This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: SSE types and structures
> On Fri, Jun 14, 2002 at 10:50:41AM +0200, Jan Hubicka wrote:
> > I guess we want to 32bit align DFmode, DImode, XFmode and their complex
> > variants, but not TFmode, SSE modes and MMX modes.
>
> Why not TFmode? I'm thinking you shouldn't change that either.
Currently no-one is using TFmode, because it simply does not work.
Our libgcc is not gneerated for TFmodes for instance, so I think it
makes sense to do the change at the moment.
GCC 3.1.0 is the first recognizing TFmodes and we don't need to repeat
the mistake yet.
>
> > Unfortunately __m64 is typdefed to long long. This makes problems for
> > x86-64 ABI, as then it will be passed in the integer register, instead
> > of memory, as specified. What kind of breakage can cause if I retype it
> > to let say... V8QI?
>
> Another possibility is
>
> typedef unsigned long long __m64 __attribute__((aligned(8)));
This still makes it impossible to distinquish __m64 from long long for
x86-64 ABI, but I guess it is not hot for 3.1.x yet.
>
> but that needs to be checked against Intel's compiler and
> how it lays out __m64 in structures. It may be that we
> cannot change it at all.
I will check, thanks.
Honza
>
>
> r~