RFR(S): 8048179: Early reclaim of large objects that are referenced by a few objects

[Thomas Schatzl]

Hi Bengt,

On Wed, 2014-12-17 at 11:23 +0100, Bengt Rutisson wrote:
> 
> Hi Thomas,
> 
> New webrev looks good.
> 
> Two questions:
> 
> What do you think about changing 
> OtherRegionsTable::occupancy_less_or_equal_than() to be something like:
> 
> bool OtherRegionsTable::occupancy_less_or_equal_than(size_t limit) const {
>    if (limit <= (size_t)G1RSetSparseRegionEntries) {
>      return occ_coarse() == 0 && _first_all_fine_prts == NULL && 
> occ_sparse() <= limit;
>    } else {
>      Unimplemented();
>      // Current uses of this method only use values less than 
> G1RSetSparseRegionEntries for the limit.
>      // If we want to use it for larger values we need to implement the 
> more expensive iteration over
>      // the remembered set to check the limit.
>      return false;
>    }
> }
> 
> I think the Unimplemented() check communicates that it is this method 
> that is having a problem, not the caller of the method.
> 

I do not see either better than the other - in any case the VM will
crash with an error message. I changed it to this version anyway.

> The second question is about the flag name. I realize that 
> G1ReclaimDeadHumongousObjectsWithStaleRefsAtYoungGC is correct and that 
> it is in analogy with G1ReclaimDeadHumongousObjectsAtYoungGC. But it is 
> very long. Can we find some way of shortening it? I know it was me who 
> suggested to include "WithStaleRefs" in the name...
> 
> Maybe the flags could be called G1EagerReclaimOfHumObjs and 
> G1EagerReclaimOfHumObjsWithStaleRefs ?

I do not like abbreviations in flag names; Bengt and me discussed
offline and came up with the following new names for the flags:

G1EagerReclaimHumongousObjects
G1EagerReclaimHumongousObjectsWithStaleRefs
G1TraceEagerReclaimHumongousObjects

I.e. remove the "Dead" because the GC only ever reclaims dead objects
anyway, and remove the "AtYoungGC" because it seems somewhat superfluous
too: we can already reclaim humongous objects at all other GCs (marking,
full gcs), there would be nothing "eager" about it if it were only done
in these situations.

I am fine with adding "AtYoungGC" again if someone objects to that.

Any objections or alternatives?

Webrev:
http://cr.openjdk.java.net/~tschatzl/8048179/webrev.2/
Webrev diff:
http://cr.openjdk.java.net/~tschatzl/8048179/webrev.1_to_2/

Thanks,
  Thomas