A couple of benchmarks got obvious performance loss about 5~6% with changeset 55956, the problem was initially reproduced at Ampere eMAG aarch64 system, later on the similar loss can be seen on a x86 system as well.

https://bugs.openjdk.java.net/browse/JDK-8225670
https://hg.openjdk.java.net/jdk/jdk/rev/155b084cf384

On the aarch64 system, a few GC logs showed "Promotion failed" and 1-2 ~50s long pauses, which is very abnormal. A simple and straight-forward change is made as an experiment below, limiting the clear_row(row) with the (k!=NULL) condition, so the profiled data without a valid receiver stored in the original ReceiverTypeData would not be destroyed accidentally. In this way, the performance can be recovered (please note this is not a fix).

src/hotspot/share/ci/ciMethodData.cpp
if (k != NULL && k->is_loader_alive()) {
  ciKlass* klass = CURRENT_ENV->get_klass(k);
  set_receiver(row, klass);
} else {
// With concurrent class unloading, the MDO could have stale metadata; override it
+ if (k != NULL)
  clear_row(row);

Another experiment is, setting -XX:TypeProfileWidth=0 could help confirm the issue can be related with this function, since the performance diffs are gone if we did so.