Re: [Tinycc-devel] Request: __attribute__((vector_size))

[uso ewin]

On Thu, Jan 9, 2020 at 3:55 PM ian <address@hidden> wrote:
>
> Hi.
>
> I stay at my position, and keep saying it has nothing to deal with a compiler 
> !
> It deals with assumptions concerning the main aim of data structures.
> It seems that you know how to code it, so if tcc can compile this it's fine 
> and enough.
> But, FOR SURE, it does not have to translate an index-by-index array elements 
> sum !!!
>
>
> Regards, ian.
>

I have a different point of view:

I saw that kind of extension as somewhat similar to what float/double are in C:
Machine don't have a float data type, they have registry, memory and
instructions,
that allow to do floating point arithmetic on registry.

Float support are more or less compiler that does assumption
concerning the aim of some data.
That's why they add a type in C I guess.

And as vector extension, you could code float operation in C, but it
would be hard to tell the
compiler to use the float specific instructions.

I see this extension as similar: it add a type that do some assumption
of some data and
allow to use some instruction on them.

The main difference is that only a few processor support instruction on matrix.
(But I guess that if intel had some kind of vector operation when C
was created this would be part of the standard)

I don't think it's a must have (far from it), nor I think tcc should
support vector extension,
but if the support is done in a non intrusive way, I don't see why not.

Matthias,

>
> Le 09/01/2020 à 12:34, Rasmus Riiner via Tinycc-devel a écrit :
>
> I would say that it is okay to generate low quality code, that it's better 
> than nothing, as long the implementation doesn't impact the rest of the 
> compiler too much... but of course that is from my biased point of view.
>
> Which is that the only thing presently keeping TCC compile times out of my 
> reach is that a bunch of code akin to this in C++:
>
>     vec2 v4 = vel + a3 * dt;
>     vec2 p4 = pos + v4 * dt;
>     vec2 a4 = entity_acceleration(p4, v4, entity);
>
>     *dpos = (v1 + (v2 + v3) * 2.0 + v4) * (dt * (1.0 / 6.0));
>     *dvel = (a1 + (a2 + a3) * 2.0 + a4) * (dt * (1.0 / 6.0));
>
> Would have to be rewritten to something like this for plain C:
>
>     vec2 v4 = v2Add(vel + v2Scale(a3 * dt));
>     vec2 p4 = v2Add(pos + v2Scale(v4 * dt));
>     vec2 a4 = entity_acceleration(p4, v4, entity);
>
>     *dpos = v2Scale(v2Add(v2Add(v1, v2Scale(v2Add(v2, v3), 2.0)), v4), dt * 
> (1.0 / 6.0));
>     *dvel = v2Scale(v2Add(v2Add(a1, v2Scale(v2Add(a2, a3), 2.0)), a4), dt * 
> (1.0 / 6.0));
>
> Operator overloading doesn't fit C, and I'm glad it's not in there, but then 
> the only way to fix this particular inefficiency is with some kind of 
> built-in array math, perhaps as an extension... which is exactly what the GCC 
> extension would provide. The extension goes a bit too far perhaps, what with 
> supporting every basic type as a vector element, I can see why they did that 
> though. For TCC, it might be sensible to limit the possible element types to 
> 32bit floats or integers, and cap the possible vector length to 4...
>
> I'm not sure what I would do in your position, implementing the extension (or 
> a subset of it) might not be what's ultimately best for the project, but 
> hopefully you can see where I'm coming from. Compile times for basic C++ 
> programs take entire seconds with the available compilers, whereas I could 
> literally tcc -run and be in-game, with immediate feedback, nearly instantly.
> I could bite the bullet and start using function syntax for basic vector 
> operations, I just wanted to confirm whether the GCC extension is outside the 
> scope of TCC, or not, first. I don't actually use any C++ features over C, 
> other than operator overloading and a little bit of function overloading 
> (which I could do just fine without). It feels like so close, yet so far, you 
> know?
>
> Rasmus.
>
> ----- Reply to message -----
>
> Sure, Rasmus asked to have this extension in TCC to be able to use it for
> writing libraries or apps; i.e. he asked if TCC could be extended to
> compile code like this:
>
> ---------------------
> typedef int v4si __attribute__ ((vector_size (16)));
> void foo (void) {
> v4si a = {1,2,3,4}, b = {5,6,7,8}, c;
> c = a + b;
> bar(c[0], c[1], c[2], c[3]);
> }
> ---------------------
>
> For that TCC would need to be extended somewhat, and I was alluding to the
> fact that this extension isn't totally trivial if it shouldn't generate
> very low quality code. If it's okay to generate low quality code and
> not adhere to the psABI for parameter passing of these types then it's not
> too much work.
>
> If you were asking if such extension is really a must have in a
> compiler: no, otherwise it wouldn't be an extension. It's a nice to have,
> and I can see why Rasmus wants it, but it comes at a non-trivial cost to
> support it in a small compiler.
>
>
> Ciao,
> Michael.
>
>
>
>
> _______________________________________________
> Tinycc-devel mailing list
> address@hidden
> https://lists.nongnu.org/mailman/listinfo/tinycc-devel
>
> --
> -- address@hidden
> -- Développeur compulsif
> _______________________________________________
> Tinycc-devel mailing list
> address@hidden
> https://lists.nongnu.org/mailman/listinfo/tinycc-devel