Re: RFA: New builtin SIMD types

[Jim Blandy <jimb at cygnus dot com>]

> BTW,
> 
> I sometimes wonder if there should be a REGISTER_VIRTUAL_NAME() so that
> the 1:1 between physical and virtual register names is broken.

Not sure what you mean.  Does REGISTER_NAME_ALIAS_HOOK do this?


> > ***************
> > *** 2968,2973 ****
> > --- 2972,2985 ----
> >     /* Build SIMD types.  */
> >     builtin_type_v4sf
> >       = init_simd_type ("__builtin_v4sf", builtin_type_float, "f", 4);
> > +   builtin_type_v4si
> > +     = init_simd_type ("__builtin_v4si", builtin_type_int32, "f", 4);
> > +   builtin_type_v8qi
> > +     = init_simd_type ("__builtin_v8qi", builtin_type_int8, "f", 8);
> > +   builtin_type_v4hi
> > +     = init_simd_type ("__builtin_v4hi", builtin_type_int16, "f", 4);
> > +   builtin_type_v2si
> > +     = init_simd_type ("__builtin_v2si", builtin_type_int32, "f", 2);
> 
> Looks like it's time for float types equivalent to the int32 et.al.
> types.  I've got to add some address types so I might add those as well
> :-)

If you do this, it will make some upcoming x86 family work much
cleaner.  If the host can handle x86 80-bit longs, then
REGISTER_CONVERT_TO_{VIRTUAL,RAW} don't need to lose information.
So I encourage you to hack this in.