CAS is slow because it acts as a serializing instruction. If you depend on success or fail the instructions after it have to wait until it has been resolved. Also, the update has to be pushed out to the cache at least.

So yes, slow but otherwise it would be of no use.

load_linked and store_conditional are not much better. Depending on implementation they can spuriously fail because something poked the cache, there was memory traffic etc.

How can one implement a lock on x86 without cas/ll-sc? Lamport's Bakery?

so zen: cas slow, locks slower.

CAS on platforms which provide CAS for lock-free (x86/x64), LL/SC on platforms which provide LL/SC for lock-free (ARM).

AFAIK, platforms provide one or the other; not both. So you have to use what you're given.

The java JDK lockless queue is also compare and swap. I don't know the reasoning behind it though.