The old implementation was not correct on complicated trees. If a tree was a bit heterogeneous, then the parent-child reduction would take increase the result of one sub-tree to the starting count of the next sub-tree. Thus having higher blocking than what is supposed to be.