On Wed, Oct 18, 2023 at 04:45:46PM -0400, Tom Lane wrote:
> > On Wed, Oct 18, 2023 at 12:02 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> >> Probably. Independent of that, it's fair to ask why we're still
> >> testing against xlc 12.1 and not the considerably-more-recent xlclang,
> >> or at least xlc 16.1. (I also wonder why we're still testing AIX 7.1
> >> rather than an OS version that's not EOL.)
> The machine belongs to OSU (via the gcc compile farm), and I see
> that they have another one that's POWER8 and is running AIX 7.3 [1].
> So in principle the buildfarm animals could just be moved over
> to that one.
>
> Perhaps Noah has some particular attachment to 7.1, but now that that's
> EOL it seems like we shouldn't be paying so much attention to it.
> My guess is that it's still there in the compile farm because the gcc
> people think it's still useful to have access to POWER7 hardware; but
> I doubt there's enough difference for our purposes to be worth dealing
> with a dead OS and ancient compiler.
No particular attachment. From 2019 to 2023-08, hoverfly tested xlc16 on AIX
7.2; its run ended when cfarm119's owner replaced cfarm119 with an AIX 7.3,
ibm-clang v17.1.1 machine. Since 2015, hornet and mandrill have tested xlc12
on AIX 7.1. That said, given your finding that later xlc versions have the
same code generation bug, the choice of version is a side issue. A migration
to ibm-clang wouldn't have prevented this week's xlc-prompted commits.
I feel the gravity and longevity of xlc bugs has been out of proportion with
the compiler's contribution to PostgreSQL. I would find it reasonable to
revoke xlc support in v17+, leaving AIX gcc support in place. The main
contribution of AIX has been to find the bug behind commit a1b8aa1. That
benefited from the AIX kernel, not from any particular compiler. hornet and
mandrill would continue to test v16-.
By the way, I once tried to report an xlc bug. Their system was tailored to
accept bugs from paid support customers only. I submitted it via some sales
inquiry form, just in case, but never heard back.
