Re: Binary search in fmgr_isbuiltin() is a bottleneck.

Andres Freund <andres(at)anarazel(dot)de> writes:
> On 2017-09-27 15:06:15 -0400, Tom Lane wrote:
>> Yeah, constructing an index table of that sort on top of the existing
>> FmgrBuiltin array could be done cheaply enough at startup. It irks me
>> slightly that it's not part of the read-only text segment, but I can't
>> say that there's any really measurable impact.

> I don't think this case is significant enough to make it worthwhile, but
> if we'd find one that is, we certainly could add code that builds the
> hash's array once in memory, then serializes that into a .h file, which
> then is included into the code. I can't immediately see more of these
> coming up, but who knows?

Actually ... a more defensible reason for having a precomputed constant
table is that it removes any question about where is a safe place in the
initialization sequence to inject "fmgr_startup". That would clearly
have to go before anything that could conceivably try to call a SQL
function. On the other hand, it has to go after elog.c setup (in case
you get e.g. a malloc failure), which means you've now created a reason
why it will never be safe for elog.c to include any fmgr calls. Maybe
that's unsafe anyway, but I'd just as soon not introduce constraints of
that kind just because we're too lazy to do this optimization properly.
There definitely are places in startup that assume they can call
built-in functions (relying on fmgr_isbuiltin) long before most of the
transactional infrastructure is up.

ISTM it shouldn't be that hard to get Gen_fmgrtab.pl to emit an index
array of the sort we're talking about, along with the FmgrBuiltin array
it already prints out. I'm the world's worst Perl programmer but
I'm happy to take a stab at it if you don't want to.

regards, tom lane