GC code uses CAS to exchange values to avoid real locks in a few places.

This must be done with care, as compilers must be prevented from putting local variables into registers using volatile specifiers. Issues like JDK-8033545 frequently pop up.

Look through (GC) code that uses these Atomic instructions (Atomic::cmpxchg_ptr, Atomic::cmpxchg) and make sure there might not be problems due to valid compiler optimizations.

A first look identified at least the following potential places for such issues:

void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr) rereading _dirty_cards_region_list
HeapRegion* G1CollectedHeap::pop_dirty_cards_region(), rereading _dirty_cards_region_list
PerRegionTable::bulk_free, rereading _free_list
PerRegionTable*::alloc(HeapRegion* hr), rereading _free_list
mutableNUMASpace.cpp: 850, rereading top() [end()]
MutableSpace::cas_allocate(size_t size), rereading top(), end()
DefNewGeneration::allocate()
ContiguousSpace::par_allocate_impl() ,rereading top(), end()
ConcEdenSpace::par_allocate(), probably the OrderAccess::loadload() helps?

runtime.cpp:1348: head = _named_counters;