[Bug 50292] Process-local synchronization objects use private interfaces into the Unix library

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=50292

--- Comment #8 from Zebediah Figura <z.figura12 at gmail.com> ---
(In reply to Rémi Bernon from comment #7)
> Regarding the performance issue, I think the linked-list nature of the TEB
> list could also be part of issue, especially when a large number of threads
> are created. On synthetic benchmarks, it's very clear that it induces a lot
> of CPU cache misses.
> 
> IIUC it's not really possible to avoid a global lock here. As
> NtAlertThreadByThreadId function is supposed to return a specific status if
> the targeted TID does not exist, I guess there's no other way than tracking
> all threads.
> 
> However, it should be possible to make it faster by having a static TID +
> sync object array instead, to help the CPU speculative execution do its
> magic. Of course it will then have to support a dynamic number of threads,
> but perhaps a reasonably large fixed maximum number may be acceptable and
> could still perform better than the linked list?

It should be possible to do it lock-free with the same mechanism as we use on
the PE side, i.e. a growable linked list of arrays. I hadn't though of that
when I first wrote the patch. It was my vague plan to try that optimization and
see if it helps.

-- 
Do not reply to this email, post in Bugzilla using the
above URL to reply.
You are receiving this mail because:
You are watching all bug changes.