On Sat, Jul 03, 2021 at 06:44:20PM -0400, Tom Lane wrote:
> Noah Misch <noah@leadboat.com> writes:
> > On Sat, Jul 03, 2021 at 10:45:59AM -0400, Tom Lane wrote:
> >> What I'm now thinking about is restricting the test to only be run on
> >> platforms where use of foo.a libraries is deprecated, so that we can
> >> be pretty sure that we won't hit this situation. Even if we only
> >> run the test on Linux, that'd be plenty to catch any mistakes.
>
> > Hmm. Static libraries are the rarer case on both AIX and Linux, but I'm not
> > aware of a relevant deprecation on either platform. If it comes this to, I'd
> > be more inclined to control the Makefile rule with an environment variable
> > (e.g. ENFORCE_LIBC_CALL_RESTRICTIONS) instead of reacting to the platform.
>
> That'd require buildfarm owner intervention, as well as intervention
> by users. Which seems like exporting our problems onto them. I'd
> really rather not go that way if we can avoid it.
I like that goal, though we'll have to see how difficult it proves. As of
today, a GNU/Linux user building against static OpenLDAP will get a failure,
right? That would export work onto that user, spuriously. Since the non-AIX
user count dwarfs the AIX user count, expect a user complaint from non-AIX
first.
We'd get something like 95% of the value by running the test on one Windows
buildfarm member and one non-Windows buildfarm member. If you did gate the
check on an environment variable, there would be no need to angle for broad
adoption. Still, I agree avoiding that configuration step is nice, all else
being equal. A strategy not having either of those drawbacks would be to skip
the test if libpq.so contains a definition of libpq_unbind(). If any other
dependency contains exit calls, we'd likewise probe for one symbol of that
library and skip the test if presence of that symbol reveals static linking.
(That's maintenance-prone in its own way, but a maintenance-free strategy has
not appeared.)
