Using HEAD,
create table t1 as select generate_series(1,40000000) id;
vacuum analyze t1;
explain select * from t1, t1 t1b where t1.id = t1b.id;
-- should indicate a hash join
explain analyze select * from t1, t1 t1b where t1.id = t1b.id;
... watch the process's memory consumption bloat. (It runs for
awhile before that starts to happen, but eventually it goes to
a couple of GB.)
It looks to me like the problem is that ExecHashJoinGetSavedTuple
calls ExecForceStoreMinimalTuple with shouldFree = true, and
ExecForceStoreMinimalTuple's second code branch simply ignores
the requirement to free the supplied tuple.
regards, tom lane