[PATCH] D69498: IR: Invert convergent attribute handling

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added a comment.

In D69498#1727626 <https://reviews.llvm.org/D69498#1727626>, @mehdi_amini wrote:

> In D69498#1727546 <https://reviews.llvm.org/D69498#1727546>, @jdoerfert wrote:
>
> > In D69498#1727419 <https://reviews.llvm.org/D69498#1727419>, @mehdi_amini wrote:
> >
> > > In D69498#1727080 <https://reviews.llvm.org/D69498#1727080>, @jdoerfert wrote:
> > >
> > > > Let me quote @arsenm here because this is so important: "Basically no frontend has gotten this right, including clang and non-clang frontends." For me, that statement alone is reason to change the status quo.
> > >
> > >
> > > I am not convinced by this: this seems overly simplistic. The fact that convergent wasn't rolled out properly in frontend can also been as an artifact of the past.
> >
> >
> > Evidence suggest this problem will resurface over and over again, calling it an artifact of the past is dangerous and not helpful.
>
>
> Can you provide the "evidence"? So far the only thing I have seen is historical and closely tied to how convergence was rolled out AFAICT.


I'm not sure what other evidence could exist for how something is handled in practice besides historical record. The current design requires a higher awareness at all times to avoid subtlety breaking code, which I think is fundamentally more error prone.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D69498/new/

https://reviews.llvm.org/D69498