On Mon, 6 Jul 2026, LIU Hao wrote:
在 2026-7-6 18:13, Martin Storsjö 写道:
I'm not entirely convinced that this is the right thing to do in general -
I think this would lead to destructors not being executed in cases where I
think they would be executed in e.g. an MSVC build.
Does this cause e.g. destructors for user code TLS C++ objects in DLLs to
not be executed on process exit, for builds using libstdc++/libgcc?
No. Destructors of thread-local objects are registered with
`__cxa_thread_atexit()` which is implemented in cxa_thread_atexit.c.
However, this change has an effect on emutls, so memory for thread-local
objects is no longer freed upon process exit. This might look harmless; but
frankly speaking, I didn't test it.
Oh, intresting. There are other known problems with C++ TLS dtors and
emutls freeing the memory as well - although I don't remember if that was
at thread exit, DLL unload or process exit though, so I guess this also
would avoid some aspects of that problem too.
This should at least affect fewer cases than the previous patch.
But overall, the problem is the same - destructors in DLLs on process
exist, when threads still may be running, really are tricky on Windows in
general. I think the conventional method is to fix this on the user code
side - but there's usually no single easy way of fixing it.
TL;DR - this patch shouldn't hurt me and my configurations I think, so I
won't block it, but I think it's principally slightly wrong.
`__mingwthr_key_dtor()` is called internally by `__gthread_key_create()` for
win32 thread model [1]. In the case of posix thread model,
`pthread_key_create()` is called instead. My conclusion is that these two
variants should behave identically.
Does this hold for pthread_key_create right now, that the callbacks aren't
called on process exit? Then that's a good argument for this change.
Interestingly, while `pthread_key_create()` says nothing about process
termination [2], the C11 `tss_create()` requires that '[d]estructors
associated with thread-specific storage shall not be invoked at process
termination.' [3]
Interesting - I guess that's even more arguments for this change then.
// Martin
_______________________________________________
Mingw-w64-public mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/mingw-w64-public
