CMAKE: Bring back checks for atomic CAS

[devreal]

Testing for atomic_is_lock_free() may not be enough as it may not require -latomic that is otherwise needed for 128bit atomics.

Reported by @DSMishler.

[therault]

The original code returns 0 or 1. This patch makes it return 0 (success), 1 or -1. I guess 1 and -1 can be conflated to failure in both cases. POSIX specifies only that the 8 lower bits are meaningful (https://pubs.opengroup.org/onlinepubs/9699919799.2018edition/functions/V2_chap02.html#tag_15_13), so -1 is not clear.

May I suggest we use EXIT_SUCCESS and EXIT_FAILURE for all return/exit values, since those are POSIX compliant?

[devreal]

I thought about harmonizing them but then I realized that it might be helpful to have different return values for the two error cases. That way we can distinguish (in the CMake log) whether the operation is not lock-free or whether we failed to execute it in the first place...

[abouteiller]

return 1 and 2 then, that's probably more clear than 1 and 254

[bosilca]

In what case this is necessary ? On my tests when using the libatomic is necessary, the associated intrinsics are automatically made available, and atomic_is_lock_free return the expected outcome.

[devreal]

I can only guess here based on @DSMishler issue: atomic_is_lock_free is a macro that may not need to call functions in libatomic even when CAS on 128bit integer would require libatomic to be present. So we never detect that we need to link -latomic.

Either way, these checks don't hurt and may help if a weird compiler may need -latomic for all atomic operations, but not for atomic_is_lock_free.