RFC: Accept embedded oops barriers to alleviate final mark time

[Roman Kennke]

The results are cool.

The patch is not trivial enough to quickly check for correctness, or
else hide everything behind a flag (at least not afaics) ;-) I'll take a
closer look next week, when I am back, if that is ok with you.

Thanks, and enjoy your vacation!

Roman

Am 16.06.2017 um 21:14 schrieb Aleksey Shipilev:
> Hi,
>
> This is my experimental patch for accepting barriers in the code cache, *and*
> avoiding the evacuation work at the pause. It is dirty, and maybe incorrect, but
> it passes hotspot_gc_shenandoah, and provides nice improvements for final mark
> pause time. The overheads for emitting the barrier for constants is hard to
> quantify before Roland fixes up the compiler bugs around it. Throwing the patch
> out for the record, while I am going on vacation :)
>
> Webrev:
>   http://cr.openjdk.java.net/~shade/shenandoah/codecache-barriers/webrev.01/
>
> GC stats before/after on some workloads:
>   http://cr.openjdk.java.net/~shade/shenandoah/codecache-barriers/perf.txt
>
> Init evac times, and sometimes final mark times are down to 100-500 us range!
>
> Thanks,
> -Aleksey
>
>