Re: proposal: fix corner use case of variadic fuctions usage

2013/1/19 Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>:
> Pavel Stehule <pavel(dot)stehule(at)gmail(dot)com> writes:
>> 2013/1/18 Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>:
>>> The approach is also inherently seriously inefficient. ...
>
>> What is important - for this use case - there is simple and perfect
>> possible optimization - in this case "non variadic manner call of
>> variadic "any" function" all variadic parameters will share same type.
>> Inside variadic function I have not information so this situation is
>> in this moment, but just I can remember last used type - and I can
>> reuse it, when parameter type is same like previous parameter.
>
>> So there no performance problem.
>
> Well, if we have to hack each variadic function to make it work well in
> this scenario, that greatly weakens the major argument for the proposed
> patch, namely that it provides a single-point fix for VARIADIC behavior.
>
> BTW, I experimented with lobotomizing array_in's caching of I/O function
> lookup behavior, by deleting the if-test at arrayfuncs.c line 184. That
> seemed to make it about 30% slower for a simple test involving
> converting two-element float8 arrays. So while failing to cache this
> stuff isn't the end of the world, arguing that it's not worth worrying
> about is simply wrong.
>
>>> But large arrays have a worse problem: the approach flat out does
>>> not work for arrays of more than FUNC_MAX_ARGS elements, because
>>> there's no place to put their values in the FunctionCallInfo struct.
>>> This last problem is, so far as I can see, unfixable within this
>>> approach; surely we are not going to accept a feature that only works
>>> for small arrays. So I am going to mark the CF item rejected not just
>>> RWF.
>
>> disagree - non variadic manner call should not be used for walk around
>> FUNC_MAX_ARGS limit. So there should not be passed big array.
>
> That's utter nonsense. Why wouldn't people expect concat(), for
> example, to work for large (or even just moderate-sized) arrays?
>
> This problem *is* a show stopper for this patch. I suggested a way you
> can fix it without having such a limitation. If you don't want to go
> that way, well, it's not going to happen.

>
> I agree the prospect that each variadic-ANY function would have to deal
> with this case for itself is a tad annoying. But there are only two of
> them in the existing system, and it's not like a variadic-ANY function
> isn't a pretty complicated beast anyway.
>
>> You propose now something, what you rejected four months ago.
>
> Well, at the time it wasn't apparent that this approach wouldn't work.
> It is now, though.

I have no problem rewrite patch, I'll send new version early.

Regards

Pavel

>
> regards, tom lane