RFR: Avoid evacuating filler objects in Shenandoah

[Zhengyu Gu]

On 08/30/2017 05:13 PM, Roman Kennke wrote:
> Hi Zhengyu,
> 
>>
>> Webrev:
>> http://cr.openjdk.java.net/~zgu/shenandoah/sh_filler/webrev.00/index.html
>>
>> Thanks,
>>
>> -Zhengyu
> 
> Is the motivation for this change that we can check that the objects we
> are evacuating are not filler objects?

Aleksey found that 5-10% evacuated objects are fillers. So, it is high 
cost to evacuate them, and also increases garbage in to space.

> 
> I don't really understand what you mean by:
> 
> 
>> Using fake filler array klass seems to have less side-effects,
>> especially on upstream collectors.
>>

Creating fillerArrayKlassObject (as currently implemented) is pretty 
effortless. However, creating fillerKlassObject will need corresponding 
Java class definition, which will pollute JDK.

>> The drawback is that, we still evacuate small object fillers, since fake
>> filler object is not easy.

>>
> 
> I'm probably missing some context ;-)

Another option was using invalid mark word for filler objects, which 
works well with object fillers and array fillers, and it works well with 
Shenandoah.

However, other collectors, e.g. g1, bark on invalid mark words, which we 
have to fix. Also, Aleksey pointed out that invalid mark words can hide 
some synchronization bugs or confuse people while debugging.

-Zhengyu


> 
> Roman
>