On Fri, Mar 04, 2022 at 05:03:14PM -0500, Tom Lane wrote: > hubert depesz lubaczewski <depesz@xxxxxxxxxx> writes: > > On Fri, Mar 04, 2022 at 02:09:52PM -0500, Tom Lane wrote: > >> I tried and failed to reproduce this on Fedora 35 on aarch64, > >> but that has what I think is a newer LLVM version: > > > I have suspicion that it also kinda depends on number of rows in there. > > When I deleted some schemas, which included some functions, the problem > > disappeared. > > I wasn't able to pinpoint specific thing, though, and when I called the > > pg_proc + pg_roles query for each separate row - it worked flawlessly. > > Mmm ... it might have just been that the planner chose not to use > JIT when it thought there were fewer rows involved. Did you check > with EXPLAIN that these cut-down cases still used JIT? Hi, I tore these boxes down, so can't check immediately, but I think I remember that you're right - single-row queries didn't use JIT. Best regards, depesz
