The problem is that simba's are wired so the first bus is
   used for PCI slots and the second bus is used for on-board
   devices.  Our PCI code probes buses in order, so the 
   PCI slots are probed before the on-board devices.

exactly.  and the second bus on the real psycho is also used
for on-board devices, which is pci1...
   
   The real solution to the problem would be to change
   our PCI probe code to use the PROM device tree rather
   than PCI config space to probe devices.  That would
   also solve our interrupt remapping problem, and
   resolve the need to scan the device tree multiple
   times to hunt for specific devices.


actually, somewhere, i have patches along these lines.  the idea
was to add, eg, pci_attach_args{}->pa_node, plus to probe the bus
via OFW.

the problem is that the OFW probe will find the on-board devices
last as well.  :-(

the only hacks i can think of without some real serious rototillage
(and design before hand!) would be to force the sabre driver to
swap the pci bus numbers...  or perhaps we could do a config_defer()
or whatever it is of the "primary" bus ("primary" because it is
pci1 on the APB/USIIi), so that they are probed after all "secondary"
busses.  this might even scale to multi-cpu boards where there are
several on-board hme's plus potentially more installed?


.mrg.