The future of UDAs.

[Walter Bright]

On 11/27/2012 6:42 PM, Gor Gyolchanyan wrote:
> AFAIK, currently UDAs are set at declaration site, are immutable and the
> declaration cannot get additional UDAs externally. Moreover, only global
> declarations can have UDAs, which removes some very useful possible uses
> of UDAs.

What very useful thing could be done with UDAs on local variables?


> Lack of mutable compile-time variables sometimes cripples the
> metaprogramming in D. For instance all classes, derived from a certain
> type must be dealt with in a special way, which requires a tuple of
> those types. Gathering a tuple of unrelated types is currently
> impossible, because that would require mutable compile-time variables.

I don't understand this. It is perfectly possible to create compile time 
tuple of (int, long, float, int*, S).


> Mutable compile-time variables would also be extremely useful for
> implementing a very powerful compile-time reflection library without the
> need for compiler magic. All you'd have to do is mix in a template in
> your class or your module and voila. The mixin would then add the
> introspected declarations to the central compile-time declaration
> repository.

I don't understand the difficulty here. UDAs are additive, so you could 
mixin your template in one space at the top, and it will apply to all 
the declarations nested under it, exactly like other attributes do now.


> There are also many cases when some actions need to happen in case a
> declaration gets a UDA of a specific type. For instance, a powerful RTTI
> library, which (when a type gets a dedicated RTTI-typed UDA) adds the
> run-time information about the class to the central registry at
> load-time.

That can be done now. Loop over the declarations, checking to see if 
each has a particular attribute.

> This kind of stuff could be easily achieved using the
> constructors and destructors of the structures, being added to the UDAs
> of a declaration (in this case a class). The only missing thing for this
> to work is the ability for the constructor to see the declaration it's
> being put on. I'd personally expect there to be some sort of a
> __traits(getAttributeDeclaration), which would evaluate to the symbol on
> which the enclosing type (a structure, a union or whatever) is being
> placed on as an attribute.

That could cause ugly forward reference issues, and also recall that a 
UDA can be applied to many declarations at once.