memoize (is this a fix for overloading?)

[Max Samukha]

On 01/05/2011 02:12 PM, Lars T. Kyllingstad wrote:
> On Wed, 05 Jan 2011 13:40:05 +0200, Max Samukha wrote:
>
>> On 01/05/2011 12:24 PM, Lars T. Kyllingstad wrote:
>>> On Wed, 05 Jan 2011 12:07:50 +0200, Max Samukha wrote:
>>>
>>>> On 01/05/2011 11:21 AM, Lars T. Kyllingstad wrote:
>>>>> On Tue, 04 Jan 2011 17:06:45 -0600, Andrei Alexandrescu wrote:
>>>>>> [...]
>>>>>> Overloads are compile-time entities so they should be easily
>>>>>> inspectable.
>>>>>
>>>>> Since this is likely to be useful in other cases as well, maybe it
>>>>> would be better to make a general template that selects an overload?
>>>>> So you'd use it like this:
>>>>>
>>>>>      alias memoize!(selectOverload!(sqrt, double)) msqrt;
>>>>>
>>>>> A first step towards this would be to make __traits(getOverloads)
>>>>> work with module-level functions.  Currently it only works for member
>>>>> functions.  Also it would probably be better if it accepted a
>>>>> function alias instead of a string containing the function name.
>>>>
>>>> __traits(getOverloads) does work with module level functions if you
>>>> pass a module alias to it.
>>>
>>> Cool!  I didn't know that.  Here's an updated version, where
>>> selectOverload works with both module-level and member functions:
>>>
>>> [...]
>>
>> Nice! Though I think that overloads should be selected not by type
>> equality but by regular overload resolution rules. That is the template
>> should take not parameter but argument types.
>
> I agree.  It is actually trivial to modify my example by replacing
> equalTuple!() with implicitlyConvertibleTuple!() which uses an is(T1:T2)
> test instead of is(T1==T2).  But that would make the template select the
> first possible match, instead of the "best match" resolution used by the
> compiler, which is often not what you want.  It would fail miserably for
> the sqrt() example, for instance, by always picking out sqrt(float).
>
> In the end, I think implementing full overload resolution in the library
> will be so hairy it will be better to just add __traits(selectOverload).
> This doesn't look too bad:
>
>    alias memoize!(__traits(selectOverload, sqrt, double)) msqrt;
>
> And if, one day, __traits gets replaced by the meta namespace (bug 3702,
> one of the highest-voted in Bugzilla), it actually looks pretty good:
>
>    alias memoize!(meta.selectOverload(sqrt, double)) msqrt;
>
> ;)
> -Lars

I agree with every point. Now time for a compiler patch ;)