I wrote:
> Sami Imseih <samimseih@gmail.com> writes:
>> I am attaching a patch that deals with the RTE_JOIN case.
> I'll take a look. Thanks for the test demonstrating that
> this makes a visible performance difference.
Pushed with some simplification: we don't need a new flag,
because none of the callers of set_simple_column_names need it
to do anything with join RTEs. This is better anyway because
set_relation_column_names' comment explicitly says it is not
for join RTEs, and now we don't use it on them ever.
I poked at the question of whether it's worth skipping
unique-ification for relation RTEs, and I came to the same
conclusion as you: it doesn't seem to be. The related code
is down in the noise according to "perf" once we skip join
RTEs. I think the reason the join RTEs are so expensive for
this is that the upper ones get very wide in join nests like
the example query.
regards, tom lane
