RFR: Factor out storeval barrier from read barriers

[Roman Kennke]

Am 28.08.2017 um 16:47 schrieb Aleksey Shipilev:
> On 08/28/2017 04:44 PM, Roman Kennke wrote:
>> OK. But it is not independent in this patch: turning off RB would also turn off the SVB. (Or, with
>> upcoming partial GC, same with WBs). Would it be more useful to have them more independent?
> 
> I think it would be useful to separate. RB is our major overhead dominator, would be interesting to
> see if SVB is a significant part of it, and thus an appealing optimization target (as you said in
> the previous note).

Alright. I separated them by introducing _impl() methods that actually
implement the barriers (or, in the case of c2 and runtime, reuse
existing _impls), and have the entry points only check their flags and
then call _impl. This means we can now turn off read-barriers and still
get storeval barriers, or turn off storeval barriers and still get read
barriers.

Tested by: hotspot_gc_shenandoah.

http://cr.openjdk.java.net/~rkennke/storevalbarrier/webrev.01/

Ok?

Roman