He fails to mention that Apple added support for ref counting in silicon. And.. often GC will be able to use area allocators, before falling back to "proper" G

He fails to mention that Apple added support for ref counting in silicon. And.. often GC will be able to use area allocators, before falling back to "proper" GC allocation. Which will be a lot faster than ref counting everything. And atomics can get very slow, I've had atmics show up regularly in the profiler. For my project, the combination that works great so far: unbox all types, use area allocators if the compiler can guarantee the value doesn't escape, use GC for data that changes often and ref counting for data that hardly ever changes. (luckily cycles are not possible)