Current code in G1BarrierSet::write_ref_field_post() seems racy:

inline void G1BarrierSet::write_ref_field_post(T* field) {
  volatile CardValue* byte = _card_table->byte_for(field);
  if (*byte == G1CardTable::clean_card_val()) {
    *byte = G1CardTable::dirty_card_val();
  }
}

That attempts to make sure that _card_table is not potentially read twice, because if it were, we might get a card mark in the wrong card table, potentially causing crashes.

Could be the cause for JDK-8373944, at least it would show the same symptom.