> I don't see a massive problem in having cpu dependent code in the
> kernel.  It is a bigger problem for user space code.

Nor do I, provided that the cost of indirecting through a function pointer 
(rather than just using BL) doesn't become a problem.

Perhaps the best way of doing this is to make the entry code common, and 
then when we've determined that a 'big' copy is required to dispatch to 
the CPU-specific routines.  That means we only take the cost of an 
indirect when it's likely to be a win.

R.