JNI_CreateJavaVM_inner calls create_vm and examines the result. If it indicates an error (not JNI_OK) then it checks for a pending exception and reports that. That check requires the current thread. However, in some cases create_vm deletes the main thread before returning an error code. As a result, the attempt to get and use current thread will crash (either an assert or a null pointer access).

This crash can be reproduced fairly easily by patching create_vm's handling of the result of init_globals(), using -XX:+UseNewCode to force an error return. I tripped over it while testing a fix for JDK-8240774.