wjristow wrote:

Given @andykaylor's comment:
> I continue to believe that ignoring pragmas when fast-math is enabled is a 
> bug, and I'd like to see a more general change such as I proposed in 
> `https://github.com/llvm/llvm-project/pull/105746`.

and @efriedma-quic's comment:

> Maybe time to revisit #105746; the situation in the backends looks a lot 
> better. Looks like Arm and Mips still have some problematic isel patterns; 
> everything else looks okay at first glance?
>
> We could blacklist specific backends which still have issues.

I think revisiting that is the right thing to do.  That said, if complications 
on some specific backends make that difficult, we could move forward with this 
PR, fixing this on PlayStation (in our downstream version, we're already doing 
that).

Regarding blacklisting specific backends, if we were to do this PR, we could 
also do the dual of that blacklist approach: That is, others that want the 
pragmas to always be honored could opt-in to the approach proposed here in this 
PR.  I've just added a note in Andy's PR #105746, suggesting these two 
possibilities.

https://github.com/llvm/llvm-project/pull/162549
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to