RFR: 8327097: GenShen: Align PLAB sizes down rather than up

[Y. Srinivas Ramakrishna]

On Wed, 6 Mar 2024 20:43:03 GMT, Y. Srinivas Ramakrishna <ysr at openjdk.org> wrote:

>> When adjusting LAB sizes, round down rather than rounding up.  Otherwise, we may violate the ShenandoahHumongousThreshold bound.
>
> src/hotspot/share/gc/shenandoah/shenandoahHeap.cpp line 1114:
> 
>> 1112:   future_size = MIN2(future_size, PLAB::max_size());
>> 1113:   future_size = MAX2(future_size, PLAB::min_size());
>> 1114:   future_size = align_down(future_size, CardTable::card_size_in_words());
> 
> Can one use cardsize-aligned min and max sizes (taking whetever PLAB gives us, may be by using a ShenandoahPLAB subclass). We then first align future_size up wrt card size, then floor/clamp it with the already cardsize-aligned min and max?
> 
> In particular if cur_size starts off a multiple of cardsize then any doubling keeps it a multiple of cardsize, and you can more simply perhaps just assert that it's always  a cardsize-multiple.

We also need a comment somewhere stating why we need these to be card-aligned.

I wonder also if this is entirely correct in itself, unless the shared allocation itself also always leaves us with a card-aligned top from which these PLABs are allocated.

In particular, it would be good to assert that `allocate_new_plab()` always gives us back a card-aligned start address.

Would be good to also use `align_up` or `align_down` in `allocate_new_plab()` as well. Indeed, it might be a good idea to do so in any code in Shenandoah where we are unnecessarily rolling our own custom alignment code.

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

PR Review Comment: https://git.openjdk.org/shenandoah/pull/401#discussion_r1517018057