[PATCH][libgcc-math] Vectorized intrinsics for x86_64

[Jan Hubicka]

> H. J. Lu wrote:
> 
> > Our initial analysis indicates that the current patch won't perform
> > well on Intel chips and any tuning for Intel chips will very likely
> > have negative impatch on Opteron.
> 
> That didn't answer the question I raised: will the current patch hurt
> Intel chips relative to the current compiler sources?
> 
> > So the question becomes, does the
> > current scheme support the future changes for -mtune=generic? Are
> > we going to provide multiple libraries, which will be selected by
> > -mtune=xxx?
> 
> We can provide multilibs, just like for every other library.

I would preffer just different entry points for differently tuned
libraries, so we can link together code optimized for different CPUs
into single binary.  

I am currently traveling so I won't be able to look into this in detail
before next monday, but it seems to me that the tweaks needed for Nocona
won't hurt K8 too much in this case (epsecially the movsd->movapd
changes are pretty cheap in code not having large # of instructions per
cycle), so I wouild just go for shared implementation unless we see some
important place where we want to fork.  We probably will want "generic"
implementation anyway.

Honza
> 
> -- 
> Mark Mitchell
> CodeSourcery
> mark@codesourcery.com
> (650) 331-3385 x713