> I wonder if we don't support some of the cache 'quirks' of some of
> the early-rev R4000 chips -- I know some early R4600 chips had bugs
> in this area and maybe the R4000s did as well.

it might be time for a little worlds-collision: a bunch of former sgi
engineers have dinner on wednesday nights in The Valley (usually this
mexican place in mountain view), maybe someone should join them and
pick brains. i distinctly remember irix 6.5 almost didn't support the
R4000 (but not R4400) indigo until someone checked a funky workaround
into the tree, and i'm willing to bet that this problem is closely
related, if not the same bug...