Pavel Stehule <pavel.steh...@gmail.com> writes:
> Did you do some performance checks?
This is a good question to ask ...
> I tried some worst case
> CREATE OR REPLACE FUNCTION fx(int)
> RETURNS int AS $$
> SELECT $1 + $1
> $$ LANGUAGE SQL IMMUTABLE;
... but I don't think tests like this will give helpful answers.
That function is simple enough to be inlined:
regression=# explain verbose select fx(f1) from int4_tbl;
QUERY PLAN
---------------------------------------------------------------
Seq Scan on public.int4_tbl (cost=0.00..1.06 rows=5 width=4)
Output: (f1 + f1)
(2 rows)
So functions.c shouldn't have any involvement at all in the
actually-executed PERFORM expression, and whatever difference
you measured must have been noise. (If the effect *is* real,
we'd better find out why.)
You need to test with a non-inline-able function. Looking
at the inlining conditions in inline_function(), one simple
hack is to make the function return SETOF. That'll only
exercise the returns-set path in functions.c though, so it'd
be advisable to check other inline-blocking conditions too.
regards, tom lane
