On Thu, Oct 08, 2020 at 09:15:12AM +1300, David Rowley wrote:
> On Thu, 8 Oct 2020 at 06:13, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> > Noah Misch <noah@leadboat.com> writes:
> > > gram.o was it. (The way I rebuilt gram.o also rebuilt parser.o and scan.o,
> > > but those seem far less likely.)
> >
> > This suggests that the problem is misoptimization of gram.y's
> > makeOrderedSetArgs:
>
> It would be interesting to see gram.s from both cc99baa4~1 and cc99baa4.
Attached. Both generated like this:
(cd src/backend/parser && xlc_r -D_LARGE_FILES=1 -qnoansialias -g -O2 -S -qmaxmem=33554432 -I. -I.
-I../../../src/include-c gram.c; )
(Again, the compiler is eight years old, and we're not seeing the problem on
32-bit (mandrill) or on 2019-vintage xlc (hoverfly). As soon as the situation
no longer piques your curiosity, I'm happy to make hoverfly use -O0 on v14+.)