Any MIPS assembly experts in the house?

I see the latest buildfarm result from a mipsel machine is failing:
http://www.pgbuildfarm.org/cgi-bin/show_log.pl?nm=lionfish&dt=2005-08-26%2005:30:07

and the failure is this:

TRAP: FailedAssertion("!(lock->shared > 0)", File: "lwlock.c", Line: 456)
LOG:  server process (PID 10112) was terminated by signal 6

which makes it seem highly probable that this recently committed patch
to convert the MIPS out-of-line spinlock code into inline assembler
isn't right:
http://archives.postgresql.org/pgsql-committers/2005-08/msg00319.php

Can anyone spot the problem?  If not I fear we'll have to revert this.

Also, upon looking more closely at the new inline code, it looks like
it's designed to *loop* if it fails to get the lock, which is *wrong*.
The assembly fragment should test once and fall out.  This doesn't
explain the regression test failure, it would just cause stuck-lock
detection to not work.
        regards, tom lane