RFR: Disable UseFastJNIAccessors for Shenandoah

[Roman Kennke]

Erik Österlund pointed out a nasty bug in the JNI fast-get-field code.
This code is executed in thread-native state, which means it doesn't
keep safepoints from happening. Between the forward pointer load and the
actual field load, the thread may get descheduled, and 1 or more
safepoints can happen, and the actual field access would crash. This
problem hasn't happened in all those years (at least not knowingly), but
*if* it happens, it's catastrophic.

See also:
http://mail.openjdk.java.net/pipermail/hotspot-dev/2018-June/032763.html

We may have a solution, but currently have more important fires to put out.


diff --git a/src/hotspot/share/gc/shenandoah/shenandoahArguments.cpp
b/src/hotspot/share/gc/shenandoah/shenandoahArguments.cpp
--- a/src/hotspot/share/gc/shenandoah/shenandoahArguments.cpp
+++ b/src/hotspot/share/gc/shenandoah/shenandoahArguments.cpp
@@ -172,6 +172,13 @@
                  "are observed on class-unloading sensitive workloads");
     FLAG_SET_DEFAULT(ClassUnloadingWithConcurrentMark, false);
   }
+
+  // JNI fast get field stuff is not currently supported by Shenandoah.
+  // It would introduce another heap memory access for reading the
forwarding
+  // pointer, which would have to be guarded by the signal handler
machinery.
+  // See:
+  //
http://mail.openjdk.java.net/pipermail/hotspot-dev/2018-June/032763.html
+  FLAG_SET_DEFAULT(UseFastJNIAccessors, false);
 }

 size_t ShenandoahArguments::conservative_max_heap_alignment() {