Reference processing efficiency problem #324
Description
After we enabled reference processing by default, we have observed performance problem with the current implementation of reference processing.
From the eBPF-generated timeline taken from lusearch executed with the GenImmix plan, we can see that during a nursery GC, the transitive closure time was very short. In comparison, reference processing took a long time, and was not well-parallelized.
One major problem is that the current algorithm requires soft and weak references to be re-scanned after the transitive closure has been expanded (soft reference and finalizer processing). This introduced extra GC phases, and is currently completely sequential.
The obvious solution is applying the optimizations mentioned in our porting guide, that is, merging soft reference processing with the strong closure or treating them as weak references, and merging weak reference processing with finalizer processing. This will handle soft references on the fly during non-emergency GC which can be well-parallelized, and will reduce the number of soft/weak/final/phantom reference processing phases from 4 to 2. See this section for more details.
This is related to #204, i.e. do reference processing in OpenJDK-specific ways.