[PATCH] D50426: [NFC][MustExecute] Rework API to start making better analysis

[Max Kazantsev via Phabricator via llvm-commits]

mkazantsev added inline comments.


================
Comment at: lib/Analysis/MustExecute.cpp:29
+  assert(CurLoop && "Should calculate the info first!");
+  assert(CurLoop->getHeader() == BB && "Currently we only support header!");
+  return HeaderMayThrow;
----------------
reames wrote:
> mkazantsev wrote:
> > reames wrote:
> > > Bad API.  A better adapter  implementation:
> > > if (getHeader() == BB)
> > >   return HeaderMayThrow;
> > > else
> > >  return MayThrow;
> > What you suggested contradicts the method's specification. It is not supposed to give false-positive answers.
> Er, I disagree.  "may" always allows a conservative answer.
> 
> A valid answer to the question "may X happen"? Is *always* yes.
> 
> (To avoid confusion: this is not normal english usage, it is specifically jargon used within the compiler field.  "may" and "must" have very specific meaning.  For examples, see may alias and must alias)
I would agree if there was no context. But this method has a comment which clearly states what it returns.


https://reviews.llvm.org/D50426