Normalization issue with operators or `Deref`

[lnicola]

CC #8961

let s = String::new() + &String::new(); // s is <String as Add<String>>::Output

[lnicola]

This might be related to Deref or something else. The following snipped works as expected:

use std::ops::Add;

struct S;

impl Add<&Self> for S {
    type Output = S;

    fn add(self, _rhs: &Self) -> Self::Output {
        todo!()
    }
}

fn main() {
    let s = S + &S;
}

[flodiebold]

Actually, String::new() + String::new() is a type error, isn't it?

Edit: let s = String::new() + &String::new() is valid and also a problem for us, though.

So this might actually be the same as #9968.

[flodiebold]

Yeah, I think the problem is that we don't do coercions on the right hand side in binary ops (and assignment is a binary op, so that's #9968).

[lnicola]

Actually, String::new() + String::new() is a type error, isn't it?

Well, this is awkward...

[lnicola]

Should we also coerce on the LHS?

[flodiebold]

There's nothing to really coerce to at that point. (rustc does coerce to a fresh type variable, but just to avoid too strict lifetime constraints.)

[lnicola]

The problem is that we only do the trait resolution at the end, but we need to do it before checking the right-hand side (with a variable in place of the RHS type). For String, that will tell us that the RHS type has to be &str.