RFR: 8314972: GenShen: promote_in_place needs to prepare remembered set before it enables old allocations within region

[Kelvin Nilsen]

On Fri, 25 Aug 2023 17:53:06 GMT, Kelvin Nilsen <kdnilsen at openjdk.org> wrote:

>> src/hotspot/share/gc/shenandoah/shenandoahHeapRegion.cpp line 1040:
>> 
>>> 1038:       assert(fill_size >= ShenandoahHeap::min_fill_size(), "previously allocated objects known to be larger than min_size");
>>> 1039:       ShenandoahHeap::fill_with_object(obj_addr, fill_size);
>>> 1040:       heap->card_scan()->register_object_without_lock(obj_addr);
>> 
>> On second thoughts, do we really need to register the dead objects? Does it help at all, since the preceding and succeeding live objects will be registered anyway. Not wrong, but I do wonder if it helps at all or not. Probably not a big deal either way since dead objects are likely very few in regions promoted in place.
>
> Existing code (and some existing assertions in ShenandoahVerifier) expect the filler objects to be registered.  You make a good point though.  We don't really need these to be registered.  If the first or only object in a card range is a filler, remembered set scanning can just ignore that.  The remembered set scanner wants to find the first non-filler object in the range.  We might want to address this in a separate ticket later, if we decide to keep the current remembered set representation longer term.

Note that the cost of registering filler objects is minimal because there are never multiple contiguous filler objects residing within memory.  We always coalesce them together.  At most, the cost during remembered set scanning is that we look at one extra filler object per DIRTY card.

-------------

PR Review Comment: https://git.openjdk.org/shenandoah/pull/310#discussion_r1305982797