Rethink nullingrel marking rules in build_joinrel_tlist().
The logic for when to add the current outer join's own relid
to the nullingrels sets of output Vars and PHVs was overly
complicated and underly correct. Not sure why I didn't think
of this before, but since what we want is marking per the
syntactic structure, we can just consult our records about
the syntactic structure, ie syn_righthand/syn_lefthand.
Also, tighten the rule about when to add the commute_above_r
bits, in hopes of eliminating some squishy reasoning. I do not
know of a reason to think that that's broken as-is, but this way
seems better.
Per bug #17781 from Robins Tharakan.
Discussion: https://postgr.es/m/17781-c0405c8b3cd5e072@postgresql.org
Branch
------
master
Details
-------
https://git.postgresql.org/pg/commitdiff/fee7b77b9000f35e445de9954a8cbf241f181e60
Modified Files
--------------
src/backend/optimizer/util/relnode.c | 43 +++++++++++++++++++----------------
src/test/regress/expected/join.out | 44 ++++++++++++++++++++++++++++++++++++
src/test/regress/sql/join.sql | 28 +++++++++++++++++++++++
3 files changed, 96 insertions(+), 19 deletions(-)