std.boxer vs Any

[Daniel Keep]

Bill Baxter wrote:
> Daniel Keep wrote:
>>
>> Bill Baxter wrote:
>>> Daniel Keep wrote:
>>>> Bill Baxter wrote:
>>>>> Daniel Keep wrote:
>>>>>> And just to throw a little more wood on the fire:
>>>>>> http://www.prowiki.org/wiki4d/wiki.cgi?DanielKeep/Variant
>>>>>>
>>>>>> :P
>>>>>>
>>>>>>     -- Daniel
>>>>> Maybe you could comment on what makes yours different/better than the
>>>>> alternatives, then.  Do we really need three different versions of
>>>>> this?
>>>>>  Are they suited for different purposes?
>>>>>
>>>>> --bb
>>>> Well, the primary difference between mine and std.boxer is that mine
>>>> works in Tango out of the box.
>>>>
>>>> The other reason I wrote it was because of something Pragma posted
>>>> about
>>>> on the Tango forums about making a Box that was easier to use.  The
>>>> major improvement is the use of opAssign that allows you to directly
>>>> assign to a Variant.  I also like to think that Variant is easier to
>>>> use
>>>> than std.boxer.
>>>>
>>>> The whole "operator overloading" business was done mostly because I
>>>> wanted to see if it *could* be done.  :P
>>>>
>>>> Really, I think there's only so much you can do with a variant type in
>>>> D, and Box and Variant each do about 90% of it (not the *same* 90%,
>>>> mind
>>>> you; Variant doesn't do the "call a vararg function using a Box array"
>>>> trick.)  In the case of Box and Variant, the differences are mostly
>>>> just
>>>> implementation and interface issues.
>>> Looks like maybe yours avoids the opaque storage gc problem too?
>>>
>>> --bb
>>
>> I doubt it.  One of its members an array, which should mark Variants as
>> containing pointers.  The problem is that it's just as valid to store an
>> array or pointer in a Variant as it is to store a value like an int or
>> float.  Also, since a Variant is a struct, I don't think I can use
>> gc.hasNoPointers.
>>
>> Making Variant a class would allow one to properly inform the GC what
>> kind of data its storing, but then I wanted Variant to be a struct so I
>> could use the opAssign trick, not to mention value semantics.
>>
>>     -- Daniel
> 
> 
> Ok, so that's a good raison d'etre for Aarti's 'Any' class.  It uses the
> actual type as a template parameter to an inner class that contains
> nothing but the payload, so I think it should be ok.
> 
> Seems like you should be able to use the same trick for a struct if you
> don't mind the extra heap alloc on the inside.
> 
> --bb

Well, one of the reasons Variant is a struct is to *avoid* heap
allocations, so that would somewhat defeat the purpose.  What makes it
more annoying is that, AFAIK, the default is for the stack to "contain
pointers", so it would only need a separate heap allocation to correctly
store value types, which just sucks.

	-- Daniel