Broken compilation with `&(dyn Trait + '_)`

[Cerber-Ursi]

Code

I tried this code:

pub trait AsStr {
    fn as_str(&self) -> &str;
}

fn foo(bar: &(dyn AsStr + '_)) -> &'_ str {
    bar.as_str()
}

I expected to see this happen: code compiles successfully.

Instead, this happened:

error[E0106]: missing lifetime specifier
 --> src/main.rs:5:36
  |
5 | fn foo(bar: &(dyn AsStr + '_)) -> &'_ str {
  |             -----------------      ^^ expected named lifetime parameter
  |
  = help: this function's return type contains a borrowed value, but the signature does not say which one of `bar`'s 2 lifetimes it is borrowed from
help: consider introducing a named lifetime parameter
  |
5 | fn foo<'a>(bar: &'a (dyn AsStr + 'a)) -> &'a str {
  |       ++++       ++              ~~       ~~

For more information about this error, try `rustc --explain E0106`.

Version it worked on

It most recently worked on Rust 1.63.0.

Version with regression

It is broken on current stable, i.e. 1.64.0.

Here's the output by cargo-bisect-rustc:

searched nightlies: from nightly-2022-06-24 to nightly-2022-08-05
regressed nightly: nightly-2022-07-15
searched commit range: 87588a2...c2f428d
regressed commit: f1a8854

bisected with cargo-bisect-rustc v0.6.4

Host triple: x86_64-unknown-linux-gnu
Reproduce with:

cargo bisect-rustc --end=1.64.0 -- check 

It seems to have been found previously by gluon developers and fixed in their own code in this commit, and today popped up also on URLO.

@rustbot modify labels: +regression-from-stable-to-stable -regression-untriaged

[csmoe]

Seems related to #97720
cc @cjgillot

[apiraino]

I think it's very close to #100615, so I'll close this as duplicate. Please feel free to reopen if it's not the case.

[apiraino]

Reverting the decision because I misunderstood the context and assigning priority (Zulip discussion).

@rustbot label -I-prioritize +P-high

[BGR360]

Indeed, from what I can tell, it seems that #97720 regressed this. Specifically, this change here.

My understanding of this change is that it stopped pushing lifetime resolution in function signatures to typeck, and instead made it so that every occurrence of '_, &'_, or & in the function's inputs resulted in a "new" lifetime.

So I guess rustc used to think that the '_ in &(dyn AsStr + '_) was just referring to the same lifetime as the &. But now it is being treated the same way as, e.g., the second & in (a: &str, b: &str) would be treated.

It seems a little unfortunate that we ever accepted this. I say this because having &(dyn Trait + '_) doesn't really seem any different than &Type<'_> to me. Yet we never accepted the latter in previous versions of Rust:

struct Type<'a> { a: &'a () }

fn foo(_: &Type<'_>) -> &'_ str { loop {} }

fn main() {}

$ rustc +1.63.0 test.rs
error[E0106]: missing lifetime specifier
 --> test.rs:3:26
  |
3 | fn foo(_: &Type<'_>) -> &'_ str { loop {} }
  |           ---------      ^^ expected named lifetime parameter
  |
  = help: this function's return type contains a borrowed value, but the signature does not say which one of argument 1's 2 lifetimes it is borrowed from
help: consider introducing a named lifetime parameter
  |
3 | fn foo<'a>(_: &'a Type<'_>) -> &'a str { loop {} }
  |       ++++     ++               ~~

error: aborting due to previous error

I'm interested in trying to undo the regression, so I'll claim this. It seems like what I'd need to do is make the + '_ resolve the same way the -> &'_ str gets resolved. Unsure how hard that will be yet, I haven't spent much time in rustc_resolve yet.

But perhaps there's someone who thinks we should just let it regress?

@rustbot claim
@rustbot label +A-lifetimes +A-resolve