A customer has a reproducible crash and I have a fix which is working for them. Unfortunately so far I do not have a reproducer/testcase. The crash+fix is on jdk21 but I haven't seen any changes in trunk in that code.

The hs_err backtrace is:

ConstantPool::klass_at_impl(constantPoolHandle const&, int, JavaThread*)+0x8d
        Handle loader (THREAD, this_cp->pool_holder()->class_loader());
Reflection::check_for_inner_class(InstanceKlass const*, InstanceKlass const*, bool, JavaThread*)+0x2a0
        outer=ClassName
        inner=ClassName$Inner
        caller: Klass* i = cp->klass_at(ioff, CHECK);
InstanceKlass::compute_enclosing_class(bool*, JavaThread*) const+0x14a
JVM_GetDeclaringClass+0x7f

It crashes at:

https://github.com/openjdk/jdk21u-dev/blob/04046b185003a99e93905d54964b7018ae892147/src/hotspot/share/oops/constantPool.cpp#L528
Handle loader (THREAD, this_cp->pool_holder()->class_loader());

It crashes inside class_loader():

https://github.com/openjdk/jdk21u-dev/blob/04046b185003a99e93905d54964b7018ae892147/src/hotspot/share/oops/klass.inline.hpp#L68
inline oop Klass::class_loader() const {
  return class_loader_data()->class_loader();
And class_loader_data() returns NULL here.