objects as AA keys

[captaindet]

On 2013-10-15 08:32, Daniel Davidson wrote:
> On Tuesday, 15 October 2013 at 05:44:25 UTC, captaindet wrote:
>> hi,
>>
>> i am a bit confused.
>>
>> the official language ref ( http://dlang.org/hash-map.html ) states:
>> "
>> Classes can be used as the KeyType. For this to work, the class definition must override the following member functions of class Object:
>> hash_t toHash()
>> bool opEquals(Object)
>> int opCmp(Object)
>> ...
>> "
>>
>> but now i stumbled on http://forum.dlang.org/post/mailman.2445.1354457588.5162.digitalmars-d-learn@puremagic.com
>> "
>> int[typeof(O)] rc;
>> rc[O] = 42;
>> auto O2 = O;
>> // [...]
>> if (auto r = O2 in rc)
>> return *r;
>> else
>> return rc[O2] = compute(O2);
>>
>> IOW explicitly taking the address may not be necessary when doing that kind of things.
>> "
>>
>> and i did a quick test and indeed, it seems to work out of the box - without overriding any member functions. in my use case, i wouldn't be able to modify the class anyway.
>>
>> so my questions:
>>
>> why is it working, is it just syntactic sugar for using cast(void*)Obj as key?
>>
>> what is the danger of using objects as keys? when would it fail?
>>
>> as it seems to be working against language specs, will this 'feature' eventually be removed?
>> (then maybe i should use cast(void*)Obj right away...)
>>
>>
>> thanks, det
>
> Do you have an example where it is really working? The problem of not
> overriding those functions is that maybe you are not getting what you
> think. For example, the code below will print:
>
>
> [aaa.S:42]
> [aaa.S:42, aaa.S:43]
>
> This is probably not what you want. If you don't override the
> functions how is the implementation to know what are equivalent keys?
> Below you would probably expect two S's that are default constructed
> to hit the same key spot in the AA. But this does not happen since
> the implementation does not know you want new S to equal new S
> because there is no opEquals.
>
>
> Thanks,
> Dan

in my use case i don't have multiple objects with the same state, so keying them per obj_ptr (if this is what is happening) seems to be sufficient for me. (the object classes are 3rd party so i cannot overwrite member functions anyway.)

/det