"parameter `T` is never used" with recursive type

[eefriedman]

Testcase:

enum Foo<T> {
    F(Option<Box<Foo<Box<T>>>>)
}
fn main() { }

Gives:

<anon>:16:10: 16:11 error: parameter `T` is never used [E0392]
<anon>:16 enum Foo<T> {
                   ^
<anon>:16:10: 16:11 help: consider removing `T` or using a marker such as `core::marker::PhantomData`

Not sure if this is really something rustc should actually support (I imagine trying to prove recursive properties about it is extremely awkward), but the error message makes no sense.

More related testcases:

// Crashes rustc with a stack overflow
use std::marker::PhantomData;
enum Foo<T> {
    F(Option<Box<Foo<Box<T>>>>, PhantomData<T>)
}
fn takesfoo<T>(f:&Foo<T>) {
    match f {
        &Foo::F(Some(ref b), PhantomData) => takesfoo(&**b),
        &Foo::F(None, PhantomData) => ()
    }
}
fn main() { }

// Gives "error: overflow while adding drop-check rules for Foo<i32>"
use std::marker::PhantomData;
enum Foo<T> {
    F(Option<Box<Foo<Box<T>>>>, PhantomData<T>)
}
fn main() { let _f = Foo::F::<i32>(None, PhantomData); }

[arielb1]

These are called "non-regular ADT-s" and are not supported (you should get an overflow complaint). We shouldn't stack-overflow through.

[arielb1]

Your second example somehow survived type checking wow (you really shouldn't have gone that far). I guess we stopped using type_contents for pretty much anything.

I think we should add an actual check for this somewhere (althrough it won't prevent examples using associated types, which can strike in trans, but we should have a limit for that).

[steveklabnik]

Triage: no change

[leoyvens]

Triage: Examples 2 now compiles, but if you try to use takefoo then the same error as the third example happens.

[Spoonbender]

Triage: no change from latest check