fetch_or instead of cas-loop in mutex code

[MBkkt]

I noticed in that commit function AtomicSetBits and thought it can be better written like this, because on armv8 fetch_or is "single" instruction

    intptr_t v;
    do {
        v = pv->load(std::memory_order_relaxed);
        if ((v & bits) != 0) {
            return false;
        }
    } while ((v & wait_until_clear) != 0);
    return (pv->fetch_or(bits, std::memory_order_release) & bits) == 0;

It's not the same logic (wait_until_clear not checked after achieved), but it seems like it doesn't matter and also can be used.

From other side clang on x86 cannot put condition with early-exit in the "fetch_or" cas-loop

[derekmauro]

As you said, this isn't quite the same thing. I'm not sure that this is worth optimizing since it is only used in a debug path which is already going to be slow.

/cc @dvyukov

[MBkkt]

Is it debug only? Ok, I didn't notice

[dvyukov]

} while ((v & wait_until_clear) != 0);

return (pv->fetch_or(bits, std::memory_order_release) & bits) == 0;

This looks broken. The flag may be set in between the check and the fetch_or.

[MBkkt]

I thought wait_until_clear cannot be set after it was cleared.
If it's not true, ofc, thanks