Replace alignment_of_long_double with C11 alignof

[devreal]

This removes a static variable in a static function and uses C11 provides alignof instead to get a compile-time value.

Also fixes this compiler warning:

opal_copy_functions_heterogeneous.c:495:1: warning: function declaration isn't a prototype [-Wstrict-prototypes]
  495 | alignment_of_long_double() {
      | ^~~~~~~~~~~~~~~~~~~~~~~~

I'm not sure why alignof would not be portable, since it's part of the C11 standard. All C11 compilers I tried on godbolt worked as expected (and some exotic 32bit compiler versions seemed to indicate 8B alignment for long double, not 4). Asking the compiler for the alignment seems to be the safe and reasonable choice.

[wenduwan]

Running AWS CI

[bosilca]

We're not supporting 32 bits anymore, so moving to alignof is sensible.

[rhc54]

Errr...just a question. Is OMPI now requiring C11 compiler? We historically chose to require C11 atomics, but not a C11 compiler. Just wondering if you are adding a new requirement as PMIx and PRRTE don't.

[devreal]

My understanding was that we use C11 (not just the atomics). I might be wrong though. It's been 12+ years though, so maybe it's it's time to move on from last century?

[bosilca]

The C11 operator is _Alignof, and the C23 operator is alignof according to https://en.cppreference.com/w/c/language/_Alignof

[devreal]

The C11 operator is _Alignof, and the C23 operator is alignof according to https://en.cppreference.com/w/c/language/_Alignof

Right, C11 provided alignof as a macro while C23 made it an operator. See https://en.cppreference.com/w/c/types for the macro.

[rhc54]

My understanding was that we use C11 (not just the atomics). I might be wrong though. It's been 12+ years though, so maybe it's it's time to move on from last century?

I honestly don't remember, except that this was a significant discussion (separating the two) at the time when the C11 atomic requirement was proposed. We intentionally left the base level of the C compiler lower because of defaults in distros. May well be that has changed - truly don't know. Just wanted to point it out here as we have unintentionally broken things in the past with what appeared to be an innocent/reasonable change.

Might be worth checking to see what OMPI actually does require today, and then discuss what that requirement should be going forward. FWIW: we required C99 for many, many years beyond what seemed "reasonable" just because of long-running distro defaults.

[devreal]

It looks like we're trying to enable C11 across the board (if the compiler needs a specific flag) but it is not an error if C11 support is not detected: https://github.com/open-mpi/ompi/blob/main/config/opal_setup_cc.m4#L182

[jsquyres]

After some discussion on the RM call today (let's also discuss on community call tomorrow):

• Let's force the use of C11 for v6.0.x (i.e., doing this on main is fine -- but let's not bring it to v5.0.x)
• Let's also update the m4 to enforce that we need C11 (per what @devreal cited)
• Let's also update the docs to state that C11 is a prereq (wherever we currently state C99)

[janjust]

depends on #12660