DIP6: Attributes

[Denis Koroskin]

On Mon, 03 Aug 2009 19:53:21 +0400, Steven Schveighoffer  
<schveiguy at yahoo.com> wrote:

> On Mon, 03 Aug 2009 11:42:33 -0400, Ary Borenszweig  
> <ary at esperanto.org.ar> wrote:
>
>> Don wrote:
>>> Ary Borenszweig wrote:
>>>> Don escribió:
>>>>> grauzone wrote:
>>>>>> Don wrote:
>>>>>>> Ary Borenszweig wrote:
>>>>>>>> http://www.prowiki.org/wiki4d/wiki.cgi?LanguageDevel/DIPs/DIP6
>>>>>>>
>>>>>>> This looks like a solution in search of a problem. What's the  
>>>>>>> problem being solved?
>>>>>>
>>>>>> Attaching additional data to types, that can't be specified  
>>>>>> otherwhere. This should help with metaprogramming-like stuff.
>>>>>>
>>>>>> For example serialization. How do you specify that a field  
>>>>>> shouldn't be part of the serialized data? Java has an extra keyword  
>>>>>> attribute like
>>>>>> "transient" (comes from before attributes were introduced). C# uses  
>>>>>> what we call annotation in this thread. How would you do this in D?
>>>>>
>>>>> I agree that there doesn't seem to be a nice way at present. One  
>>>>> possibility would be to establish a naming convention for transient  
>>>>> fields -- a Ruby-style solution, I guess.
>>>>>
>>>>> But are annotations actually an ideal solution for this problem?
>>>>> Presumably, you'll have to specify that somewhere else anyway. I  
>>>>> mean, during reading it will need to either be initialized  
>>>>> separately after serialisation (like opPostBlit, perhaps?), or else  
>>>>> remain uninitialized. Serialisation seems to be _extremely_ similar  
>>>>> to construction. I'm not sure that annotations capture that.
>>>>>
>>>>> D has much more powerful metaprogramming than C# or Java, so my  
>>>>> intuition and hope is that we shouldn't need to adopt hacks from  
>>>>> those weaker languages. The annotation syntax in C# and Java looks  
>>>>> like an ugly hack to me. Purely a subjective opinion, of course, but  
>>>>> it seems really out of place in a C-family language.
>>>>
>>>> Attributes has many, many other uses. Appart from serialization, you  
>>>> could specify how a field is stored in a database. How a method maps  
>>>> to an http request (post, get, which parameters to bind to the  
>>>> request, etc.). Whether a method should do security checks before  
>>>> executing. Whether a method should be run as a test, and what's the  
>>>> expected exception to be thrown. [insert your usage here]
>>>  Great, you've answered my question. That should be in the DIP,  
>>> instead of the vague stuff that's in there now -- the existing DIP is  
>>> about replacing keywords, which is very unconvincing. (It doesn't  
>>> work, actually -- the name mangling is important for most of the  
>>> keywords mentioned).
>>
>> But the DIP I wrote isn't about general-purpose annotations. It's just  
>> the first step. Are "pure" and "nothrow" part of the mangling? Or which  
>> are? I thought not. Can you overload a pure and a not-pure function  
>> with the same parameter count and types?
>
> Yes, they have to be.  There are reasons besides overloading for  
> including other attributes in the naming.
>
> For example, if a function is pure, then becomes unpure, you don't  
> existing code that is expecting a pure function to link against it.
>
> In other words, the linker is dumb.  It only knows how to match symbols,  
> so you have to embed into the symbols the important pieces of the  
> interface that you want the linker to consider important.
>
> To answer Don's point, there is nothing saying that the compiler can't  
> read attributes and change its behavior.  Of course, those would have to  
> be builtin attributes.
>
> My opinion on removing existing keywords is -- don't.  There's little to  
> no gain.  Let that ship sail, and concentrate on future keyword  
> proposals.
>
> -Steve

Compiler could embed some kind of a hash from attributes in a mangled name.