At this stage, I simply set OpenJDK::VMEdge to Address so that it does not change the behaviour.

Microbenchmark shows the performance impact is negligible.

Benchmark: Create a 22-level binary search tree, run System.gc() 10 times for warmup, and run System.gc() 100 times to record the GC time. Repeat the experiment for 5 times, alternating between two builds (master and this PR) and two plans (SemiSpace and Immix). Collect all 500 GC times for each build+plan combination and see their distributions. The code is available at https://github.com/wks/tracetest

It is run on elk.moma. The command is: MMTK_THREADS=1 RUST_BACKTRACE=1 MMTK_PLAN=$plan /path/to/java -XX:+UseThirdPartyHeap -server -XX:-TieredCompilation -Xcomp -XX:MetaspaceSize=100M -Xm{s,x}500M TraceTest $depth $gcs $warmups.

Distribution:

Given that there is no performance degradation, I'll continue with other bindings. I'll also confirm the performance results by running DaCapo benchmarks.

And the results on DaCapo Chopin. Run on elk.moma, 20 iterations.

Results for SemiSpace (on squirrel.moma)

Results for Immix (on squirrel.moma)

The mean STW times of most benchmarks are less than 1% different from master. Batik has huge error bars because GC only happened on average 3 times for SemiSpace and 1.6 times for Immix. H2 is a bit strange because in the 20 invocations, only 4 has results for SemiSpace and 2 has results for Immix, and I think the rest of the invocations failed silently for some reasons.

STW time distribution: