We're leaking predicate locks in HEAD

After running the core regression tests with installcheck-parallel,
the pg_locks view sometimes shows me apparently-orphaned SIReadLock
entries. They accumulate over repeated test runs. Right now,
for example, I see

regression=# select * from pg_locks;
locktype | database | relation | page | tuple | virtualxid | transactionid | classid | objid | objsubid | virtualtransaction | pid | mode | granted | fastpath
------------+----------+----------+------+-------+------------+---------------+---------+-------+----------+--------------------+------+-----------------+---------+----------
relation | 130144 | 12137 | | | | | | | | 3/7977 | 8924 | AccessShareLock | t | t
virtualxid | | | | | 3/7977 | | | | | 3/7977 | 8924 | ExclusiveLock | t | t
relation | 130144 | 136814 | | | | | | | | 22/536 | 8076 | SIReadLock | t | f
relation | 111195 | 118048 | | | | | | | | 19/665 | 6738 | SIReadLock | t | f
relation | 130144 | 134850 | | | | | | | | 12/3093 | 7984 | SIReadLock | t | f
(5 rows)

after having done a couple of installcheck iterations since starting the
postmaster.

The PIDs shown as holding those locks don't exist anymore, but digging
in the postmaster log shows that they were session backends during the
regression test runs. Furthermore, it seems like they usually were the
ones running either the triggers or portals tests.

I don't see this behavior in v11 (though maybe I just didn't run it
long enough). In HEAD, a run adds one or two new entries more often
than not.

This is a pretty bad bug IMO --- quite aside from any ill effects
of the entries themselves, the leak seems fast enough that it'd run
a production installation out of locktable space before very long.

I'd have to say that my first suspicion falls on bb16aba50 ...

regards, tom lane