> And old code has
>
> while (unlikely(InterlockedExchangePointer((PVOID*)lk, 0) == 0))
> YieldProcessor();
>
> which effects a huge amount of cache pressure. This can be much better:
>
> while (unlikely(InterlockedExchangePointer((PVOID*)lk, 0) == 0))
> for (int i = 0; i != 100; ++i)
> YieldProcessor();
>
> The magic number 100 is, of course, an arbitrary choice. I believe you can
> find a better one basing on
> benchmark results on your hardware.
Usually spin locks read the value without fences, then attempt a swap when the
retrieved value is "good". This ensures that cache lines aren't locked in
exclusive mode for no reason.
retry:
while (ReadPointerNoFence((PVOID*)lk) == 0))
YieldProcessor();
If (InterlockedExchangePointer((PVOID*)lk, 0) == 0)
goto retry;
We could implement ReadPointerNoFence in winnt.h, however for the time being a
volatile access will work just fine:
retry:
while (*(void * volatile *)lk == 0)
YieldProcessor();
If (InterlockedExchangePointer((PVOID*)lk, 0) == 0)
goto retry;
Best Regards,
Luca
________________________________
Da: Kirill Makurin <[email protected]>
Inviato: lunedì 11 maggio 2026 16:25
A: [email protected]
<[email protected]>
Oggetto: Re: [Mingw-w64-public] winpthreads: fix XFAIL tests and provide new
spinlock implementation
> I think this spinlock implementation is over-complicated. Since
> `pthread_spinlock_t` is just an integer,
> you can store the ID of the owner thread in there.
I wanted to finalize and send these patches on Saturday, but other things
kicked in. Earlier today while double-checking everything, I realized that we
really just can store thread ID in `pthread_spinlock_t`. Anyway, I sent the
patches with implementation that I had ready. I'll send updated implementation
in the coming days.
> And old code has
>
> while (unlikely(InterlockedExchangePointer((PVOID*)lk, 0) == 0))
> YieldProcessor();
>
> which effects a huge amount of cache pressure. This can be much better:
>
> while (unlikely(InterlockedExchangePointer((PVOID*)lk, 0) == 0))
> for (int i = 0; i != 100; ++i)
> YieldProcessor();
>
> The magic number 100 is, of course, an arbitrary choice. I believe you can
> find a better one basing on
> benchmark results on your hardware.
I'm not sure about it... I feel like number which gives good results for one
hardware may give completely different results on another.
- Kirill Makurin
_______________________________________________
Mingw-w64-public mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/mingw-w64-public
_______________________________________________
Mingw-w64-public mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/mingw-w64-public
