diff --git a/lib/Analysis/ScalarEvolution.cpp b/lib/Analysis/ScalarEvolution.cpp index 8def10c2030..3cf2815e7bc 100644 --- a/lib/Analysis/ScalarEvolution.cpp +++ b/lib/Analysis/ScalarEvolution.cpp @@ -5646,6 +5646,13 @@ const SCEV *ScalarEvolution::getBECount(const SCEV *Start, "This code doesn't handle negative strides yet!"); const Type *Ty = Start->getType(); + + // When Start == End, we have an exact BECount == 0. Short-circuit this case + // here because SCEV may not be able to determine that the unsigned division + // after rounding is zero. + if (Start == End) + return getConstant(Ty, 0); + const SCEV *NegOne = getConstant(Ty, (uint64_t)-1); const SCEV *Diff = getMinusSCEV(End, Start); const SCEV *RoundUp = getAddExpr(Step, NegOne); @@ -5768,7 +5775,16 @@ ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS, // The maximum backedge count is similar, except using the minimum start // value and the maximum end value. - const SCEV *MaxBECount = getBECount(MinStart, MaxEnd, Step, NoWrap); + // If we already have an exact constant BECount, use it instead. + const SCEV *MaxBECount = isa(BECount) ? BECount + : getBECount(MinStart, MaxEnd, Step, NoWrap); + + // If the stride is nonconstant, and NoWrap == true, then + // getBECount(MinStart, MaxEnd) may not compute. This would result in an + // exact BECount and invalid MaxBECount, which should be avoided to catch + // more optimization opportunities. + if (isa(MaxBECount)) + MaxBECount = BECount; return BackedgeTakenInfo(BECount, MaxBECount); } diff --git a/test/Analysis/ScalarEvolution/2011-03-09-ExactNoMaxBECount.ll b/test/Analysis/ScalarEvolution/2011-03-09-ExactNoMaxBECount.ll new file mode 100644 index 00000000000..9f17e27577c --- /dev/null +++ b/test/Analysis/ScalarEvolution/2011-03-09-ExactNoMaxBECount.ll @@ -0,0 +1,34 @@ +; RUN: opt -indvars %s +; PR9424: Attempt to use a SCEVCouldNotCompute object! +; The inner loop computes the Step and Start of the outer loop. +; Call that Vexit. The outer End value is max(2,Vexit), because +; the condition "icmp %4 < 2" does not guard the outer loop. +; SCEV knows that Vexit has range [2,4], so End == Vexit == Start. +; Now we have ExactBECount == 0. However, MinStart == 2 and MaxEnd == 4. +; Since the stride is variable and may wrap, we cannot compute +; MaxBECount. SCEV should override MaxBECount with ExactBECount. + +define void @bar() nounwind { +entry: + %. = select i1 undef, i32 2, i32 1 + br label %"5.preheader" + +"4": ; preds = %"5.preheader", %"4" + %0 = phi i32 [ 0, %"5.preheader" ], [ %1, %"4" ] + %1 = add nsw i32 %0, 1 + %2 = icmp sgt i32 %., %1 + br i1 %2, label %"4", label %"9" + +"9": ; preds = %"4" + %3 = add i32 %6, 1 + %4 = add i32 %3, %1 + %5 = icmp slt i32 %4, 2 + br i1 %5, label %"5.preheader", label %return + +"5.preheader": ; preds = %"9", %entry + %6 = phi i32 [ 0, %entry ], [ %4, %"9" ] + br label %"4" + +return: ; preds = %"9" + ret void +}