the type of ProjectionElem::Field is not correct wrt subtyping

[lcnr]

struct Foo<T>(T); // `T` is covariant.

fn foo<'b>(x: Foo<for<'a> fn(&'a ())>) {
    let Foo(y): Foo<fn(&'b ())> = x;
}

fn main() {}

should compile but errors with

error[E0308]: mismatched types
 --> src/main.rs:4:13
  |
4 |     let Foo(y): Foo<fn(&'b ())> = x;
  |             ^ one type is more general than the other
  |
  = note: expected fn pointer `for<'a> fn(&'a ())`
             found fn pointer `fn(&())`

looking at the mir of foo it becomes clear why that happens:

fn foo(_1: Foo<for<'a> fn(&'a ())>) -> () {
    debug x => _1;                       // in scope 0 at src/main.rs:3:12: 3:13
    let mut _0: ();                      // return place in scope 0 at src/main.rs:3:40: 3:40
    let _2: fn(&()) as UserTypeProjection { base: UserType(0), projs: [Field(field[0], ())] }; // in scope 0 at src/main.rs:4:13: 4:14
    scope 1 {
        debug y => _2;                   // in scope 1 at src/main.rs:4:13: 4:14
    }

    bb0: {
        AscribeUserType(_1, +, UserTypeProjection { base: UserType(1), projs: [] }); // scope 0 at src/main.rs:4:17: 4:32
        StorageLive(_2);                 // scope 0 at src/main.rs:4:13: 4:14
        _2 = (_1.0: fn(&()));            // scope 0 at src/main.rs:4:13: 4:14
        _0 = const ();                   // scope 0 at src/main.rs:3:40: 5:2
        StorageDead(_2);                 // scope 0 at src/main.rs:5:1: 5:2
        return;                          // scope 0 at src/main.rs:5:2: 5:2
    }
}

notice that in _2 = (_1.0: fn(&())); the type of _1.0 is actually for<'a> fn(&'a ()).

Possible ways to fix this are:

• correctly deal with subtyping when sanitizing field projections, note that they cannot just always be covariant.
• change the type of the field to be the actual one during mir building, so that subtyping actually happens at the assignment
• always lazily recompute the type of the field instead of storing it in the mir

[lcnr]

@rustbot claim

intend to work on this myself

[lcnr]

also add a test for

struct Foo<T>(T); // `T` is covariant.

fn foo<'b>(x: Foo<Foo<for<'a> fn(&'a ())>>) {
    let Foo(Foo(y)): Foo<Foo<fn(&'b ())>> = x;
}

[lcnr]

change the type of the field to be the actual one during mir building, so that subtyping actually happens at the assignment

That's imo the correct solution. Open to mentor someone here, ideally with some experience with working with ty::Ty and mir building.

If you're interested and need help, contact me on zulip and we can either chat about it there or on zoom.

[b-naber]

@rustbot claim

[trevyn]

This needs a reopen since 696563e landed.

[lqd]

Reopening, as I don't think @RalfJung expected to close this with the rust-lang/miri#2738 miri rustup.

[RalfJung]

Yeah that was some rustc commit with a "Fixes #96514" in it being mirrored. This might happen a few more times, since github will re-execute these commit commands when commits move around repositories.