4 ms·

There are already specialized instructions in the Apple Silicon chips. IIRC there's something tailored for the Objective-C runtime, and something useful for Jav

• favorited 2y ago
  Uncontended acquire-release atomic operations are basically free on Apple Silicon, which synergizes with the Objective-C (and Swift!) runtimes, where every retain/release is an atomic increment/decrement. https://web.archive.org/web/20201119143547/https://twitter.com/Catfish_Man/status/1326238434235568128 https://web.archive.org/web/20201119143547/https://twitter.c...
  • throwaway2037 2y ago
    > Uncontended acquire-release atomic operations are basically free on Apple Silicon While I don't doubt you, the poster, specifically, how is this possible? To be clear, my brain is x86-wired, not ARM-wired, so I may have some things wrong. Most of the expense of atomic inc/dec is "happens before", which essentially says before the current core reads that memory address, it will be guaranteed to be updated to the latest shared value. How can this be avoided? Or is it not avoided, but just much, much faster than x86? If the shared value was updated in a different core, some not-significant CPU cycles are required to update L1 cache on current current with latest shared value.
    • throwaway2037 2y ago
      EDIT: > some not-significant CPU cycles should say: > some not-insignificant CPU cycles