As specified here:
 http://mail.openjdk.java.net/pipermail/jmm-dev/2015-August/000223.html

I don't really like piggybacking on inlining policy to achieve the desired effect here. It's somewhat okay for scope-extending techniques a la reachabilityFence, since it is not supposed to be on a hotpath, but for (get|set)Opaque we need more performance, so MemBarCPUOrder for C2 it is.

Method call costs around 10 additional insns, at least because of JDK-8130398. Non-inlineable method is still a good fallback for C1, which does not have explicit compiler barriers. Breaking value numbering and similar optimizations is not really reliable, given how hard it is to test (get|set)Opaque.