Attached is a patch that exports a new function from logtape.c:
extern LogicalTapeSet *LogicalTapeSetExtend(
LogicalTapeSet *lts, int nAdditional);
The purpose is to allow adding new tapes dynamically for the upcoming
Hash Aggregation work[0]. HashAgg doesn't know in advance how many
tapes it will need, though only a limited number are actually active at
a time.
This was proposed and originally written by Adam Lee[1] (extract only
the changes to logtape.c/h from his posted patch). Strangely, I'm
seeing ~5% regression with his version when doing:
-- t20m_1_int4 has 20 million random integers
select * from t20m_1_int4 order by i offset 100000000;
Which seems to be due to using a pointer rather than a flexible array
member (I'm using gcc[2]). My version (attached) still uses a flexible
array member, which doesn't see the regression; but I repalloc the
whole struct so the caller needs to save the new pointer to the tape
set.
That doesn't entirely make sense to me, and I'm wondering if someone
else can repro that result and/or make a suggestion, because I don't
have a good explanation. I'm fine with my version of the patch, but it
would be nice to know why there's such a big difference using a pointer
versus a flexible array member.
Regards,
Jeff Davis
[0]
https://postgr.es/m/6e7c269b9a84ff75fefcad8ab2d4758f03581e98.camel%40j-davis.com
[1] https://postgr.es/m/20200108071202.GA1511@mars.local
[2] gcc (Ubuntu 7.4.0-1ubuntu1~18.04.1) 7.4.0