Found this while micro benchmarking Generational ZGC.

See how the following code uses __ATOMIC_RELEASE and FULL_MEM_BARRIER even when order == memory_order_relaxed:

template<size_t byte_size>
struct Atomic::PlatformAdd {
  template<typename D, typename I>
  D add_and_fetch(D volatile* dest, I add_value, atomic_memory_order order) const {
    D res = __atomic_add_fetch(dest, add_value, __ATOMIC_RELEASE);
    FULL_MEM_BARRIER;
    return res;
  }

This causes a noticeable slowdown in parts of the Generational ZGC code that walks over large C++ arrays, and claim indices/chunks with fetch_and_add.

If I change the code to be:
if (order == memory_order_relaxed) {
  return __atomic_add_fetch(dest, add_value, __ATOMIC_RELAXED);
} else {
  ... old code ...
}

then I see a significant enough speed up.