commit 5224d8c02f154af0c9ad8d2e627134c4e24a6ca6
author bjacob <benoitjacob@google.com> Tue Jul 19 14:02:46 2022 -0400
committer GitHub <noreply@github.com> Tue Jul 19 11:02:46 2022 -0700
tree 52a736d10568441b43159e4a8e8c6710a3e10e45
parent 1fe95b6b016de4dea3cd006188d7a097e66a82f2

Finish rationalizing atomics in remaining runtime/ code (#9837)

Similar to other PRs from today (#9834, #9835, #9836) this is the rest of the "audit" on all remaining runtime/ code.

With this, there are no remaining sequentially-consistent atomics in non-test IREE code. That wasn't a goal in itself: seq_cst atomics are fine (though expensive). But it happened that as many accesses were acquire/release anyway, making remaining accesses acquire/release was not only a potential optimization, it also makes the overall code easier to reason about and more self-documenting. One good thing about acquire and release orders is that they convey intent. By contrast, seq_cst and relaxed don't convey as much information about intent, so an audit of all uses was sure to find at least some cases that were using them for no particular reason.

Some places are changed all the way from seq_cst to relaxed, when that was correct. Comments are added to justify relaxed, also at existing places that were already using relaxed. Relaxed is the most aggressive order and doesn't generally self-document intent (unlike acquire/release), so uses should generally be justified by a comment.