove unnecesssary instances of the
 /// test. In other words, control may exit the loop prior to TripCount
 /// iterations via an early branch, but control may not exit the loop from the
 /// LatchBlock's terminator prior to TripCount iterations.
 ///
 /// PreserveCondBr indicates whether the conditional branch of the LatchBlock
 /// needs to be preserved.  It is needed when we use trip count upper bound to
-/// fully unroll the loop.
+/// fully unroll the loop. If PreserveOnlyFirst is also set then only the first
+/// conditional branch needs to be preserved.
 ///
 /// Similarly, TripMultiple divides the number of times that the LatchBlock may
 /// execute without exiting the loop.
 ///
 /// If AllowRuntime is true then UnrollLoop will consider unrolling loops that
 /// have a runtime (i.e. not compile time constant) trip count.  Unrolling these
 /// loops require a unroll "prologue" that runs "RuntimeTripCount % Count"
 /// iterations before branching into the unrolled loop.  UnrollLoop will not
 /// runtime-unroll the loop if computing RuntimeTripCount will be expensive and
 /// AllowExpensiveTripCount is false.
 ///
 /// The LoopInfo Analysis that is passed will be kept consistent.
 ///
 /// This utility preserves LoopInfo. It will also preserve ScalarEvolution and
 /// DominatorTree if they are non-null.
 bool llvm::UnrollLoop(Loop *L, unsigned Count, unsigned TripCount, bool Force,
                       bool AllowRuntime, bool AllowExpensiveTripCount,
-                      bool PreserveCondBr, unsigned TripMultiple, LoopInfo *LI,
-                      ScalarEvolution *SE, DominatorTree *DT,
-                      AssumptionCache *AC, OptimizationRemarkEmitter *ORE,
-                      bool PreserveLCSSA) {
+                      bool PreserveCondBr, bool PreserveOnlyFirst,
+                      unsigned TripMultiple, LoopInfo *LI, ScalarEvolution *SE,
+                      DominatorTree *DT, AssumptionCache *AC,
+                      OptimizationRemarkEmitter *ORE, bool PreserveLCSSA) {
   BasicBlock *Preheader = L->getLoopPreheader();
   if (!Preheader) {
     DEBUG(dbgs() << "  Can't unroll; loop preheader-insertion failed.\n");
     return false;
   }
 
   BasicBlock *LatchBlock = L->getLoopLatch();
   if (!LatchBlock) {
if (CompletelyUnroll) {
       if (j == 0)
         Dest = LoopExit;
       // If using trip count upper bound to completely unroll, we need to keep
       // the conditional branch except the last one because the loop may exit
       // after any iteration.
       assert(NeedConditional &&
              "NeedCondition cannot be modified by both complete "
              "unrolling and runtime unrolling");
-      NeedConditional = (PreserveCondBr && j);
+      NeedConditional = (PreserveCondBr && j && !(PreserveOnlyFirst && i != 0));
     } else if (j != BreakoutTrip && (TripMultiple == 0 || j % TripMultiple != 0)) {
       // If we know the trip count or a multiple of it, we can safely use an
       // unconditional branch for some iterations.
       NeedConditional = false;
     }
 
     if (NeedConditional) {
       // Update the conditional branch's successor for the following

test/Analysis/ScalarEvolution/trip-count13.ll

 ; RUN: opt -S -analyze -scalar-evolution < %s | FileCheck %s
 
 define void @u_0(i8 %rhs) {
 ; E.g.: %rhs = 255, %start = 99, backedge taken 156 times
 entry:
   %start = add i8 %rhs, 100
   br label %loop
 
 loop:
   %iv = phi i8 [ %start, %entry ], [ %iv.inc, %loop ]
   %iv.inc = add nuw i8 %iv, 1  ;; Note: this never unsigned-wraps
   %iv.cmp = icmp ult i8 %iv, %rhs
   br i1 %iv.cmp, label %loop, label %leave
 
 ; CHECK-LABEL: Determining loop execution counts for: @u_0
 ; CHECK-NEXT: Loop %loop: backedge-taken count is (-100 + (-1 * %rhs) + ((100 + %rhs) umax %rhs))
-; CHECK-NEXT: Loop %loop: max backedge-taken count is -100
+; CHECK-NEXT: Loop %loop: max backedge-taken count is -100, actual taken count either this or zero.
 
 leave:
   ret void
 }
 
 define void @u_1(i8 %start) {
 entry:
 ; E.g.: %start = 99, %rhs = 255, backedge taken 156 times
   %rhs = add i8 %start, -100
   br label %loop
 
 loop:
   %iv = phi i8 [ %start, %entry ], [ %iv.inc, %loop ]
   %iv.inc = add nuw i8 %iv, 1  ;; Note: this never unsigned-wraps
   %iv.cmp = icmp ult i8 %iv, %rhs
   br i1 %iv.cmp, label %loop, label %leave
 
 ; CHECK-LABEL: Determining loop execution counts for: @u_1
 ; CHECK-NEXT: Loop %loop: backedge-taken count is ((-1 * %start) + ((-100 + %start) umax %start))
-; CHECK-NEXT: Loop %loop: max backedge-taken count is -100
+; CHECK-NEXT: Loop %loop: max backedge-taken count is -100, actual taken count either this or zero.
 
 leave:
   ret void
 }
 
 define void @s_0(i8 %rhs) {
 entry:
 ; E.g.: %rhs = 127, %start = -29, backedge taken 156 times
   %start = add i8 %rhs, 100
   br label %loop
 
 loop:
   %iv = phi i8 [ %start, %entry ], [ %iv.inc, %loop ]
   %iv.inc = add nsw i8 %iv, 1  ;; Note: this never signed-wraps
   %iv.cmp = icmp slt i8 %iv, %rhs
   br i1 %iv.cmp, label %loop, label %leave
 
 ; CHECK-LABEL: Determining loop execution counts for: @s_0
 ; CHECK-NEXT: Loop %loop: backedge-taken count is (-100 + (-1 * %rhs) + ((100 + %rhs) smax %rhs))
-; CHECK-NEXT: Loop %loop: max backedge-taken count is -100
+; CHECK-NEXT: Loop %loop: max backedge-taken count is -100, actual taken count either this or zero.
 
 leave:
   ret void
 }
 
 define void @s_1(i8 %start) {
 entry:
 ; E.g.: start = -29, %rhs = 127, %backedge taken 156 times
   %rhs = add i8 %start, -100
   br label %loop
 
 loop:
   %iv = phi i8 [ %start, %entry ], [ %iv.inc, %loop ]
   %iv.inc = add nsw i8 %iv, 1
   %iv.cmp = icmp slt i8 %iv, %rhs
   br i1 %iv.cmp, label %loop, label %leave
 
 ; CHECK-LABEL: Determining loop execution counts for: @s_1
 ; CHECK-NEXT: Loop %loop: backedge-taken count is ((-1 * %start) + ((-100 + %start) smax %start))
-; CHECK-NEXT: Loop %loop: max backedge-taken count is -100
+; CHECK-NEXT: Loop %loop: max backedge-taken count is -100, actual taken count either this or zero.
 
 leave:
   ret void
 }

test/Analysis/ScalarEvolution/trip-count14.ll

do.body:
   %arrayidx = getelementptr i32, i32* %p, i32 %i.0
   store i32 %i.0, i32* %arrayidx, align 4
   %inc = add i32 %i.0, 1
   %cmp = icmp slt i32 %i.0, %add
   br i1 %cmp, label %do.body, label %do.end ; taken either 0 or 1 times
 
 ; CHECK-LABEL: Determining loop execution counts for: @s32_max1
 ; CHECK-NEXT: Loop %do.body: backedge-taken count is ((-1 * %n) + ((1 + %n) smax %n))
-; CHECK-NEXT: Loop %do.body: max backedge-taken count is 1
+; CHECK-NEXT: Loop %do.body: max backedge-taken count is 1, actual taken count either this or zero.
 
 do.end:
   ret void
 }
 
 define void @s32_max2(i32 %n, i32* %p) {
 entry:
   %add = add i32 %n, 2
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
   %arrayidx = getelementptr i32, i32* %p, i32 %i.0
   store i32 %i.0, i32* %arrayidx, align 4
   %inc = add i32 %i.0, 1
   %cmp = icmp slt i32 %i.0, %add
   br i1 %cmp, label %do.body, label %do.end ; taken either 0 or 2 times
 
 ; CHECK-LABEL: Determining loop execution counts for: @s32_max2
 ; CHECK-NEXT: Loop %do.body: backedge-taken count is ((-1 * %n) + ((2 + %n) smax %n))
-; CHECK-NEXT: Loop %do.body: max backedge-taken count is 2
+; CHECK-NEXT: Loop %do.body: max backedge-taken count is 2, actual taken count either this or zero.
 
 do.end:
   ret void
 }
 
 define void @s32_maxx(i32 %n, i32 %x, i32* %p) {
 entry:
   %add = add i32 %x, %n
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
   %arrayidx = getelementptr i32, i32* %p, i32 %i.0
   store i32 %i.0, i32* %arrayidx, align 4
   %inc = add i32 %i.0, 1
   %cmp = icmp slt i32 %i.0, %add
   br i1 %cmp, label %do.body, label %do.end ; taken either 0 or x times
 
 ; CHECK-LABEL: Determining loop execution counts for: @s32_maxx
 ; CHECK-NEXT: Loop %do.body: backedge-taken count is ((-1 * %n) + ((%n + %x) smax %n))
-; CHECK-NEXT: Loop %do.body: max backedge-taken count is -1
+; CHECK-NEXT: Loop %do.body: max backedge-taken count is -1{{$}}
 
 do.end:
   ret void
 }
 
 define void @s32_max2_unpredictable_exit(i32 %n, i32 %x, i32* %p) {
 entry:
   %add = add i32 %n, 2
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %if.end ]
   %cmp = icmp eq i32 %i.0, %x
   br i1 %cmp, label %do.end, label %if.end ; unpredictable
 
 if.end:
   %arrayidx = getelementptr i32, i32* %p, i32 %i.0
   store i32 %i.0, i32* %arrayidx, align 4
   %inc = add i32 %i.0, 1
   %cmp1 = icmp slt i32 %i.0, %add
   br i1 %cmp1, label %do.body, label %do.end ; taken either 0 or 2 times
 
 ; CHECK-LABEL: Determining loop execution counts for: @s32_max2_unpredictable_exit
 ; CHECK-NEXT: Loop %do.body: <multiple exits> Unpredictable backedge-taken count.
-; CHECK-NEXT: Loop %do.body: max backedge-taken count is 2
+; CHECK-NEXT: Loop %do.body: max backedge-taken count is 2{{$}}
 
 do.end:
   ret void
 }
 
 define void @u32_max1(i32 %n, i32* %p) {
 entry:
   %add = add i32 %n, 1
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
   %arrayidx = getelementptr i32, i32* %p, i32 %i.0
   store i32 %i.0, i32* %arrayidx, align 4
   %inc = add i32 %i.0, 1
   %cmp = icmp ult i32 %i.0, %add
   br i1 %cmp, label %do.body, label %do.end ; taken either 0 or 1 times
 
 ; CHECK-LABEL: Determining loop execution counts for: @u32_max1
 ; CHECK-NEXT: Loop %do.body: backedge-taken count is ((-1 * %n) + ((1 + %n) umax %n))
-; CHECK-NEXT: Loop %do.body: max backedge-taken count is 1
+; CHECK-NEXT: Loop %do.body: max backedge-taken count is 1, actual taken count either this or zero.
 
 do.end:
   ret void
 }
 
 define void @u32_max2(i32 %n, i32* %p) {
 entry:
   %add = add i32 %n, 2
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
   %arrayidx = getelementptr i32, i32* %p, i32 %i.0
   store i32 %i.0, i32* %arrayidx, align 4
   %inc = add i32 %i.0, 1
   %cmp = icmp ult i32 %i.0, %add
   br i1 %cmp, label %do.body, label %do.end ; taken either 0 or 2 times
 
 ; CHECK-LABEL: Determining loop execution counts for: @u32_max2
 ; CHECK-NEXT: Loop %do.body: backedge-taken count is ((-1 * %n) + ((2 + %n) umax %n))
-; CHECK-NEXT: Loop %do.body: max backedge-taken count is 2
+; CHECK-NEXT: Loop %do.body: max backedge-taken count is 2, actual taken count either this or zero.
 
 do.end:
   ret void
 }
 
 define void @u32_maxx(i32 %n, i32 %x, i32* %p) {
 entry:
   %add = add i32 %x, %n
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
   %arrayidx = getelementptr i32, i32* %p, i32 %i.0
   store i32 %i.0, i32* %arrayidx, align 4
   %inc = add i32 %i.0, 1
   %cmp = icmp ult i32 %i.0, %add
   br i1 %cmp, label %do.body, label %do.end ; taken either 0 or x times
 
 ; CHECK-LABEL: Determining loop execution counts for: @u32_maxx
 ; CHECK-NEXT: Loop %do.body: backedge-taken count is ((-1 * %n) + ((%n + %x) umax %n))
-; CHECK-NEXT: Loop %do.body: max backedge-taken count is -1
+; CHECK-NEXT: Loop %do.body: max backedge-taken count is -1{{$}}
 
 do.end:
   ret void
 }
 
 define void @u32_max2_unpredictable_exit(i32 %n, i32 %x, i32* %p) {
 entry:
   %add = add i32 %n, 2
   br label %do.body
 
 do.body:
   %i.0 = phi i32 [ %n, %entry ], [ %inc, %if.end ]
   %cmp = icmp eq i32 %i.0, %x
   br i1 %cmp, label %do.end, label %if.end ; unpredictable
 
 if.end:
   %arrayidx = getelementptr i32, i32* %p, i32 %i.0
   store i32 %i.0, i32* %arrayidx, align 4
   %inc = add i32 %i.0, 1
   %cmp1 = icmp ult i32 %i.0, %add
   br i1 %cmp1, label %do.body, label %do.end ; taken either 0 or 2 times
 
 ; CHECK-LABEL: Determining loop execution counts for: @u32_max2_unpredictable_exit
 ; CHECK-NEXT: Loop %do.body: <multiple exits> Unpredictable backedge-taken count.
-; CHECK-NEXT: Loop %do.body: max backedge-taken count is 2
+; CHECK-NEXT: Loop %do.body: max backedge-taken count is 2{{$}}
 
 do.end:
   ret void
 }

test/Transforms/LoopUnroll/full-unroll-keep-first-exit.ll

+; RUN: opt -S -loop-unroll < %s | FileCheck %s
+
+; Unroll twice, with first loop exit kept
+; CHECK-LABEL: @s32_max1
+; CHECK: do.body:
+; CHECK:  store
+; CHECK:  br i1 %cmp, label %do.body.1, label %do.end
+; CHECK: do.end:
+; CHECK:  ret void
+; CHECK: do.body.1:
+; CHECK:  store
+; CHECK:  br label %do.end
+define void @s32_max1(i32 %n, i32* %p) {
+entry:
+  %add = add i32 %n, 1
+  br label %do.body
+
+do.body:
+  %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
+  %arrayidx = getelementptr i32, i32* %p, i32 %i.0
+  store i32 %i.0, i32* %arrayidx, align 4
+  %inc = add i32 %i.0, 1
+  %cmp = icmp slt i32 %i.0, %add
+  br i1 %cmp, label %do.body, label %do.end ; taken either 0 or 1 times
+
+do.end:
+  ret void
+}
+
+; Unroll thrice, with first loop exit kept
+; CHECK-LABEL: @s32_max2
+; CHECK: do.body:
+; CHECK:  store
+; CHECK:  br i1 %cmp, label %do.body.1, label %do.end
+; CHECK: do.end:
+; CHECK:  ret void
+; CHECK: do.body.1:
+; CHECK:  store
+; CHECK:  store
+; CHECK:  br label %do.end
+define void @s32_max2(i32 %n, i32* %p) {
+entry:
+  %add = add i32 %n, 2
+  br label %do.body
+
+do.body:
+  %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
+  %arrayidx = getelementptr i32, i32* %p, i32 %i.0
+  store i32 %i.0, i32* %arrayidx, align 4
+  %inc = add i32 %i.0, 1
+  %cmp = icmp slt i32 %i.0, %add
+  br i1 %cmp, label %do.body, label %do.end ; taken either 0 or 2 times
+
+do.end:
+  ret void
+}
+
+; Should not be unrolled
+; CHECK-LABEL: @s32_maxx
+; CHECK: do.body:
+; CHECK: do.end:
+; CHECK-NOT: do.body.1:
+define void @s32_maxx(i32 %n, i32 %x, i32* %p) {
+entry:
+  %add = add i32 %x, %n
+  br label %do.body
+
+do.body:
+  %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
+  %arrayidx = getelementptr i32, i32* %p, i32 %i.0
+  store i32 %i.0, i32* %arrayidx, align 4
+  %inc = add i32 %i.0, 1
+  %cmp = icmp slt i32 %i.0, %add
+  br i1 %cmp, label %do.body, label %do.end ; taken either 0 or x times
+
+do.end:
+  ret void
+}
+
+; Should not be unrolled
+; CHECK-LABEL: @s32_max2_unpredictable_exit
+; CHECK: do.body:
+; CHECK: do.end:
+; CHECK-NOT: do.body.1:
+define void @s32_max2_unpredictable_exit(i32 %n, i32 %x, i32* %p) {
+entry:
+  %add = add i32 %n, 2
+  br label %do.body
+
+do.body:
+  %i.0 = phi i32 [ %n, %entry ], [ %inc, %if.end ]
+  %cmp = icmp eq i32 %i.0, %x
+  br i1 %cmp, label %do.end, label %if.end ; unpredictable
+
+if.end:
+  %arrayidx = getelementptr i32, i32* %p, i32 %i.0
+  store i32 %i.0, i32* %arrayidx, align 4
+  %inc = add i32 %i.0, 1
+  %cmp1 = icmp slt i32 %i.0, %add
+  br i1 %cmp1, label %do.body, label %do.end ; taken either 0 or 2 times
+
+do.end:
+  ret void
+}
+
+; Unroll twice, with first loop exit kept
+; CHECK-LABEL: @u32_max1
+; CHECK: do.body:
+; CHECK:  store
+; CHECK:  br i1 %cmp, label %do.body.1, label %do.end
+; CHECK: do.end:
+; CHECK:  ret void
+; CHECK: do.body.1:
+; CHECK:  store
+; CHECK:  br label %do.end
+define void @u32_max1(i32 %n, i32* %p) {
+entry:
+  %add = add i32 %n, 1
+  br label %do.body
+
+do.body:
+  %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
+  %arrayidx = getelementptr i32, i32* %p, i32 %i.0
+  store i32 %i.0, i32* %arrayidx, align 4
+  %inc = add i32 %i.0, 1
+  %cmp = icmp ult i32 %i.0, %add
+  br i1 %cmp, label %do.body, label %do.end ; taken either 0 or 1 times
+
+do.end:
+  ret void
+}
+
+; Unroll thrice, with first loop exit kept
+; CHECK-LABEL: @u32_max2
+; CHECK: do.body:
+; CHECK:  store
+; CHECK:  br i1 %cmp, label %do.body.1, label %do.end
+; CHECK: do.end:
+; CHECK:  ret void
+; CHECK: do.body.1:
+; CHECK:  store
+; CHECK:  store
+; CHECK:  br label %do.end
+define void @u32_max2(i32 %n, i32* %p) {
+entry:
+  %add = add i32 %n, 2
+  br label %do.body
+
+do.body:
+  %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
+  %arrayidx = getelementptr i32, i32* %p, i32 %i.0
+  store i32 %i.0, i32* %arrayidx, align 4
+  %inc = add i32 %i.0, 1
+  %cmp = icmp ult i32 %i.0, %add
+  br i1 %cmp, label %do.body, label %do.end ; taken either 0 or 2 times
+
+do.end:
+  ret void
+}
+
+; Should not be unrolled
+; CHECK-LABEL: @u32_maxx
+; CHECK: do.body:
+; CHECK: do.end:
+; CHECK-NOT: do.body.1:
+define void @u32_maxx(i32 %n, i32 %x, i32* %p) {
+entry:
+  %add = add i32 %x, %n
+  br label %do.body
+
+do.body:
+  %i.0 = phi i32 [ %n, %entry ], [ %inc, %do.body ]
+  %arrayidx = getelementptr i32, i32* %p, i32 %i.0
+  store i32 %i.0, i32* %arrayidx, align 4
+  %inc = add i32 %i.0, 1
+  %cmp = icmp ult i32 %i.0, %add
+  br i1 %cmp, label %do.body, label %do.end ; taken either 0 or x times
+
+do.end:
+  ret void
+}
+
+; Should not be unrolled
+; CHECK-LABEL: @u32_max2_unpredictable_exit
+; CHECK: do.body:
+; CHECK: do.end:
+; CHECK-NOT: do.body.1:
+define void @u32_max2_unpredictable_exit(i32 %n, i32 %x, i32* %p) {
+entry:
+  %add = add i32 %n, 2
+  br label %do.body
+
+do.body:
+  %i.0 = phi i32 [ %n, %entry ], [ %inc, %if.end ]
+  %cmp = icmp eq i32 %i.0, %x
+  br i1 %cmp, label %do.end, label %if.end ; unpredictable
+
+if.end:
+  %arrayidx = getelementptr i32, i32* %p, i32 %i.0
+  store i32 %i.0, i32* %arrayidx, align 4
+  %inc = add i32 %i.0, 1
+  %cmp1 = icmp ult i32 %i.0, %add
+  br i1 %cmp1, label %do.body, label %do.end ; taken either 0 or 2 times
+
+do.end:
+  ret void
+}