commit f36ec2915aae072f70273c48a0703af189a5facd
author bjacob <benoitjacob@google.com> Thu Jul 07 11:39:54 2022 -0400
committer GitHub <noreply@github.com> Thu Jul 07 08:39:54 2022 -0700
tree d7348d18d8c43088a0d54a02d98932ae111f14aa
parent 202436e60e2874a8d9a5adaac296d330f0071b1f

Consistently treat worker_*_masks as loose hints with relaxed accesses (#9722)

This makes worker_*_mask accesses use memory_order_relaxed, making them
be only fast and loose hints, and not play a role in ordering other
accesses.

Some of these used to be relaxed until
https://github.com/iree-org/iree/pull/8854
which made worker_live_mask use acquire/release to fix a specific race
reported by TSan. This PR re-solves that in a different way: the missing
ordering is now provided instead by an atomic access to worker->state.

Some accesses (in worker.c) were seq_cst even before
https://github.com/iree-org/iree/pull/8854,
so this new PR makes more things relaxed than were even back then.

This is part of a more general, consistent move towards this design:
* worker->state is the only memory-ordering, coherent view of workers'
  state.
* worker_*_mask is only fast and loose hints.

In https://github.com/iree-org/iree/pull/9560 I was questioning the need
for those hints, but they really serve a purpose: for callers that
favor performance over getting the most up-to-date information and that
need to iterate over all workers, worker_*_mask allows to get that
information in just 1 atomic load instead of N, and that atomic load is
now relaxed.