It's often not possible to implement ToOwned for user composite types

[uzytkownik]

&str and &[u8] seems to work by internal magic but it is hard to implement ToOwned for user composite types. Consider:

struct Foo<'a> {
    foo: &'a str,
    bar: &'a str,
    foobar: &'a str
}

struct FooBuf {
    foo: String,
    bar: String,
    foobar: String
}

It is not possible to mark that FooBuf is an owned version of Foo as it would require implementing Borrow for FooBuf. However it is not possible to implement fn borrow(&FooBuf) -> &Foo as we don't have anywhere we can store Foo. Built-in types workaround it by handling &str and similar types separately from &T and packing the actual structure there.

(I know too little about type system to create any RFR which would fix it)

[retep998]

The problem with Borrow enforcing the return type to be a &T is similar to the situation with Index where it is impossible to return any sort of actual value which isn't a &T.

[bluss]

There's sort of a design for an extension of the "Borrow" mechanism for hash tables here: indexmap-rs/indexmap#10 (not in any packaged version yet). It's backwards compatible modulo type inference regressions.. :-/

The new trait allows this kind of key equivalence of Foo and FooBuf.