RFR: Avoid indirection to next-mark-context

[Roman Kennke]

Am 17.07.2018 um 22:53 schrieb Aleksey Shipilev:
> On 07/17/2018 10:26 PM, Roman Kennke wrote:
>> On top of the ShenandoahMarkingContext patch, this change avoids the
>> extra indirection that we always need to take in hot marking path to get
>> to the next-mark-bitmap and next-TAMS-structures. It does so by picking
>> the next-marking-context at the beginning of the phase, and then keep it
>> around in the closure.
>>
>> Testing: tier3_gc_shenandoah ok
>>
>> http://cr.openjdk.java.net/~rkennke/direct-marking/webrev.00/
> 
> Looks good. Does it actually help marking performance?

My little bit of perf testing was inconclusive. If it helps, then not by
much. I think the marking/traversal loops are drowned by fwd ptr
reads+stalls.

Will fix the rest and RFR tomorrow.

Cheers,
Roman

> Stylistic:
> 
>  *) It seems odd that in these blocks two class members are on the same line:
> 
>   65   ShenandoahInitMarkRootsClosure(ShenandoahObjToScanQueue* q) :
>   66     _queue(q), _heap(ShenandoahHeap::heap()),
>   67     _mark_context(_heap->next_marking_context()) {};
> 
> ...but here, they are split:
> 
> 
>   73 ShenandoahMarkRefsSuperClosure::ShenandoahMarkRefsSuperClosure(...) :
>   74   MetadataVisitingOopIterateClosure(rp),
>   75   _queue(q),
>   76   _heap(ShenandoahHeap::heap()),
>   77   _mark_context(_heap->next_marking_context())
>   78 { }
> 
>  Pick a style and stick to it?
> 
> Thanks,
> -Aleksey
>