We improve on 2 things:
1. Only ARC HS family has "dmb" instructions so do compile-time
check for automatically defined macro __ARCHS__.
Previous check for ARCv2 ISA was not good enough because ARC EM
family is v2 ISA as well but still "dmb" instaruction is not
supported in EM family.
2. Still if there's no dedicated instruction for memory barrier
let's at least insert compile-time barrier to make sure
compiler deosn't reorder critical memory operations.