The unlocking code generated in a synchronized native wrapper will
use an oop that is in a register across the jni call. This means
that the register can be missed by gc because of the infamous
register-window window that exists during native transitions.


###@###.### 2004-02-02

NSK tests:

runtime/threads/PipedThreadsStressTest
runtime/threads/ThreadsUsePipedIO

fail with

# Internal Error (/nightly/ws/src/share/vm/runtime/synchronizer.cpp, 107), pid=963, tid=13326
#
# Java VM: Java HotSpot(TM) 64-Bit Server VM (1.5-nightly_2004.01.31-debug mixed mode)
#
# Error: assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()),"monitor is invalid")

and

# Internal Error (/nightly/ws/src/share/vm/runtime/synchronizer.cpp, 164), pid=6344, tid=17426
#
# Java VM: Java HotSpot(TM) 64-Bit Server VM (1.5-nightly_2004.01.31-debug mixed mode)
#
# Error: assert(test != __null,"Unexpected object header as 0!")

with this bug.

###@###.### 2004-02-05
atg failed on Itanium 2 RedHat AS3 ( test machine jtg-it7.sfbay) after 6 hours
in b37 bigapps run. Steve Goldman thought the failure could posisble
be bug 4954584.