Joerg Sonnenberger <joerg@bec.de> writes:
> On Sun, Jan 06, 2019 at 02:29:05PM -0500, Tom Lane wrote:
>> So we probably can't have inlined hashing code --- I imagine the
>> hash generator needs the flexibility to pick different values of
>> those multipliers.
> Right now, only the initial values are randomized. Picking a different
> set of hash functions is possible, but someone that should be done only
> if there is an actual need. That was what I meant with stronger mixing
> might be necessary for "annoying" keyword additions.
Hmm. I'm still leaning towards using generated, out-of-line hash
functions though, because then we could have a generator switch
indicating whether to apply the |0x20 case coercion or not.
(I realize that we could blow off that consideration and use a
case-insensitive hash function all the time, but it seems cleaner
to me not to make assumptions about how variable the hash function
parameters will need to be.)
> There are two ways for dealing with it:
> (1) Have one big hash table with all the various keywords and a class
> mask stored. If there is enough overlap between the keyword tables, it
> can significantly reduce the amount of space needed. In terms of code
> complexity, it adds one class check at the end, i.e. a bitmap test.
No, this would be a bad idea IMO, because it makes the core, plpgsql,
and ecpg keyword sets all interdependent. If you add a keyword to any
one of those and forget to rebuild the other components, you got trouble.
Maybe we could make that reliable, but I don't think it's worth fooling
with for hypothetical benefit. Also, it'd make the net space usage more
not less, because each of those executables/shlibs would contain copies
of all the keywords for the other ones' needs.
regards, tom lane