[master] RFR: 8346011: [Lilliput] Compact Full-GC Forwarding [v3]

[Thomas Stuefe]

On Tue, 21 Jan 2025 11:41:47 GMT, Roman Kennke <rkennke at openjdk.org> wrote:

>> The current forwarding scheme that is used during full-GC overrides much of the mark-word. This is a problem for current compact headers when the heap is larger than 8TB (in which case we currently disable compact headers altogether) and becomes a much worse issue with 4-byte-compact-headers, because it would override the compressed class-pointer in the upper header bits.
>> 
>> This implementation uses a side-table (currently 1/512th of the heap size) to store part of the target addresses, and 9 bits in the header for the rest of it. For the full description of the algorithm, see top of fullGCForwarding.hpp.
>> 
>> Some performance testing results: https://gist.github.com/rkennke/5a53d21337fc6e696041062d6b972dd6
>> Interpretation: The new full-GC forwarding is slightly slower across the board. The different is almost completely a constant ~10ms. This corresponds to the setup costs of allocating and clearing the side-table. That cost could be reduced by allocating and clearing the table up-front, when the GC gets initialized, and then only ever clearing the table when full-GC is finished. That brings down the numbers to almost exactly baseline level, but I think we wouldn't want to hold on to that memory all the time, especially not in G1 and Shenandoah, where full-GC is an exceptional mode that is not intended to run frequently.
>> 
>> Testing:
>>  - [x] hotspot_gc
>>  - [x] tier1
>>  - [x] tier1 +UseSerialGC
>>  - [x] tier1 +UseParallelGC
>>  - [x] tier1 +UseShenandoahGC
>
> Roman Kennke has refreshed the contents of this pull request, and previous commits have been removed. The incremental views will show differences compared to the previous content of the PR. The pull request contains one new commit since the last revision:
> 
>   8346011: [Lilliput] Compact Full-GC Forwarding

Looks reasonable.

src/hotspot/share/gc/shared/fullGCForwarding.hpp line 140:

> 138:   // block.
> 139:   static const uintptr_t FALLBACK_PATTERN = right_n_bits(NUM_OFFSET_BITS);
> 140:   static const uintptr_t FALLBACK_PATTERN_IN_PLACE = FALLBACK_PATTERN << OFFSET_BITS_SHIFT;

I would make all constants constexpr

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

Marked as reviewed by stuefe (Reviewer).

PR Review: https://git.openjdk.org/lilliput/pull/191#pullrequestreview-2564414971
PR Review Comment: https://git.openjdk.org/lilliput/pull/191#discussion_r1923603764