Thread: Shorter iterations of join_info_list

Shorter iterations of join_info_list

From
Antonin Houska
Date:
As far as I understand, deconstruct_recurse() ensures that 
SpecialJoinInfo of a new join always gets added to higher position in 
join_info_list than SJ infos of all joins located below the new join in 
the tree. I wonder if we can rely on that fact sometimes.

One possible use case could be 
placeholder.c:update_placeholder_eval_levels():
  1. The first (in terms of position in join_info_list) join above 
phinfo->ph_var->phrels can be marked as the (exclusive) upper bound for 
all iterations.
  2. The first join for which particular iteration of SJ infos 
identifies the necessity to extend eval_at can be marked in 
join_info_list as (exclusive) lower bound for the next iteration. This 
is because that addition can only affect parents of the join whose 
relations we just added to eval_at. And these essentially can't be 
located at lower positions in join_info_list.

The lower limit could also be used in initsplan.c:check_outerjoin_delay().

Is this worth a patch? It's not much coding but I'd appreciate some 
feedback before I try to do anything.

Thanks,
Antonin Houska (Tony)



Re: Shorter iterations of join_info_list

From
Tom Lane
Date:
[ sorry for slow response, this month has been mostly crazy ]

Antonin Houska <antonin.houska@gmail.com> writes:
> As far as I understand, deconstruct_recurse() ensures that 
> SpecialJoinInfo of a new join always gets added to higher position in 
> join_info_list than SJ infos of all joins located below the new join in 
> the tree. I wonder if we can rely on that fact sometimes.

FWIW, I think of most of those planner lists as being unordered sets.
Depending on a particular ordering definitely adds fragility; so I'd
not want to introduce such a dependency without solid evidence of a
substantial speed improvement.  The cases you mention don't seem very
likely to offer any noticeable gain at all --- at least, I don't recall
seeing any of those functions show up high in profiles.
        regards, tom lane