Github Add an impl of Error on `Arc<impl Error>`. by derekdreery · Pull Re...

Member

unlike Box we should be able to add the ?Sized bound.

Contributor

That would prevent a From impl for E: Error in the future, we okay with that? Seems like an inconsistency that would probably bother a lot of people... (not sure this is a good reason to not do it)

Author

Contributor

I'm not sure there would be many ?Sized errors in practice. Could this bound be relaxed (with ?Sized) at a later date and still be backwards-compatible?

Author

Contributor

I mean, it might be that it does no harm if we assume a future specialization feature would allow From impl for E: Error in the future.

Contributor

we couldn't put a ?Sized on in the future backwards compatibly. This is a problem with Box right now. This causes overlap for the impls From<T> for T and From<E: Error> for Arc<dyn Error + ...>. It would become ambiguous when converting Arc<dyn Error> into an Arc<dyn Error> whether or not you should do a unit conversion or wrap the Arc in another Arc.

As for specialization, I'm not as sure but as far as I know it won't help here. I've looked this up before and my understanding is that this would require rustc to implement the lattice rule, but afaik there are no plans to implement this. That said, I've never had this confirmed by @rust-lang/wg-traits, so this might be on the roadmap even if it's not in the RFC, or I could be misunderstanding the RFC.

Author

Contributor

This stuff is extremely nuanced, thankyou for taking the time to work though it all.

Contributor

I think ideally we'd want the ?Sized bound on there. The fact that Box<dyn Error>, which is a very common pattern, doesn't implement Error is something we regret. So I think aligning this impl to the one for &'a E: Error + ?Sized is better than aligning it to Box<E: Error>.

Contributor

Agreed, @derekdreery, could you go ahead and apply this suggestion?