G1 ConcurrentG1RefineThread::stop delays JVM shutdown for >150ms

G1 introduces >150 ms shutdown lag, allegedly because of Terminator_lock wait in ConcurrentG1RefineThread::stop. This can be demonstrated with a trivial HelloWorld application and G1 instrumentation:
  http://cr.openjdk.java.net/~shade/8136854/Hello.java
  http://cr.openjdk.java.net/~shade/8136854/tracing.patch

Timing on 1x4x2 i7-4790K, Linux x86_64, jdk9/dev at Sep 21, 2015.

$ time java -XX:+UseParallelGC Hello
Hello World

real 0m0.030s
user 0m0.036s
sys 0m0.000s

$ time java -XX:+UseConcMarkSweepGC Hello
Hello World

real 0m0.032s
user 0m0.028s
sys 0m0.008s

$ time java Hello
Hello World
ConcurrentG1RefineThread waited for Terminator_lock for 168 msec
ConcurrentG1RefineThread waited for Terminator_lock for 0 msec
ConcurrentG1RefineThread waited for Terminator_lock for 0 msec
ConcurrentG1RefineThread waited for Terminator_lock for 0 msec
ConcurrentG1RefineThread waited for Terminator_lock for 0 msec
ConcurrentG1RefineThread waited for Terminator_lock for 0 msec
ConcurrentG1RefineThread waited for Terminator_lock for 0 msec
ConcurrentG1RefineThread waited for Terminator_lock for 0 msec
ConcurrentG1RefineThread waited for Terminator_lock for 0 msec

real 0m0.208s
user 0m0.040s
sys 0m0.004s

We can clearly see that waiting on Terminator_lock wastes ~150msec. This issue blocks the development of startup performance-sensitive JDK features, as the startup overheads drown in this G1 delay.