Regression in required trait bounds in Rust 1.78.0 about rust

[alexcrichton]

Ah no worries! I was able to adapt this and get something working, albeit not as egonomically as desired.

In any case thanks for the quick responses here, it's definitely appreciated!

[lcnr]

The hidden type of impl Host ends up as &'a U and we currently don't have an implied U: 'a bound here. This breaks the invariants of our type system, using a pattern which can result in actual unsoundnesses. I don't think your snippet is unsound as is, but also don't know how we would fix #114936 without breaking your example

[lcnr]

your code snippet does not seem to compile in 1.77 or 1.76? 🤔 https://rust.godbolt.org/z/vzhn87bfz

[alexcrichton]

That may be a godbolt issue perhaps? https://rust.godbolt.org/z/e6nxK5oaq shows it passing with 77/76 and locally 1.77.0 compiles the example for me.

[compiler-errors]

@alexcrichton: The code you shared in the issue has an extra + 'a, causing it to fail both before and after 1.78.

[alexcrichton]

Er apologies, that's my mistake. I was trying various versions locally and got the wires crossed. I've updated the above which also explains why the two godbolt links are differing in results.

I was surprised that this actually worked on 1.77.0, so is this actually a soundness issue fixed by #118882?

[alexcrichton]

Ah ok, if this is expected no worries, I just wanted to confirm!

Would you happen to know off the top of your head if there's a way we could somehow get this working? The ideal thing we want is:

fn foo<T>(f: impl Fn(&mut T) -> impl Host) { ... }

and we're trying to hack around it of sorts with a custom GetHost trait, but I'm having a tough time figuring out how to model this use case.

[alexcrichton]

Ah well in any case this appears to be expected so there's no issue here I believe. We'll work on getting something alternative working instead.

[lcnr]

Would you happen to know off the top of your head if there's a way we could somehow get this working? The ideal thing we want is:

It's really hard as you need a for<'a> U: 'a bound which currently forces U to be 'static while ideally it would only allow instantiating 'a with lifetimes which outlive 'a. We will support this in a few years as these kinds of bounds are necessary to fix e.g. #25860

but up until then I also don't have a nice fix for this :/