在 2024-04-24 20:11, Martin Storsjö 写道:
Agreed, although it's less of a regression if we require windows headers for pthread_cleanup_push only when built with MSVC, than making it required for everybody, as the library wasn't usable at all with MSVC until recently.However, I looked at the output of _ReadWriteBarrier(), and in most cases, it seems like this doesn't seem like something equivalent to __sync_synchronize. In the mingw headers, it's just defined as __asm__ __volatile__ ("" ::: "memory"), i.e. something that stops the compiler from reordering code across this point, but does nothing for synchronizing with other threads. See e.g. https://godbolt.org/z/ccGdd7P44.I.e., for now, I still think this patch in the shape I posted it, is the best candidate for unbreaking this case.
This reminds me that `MemoryBarrier()` is a full barrier (`atomic_seq_cst`) instead of a load-store barrier (`atomic_acq_rel`)., and the barriers are there to prevent partial writes in case of asynchronous cancellation.
The patch looks correct to me. Please go ahead and apply. Thanks. -- Best regards, LIU Hao
OpenPGP_signature.asc
Description: OpenPGP digital signature
_______________________________________________ Mingw-w64-public mailing list Mingw-w64-public@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/mingw-w64-public
