As far as I understand, performance is not a valid concern.

For example, from dvyukov's KTSAN wiki [1]:

  Given a sufficiently expressive atomic API and a good
  implementation, you pay only for what you really need (if
  you pay just a bit less, generated code becomes incorrect).
  So performance is not an argument here. 

[1] https://github.com/google/ktsan/wiki/READ_ONCE-and-WRITE_ONC...