Existensial types in Scala

Question

Little confusing about existential types.

This works for me:

def valueOf(c: Class[_], name: String) {
  type C = Class[T] forSome {type T <: Enum[T]}
  Enum.valueOf(c.asInstanceOf[C], name)
} 

but this does not:

def valueOf(c: Class[_], name: String) {
  type T = T forSome {type T <: Enum[T]}
  Enum.valueOf(c.asInstanceOf[Class[T]], name)
}

In my mind both expressions are equivalent to:

Enum.valueOf(z.asInstanceOf[Class[T] forSome {type T <: Enum[T]}], name)

But Scala says that it's in my mind only:

inferred type arguments [T] do not conform to method valueOf's type parameter bounds [T <: Enum[T]]
         Enum.valueOf(c.asInstanceOf[Class[T]], name)
              ^

Answer 1

Consider the difference between the following two expressions:

Enum.valueOf(x.asInstanceOf[Class[X] forSome { type X <: Enum[X] }], name)

And:

Enum.valueOf(x.asInstanceOf[Class[X forSome { type X <: Enum[X] }]], name)

Now think about how the type parameter T of valueOf will be inferred in each of these cases. In the first case, we've got an X that we know is a subtype of Enum[X], and we're all set. In the second case, on the other hand, T would have to be X forSome { type X <: Enum[X] }, and crucially this type is not a subtype of Enum[X forSome { type X <: Enum[X] }], so we haven't satisfied the constraint on T.

The problem is that your second example is equivalent to the latter.

---

As a footnote, this would work just fine if Enum were covariant in its type parameter. Take the following simplified example:

trait Foo[A]
trait Bar[A]

def foo[A <: Bar[A]](f: Foo[A]) = f

def x: Foo[X] forSome { type X <: Bar[X] } = ???
def y: Foo[Y forSome { type Y <: Bar[Y] }] = ???

Now foo(x) will compile, but foo(y) won't, just as in your code. But change Bar a bit:

trait Foo[A]
trait Bar[+A]

def foo[A <: Bar[A]](f: Foo[A]) = f

def x: Foo[X] forSome { type X <: Bar[X] } = ???
def y: Foo[Y forSome { type Y <: Bar[Y] }] = ???

Now they'll both compile. I'd guess that this has something to do with the reason that we have such strong intuitions about your two examples being equivalent.

---

As another footnote (in response to gzmo's comment below), consider the following:

scala> trait Foo[A <: Foo[A]]
defined trait Foo

scala> class MyFoo extends Foo[MyFoo]
defined class MyFoo

scala> val myFoo = new MyFoo
myFoo: MyFoo = MyFoo@3ee536d

scala> myFoo: (X forSome { type X <: Foo[X] })
res0: X forSome { type X <: Foo[X] } = MyFoo@3ee536d

scala> myFoo: Foo[MyFoo]
res1: Foo[MyFoo] = MyFoo@3ee536d

Let's suppose that X forSome { type X <: Foo[X] } were a subtype of Foo[X forSome { type X <: Foo[X] }] (ignoring for a moment the fact that the latter isn't even a valid type). Then we'd be able to write the following:

myFoo: Foo[X forSome { type X <: Foo[X] }]

But Foo is invariant, so if we have some thing that's an instance of both Foo[A] and Foo[B], then it must be the case that A =:= B. But it's definitely not the case that MyFoo =:= (X forSome { type X <: Foo[X] }). Not sure all of that is any less confusing, but it's how I convinced myself that the compiler knows what it's doing here.