Explain why Self is invalid in generic parameters

[hkmatsumoto]

Close #89985.

r? @estebank

[estebank]

This should likely be something like:

    Ok(GenericParam {
            ident: Ident::new(kw::SelfUpper, this.prev_token.span),
            id: ast::DUMMY_NODE_ID,
            attrs: vec![],
            bounds: vec![], //maybe even parse/skip bounds?
            kind: GenericParamKind::Type { default: None },
            is_placeholder: false,
        })

so that we still take Self as a type param (although this might actually bring worse diagnostics if valid uses of Self are seen as the param 🤔).

[hkmatsumoto]

(although this might actually bring worse diagnostics if valid uses of Self are seen as the param thinking).

I wonder if there could be a valid use of Self as a type parameter. Yes, one could write let a: Option<Self> = ... in assoc items, but that is a generic argument and will be parsed by a function other than parse_generic_params.

... so that we still take Self as a type param

Recovering by treating as if Self is a valid ident emits these diagnostics below, which I'm not very satisfied with; given Self is unusable here, parameter `Self` is never used sounds like we should use parameter Self, which isn't right.

error: unexpected keyword `Self` in generic parameters
  --> $DIR/keyword-self-as-type-param.rs:3:12
   |
LL | struct Foo<Self>(Self);
   |            ^^^^
   |
   = note: you cannot use `Self` as a generic parameter because it is reserved for associated items

error[E0392]: parameter `Self` is never used
  --> $DIR/keyword-self-as-type-param.rs:3:12
   |
LL | struct Foo<Self>(Self);
   |            ^^^^ unused parameter

The best-ish solution I come up with is that, when we see Self, emit the diagnostic and skip Self and the following comma as if the generic param never existed.

[jackh726]

@bors r+

This seems reasonable.

[bors]

📌 Commit a72dd4a has been approved by jackh726