I'm thinking about the way to improve add_outer_joins_to_relids() and
here is what I come up with. When we've completely formed a pushed up
outer join, actually we only need to consider the pushed-down joins that
are in its commute_above_l. But note that this process should be
recursive, meaning that if an outer join in its commute_above_l is also
considered qualified to be added to the relid set, we need to consider
that outer join's commute_above_l too. By this way we only need to
check the relevant SpecialJoinInfos, rather than scan the whole
join_info_list.
To do it we need a way to fetch SpecialJoinInfo by ojrelid. So 0001
patch introduces join_info_array for direct lookups of SpecialJoinInfo
by ojrelid. I find that it also benefits some existing functions, such
as clause_is_computable_at() and have_unsafe_outer_join_ref(). So I
started a new thread for it at [1].
0002 is the original patch that introduces add_outer_joins_to_relids().
0003 patch implements the improvement to add_outer_joins_to_relids().
All the pushed down joins that are needed to check are kept in
commute_above_relids. Each time we fetch and remove the first outer
join from commute_above_relids and check if that outer join is qualified
to be added to the relid set. If so we add it and also add its
commute_above_l to commute_above_relids. This process continues until
commute_above_relids becomes empty.
0004 patch adjusts outer join's target list as described before, but
leverages the new join_info_array to fetch relevant SpecialJoinInfos.
[1]
https://www.postgresql.org/message-id/flat/CAMbWs4_EyKimsqkkBAddEW8n1YyPjQd4gmnwYqqHHAUjKkBVQw%40mail.gmail.comThanks
Richard