First of all sorry for not saying anything the whole day ... but I don't have internet connection at work. I hope I'll be excused though ;)

Now it will be a difficult for me to address all the issues raised ... however I will try

First of all, IMHO any FFI should facilitate the task of the person writing the library wrapper.

This comes the issue of (hopefully automatically) parsing C headers and generating the calls.�

Why?... well because if I take for instance the BLAS/LAPACK case (which was cited today) have over 1000 functions exported.�

Even if it is not 1000 lets say that you have a smaller library from which you want to use only one function X but in order to use that precise function you will maybe have to initialize some C context data, or a specific data structure.�

Of course you can do that by hand, but what if you really want to benefit from the library without diving into the details... well in that case you are a little bit stuck, either you write by hand a bunch of wrappers or you quit.�

The problem is that nice libraries that you want to use they usually have horrible data structures behind that you cannot set up easily. This happened to me several times in the future while trying to use different external library calls (graphviz, metis, abc, vpr to name a few) in Pharo (with FFI, with Alien, etc). In almost all cases I ended up generating text files and calling some bloody c program using these files as input.�The problem is that has a huge impact on performances...�

Now with nativeboost I found it pretty easy to automatically generate some usable binding... I parsed the C headers with srcml, the I have parsed the XML looking for function definitions... For these definitions I've generated the wrappers. Only that for over 15 argument functions that did not work.

Now, I do not like the idea of hacking the compiler, or even subclassing it like Nicolas did for Smallapack. And luckily we do not need to do that with the arguments in an array trick... Moreover, I completely agree with Igor on the fact that 15 arguments are to many, however Pharo really needs to have an automatable FFI generation.�

Igor, I have seen that you have mentioned some work-arounds using instance variables, and/or�NBExternalStructures. Both these ideas are great, especially the use of the NBExternalStructure, however in my opinion if you want to generate them automatically you've got yourself an even bigger problem.�

Will you generate a new class for each function call that you have?�

What about the FFI being a wormhole to an ugly and mean world?�

I think we should try to reduce the size of that hole ;)�

Though, you have a point with using named arguments, and I think that is a cool solution too.�That is why I was speaking about directly accessing the instance variables by name.

I simply didn't know enough about the hidden powers of�NBExternalStructure. �;)

As for extending the nativeboost signatures and accessing object fields by name, I completely agree with you that it is not a good idea.�

However, using the indices of arrays I think is only a small addition that can have a huge impact on the way people use NativeBoost FFI without adding to much extra-overhead.�

By the way, I think we definitely have to do the bound-checking trick by default.�

Moreover, having the�NBSTInObjectArgument as a wrapper over another loader is a great idea, giving us another degree of "controlled" freedom. ;)

Thanks Henrik for the good joke (I didn't see it coming :)), It makes me hate this idea now. :P

self callFn: {x . y}

Looks kinda familiar doesnt it? (hint: swap { for ( and . for , )

I for one welcome our new syntactic overlords!

Now let me come back to practical issues. My squeaksource id is�cip.t (you can find me searching "ciprian" in the members list).

On Mon, Oct 22, 2012 at 1:09 PM, Igor Stasenko <siguctua@gmail.com> wrote:

On 22 October 2012 12:00, Henrik Sperre Johansen

<henrik.s.johansen@veloxit.no> wrote:
> On 22.10.2012 02:37, Igor Stasenko wrote:
>>
>> On 22 October 2012 01:59, Nicolas Cellier
>> <nicolas.cellier.aka.nice@gmail.com> wrote:
>>>
>>> 2012/10/22 Igor Stasenko <siguctua@gmail.com>:
>>>>
>>>> On 22 October 2012 01:20, Nicolas Cellier
>>>> <nicolas.cellier.aka.nice@gmail.com> wrote:
>>>>>
>>>>> 5 parameters???? Igor, you're a dictator ;)
>>>>> Those theories are nice, but unhelpfull when applied to FFI
>>>>> Pragmatically there's not any chance I rewrite LAPACK+BLAS for the
>>>>> sake of purity.
>>>>>
>>>>> And your workaround (creating a class) is very poor, because maybe
>>>>> classes themselves should not have more than 5 instance variables ;)
>>>>>
>>>> Hehe.
>>>> This is same thing like increasing number of literals for methods (and
>>>> max distances between jumps).
>>>> It just makes sense where you deal with external chaotic world.
>>>> My ideology is simple: prevent that chaos from entering our little
>>>> peaceful bay.
>>>>
>>> That's not exactly the philosophy behind FFI.
>>> FFI is here to let the user manage the external chaotic world.
>>> OK, external peels of the onion should have 5 parameters or less.
>>> Near the sprout, you can't raise such barriers, or there is no onion at
>>> all.
>>>
>> Well, that's part of developer's responsibility, how to prevent chaos.
>> Needless to say, nobody wants to deal with so many arguments at once
>> (too much space for mistakes).
>> As for my workaround: this mainly, how you tame the complexity in case
>> it is inevitable?
>> Look how code to call that function will look like:
>>
>> 1. passing as array
>>
>> args := Array new: 100.
>> args at:1 put: x;
>> at:2 put: y;
>> � ...
>> at: 100 put: zork
>>
>> self callFn: args.
>
> self callFn: {x . y}
> Looks kinda familiar doesnt it? (hint: swap { for ( and . for , )
> I for one welcome our new syntactic overlords!
>

Yes, it looks familiar, but cannot tell where i seen it.
Gah.. how i could forget about it?

>>
>> 2. passing as instance of class, or external structure:
>>
>> args := MyFunctionArgs new.
>> args
>> � firstArgumentName: x;
>> � secondArgumentName: y;
>> � ...
>> � hundrethArgumentName: zork.
>> self callFn: args.
>>
>> admit that dealing with names instead of numbers leaves much less
>> space for mistakes
>> and serves for better clarity at same time.
>> So, even if it is more cumbersome because requires defining extra
>> class, at the end you win much more.
>>
>> Anyways, if people think it is worth adding indirect argument loader (
>> in form of param@<index>, but not param@ivar), we can introduce that.
>>
> While I often find this a good idea for maintainability, it sorta flies in
> the face of another of ST's strengths, iterative/exporatory programming.
> If you are forced to think up front about which parameter classes you need
> due to a small limit, rather than introduce them ad-hoc when the code really
> needs the refactoring to remain legible, it slows you down.
>
> Not thinking of FFI specifically, but I have seen lots of evolved
> mathematical models where a 5-parameter limit upfront would probably lead to
> either:
> a) *Really* bad code, ie. making the calculation object stateful by storing
> in instvars instead. (and in the process, make it really hard to know which
> instvars are actually part of object state and not temp vars of some
> calculation)
> b) Switching to another programming language out of frustration.
>

keyword message syntax is bad for many arguments..
for such cases i find a positional argument notation more appeal
because it is more compact.
In any case, a complex math formulas smell equally bad in any
programming language
(sometime even if written by hand on paper using math notation(s) ;)

> Cheers,
> Henry

>

--
Best regards,
Igor Stasenko.

--
Dr. Ciprian TEODOROV
Ing�nieur�D�veloppement�CAO

t�l : 06 08 54 73 48
mail :�ciprian.teodorov@gmail.com
www.teodorov.ro