tree f42f165955828a8627779a6d3ac7cae696d864ee
parent b4c09637a67492be2fdb125054cf4535947ea3ea
author Ben Vanik <ben.vanik@gmail.com> 1649884397 -0700
committer GitHub <noreply@github.com> 1649884397 -0700
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJiVzztCRBK7hj4Ov3rIwAACo8IAISSMb7OX/IYgCXAim44mrFu
 NgL+8AFzHWACR638OiimQ99sqqez/+w0D2hudMs04fIUosALvz3uNK76ZVEXWr7o
 W1x9o02xK7UwYrYU0ADDmLSHt2hoV3naiQUoEbJkoXytDuuaN5h6qa7jSpxat9jO
 1w0wo/6FdWmiinSdK/3JFlMPLfddqsFQl5XlYdJPz6k2GctSiIa1nnnLvIGPqxLz
 06XpH5oAPURjOiiH/67lumI+fDRyqSPr2TMm5SGXmud+qa/ZE/PW8pTMuzyalf7M
 pg9kpfLdrbS7FgI/qDVI9m+izoocyyTR4jFQ6+NlUathPVA53RQY8N/DvVlMHWI=
 =e5eo
 -----END PGP SIGNATURE-----
 

Rearranging iree_notification_t code to make it easier to follow. (#8877)

The mix of atomic/futex and pthreads operations made it hard to know
when the pthread mutex should be locked and when atomics were ok. The
disabled synchronization path that no-op'ed also obscured the intended
behavior. Now we have the divergent code split up more cleanly so that
we can more easily inspect it for correctness.

iree_notification_await is left shared across the implementations to
preserve program behavior (that the condition is not checked with
a lock held) though we could probably tweak that if we wanted. Erring
on the side of keeping things the same for now.

Integrates the fix from #8876 for a hang/deadlock when using pthreads.