`mem::size_of_val` returns wrong value for DSTs with certain field layouts

[apasel422]

On stable, beta, and nightly, the follow code prints 16 followed by 24:

struct Foo<T: ?Sized> {
    a: i64,
    b: bool,
    c: T,
}

fn main() {
    let a = Box::new(Foo {a: 1, b: false, c: 1u32});
    println!("{}", std::mem::size_of_val(&*a));

    let a: Box<Foo<Send>> = a;
    println!("{}", std::mem::size_of_val(&*a));
}

CC #27023

[eddyb]

This is probably for drop flags, in which case it should get fixed by #35764

[apasel422]

Reduced:

struct Foo<T: ?Sized> {
    a: i64,
    b: bool,
    c: T,
}

fn main() {
    let a: &Foo<i32> = &Foo { a: 1, b: false, c: 2i32 };
    println!("{}", std::mem::size_of_val(a));

    let a: &Foo<Send> = a;
    println!("{}", std::mem::size_of_val(a));
}

[durka]

@eddyb where would there be a drop flag here?

[eddyb]

@durka My point isn't that there is, but that the code assumes all trait objects have drop flags. I could be wrong though, it might be a bug in the alignment logic.

[eddyb]

Not fixed by #35764, will have to investigate further.

[eddyb]

Uhhhh, the computation starts with:

rustc_trans::glue: DST Foo<std::marker::Send> sizing_type: { i64, i8 }
rustc_trans::glue: DST Foo<std::marker::Send> statically sized prefix size: 16 align: 8

Which is completely wrong, because you want no trailing padding, i.e. size should be 9.

EDIT: getting a field pointer might also be broken, needs looking into.
EDIT2: there's code to handle the "minimum size" needed to compute the field offset, so that works.