data races in openmp on shared boolean flags

[ben-schwen]

The openmp documentation states our common used idiom e.g. setting shared boolean flag is UB in openmp.

Two memory operations are considered unordered if the order in which they must complete, as seen by their affected threads, is not specified by the memory consistency guarantees listed in Section 1.4.6. If multiple threads write to the same memory unit (defined consistently with the above access considerations) then a data race occurs if the writes are unordered. Similarly, if at least one thread reads from a memory unit and at least one thread writes to that same memory unit then a data race occurs if the read and write are unordered. If a data race occurs then the result of the program is unspecified.

#pragma omp parallel for
for (int i = 0; i < n; i++) {
    shared_flag = true;   // UB: concurrent unsynchronized writes
}

Originally posted by @ben-schwen in #7361 (comment)

[jangorecki]

We never update flag to false, only to true, so it is one way update.
Isn't it like this, that in the worst case some of threads will simply unnecessarily do extra iterations? (if read was happening during write of another thread, till flag refreshed from memory)
That's how I would imagine data races in this context.
In such case we don't mind.

In the past this variable was declared volatile and then for omp also wrapped into shared clause: 055047b

as of now it seems that atomic write is the best way to address it