Problems with "dyn* Trait"

[quinedot]

Most of this comment is just nits.

@Blog

In Rust today, the type dyn Trait is guaranteed to implement the trait Trait , so long as Trait is dyn safe.

Probably irrelevant to this topic, but this has never been true.

@Blog

Box<dyn Trait> would become dyn* Trait (note that a ’static bound is implied today; this might be worth reconsidering, but that’s a separate question).

Box<dyn Trait> bounds are more nuanced than that.

• The lifetime is inferred within a function body
  • RFC 1156
• The lifetime is bound in generic structs with an explicit bound
  • RFC 2093 (maintaining the RFC 1156 rule for this case)

@withoutboats

In particular, if you have &'a mut dyn Trait + 'b , you are covariant in 'a and invariant in 'b .

While true from a subtype perspective, Unsize coercion can make 'b act covariantly (search for "Interaction with object coercion"). ^[1] This doesn't matter when the lifetimes are intertwined with others in the signature, as you note.

@tema2

Not a problem, but an alternative direction we can take here. We can make dyn Trait a sized type:

• allow various owning pointers [of certain layout] to be coerced to an owned dyn Trait type
• this notion can be extended to allow arbitrary unsized types, so that [T] type is sized, an actual storage is determined at creation site.

How would that be phased in considering these currently-non-overlapping implementations?

impl<T /* : Sized */> Trait for [T] {}
impl<T /* : Sized */> Trait for T {}
impl Trait for dyn Display {}

---

1. And maybe the coercion makes the subtype distinction meaningless. ↩︎