In atomic_bsd_zero.hpp and atomic_linux_zero.hpp there are uses of __sync_synchronize(). However, orderAccess_*_zero.hpp calls the kernel helper, because:

/*
 * ARM Kernel helper for memory barrier.
 * Using __asm __volatile ("":::"memory") does not work reliable on ARM
 * and gcc __sync_synchronize(); implementation does not use the kernel
 * helper for all gcc versions so it is unreliable to use as well.
 */

We need to clean this up to use OrderAccess::fence().