Formally pair store_rel(&smp_started) with load_acq(&smp_started).
The expected semantic is to have misc. data, e.g. CPU bitmaps, visible
in the BSP after smp_started is written by the last started AP, which
formally requires acquire barrier on the load. The change is mostly
nop due to the ordered behaviour of the x86 CPUs.

Reviewed by: alc
Sponsored by: The FreeBSD Foundation
MFC after: 2 weeks