I was looking at profiles running JDK tests under fastdebug, and noticed we are spending quite a bit of time in walking the CodeCache looking for deoptimization candidates. See attached screenshot from langtools:tier1 run.

I suspect this is a variant of JDK-8325613, but for jtreg AgentVMs. As Agent VMs keep alived, they apparently slowly accumulate the cruft from the previous tests. I think we need to see if we need to clean up.