Re: reducing the footprint of ScanKeyword (was Re: Large writable variables)

Andres Freund <andres@anarazel.de> writes:
> On 2019-01-08 13:41:16 -0500, John Naylor wrote:
>> Do you mean the fmgr table?

> Not the entire fmgr table, but just the builtin oid index, generated by
> the following section:
> ...
> The generated fmgr_builtin_oid_index is pretty sparse, and a more dense
> hashtable might e.g. more efficient from a cache perspective.

I experimented with this, but TBH I think it's a dead loss.  We currently
have 2768 built-in functions, so the perfect hash table requires 5537
int16 entries, which is not *that* much less than the 10000 entries
that are in fmgr_builtin_oid_index presently.  When you consider the
extra cycles needed to do the hashing, and the fact that you have to
touch (usually) two cache lines not one in the lookup table, it's hard
to see how this could net out as a win performance-wise.

Also, I fail to understand why fmgr_builtin_oid_index has 10000 entries
anyway.  We could easily have fmgrtab.c expose the last actually assigned
builtin function OID (presently 6121) and make the index array only
that big, which just about eliminates the space advantage completely.

BTW, I found out while trying this that Joerg's fear of the hash
multipliers being too simplistic is valid: the perfect hash generator
failed until I changed them.  I picked a larger value that should be
just as easy to use for shift-and-add purposes.

			regards, tom lane