fvdl@netbsd.org said:
> The 'irq' value is really only for reference in ioapic mode, you can
> ignore it for all practical purposes. 

Ah - thanks for the clarification.

> Also, the interrupt code is virtually the same between amd64 and
> i386, so I don't think it's an interrupt sharing issue.

So are there i386 boxes with more than 1 ioapic? Perhaps this
is what makes my case different?

> Perhaps you can try using MPACPI
> instead of MPBIOS and see if that gives a different configuration.

This changed one unrelated interrupt, and the box still hung
after a while as before.
Without the SCSI controller the box runs without problems now.
(There were crashes when running i386 binaries due to the missing
kernel lock, but haven't seen any crash anymore after the fix.
Only that the kernel hangs solidly after the initial copyright
printout when compiled with LOCKDEBUG...)

I'll investigate more later this week (after a 2-day business trip).

best regards
Matthias