Weak reference processing

[qinsoon]

[qinsoon]

#564 adds a functional implementation for weak reference. Here lists the performance results for the PR: #564 (comment)

Remaining work based on #564:

• Should we keep Java's weak reference semantics? (soft/weak/phantom)
  • If most languages just use weak reference and it is similar to Java, maybe we should make soft/phantom as an optional feature.
  • Java's semantics require references to be enqueue'd if their referent is cleared. Maybe we should make this an optional feature as well.
• Store weak refs in buffer during object tracing. This will speed up the process edges time shown in the performance results. It cuts the overhead from 5.4% down to 2.7%. *
• Investigate GC time slowdown. We have a 10% slowdown in GC time. Process edges only have a 5% slowdown. That means we have slowdown somewhere else. It could be the weak reference work itself. Adding extra work bucket has a slowdown of ~1%. *
• Investigate batik GC time (40% slowdown in GC time). Batik only does 1 GC in the measurement.

* Performance links are not publicly visible.