6994628: G1: Test gc/gctests/FinalizeTest05 fails (one live object is finalized)
Summary: The Solaris Studio 12 update 1 C++ compiler was incorrectly re-ordering the reads of an object's mark word in oopDesc::forward_to_atomic(). This opened a small window where one thread could execute the successful CAS path even though another thread had already successfully forwarded the object. This could result in an object being copied twice. The code in oopDesc::forward_to_atomic() was changed to read the mark word once.
Reviewed-by: ysr, tonyp
--- a/hotspot/src/share/vm/oops/oop.pcgc.inline.hpp Thu Dec 02 13:20:39 2010 -0500
+++ b/hotspot/src/share/vm/oops/oop.pcgc.inline.hpp Tue Dec 07 16:18:45 2010 -0800
@@ -118,12 +118,15 @@
assert(forwardPtrMark->decode_pointer() == p, "encoding must be reversable");
assert(sizeof(markOop) == sizeof(intptr_t), "CAS below requires this.");
- while (!is_forwarded()) {
+ while (!oldMark->is_marked()) {
curMark = (markOop)Atomic::cmpxchg_ptr(forwardPtrMark, &_mark, oldMark);
+ assert(is_forwarded(), "object should have been forwarded");
if (curMark == oldMark) {
- assert(is_forwarded(), "the CAS should have succeeded.");
return NULL;
}
+ // If the CAS was unsuccessful then curMark->is_marked()
+ // should return true as another thread has CAS'd in another
+ // forwarding pointer.
oldMark = curMark;
}
return forwardee();