> Now only the problem with bad I/O addresses remain. This is what I get
> (and change) in my SX at boot time. I have to take back that about
> overlapping addresses, I was masking too much (different masks for
> mem/io). It seems the firmware simply thinks that IO addresses above
> 0x10000 are ok (which they probably are in the alpha), but NetBSD wraps IO
> addresses higher than 0xffff.

It looks to me like this probably gets broken in the CIA/BWX bus_space
backend.  I don't see any code in dev/pci, dev/ic, or dev/ata that
deals with anything other than a 32-bit I/O port number or a
bus_space_handle_t.