SSE fix 3
Jan Hubicka
jh@suse.cz
Mon Oct 14 09:46:00 GMT 2002
> On Mon, Oct 14, 2002 at 05:22:35PM +0200, Jan Hubicka wrote:
> > Also there seems to be ICC compatibility problem - ICC requires emmintrin.h
> > to be included for SSE2 intrin and we don't have such a header.
>
> Huh. The last spec that I saw didn't have that.
>
> > Aditionally ICC uses register passing conventions for SSE operands...
>
> They do? That's a definite ABI change from before.
ICC compiles my testsuite into:
t1:
.B1.1: # Preds .B1.0
pushl %ebx #15.1
movl %esp, %ebx
andl $-16, %esp
pand %xmm1, %xmm0
movl %ebx, %esp
popl %ebx
ret
Looks like register passing convention :)
>
> > *** xmmintrin.h Mon Oct 14 18:09:21 2002
> > --- /p1/ssediv/egcs/gcc/config/i386/xmmintrin.h Wed Sep 18 00:58:59 2002
> > *************** typedef int __v4si __attribute__ ((mode
> > *** 1066,1072 ****
> > typedef int __v8hi __attribute__ ((mode (V8HI)));
> > typedef int __v16qi __attribute__ ((mode (V16QI)));
> >
> > - #define __m128i __v2di
> > + #define __m128i __m128
> > #define __m128d __v2df
> >
> > static __inline __m128d
>
> Is the patch reversed?
Yes, sorry for that.
There is also bug in the fabs patch, I will send updated version of that
one.
Honza
>
>
> r~
More information about the Gcc-patches
mailing list