https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126399
--- Comment #4 from Vedran Miletic <vedran at miletic dot net> ---
(In reply to andy from comment #3)
> Confirming this on x86_64-pc-solaris2.11 (illumos, OmniOS r151059) with
> GCC 15.2.0.
>
> We first noticed that when LLVM/clang began to abort on every invocation
> after changing to build against gcc15 rather than gcc14.
>
> terminate called after throwing an instance of 'std::system_error'
> what(): Unknown error
> ...
> 13 libstdc++.so.6.0.34 std::__throw_system_error(int) + 75
> 14 libLLVM.so.19.1 llvm::initializeIRTranslatorPass(...) + 283
> ...
> 17 clang-19 llvm::InitializeAllTargets() + 14
>
> That is the first std::call_once executed inside libLLVM.so. The
> library's __gthread_active_p() binds __gthread_trigger to the
> executable's exported copy, so the executable's hidden flag is set,
> libLLVM's own stays -1, is clamped to 0, and __gthread_once() returns
> -1 from then on (hence strerror(-1) = "Unknown error").
>
> The silent version is worse than the abort - __gthread_mutex_lock()
> consults the same flag, so std::mutex locks become no-ops in the
> affected object.
>
> Here is our portable reproducer with two shared libraries, which does
> not depend on the executable exporting its globals (Solaris ld does;
> GNU ld does not), so it should reproduce on FreeBSD with the same headers:
>
> $ cat liba.cc
> #include <mutex>
> static std::once_flag f;
> extern "C" void a_once() { std::call_once(f, []{}); }
>
> $ cat libb.cc
> #include <mutex>
> #include <cstdio>
> static std::once_flag f;
> extern "C" void b_once() { std::call_once(f, []{ std::puts("libb once
> ran"); }); }
>
> $ cat main.cc
> #include <cstdio>
> extern "C" void a_once();
> extern "C" void b_once();
> int main() { b_once(); std::puts("ok"); return 0; }
>
> $ g++ -std=c++17 -shared -fPIC liba.cc -o liba.so
> $ g++ -std=c++17 -shared -fPIC libb.cc -o libb.so
> $ g++ -std=c++17 main.cc -o main -L. -la -lb -Wl,-R$PWD
> $ ./main
> terminate called after throwing an instance of 'std::system_error'
> what(): Unknown error
> (exit status 134)
>
> OmniOS is shipping the equivalent of attachment 65126 [details] in its GCC 15
> since 2026-09-01 and has rebuilt the affected packages with it.
>
> As nothing seems to have reached gcc-patches yet, we are happy to
> submit the patch for trunk and the 15/16 branches with
>
> libgcc/ChangeLog:
>
> PR libstdc++/126399
> * gthr-posix.h (__gthread_trigger): Move into the hidden
> visibility region alongside __gthread_active.
>
> unless the reporter would prefer to.
I am glad to get an independent confirmation. Please go ahead.
