[PATCH] D147932: ValueTracking: Implement computeKnownFPClass for various rounding intrinsics

[Jay Foad via Phabricator via llvm-commits]

foad added inline comments.


================
Comment at: llvm/lib/Analysis/ValueTracking.cpp:4873-4874
         FPClassTest InterestedSrcs = InterestedClasses;
         if (InterestedClasses & fcZero)
-          InterestedClasses |= fcNormal | fcSubnormal;
+          InterestedSrcs |= fcNormal | fcSubnormal;
 
----------------
arsenm wrote:
> foad wrote:
> > This doesn't seem to match the code any more? I guess you want:
> > ```
> >   if (InterestedClass & fcPosFinite)
> >     InterestedSrcs |= fcPosFinite;
> > ```
> > etc.
> > 
> > (Strictly I think you ought to do this for fcNan too, throughout computeKnownFPClass, in case InterestedClasses only contains one of fcSNan or fcQNan.)
> I don't see the point of enabling more bits here, not sure why I put this here in the first place 
The point is that if you're only interested in whether the result is (say) negzero, you can't prove that unless you know the source is none of negzero|negsubnormal|negfinite. Isn't that the way InterestedClasses is supposed to work?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D147932/new/

https://reviews.llvm.org/D147932