Re: RFC: MIPS paired single vector support

[Nigel Stephens <nigel at mips dot com>]

James E Wilson wrote:

> > I guess that we need to agree which to go forward with, or whether we 
> > can merge our efforts in any way.
> >    
>
> I am certainly willing to cooperate, but I am on a deadline.  I am
> hoping to have the initial support done within a few weeks.  I am
> already starting to post preliminary patches.
>
> Are you planning to post patches anytime soon?
>
>  

Jim

Sorry for the delay, I've been offline since Thursday, so haven't kept 
up my end of the conversaion.

We have now got a complete set of patches for mainline. Their author is 
my colleague Chao-ying, and he should be able to post them on Monday. 
He'll also be better able to comment on the details and why things were 
done the way they were.

> It looks like you have put a lot more effort into builtin functions than
> I have so far.  And you have certainly put more effort into the docs.  I
> haven't written any as yet.
>  

Well, they still need translating into texinfo format, but at least 
they're a start... I think that I volunteered to do the translation.

> If I drop the compare and builtin function support from my initial
> patch, then that probably eliminates most of the conflict.  My compare
> and builtin function support is incomplete and broken.
>
> This leaves the arithmetic and move instructions which are pretty
> straight forward and unlikely to be much different.  I also have
> patterns for the named vector operations which are pretty straight
> forward.
>
>  
Good, let's hope that they're easy to merge.

> > We won't have the SB-1 extensions, of course, but we have got MIPS-3D.
> > Have you done anything about MDMX support in the compiler?
> >    
>
> I don't have any MIPS-3D or MDMX support, and wasn't planning on
> implementing it at this time.
>  

OK, we've not done MDMX either, I just wondered if you had. We did do 
MIPS-3D though, which is important for us because we have one CPU which 
implements it.

> The SB-1 support I could always split out as a separate patch.
>
>  
>
> > 1.4.5 Floating Point Convert Pair to Paired Single (CVT.PS.S)
> > v2sf __builtin_mips_cvt_ps_s (float, float);
> > c = __builtin_mips_cvt_ps_s (a, b);
> >    
>
> This isn't necessary.  You can just do
>  (v2sf) {a, b}
> This requires that you have the vec_initv2sf pattern defined though.
>
>  

Good point. I think we do support that construct, so the builtin can 
just be deleted.

> > 1.4.9 Floating Point Align Variable (ALNV.PS)
> > v2sf __builtin_mips_alnv_ps (v2sf, v2sf, int);
> >    
>
> This isn't very useful without the lux/sux instructions.  They are
> currently not supported by mips.md, and you didn't add a builtin for
> them either.  It doesn't hurt to have this builtin though.  We just need
> to remember to add lux/sux support later.
>  

Good point.

Nigel