language: disallow type literals in annotations

[kasperl]

Right now, the spec explicitly allows:

   class Foo { ... }
   
   @­Foo
   bar() { ... }

We should disallow that so we have time to figure out what it is supposed to mean. We should change the wording in section 11 to:

---

Metadata consists of a series of annotations, each of which begin with the
character @­, followed by a constant expression that starts with an identifier. It is a compile time error if the expression is not one of the following:

� - A reference to a compile-time constant variable.
�� - A call to a constant constructor (without the leading 'const').

[lrhn]

Agree completely.

The current way it is working is error prone.

Users are expecting @­Deprecated to mean something that it doesn't, but it doesn't trigger any warnings or errors.

[gbracha]

I disagree of course. This all feels rather Javanese. But I'm happy to accommodate.

---

Added Accepted label.

[gbracha]

Would it be acceptable to make this a warning?

[kasperl]

I'd say that this has to be a compile-time error (not a static warning) for two reasons:

First (and most importantly), only an error allows us to decide on the semantics later. If we make it a warning, we cannot really change what it means at runtime later without breaking stuff.

Second, it is much more consistent if everything that isn't explicitly allowed in an annotation leads to a compile-time error.

[gbracha]

Ok. However, see my comments on 14652. These things won't be detected by the VM unless someone accesses the metadata.

[lrhn]

Many errors won't be detected by the VM unless the code is run, so that should be fine.

The only potentially surprising part is that the "compile time" of an annotation isn't until it's accessed through mirrors, but as long as that's stated somewhere in the spec, I'm fine with it. The analyzer will report the error, which is what developers running on the VM should be using anyway.

[gbracha]

Fixed in 0.8 spec (revision 29775).

---

Added Done label.