From: Tobias Schwinger (tschwinger_at_[hidden])
Date: 2007-12-09 03:38:48
Joel de Guzman wrote:
> Joel wrote:
>
>>> I'd like to see more tweakable
>>> parameters for specific uses. I frequently deal with
>>> different number expansions in a case to case basis.
>>> For example, is it possible to have arity 3 for function
>>> X and arity-5 for function Y? A single number seems
>>> wasteful if I'm sure that a certain use needs only a few.
>>> Is that possible?
>> Yes, I think so. And it turns out to be a very good idea, as it
>> potentially makes overload resolution *a lot* faster. Here comes the
>> benchmark compiling do_the_bind.cpp, ported to use the forward adapter
>> with gcc:
>>
>> Arity=8
>> preprocess
>>
>> real 0m1.384s
>> user 0m1.270s
>> sys 0m0.109s
>>
>> preprocess, compile, link
>>
>> real 0m12.405s
>> user 0m11.769s
>> sys 0m0.540s
>>
>>
>> Arity=4
>> preprocess
>>
>> real 0m0.918s
>> user 0m0.454s
>> sys 0m0.081s
>>
>> preprocess, compile, link
>>
>> real 0m1.949s
>> user 0m1.653s
>> sys 0m0.268s
>
> Wow that's interesting. Is the trend exponential?
> Can't really tell with just 2 sample points.
Yeah, seems so:
4 -> 1.4
5 -> 1.5
6 -> 2
7 -> 3
8 -> 11.2
(compile & link time only)
And these are just a few calls - we can easily have a higher constant in
practice...
>> Interestingly preprocessing is a minor concern and so we could even
>> provide means to take our pick at compile time (as opposed to
>> preprocessing time).
>
> Not so fast. Don't forget EDG.
No panic, I will also allow fractional preprocessing :-). The idea to
(almost) mostly pay for what one knows to use at compile time (rather
than what one assumes one might need at preprocessing time) makes
compile time control a tempting feature to have, however.
Regards,
Tobias