[PATCH] D148544: ValueTracking: Handle fptrunc_round in computeKnownFPClass

[Jay Foad via Phabricator via llvm-commits]

foad added inline comments.


================
Comment at: llvm/lib/Analysis/ValueTracking.cpp:4078
+  KnownFPClass KnownSrc;
+  computeKnownFPClass(Op->getOperand(0), DemandedElts, InterestedClasses,
+                      KnownSrc, Depth + 1, Q, TLI);
----------------
arsenm wrote:
> foad wrote:
> > I think there might be subtle bugs lurking in the way you pass InterestedClasses into recursive calls. E.g if InterestedClasses was fcQNan here, and the recursive call returned fcNone, you would conclude the operand is not nan - but in reality it could still be snan.
> > 
> > If I'm right that that's a bug then you could fix this case by passing fcNan into the recursive call.
> > 
> > Anyway this is pre-existing.
> InterestedClasses shouldn't be an issue, it's only used for early exiting as it's a compile time hint. The nan propagation could be better, I started introducing a helper for this in D148599
OK, if we always return conservatively correct info the for the un-interested-classes then I guess there's no bug.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D148544/new/

https://reviews.llvm.org/D148544