All atomic operations use the default memory_order_seq_cst

[JBontes]

Your atomic operations all use the sequentially-consistent ordering. This imposes a hard barrier on all atomic operations. For loads this can be more efficiently replaced with a memory_order_acquire and for stores this should be replaced with memory_order_release. On load/stores (aka CAS operations) you can do: memory_order_acq_rel for success and memory_order_relaxed for failure.
If you happen to have a store after the CAS, you can downgrade the success order to memory_order_acquire.

See: https://en.cppreference.com/w/cpp/atomic/memory_order

[frankist]

Are there plans to implement this @erez-strauss? I would help, but I don't feel qualified enough to do this.