[PATCH] D159002: Attributor: Fix not propagating nofpclass arguments through transitive callers

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: llvm/lib/Transforms/IPO/AttributorAttributes.cpp:10379
     State.addKnownBits(~KnownFPClass.KnownFPClasses);
-
-    bool TrackUse = false;
-    return TrackUse;
+    return true;
   }
----------------
jdoerfert wrote:
> jdoerfert wrote:
> > arsenm wrote:
> > > jdoerfert wrote:
> > > > Do we really follow any use?
> > > The only uses I could think of being worth skipping would be stores or void calls, which would naturally have no uses.
> > > 
> > > Maybe unrecognizable generic calls? Filtering out those is more effort if you want to consider libcalls
> > I mean, this could just flow into a call, what has the call return to do with the value we are tracking, right?
> > This could also flow into a fp2int and then into all sorts of things.
> So, we want it to follow fadd and friends, is that correct?
> Does the qualification of the operation result apply to the operands all the time?
Should be anything that computeKnownFPClass is implemented for, which ideally would be all recognized FP operations


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D159002/new/

https://reviews.llvm.org/D159002