We have seen several previous reports of regression test failures
due to division by zero causing SIGFPE, even though the code
should never reach the division command:
http://archives.postgresql.org/pgsql-bugs/2006-11/msg00180.php
http://archives.postgresql.org/pgsql-bugs/2007-11/msg00032.php
http://archives.postgresql.org/pgsql-bugs/2008-05/msg00148.php
http://archives.postgresql.org/pgsql-general/2009-05/msg00774.php
It's always been on non-mainstream architectures so it was hard
to investigate. But I have finally been able to reproduce this:
https://bugzilla.redhat.com/show_bug.cgi?id=520916
While s390x is still not quite mainstream, at least I can get
access to one ;-). What turns out to be the case is that
"simple" test cases likeif (y == 0) single_function_call(...);z = x / y;
do not show the problem; you need something pretty complex in the
if-command. Like, say, an ereport() construct. So that's why the gcc
boys haven't already had visits from mobs of villagers about this.
I hope that the bug will get fixed in due course, but even if they
respond pretty quickly it will be years before the problem disappears
from every copy of gcc in the field. So I'm thinking that it would
behoove us to install a workaround, now that we've characterized the
problem sufficiently. What I am thinking is that in the three
functions known to exhibit the bug (int24div, int28div, int48div)
we should do something like this:
if (arg2 == 0)
+ { ereport(ERROR, (errcode(ERRCODE_DIVISION_BY_ZERO), errmsg("division by zero")));
+ /* ensure compiler realizes we don't reach the division */
+ PG_RETURN_NULL();
+ }/* No overflow is possible */PG_RETURN_INT64((int64) arg1 / arg2);
Thoughts?
regards, tom lane
