allow sorted builds for btree_gist

Hi,

I've been looking at GiST to see if there could be a good way to do
parallel builds - and there might be, if the opclass supports sorted
builds, because then we could parallelize the sort.

But then I noticed we support this mode only for point_ops, because
that's the only opclass with sortsupport procedure. Which mostly makes
sense, because types like geometries, polygons, ... do not have a well
defined order.

Still, we have btree_gist, and I don't think there's much reason to not
support sorted builds for those opclasses, where the gist opclass is
defined on top of btree ordering semantics.

So this patch does that, and the difference (compared to builds with
buiffering=on) can easily be an order of magnitude - at least that's
what my tests show:


test=# create index on test_int8 using gist (a) with (buffering = on);
CREATE INDEX
Time: 578799.450 ms (09:38.799)

test=# create index on test_int8 using gist (a) with (buffering = auto);
CREATE INDEX
Time: 53022.593 ms (00:53.023)


test=# create index on test_uuid using gist (a) with (buffering = on);
CREATE INDEX
Time: 39322.799 ms (00:39.323)

test=# create index on test_uuid using gist (a) with (buffering = auto);
CREATE INDEX
Time: 6466.341 ms (00:06.466)


The WIP patch adds enables this for data types with a usable sortsupport
procedure, which excludes time, timetz, cash, interval, bit, vbit, bool,
enum and macaddr8. I assume time, timetz and macaddr8 could be added,
it's just that I didn't find any existing sortsupport procedure. Same
for cash, but IIRC it's mostly deprecated.

Of course, people probably don't use btree_gist with a single column,
because they could just use btree. It's useful for multi-column GiST
indexes, with data types requiring GiST. And if the other opclasses also
allow sorted builds, this patch makes that possible. Of course, most
"proper GiST opclasses" don't support that, but why not - it's easy.

FWIW this is also why sorted builds likely are not a very practical way
to do parallel builds for GiST - it would help only with a small part of
cases, I think.


regards

-- 
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company