JDK-8279016 shows there is a problem with at least Shenandoah when JFR is enabled. Arguably, it is not JFR's business to touch any mark words, since that might accidentally interfere with other subsystems, not only the GC, but also locking, identity hash codes, etc -- e.g. the uses that are described in markWord.hpp.

A separate data structure that tracks JFR-specific state would be preferred. This would be similar to e.g. the improvement to JVMTI: JDK-8283710. After we do this, we can revert parts of JDK-8279016.