For the performance improvements, I think I'm missing something but why does coalesced reference counting have a high overhead? Is it normally implemented alon

For the performance improvements, I think I'm missing something but why does coalesced reference counting have a high overhead? Is it normally implemented along with buffered reference counting? It feels like those fit together very neatly, one thread managing the counts and receiving updates from the other threads, and each other thread tries to only send updates it needs to. Is it simply a case of doing something basic a lot of times is faster than doing something smarter a few times because computers are just really fast at basic things? Or is there something more to this?