1. The VMObjectAlloc provides the callbacks with a JNI ref to the object and its class. But if the object itself is a java.lang.Class then we assert in oop_to_klassOop.

2. Only Java-level objects should be reported to agent. This "visible oop" check is done at the time the event is recorded so that non-visible oops aren't recorded. This doesn't work in the case of classes because the mirror <--> klassOop references aren't setup leading to a crash or assertion in visible_oop(oop).

3. When an allocation from TLAB fails the remaining space in the TLAB is allocated with a dummary object and a new TLAB is allocated. The allocation of this dummy object shouldn't be reported to the agent. To fix we stack allocated/marker class to disable the reporting of these allocations.

Related to issue 3 is that the GC always fills the TLABs before collection. As this happens in the VM thread it means there isn't a JvmtiThreadState so JvmtiExport::post_vm_internal_object asserts when calling JavaThread::current().
 
###@###.### 2004-01-26