Why the code is forbidden

help
1

The code is taken from learing rust by @quinedot

As it says,

fn foo<'s, 'l: 's>(v: std::cell::Cell<Box<dyn Display + 'l>>) -> std::cell::Cell<Box<dyn Display + 's>> {
    v
}

can be compiled, but

fn foo<'s, 'l: 's>(v: std::cell::Cell<Box<Box<dyn Display + 'l>>>) -> std::cell::Cell<Box<Box<dyn Display + 's>>> {
    v
}

can't.

It does, but I am thinking is the 2nd one really dangerous?

The Variance is so called unsized coercions in the book, though is really nothing to do with the sizeness.

I think the reason underhook it because dyn Display + 's is just an unknown type, you can't assign some reference to the type under dyn Display, you can only annotate the type's lifetime as whole.

But when it's a member of struct, it is different; because T<dyn Trait + 'long> means T maybe as long as 'long; Cell<Box<dyn Display + 's>>, for example, dyn Display has a life 's, then Box<<dyn Display + 's>> can't exceed the 's too, then Cell<Box<dyn Display + 's>> is same.

Why

fn foo<'s, 'l: 's>(v: std::cell::Cell<Box<dyn Display + 'l>>) -> std::cell::Cell<Box<dyn Display + 's>> {
    v
}

is accepted because the cell is passed by value, after Cell<'l> was considered as Cell<'s>, it can't be move out to outer scope of 's.

But I think the same applies to Cell<Box<Box<dyn Display + 's>>>, there is no chance for it to be moved out of 's, so seems there is no chance for it to result in dangerous things.

2

If we accept the existing "ultra-covariance" coercions (that aren't really variance) are sound, I can't think of why the nested one involving only lifetimes would not be sound.

But allowing such a coercion would have to be a special case instead of utilizing the same trait-based unsizing coercions as, say, supertrait upcasts (as those can change the vtable pointer). Which is how things work today.

2 Likes