Optimized Reference ArrayCopy on X86

[0dvictor]

Reference ArrayCopy has been optimized that reduces the code-path length,
and potentially improves performance.

Part of Issue: #3054

Signed-off-by: Victor Ding dvictor@ca.ibm.com

[andrewcraik]

@0dvictor can you please expand on what has changed and the performance testing this has received please?

[0dvictor]

This PR is based on OMR's arraycopy evaluator and OpenJ9's own "VMarrayStoreCheckArrayCopyEvaluator".
The three changes are:

1. J9 takes over the logic related to reference array copy even when store check is not required;
2. Call reference array copy helper when store check is not required but Concurrent Scavenge is active;
3. Create a JIT helper wrapper of GC's reference array copy to reduce JIT code register pressure.

[0dvictor]

Performance impact were analyzed in two ways:

1. IBM's internal benchmarks showed no performance impact;
2. Generate assemblies are verified to be identical as what OMR generates under GC policy GenCon, Concurrent Scavenge disabled, when array store check is not required.

[andrewcraik]

Jenkins test sanity xlinux,win jdk8,jdk11

[andrewcraik]

LGTM

[gacholio]

@andrewcraik I don't like the fact this was merged without involving me. I would certainly have rejected the formatting in cnathelp.cpp and xnathelp.m4. The C code also contains unnecessary (though not strictly incorrect) code. Much like the new CS code, I would likely have insisted on a more sensible way of dealing with the differences in calling conventions.

[0dvictor]

@gacholio I apologize for forget to let you know. I've created #3112 base on your comments on the CS code. Could you take a look?

[andrewcraik]

@gacholio my sincere apologies - I was reading for correctness and did not notice those changes were in files you normally review - @0dvictor thank you for opening #3112.