impl trait in type aliases does not trigger "private type in public interface" errors

[Nemo157]

The current implementation of existential_type does not trigger any "private type in public interface" errors, for the following code (playground):

#![feature(existential_type)]

existential type Foo: std::fmt::Debug;

pub fn foo() -> Foo {
    5usize
}

I would expect it to give an error like

error[E0446]: private type alias `Foo` in public interface
  --> src/lib.rs:8:1
   |
5  |   existential type Foo: std::fmt::Debug
   |   - `Foo` declared as private
...
8  | / pub fn foo() -> Foo {
9  | |     5usize
10 | | }
   | |_^ can't leak private type alias

(although, because normal type aliases are transparent to the visibility checking, I'm not certain exactly what the error message should say; it just seems bad to introduce another way to accidentally forget to make part of your API publicly nameable)

cc #63063

[Nemo157]

@rustbot modify labels to A-impl-trait, C-bug, F-impl_trait_type_aliases, T-lang

[Centril]

cc @petrochenkov

[petrochenkov]

This is intentional, impl Trait is opaque and equivalent to dyn Trait from the privacy checking point of view.
For other crates the return type of foo looks like public std::fmt::Debug and doesn't leak any info specific to the private Bar.
(Except for Send/Sync which were explicitly decided to be leaked by impl Trait, unfortunately or not).

A motivating example:

pub fn f() -> impl Fn() { // OK
    || {} // It's hard to come up with something more private than a nameless fn-local closure.
}

RFC: https://github.com/rust-lang/rfcs/blob/master/text/2145-type-privacy.md

[Nemo157]

The problem is that with RFC 2515 the impl Trait alias essentially becomes a new nominal type, by having a public API mentioning this unreachable alias you create APIs that cannot have their values stored and retrieved from non-inferred locations.

As an extended example this playground takes the above API and puts it into a local module (foo). A user of that module then attempts to use their own local impl Trait alias (MyFoo) to be able to store an instance of foo::Foo in a struct (since foo::Foo is private and cannot be named), that creates a new opaque type that doesn't unify with foo::Foo and so cannot be used with foo::bar.

[Nemo157]

(btw Bar wasn't really anything to do with the bug, just leftover from me trying stuff out in that playground, I've removed it from the example).

[Centril]

The problem is that with RFC 2515 the impl Trait alias essentially becomes a new nominal type, by having a public API mentioning this unreachable alias you create APIs that cannot have their values stored and retrieved from non-inferred locations.

I'm not sure that's so much of a problem; notably you can still use foo::bar(foo::foo()); and inferred locals are still available if you need to store in a local.

What's nice about the current semantics is that type Alias = TypeExpr; will not need a special rule for Visibility X WasThereImplTrait which should simplify both the mental model and the compiler.