[PATCH] D73978: [WIP][FPEnv] Don't transform FSUB(-0.0,X)->FNEG(X) when flushing denormals

Thu Jul 16 14:11:09 PDT 2020

arsenm added a comment.

In D73978#2153300 <https://reviews.llvm.org/D73978#2153300>, @cameron.mcinally wrote:

> In D73978#2151964 <https://reviews.llvm.org/D73978#2151964>, @arsenm wrote:
>
> > The thing I'm somewhat worried about is a subtlety with constant folding. Constant folding will blindly fold unaware of whatever canonicalization needed to happen. willCanonicalize may have lied if something happened later that caused the canonicalizing operation to constant fold away
>
>
> Ah, good point. I remember you saying that before, but I didn't absorb it at the time.
>
> That's a sticky problem. We could wait until the MachineInstr level to do the FSUB->FNEG transform, to ensure that constant folding completed. But I suspect (pretty certain) that we'll have missed other FNEG peeps we'd want by then. So that won't work.
>
> In general, it would be good to go for functional correctness first, and then try to optimize. That's kind of a problem for this specific project though, since so many existing tests would need to be updated. I'm not sure what to do. Will need to think about it...

If we modeled everything correctly, the correct thing to do would be to insert canonicalizes whenever occurs (but that's a massive change). I'm not a huge fan of getNode doing constant folding, so maybe eliminating that at least would help?

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D73978/new/

https://reviews.llvm.org/D73978