12 RFR(M) 8195635: [Graal] nsk/jvmti/unit/ForceEarlyReturn/earlyretbase crashes with assertion "compilation level out of bounds"

[dean.long at oracle.com]

http://cr.openjdk.java.net/~dlong/8195635/webrev.5/
https://bugs.openjdk.java.net/browse/JDK-8195635

Please see the bug report for all the gory details.  Here's the short 
version:

If we allow any safepoint to be a suspend point, we run into trouble 
with PopFrame and ForceEarlyReturn, which reasonably expect the top 
frame not to change between the suspend and when the 
PopFrame/ForceEarlyReturn is executed.  Normally this is not an issue, 
but certain safepoints cause problems, when we are about to call a new 
Java method.  In particular, if we safepoint and suspend in 
JavaCallWrapper, the top frame will still be the caller, but when we 
execute the PopFrame/ForceEarlyReturn we will be in the callee.

The solution this patch takes is to block suspend around troublesome VM 
code using a new "allow_suspend" thread flag.  This means 
JavaThread::java_suspend can't just ask the VMThread to safepoint and be 
done.  Instead it has wait and allow threads to roll forward to an 
allowed suspend point.

dl