In the course of working on JDK-8365493 and JDK-8370848, we noticed that there is a ~20% regression on Dacapo sunflow --size large on the OCI A1.160 shape. On this shape we run the jobs on node 0 so this one runs with 80 threads.

baseline:
-Xms8g -Xmx8g -XX:+UseLargePages -XX:+UnlockDiagnosticVMOptions -XX:+UseObjectMonitorTable

test:
-Xms8g -Xmx8g -XX:+UseLargePages -XX:+UnlockDiagnosticVMOptions -XX:+UseCompactObjectHeaders

Using Dacapo flags:
--size large --iterations 4 sunflow

AFAICT this affects G1 more than ParGC but I will update if I find otherwise.
I have not tried to find the earliest version where this started.