One practical reason for an interrupt stack is to reduce the size of
per-process kernel stacks.  If processes only need a 4k kernel stack but
you need a 32k stack to handle deep interrupt nesting or long interrupt
call chains, then it is better to only have one 32k stack rather than one
for every process (since any process might be active when an interrupt
sequence begins).

This never proved to be a problem on the 68k (hp300), so I never felt
compelled to change it.