DISPATCH-1344 Relaxing C11 sys_atomic_ operations

[jiridanek]

This is a revival of DISPATCH-1344. The issue that was never meant to have been abandoned.

@astitcher The shared_ptr article about atomics is actually not an article, but a SO question, https://stackoverflow.com/questions/28199212/why-does-libcs-implementation-of-shared-ptr-use-full-memory-barriers-instead

Regarding the question "What use are the relaxed atomics on x86_64?", as far as I can tell, the hardware recognizes two "levels": seq_cst and everything else. By that I mean that the generated code for seq_cst store is actually different from non-cst_seq store. But, the relaxation also influences what code will be generated by the compiler. Even if the CPU does not do reorders, the compiler will, and that may be significant for performance, sometimes. (https://stackoverflow.com/questions/61719680/are-memory-orderings-consume-acq-rel-and-seq-cst-ever-needed-on-intel-x86)

So, IMO if this shows promise in benchmark, then there may be value in thinking about this. Ultimately, the atomics, as a way to silence TSan, were not all that great a solution and hopefully they will go away, eventually.

[jiridanek]

The test fails are all already known skupper-router issues, with the exception of one qpid-dispatch issue not yet logged here, I think

• https://issues.apache.org/jira/browse/DISPATCH-2096 Did not get expected Aborted_Delivery in system_tests_delivery_abort
  • https://github.com/skupperproject/skupper-router/runs/5695037180?check_suite_focus=true#step:26:996

[jiridanek]

This does not appreciably increase performance for me. I still hope that it's because I don't have the right benchmark for it (this might do better when there is some fanout, these things; so far I only had one sender/receiver pair).

[jiridanek]

There is no future in this PR, let's close it.