[Openmp-commits] [PATCH] D142008: [libomptarget] Delay recursive shared library registrations until offload RTLs are loaded

[Jon Chesterfield via Phabricator via Openmp-commits]

JonChesterfield added a comment.

In D142008#4065590 <https://reviews.llvm.org/D142008#4065590>, @jdoerfert wrote:

> Instead of this local atomic and thread Id stuff, why don't we lock the entire thing? One thread at a time can register or unregister. It's not like we want to make this part super fast.

Lock sounds good here. I can't picture a use case for trying to do lots of register/unregister calls from lots of threads. If that deadlocks immediately (which it might, it's not clear to me whether the control flow here can go back in and try to re-take the same lock) that's useful information to have.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D142008/new/

https://reviews.llvm.org/D142008