Make Sized coinductive, again

[compiler-errors]

A revival of #83647

---

What exactly makes co-induction sound? Better question: are there any unsoundness risks from this? Sized can't be implemented by custom impl blocks, nor can it be conditionally implemented based on anything other than child fields being Sized, right?

r? @nikomatsakis for whenever he gets back from vacation

[bors]

☔ The latest upstream changes (presumably #98807) made this pull request unmergeable. Please resolve the merge conflicts.

[nikomatsakis]

I think we should do this, but I think it requires an FCP.

The soundness challenges with coinduction are tied to elaborated bounds, which don't apply to Sized as it has no supertraits.

Ideally we'd have an RFC for moving all trait matching to coinduction.

[compiler-errors]

Since exploring moving all traits to be coinductive is significantly harder and likely requires an RFC/opt-in/migration/etc., I'm following Nikos'a advice above and proposing merging just this for now.

@rfcbot merge

[rfcbot]

Team member @compiler-errors has proposed to merge this. The next step is review by the rest of the tagged team members:

• @compiler-errors
• @jackh726
• @lcnr
• @nikomatsakis
• @oli-obk
• @spastorino

Concerns:

• does this limit Sized having supertrait predicates (like trait Thin =Pointee<Metadata = ()>) in the future? resolved by Make Sized coinductive, again #100386 (comment)

Once a majority of reviewers approve (and at most 2 approvals are outstanding), this will enter its final comment period. If you spot a major issue that hasn't been raised at any point in this process, please speak up!

See this document for info about what commands tagged team members can give me.

[nikomatsakis]

@rfcbot reviewed

[rfcbot]

🔔 This is now entering its final comment period, as per the review above. 🔔

[bors]

☔ The latest upstream changes (presumably #96709) made this pull request unmergeable. Please resolve the merge conflicts.

[madsmtm]

don't apply to Sized as it has no supertraits.

Just wanted to note that it is being considered in #81513 to add Thin as a supertrait of Sized.

I don't know anything about coinduction, but if the only reason this is sound is that Sized has no supertraits, then you may want to reconsider doing this?

[compiler-errors]

@madsmtm:

I'm not sure if it makes sense to add Thin as a supertrait of Sized nowadays -- the Pointee trait is implemented for all types, but there is special logic in the compiler to make sure that T: Sized => T: Thin without needing to enforce that relationship as a supertrait (#93977).

I'm not aware of any benefits of that relationship being reflected as a supertrait relationship, other than simplifying some of the builtin trait code... I can put that in as a concern and raise it up at a T-types meeting though.

[compiler-errors]

@rfcbot concern does this limit Sized having supertrait predicates (like trait Thin =Pointee<Metadata = ()>) in the future?

[compiler-errors]

(Presumably not -- we could work around this with the same modifications to the trait system as we need to enable coinduction for all traits, which we eventually do want. But it would be a shame to block a library feature that many people want on a trait system modification.)

[Jules-Bertholet]

rust-lang/rfcs#3319 would add Aligned as a supertrait of Sized.

[nikomatsakis]

In general, having supertrait predicates is not a concern, but we will have to do some work on our trait system implementation before we can properly handle it.

[compiler-errors]

@nikomatsakis, do you think we should block this due to possible RFCs adding supertrait bounds to Sized? I'm not sure if I have a strong opinion on this.

[lcnr]

I don't think we should block this on adding a super trait. Making Sized coinductive rn has a noticeable positive impact as it is necessary to abstract over collections for recursive structures, see src/test/ui/sized/coinductive-2.rs added in this PR.

We intend to support super traits for coinductive types in the future anyways and I expect us to be able to deal with this once necessary.

[compiler-errors]

@rfcbot resolve does this limit Sized having supertrait predicates (like trait Thin =Pointee<Metadata = ()>) in the future?

[lcnr]

r? @lcnr

[lcnr]

Should mention this in the full relnotes. This PR changes Sized obligations to be coinductive. This means that if the trait solver encounters a cycle which only includes auto traits and now Sized, it returns "Success" instead of an error.

An example would be

struct Node<C: Trait<Self>>(C::Assoc);

trait Trait<T> {
    type Assoc;
}

impl<T> Trait<T> for Vec<()> {
    type Assoc = Vec<T>;
}

fn main() {
    let _ = Node::<Vec<()>>(Vec::new());
}

Node<Vec<()>> is Sized if <Vec<()> as Trait<Node<Vec<()>>>::Assoc is Sized. After normalization, this requires Vec<Node<Vec<()>> to be Sized. Vec<Node<Vec<()>> again requires Node<Vec<()>> to be Sized, resulting in a cycle. As Sized is now coinductive, this does not result in an error anymore.

[compiler-errors]

whoops I forgot to approve this after I rebased

@bors r=lcnr

[bors]

📌 Commit fea8d0e has been approved by lcnr

It is now in the queue for this repository.