Re: arch-specific template code

[Marc Glisse <marc dot glisse at inria dot fr>]

On Sat, 8 Sep 2012, Ulrich Drepper wrote:

> On Sat, Sep 8, 2012 at 6:12 AM, Marc Glisse <marc.glisse@inria.fr> wrote:
> > PR 54400: horizontal add
> > PR 54408: sqrt
> > PR 54422: unaligned load/store
> > PR 54427: operator< and others in C++
>
> You'll also need various shifting:operations

Thanks for the precision.

>  - based on sub-words, measured in bits

So this one is equivalent to doing >> or << on each element independently? 
Should already work in C, and 54427 should add it to C++ then :-)

>  - a whole vector, byte-wise.
>
> Those are essential for the code like the SFMT implementation.  Maybe
> the former you have in PR 54427 but the second is something that
> doesn't exist in C or C++.  You can do that via a byte-wise shuffle.
> But in that case the compiler has to recognize the shuffle patterns
> like { 6,5,4,3,2,1,0,15,14,13,12,11,10,9,8,7} and then don't use
> shuffle but instead use the rotate instruction.

I believe every __builtin_shuffle that can be done in a single instruction 
is already properly expanded on x86. For this 16 byte vector shuffle, it 
uses pshufb. Is there a better instruction? (one that doesn't use a mask 
from memory?) Or do you have a clever way to implement those 
reverse-cyclic permutations on platforms that didn't have pshufb? (it is 
probably not hard to do better than the sse3 code gcc generates for it, 
but pre-sse4 PCs are getting less common, so I am not sure it is worth 
spending too much effort there)


A harder one I saw you use is some kind of clz/ctz that acts on the whole 
vector. I don't think ARM has a similar instruction, and it is not obvious 
to me what the best interface would be. It makes sense that not everything 
will fit and some target-specific code will have to remain, I just hope 
that's not the general case.

--
Marc Glisse