I wrote:
>> Is anyone else seeing major breakage of the regression tests with
>> today's (Monday's) CVS checkins? Or did I break something myself?
Nope, Vadim broke something. It looks like anything with a subplan
will coredump in Monday's sources. executor/nodeSubPlan.c has
bool
ExecInitSubPlan(SubPlan *node, EState *estate, Plan *parent)
{ ... ExecCheckPerms(CMD_SELECT, 0, node->rtable, (Query *) NULL);
^^^^^^^^^^^^^^
(and has had that for a long time, evidently). One of the additions
Vadim checked in yesterday extends ExecCheckPerms() to try to use
its parseTree argument --- unconditionally. Guaranteed null-pointer
dereference.
Perhaps ExecInitSubPlan is in error to pass a null parseTree; if not,
then ExecCheckPerms needs to be modified to cope. I don't understand
either routine enough to fix it correctly.
This bug is the cause of the opr_sanity coredump I'm seeing.
I don't have time to investigate the other test failures right now,
but very possibly they are the same thing.
BTW, anyone who is *not* seeing regression test coredumps with the
current CVS sources must have their compile/link options set so that
dereferencing a null pointer isn't fatal. I think that's a very bad
choice for software development --- you want to hear about it, loud
and clear, if your code tries to use a null pointer.
regards, tom lane
