[SCEV] Add a clarifying comment in howManyLessThans

Wrap semantics are subtle when combined with multiple exits.  This has caused several rounds of confusion during recent reviews, so try to document the subtly distinction between when wrap flags provide <u and <=u facts.

diff --git a/llvm/lib/Analysis/ScalarEvolution.cpp b/llvm/lib/Analysis/ScalarEvolution.cpp
index cb37958..d502ac2 100644 (file)
--- a/llvm/lib/Analysis/ScalarEvolution.cpp
+++ b/llvm/lib/Analysis/ScalarEvolution.cpp
@@ -11587,6 +11587,16 @@ ScalarEvolution::howManyLessThans(const SCEV *LHS, const SCEV *RHS,
   if (!IV || IV->getLoop() != L || !IV->isAffine())
     return getCouldNotCompute();
 
+  // A precondition of this method is that the condition being analyzed
+  // reaches an exiting branch which dominates the latch.  Given that, we can
+  // assume that an increment which violates the nowrap specification and
+  // produces poison must cause undefined behavior when the resulting poison
+  // value is branched upon and thus we can conclude that the backedge is
+  // taken no more often than would be required to produce that poison value.
+  // Note that a well defined loop can exit on the iteration which violates
+  // the nowrap specification if there is another exit (either explicit or
+  // implicit/exceptional) which causes the loop to execute before the
+  // exiting instruction we're analyzing would trigger UB.
   auto WrapType = IsSigned ? SCEV::FlagNSW : SCEV::FlagNUW;
   bool NoWrap = ControlsExit && IV->getNoWrapFlags(WrapType);
   ICmpInst::Predicate Cond = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;