8261649: AArch64: Optimize LSE atomics in C++ code

[Andrew Haley]

[Fix: subject line]

This is the second part of a pair with JDK-8261027.

It generates near-optimal LSE instruction sequences for C++
atomic functions. The patch from JDK head applied with no changes,
but there's a webrev at http://cr.openjdk.java.net/~aph/8261649-jdk11u/.

We need this in 11u for performance reasons on new AArch64 hardware,
particularly chips like the Apple M1 and Arm's Neoverse N1.

-- 
Andrew Haley  (he/him)
Java Platform Lead Engineer
Red Hat UK Ltd. <https://www.redhat.com>
https://keybase.io/andrewhaley
EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671