Reject bounds in type aliases a future edition

[RalfJung]

I propose to make the type_alias_bounds lint deny-by-default with the next edition. See #21903 for more information on the issue being linted, and #48326 and #48909 for the PRs implementing the lint.

Cc @nikomatsakis @Manishearth (and Manish told me to also Cc someone else but I forgot whom)

[RalfJung]

Also maybe in the light of this, the lint message should be extended to also link to #21903 or something else providing further information?

[Manishearth]

cc @withoutboats

We can make it a hard error instead, fwiw.

[RalfJung]

We can make it a hard error instead, fwiw.

I thought deny-by-default-lint was as hard as we wanted to make things in the next edition, but yeah -- there is no good reason to still allow this code, other than to ease porting to the next edition by allowing the lint.

Would that mean this is a topic for rustfix?

[Manishearth]

Yes -- edition breakages work by taking a warn lint in a certain group and making it a hard error; and having a good suggestion that rustfix can apply

[RalfJung]

@nikomatsakis you mentioned some plans to fix where clauses in type aliases properly, so that they are added as obligations when type aliases are expanded. What is the status of that? I think this lint should become an error in the next edition, but whether it becomes Deny-by-default or a hard error I guess depends on whether we need this to be rule out for a transition in the 2021 edition.

[RalfJung]

After some talking with @Centril, I think ideally we want a lint that

• complains about type SendVec<T: Send> = Vec<T> because the bound is ignored
• complains about type Baz<I> = <I as Iterator>::Item because associated items should have a bound
• does not complain about type Baz<I: Iterator> = <I as Iterator>::Item because the bound is, de facto, enforced after expansion.

What makes this hard is that associated items are not the only way that a bound can be used; using a type like struct Foo<T: Send>(T) is another way -- but @eddyb said he had ideas.

[eddyb]

My idea is to diff what WF needs (i.e. the things that would error if not satisfied) with the predicates (bounds) written by the user.

Without Chalk this will be tricky, but we can manually run trait selection (or just plug something optional into the fulfillment context) to peek at how each WF requirement is satisfied, and mark every predicate that actually ends up used as part of that process, as useful - everything else, we'll warn (we can even have a consistency check, that if we remove the "unused bounds", WF still passes).

EDIT: Hang on, we might be able to do "difference" much easier by trying every declared bound with WF bounds as assumed: if they fail, that means the WF conditions enforced onto the user of the type alias aren't enough to subsume it, which is really what we want to check for (equivalence)!

[eddyb]

We actually have a very good reason to check "WF of RHS implies bounds" and not "used":

type DEIItem<T: DoubleEndedIterator> = T::Item;

The DoubleEndedIterator bound won't be checked in uses, only WF(T::Item) i.e. T: Iterator will.
And whenever we can start checking uses, then we don't really care if the bounds are used.

[eddyb]

I guess this is the same problem I recognized last year (#44075 (comment)), but now I have a plan.