Fix SPEC2000 GCC misscopilation and SIMD support

[Richard Henderson]

On Mon, Dec 09, 2002 at 08:43:32PM +0100, Jan Hubicka wrote:
> this implements the plan I described previously - add BYTE operand to
> REG_CANNOT_CHANGE_MODE_P so we can declare that (subreg:DF (reg:V2DF) 0)
> is valid, but (subreg:DF (reg:V2DF) 8) is not.

I don't really like this approach.

I thought you were going to come up with some new way to 
expand accesses to the vector elements, and handle this
largely at expand time.  I still like that idea.

I don't see anything wrong with using CANNOT_CHANGE_MODE_CLASS
to avoid the ICE for now, but I don't really want to add the
byte offset argument.


r~