While working on apps with a large metaspace size (over 100k classes for example) I saw some regressions that I eventually tracked back to JDK-8338526.
I wrote a JMH to show the problem based on the existing vm/compiler/CodeCacheStress.java which I pushed on a branch here:
https://github.com/ericcaspole/jdk/blob/JDK-8338526/test/micro/org/openjdk/bench/vm/runtime/MethodHandleStress.java
This JMH can probably simplified and still show the problem, but it's a start.
On our aurora harness it shows 4-18% regressions depending on the platform.