RFR (M): Rehash barrier disable flags, add Shenandoah{CAS,Acmp,Clone}Barrier flags

[Roman Kennke]

Even better! Yep!


Am 20.02.2017 um 17:03 schrieb Aleksey Shipilev:
> Forgot to handle SATB pre-barrier. Also store-val barriers should be handled by
> ShenandoahWB, not RB. Fixed both:
>   http://cr.openjdk.java.net/~shade/shenandoah/barrier-disable/webrev.02/
>
> Tests still pass.
>
> Thanks,
> -Aleksey
>
> On 02/20/2017 03:25 PM, Roman Kennke wrote:
>> Yes! Great!
>>
>>
>> Am 20.02.2017 um 15:12 schrieb Aleksey Shipilev:
>>> Hi,
>>>
>>> Running with "passive" heuristics that frees us from requiring the operational
>>> runtime barriers, and -XX:Shenandoah{Read,Write}Barriers are useful to dissect
>>> runtime performance. However, those are handled inconsistently in the code, and
>>> also we miss other three flavors of barriers: CAS, acmp, and clone barriers.'
>>>
>>> This change rehashes the handling for these barrier disable flags:
>>>   http://cr.openjdk.java.net/~shade/shenandoah/barrier-disable/webrev.01/
>>>
>>> Disabling all barriers brings runtime performance very close to ParallelGC,
>>> which seems to validate the approach.
>>>
>>> Testing: hotspot_gc_shenandoah (x86), eyeballing assembly (x86), targeted
>>> benchmarks (x86).
>>>
>>> Thanks,
>>> -Aleksey
>>>
>>