This code in ReferenceProcessor::discover_reference enqueues a reference during discovery looks like this:

     if (_discovery_is_mt) {
[1] add_to_discovered_list_mt(*list, obj, discovered_addr);
     } else {
[2] // We do a raw store here: the field will be visited later when processing
       // the discovered references.
       oop current_head = list->head();
       // The last ref must have its discovered field pointing to itself.
       oop next_discovered = (current_head != NULL) ? current_head : obj;

       assert(discovered == NULL, "control point invariant");
       RawAccess<>::oop_store(discovered_addr, next_discovered);
       list->set_head(obj);
       list->inc_length(1);

       log_develop_trace(gc, ref)("Discovered reference (" INTPTR_FORMAT ": %s)", p2i(obj), obj->klass()->internal_name());
     }

[2] argues that there is no need for a barrier because we visit this later when processing the discovered references *during gc*.

However, some G1 code sets _discovery_is_mt to false for concurrent precleaning (in G1ConcurrentMark::preclean()). This causes missing barriers in G1.

[1] executes a barrier always, just because we are using multiple threads, even during gc when it is not required.

This adds overhead for parallel gc and G1 during GC.

I.e. there is a confusion between "atomicity" and "degree of parallelism" in this code.