Re: [patch x86] use always_inline on all vector intrinsics

[Stuart Hastings <stuart at apple dot com>]

On Jun 28, 2005, at 10:51 PM, Uros Bizjak wrote:

> Hello Stuart!
>
>
> > A mechanical patch to add "__attribute__((__always_inline__))" to  
> > every vector
> > intrinsic in i386/*mmintrin.h.
>
>
> > Bootstrapped and DejaGnu on Darwin/x86, no regressions.
>
>
> > 2005-06-28 Stuart Hastings <stuart@apple.com>
>
>
> > * gcc/config/i386/mmintrin.h: Mark vector intrinsics always_inline.
> > * gcc/config/i386/emmintrin.h: Likewise.
> > * gcc/config/i386/pmmintrin.h: Likewise.
> > * gcc/config/i386/xmmintrin.h: Likewise.
>
> There is a PR regarding inlining of SSE intrinsics.
> (http://gcc.gnu.org/bugzilla/show_bug.cgi?id=21195). Is your patch  
> enough to fix
> a testcase in  the bugreport?

Apparently yes; my patched compiler inlined every vector intrinsic in  
both testcases.  I'll add a comment to Bugzilla.

> BTW: A related problem in area of *mmintrin.h headers is PR21408:  
> Some #defines
> are improperly guarded with #ifdef __SSE3__. Could someone with an  
> access to
> Intel's pmmintrin.h kindly check if there is an inconsistency in  
> gcc's headers
> compared to Intel one?

For ICC 8.0 on Linux/x86, the answer seems to be 'yes':

    stuart@citroen:~$ egrep 'cvs_id|SSE3' /opt/intel_cc_80/include/ 
pmmintrin.h
    *** cvs_id[] = "$Id: pmmintrin.h,v 1.2 2003/07/24 17:47:39  
kbsmith1 Exp $";
    stuart@citroen:~$

I've yet to install the ICC 8.1 that Intel gave to everyone at the  
recent GCC Summit.

> Thanks,

You're very welcome.

stuart