Robert Haas <robertmhaas@gmail.com> writes:
> On Wed, Aug 11, 2010 at 5:12 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Yeah, possibly. It would probably be difficult for the planner to
>> figure out where the cutover point is to make that worthwhile, though;
>> the point where you'd need to make the transformation is long before we
>> have any rowcount estimates.
> This may be a stupid question, but why does the transformation have to
> be done before we have the row count estimates?
Well, I was thinking in terms of doing it when we do the SRF inlining.
It might be that we could get away with just having an arbitrary cost
limit like 100*cpu_operator_cost, and not think about how many rows
would actually be involved.
> I think we're just
> looking for a scan node with a filter condition that contains a stable
> subexpression that's expensive enough to be worth factoring out,
I do *not* want to grovel over every subexpression (and
sub-sub-expression, etc) in a query thinking about whether to do this.
That gets O(expensive) pretty quickly. My idea of the appropriate scope
of a hack like this is just to prevent any performance loss from SRF
inlining.
Another approach we could take is to fix the implementation limitation
in inline_set_returning_function() about punting when there's a
sub-select in the arguments. Then users could make this happen for
themselves when it matters.
regards, tom lane