Switch to optimized lock

[terrarier2111]

Based on #57
Please squash this, thanks!

[Amanieu]

Could you explain why this is optimized? Your implementation is just a spinlock. I don't see any need to change this code path since it is rarely called (only once per thread).

[Amanieu]

And as a general rule, I am not a fan of spinlocks and try to avoid them as much as possible. They should only be used in an environment where the current thread will not be preempted (i.e. in a kernel with interrupts disabled, or small embedded code). The problem is that if a thread get scheduled out while holding a spinlock, other threads can spend a long time needlessly spinning instead of making the OS switch back to the thread owning the lock.

[terrarier2111]

Could you explain why this is optimized? Your implementation is just a spinlock. I don't see any need to change this code path since it is rarely called (only once per thread).

Sorry for the long delay, my thinking process was that since the codepath is so rarely called, it would make sense to optimize exactly for that (that there aren't multiple threads in this code path at once) to save a few cpu cycles (without impacting any fast path) since you are apparently very opposed to spin locks i would be okay with removing it but keeping the other changes since they eliminate calls to the get_tls_addr function in the fast path which can be quite expensive. I am doing this by tricking the compiler into addressing into the Thread struct through the pointer to it which is being stored right beside it in the ThreadWrapper struct, so the compiler doesn't understand that the pointer points to data right beside it and instead of recalculating the address to Thread through get_tls_addr it just follows the pointer, so we only need 1 tls access per get call which is quite nice. Why the heck this is necessary and the compiler isn't just getting rid of this access by itself, that i don't know.

[Amanieu]

What is the purpose of self_ptr? It seems to always point to the thread field. Is this just a bool that indicates whether the thread has been initialized? How is this better than just using an Option?

[terrarier2111]

That is what i was describing earlier, this is the hack that tricks the compiler into actually only ever generating one get_tls_addr call and reusing the calculated address, if the compiler was smart enough, this self_ptr wouldn't be necessary at all.