Hi,
On 2022-02-15 08:06:58 -0800, Andres Freund wrote:
> The more I think about it the more I'm convinced that if we want to do this,
> we should do it for variables and functions.
Btw, if we were to do this, we should just use -fvisibility=hidden everywhere
and would see the same set of failures on unixoid systems as on windows. Of
course only in in-core extensions, but it'd still be better than nothing.
I proposed using -fvisibility=hidden in extensions:
https://postgr.es/m/20211101020311.av6hphdl6xbjbuif@alap3.anarazel.de
Mostly because using explicit symbol exports makes it a lot easier to build
extensions libraries on windows (with meson, but also everything built outside
of postgres with msvc). But also because it makes function calls *inside* an
extension have noticeably lower overhead. And it makes the set of symbols
exported smaller.
One problem I encountered was that it's actually kind of hard to set build
flags only for extension libraries:
https://postgr.es/m/20220111025328.iq5g6uck53j5qtin%40alap3.anarazel.de
But if we added explicit exports to everything we export, we'd not need to
restrict the use of -fvisibility=hidden to extension libraries anymore. Would
get decent error messages on all platforms.
Subsequently we could yield a bit of performance from critical paths by
marking selected symbols as *not* exported. That'd make them a tad cheaper to
call and allow the optimizer more leeway.
Greetings,
Andres Freund
