Re: Adding a new thread model to GCC

[LIU Hao <lh_mouse@126.com>]

[-- Attachment #1.1: Type: text/plain, Size: 3867 bytes --]

在 2022/10/21 18:09, i.nixman@autistici.org 写道:
> On 2022-10-21 09:58, Jonathan Wakely via Libstdc++ wrote:
>> How does this compare with Eric B's proposal at
>> https://gcc.gnu.org/legacy-ml/gcc-patches/2019-06/msg01840.html ?
>>
>> It would be good if we can accept one of them for GCC 13, but I don't
>> know Windows well enough to determine which is better.
> 
> I had the same question...
> I would like to understand what is the difference?
> Moreover I would like to understand what is the difference with the already added support for the 
> winpthreads library?
> 
> @LIU Hao, could you explain please?
> 
> 
> 

Thank you for your interest. I'm glad to make an introduction of it.


I have read this patch before. Let's take the mutex as an example:

There are a lot of ways to implement a mutex on Windows. Basically, a non-recursive mutex can be 
implemented with an atomic counter + a binary semaphore / auto-reset event. This proposed patch 
contains a `__gthr_win32_CRITICAL_SECTION` definition that I think is a duplicate of the internal 
`CRITICAL_SECTION` structure, so should also work the same way as it.

The problem about this approach is that, semaphores are valuable kernel objects, and the maximum 
number of HANDLEs that a process can open concurrently has a limit (like FDs on Linux), while 'many 
critical sections are used only occasionally (or never at all), meaning the auto-reset event often 
isn’t even necessary' [1], the semaphores are actually allocated on demand. This means that locking 
can fail. There is a story in article [1] which also explains the origination of keyed events; it's 
worth reading.

And, since Vista we also have native win32 condition variables, also implemented basing on keyed events.


The keyed events are undocumented and are only exposed via syscalls. However, as with other 
documented syscalls, available from Windows Drivers Kit, there are several advantages:

   * There is a global keyed event, which requires no initialization, but
     can be utilized by all processes. Basing on that, mcfgthread provides
     mutexs, condition variables, once flags, etc. that are all one-pointer
     size structs, consume absolutely no additional resource, allow
     constexpr initialization, and require no cleanup, much like on Linux.

   * The wait syscalls take a 64-bit integer, whose positive value denotes
     the number of 10^-7 seconds since 1600-01-01 00:00:00 Z, and whose
     negative value denotes a relative timeout. Hence it's much more simpler
     to implement `__gthread_mutex_timedlock()` and `__gthread_cond_wait()`
     which take absolute timeouts. On the other hand, Win32 APIs generally
     take a 32-bit relative timeout in milliseconds, which not only requires
     translation from an absolute timepoint argument, but can also easily
     get overflown.

   * Building mutexes on top of syscalls allows a better designed algorithm
     [2], and sometimes it can even outperform native `SRWLOCK`s [3].

   * mcfgthread also provides standard-conforming `__cxa_atexit()` and
     `__cxa_thread_atexit()` functions, for working around some strange,
     weird, and broken behaviors [4][5][6]. On Linux it's glibc that
     provides them, so this as a whole requires a little modification in
     mingw-w64. I am working on it however; hopefully we can land it soon.


[1] 
http://joeduffyblog.com/2006/11/28/windows-keyed-events-critical-sections-and-new-vista-synchronization-features/

[2] https://github.com/lhmouse/mcfgthread/blob/master/MUTEX.md
[3] https://github.com/lhmouse/mcfgthread#benchmarking

[4] https://sourceforge.net/p/mingw-w64/mailman/message/37268447/
[5] https://reviews.llvm.org/D102944
[6] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80816

-- 
Best regards,
LIU Hao

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 840 bytes --]