Re: arch-specific template code

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

On Sat, 8 Sep 2012, Ulrich Drepper wrote:

> > 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?
>
> The shuffle operations need a memory or [xy]mm register parameter.
> That's expensive to set up.  The shuffling which doesn't rearrange
> bytes but just rotates them should use the equivalent of the
>
>  _mm_srli_si128
>
> and
>
>  _mm_slli_si128
>
> intrinsics.

I did a quick test, and psrldq+pslldq+por was slower than pshufb with a 
memory operand. True, the test isn't representative, in a program that 
uses more memory the value won't be in cache and the load can be much 
slower. I don't know how people usually determine what the best code is. 
> From your post it seems that psrldq+pslldq+por should always be preferred, 
even if in some cases it can be a loss. That makes sense, I am just more 
used to "obvious" optimizations. Is there a way to decide when a load 
becomes worth it, compared to a large number of pure logical instructions?

With -Os, is the choice the same?

(libstdc++ is not really the right list anymore...)

--
Marc Glisse