It is a atomic builtin for full memory barrier.
No memory operand will be moved across the operation, either forward
or backward. Further, instructions will be issued as necessary to
prevent the processor from speculating loads across the operation and
from queuing stores after the operation.
Check details on the link above.