Lint on overly-complex lifetime annotations when lifetime subtyping would handle it

[scottmcm]

A common mistake for beginners writing lifetime annotations is to write something like this:

pub fn min<'a: 'c, 'b: 'c, 'c, T: Ord>(a: &'a T, b: &'b T) -> &'c T {
    if b < a { b } else { a }
}

which while it's not wrong, is rather more complex than necessary.

It would be nice for the compiler to use its own knowledge of subtyping to suggest that the signature change to

pub fn min<'c, T: Ord>(a: &'c T, b: &'c T) -> &'c T {
    if b < a { b } else { a }
}

since that's equivalent.

[QuineDot]

They're not exactly the same since the lifetimes are early bound in the first case and late bound in the second case.

If one applies the suggestion to a trait fn, it will break all implementors, and if to an implementation of a trait, it breaks that implementation (see #135350 or the linked issues). I.e. all implementors need to match the early or lateness of the trait definition.

It will break any call site with lifetime turbofish, whether the fn is in a trait or not.

Playground examples.

Playground code inline

pub fn min_1<'a: 'c, 'b: 'c, 'c, T: Ord>(a: &'a T, b: &'b T) -> &'c T {
    if b < a { b } else { a }
}

pub fn min_2<'c, T: Ord>(a: &'c T, b: &'c T) -> &'c T {
    if b < a { b } else { a }
}

pub trait Trait {
    fn min_1<'a: 'c, 'b: 'c, 'c, T: Ord>(a: &'a T, b: &'b T) -> &'c T;
    fn min_2<'c, T: Ord>(a: &'c T, b: &'c T) -> &'c T;
}

impl Trait for () {
    // E0195
    fn min_1<'c, T: Ord>(a: &'c T, b: &'c T) -> &'c T {
        if b < a { b } else { a }
    }
    // E0195
    fn min_2<'a: 'c, 'b: 'c, 'c, T: Ord>(a: &'a T, b: &'b T) -> &'c T {
        if b < a { b } else { a }
    }
}

fn main() {
    let _ = min_1::<'static, '_, '_>(&0, &1);
    // E0794
    let _ = min_2::<'static, '_, '_>(&0, &1);
}