In G1AllocRegion::attempt_allocation(size_t, size_t, size_t) it looks like the read to _alloc_region misses a load_acquire, as the subsequent G1HeapRegion::par_allocate() reads `_top` and `_end` from the `G1HeapRegion` instance. These may be a mix of the new values and the ones from the dummy region (or earlier regions?), and in case we get an `end()` < `top()`, g1 will allocate wrongly.

There is a storestore when doing the `G1AllocRegion::update_alloc_region()` call, but the reader side seems to miss synchronization.

There may be a similar argument to accesses to `MutatorAllocRegion::_retained_alloc_region`.

There are no known failures due to that (or assertion failures that would trigger in that case).

Does not affect x86.