Hartmut Kaiser <hartmut.kaiser <at> gmail.com> writes:

Since a couple of days (2-3 days) the Wave tests started to fail because of a crash during application cleanup in the thread library (msvc-8, gcc3.4.5/Mingw - this problem seems to be constrained to the windows platform).

The point of the segfault is here: boost/thread/win32/tss.hpp:+30,

void cleanup(T* data) const { if(func) { func(data); // <-- here } else { delete data; } }

Where data is a valid pointer, but func points somewhere.

No code changes have been done lately in Wave (or in Spirit/Phoenix for that matter, because the tss is used in Spirit code). Any clues what might be the reason?

I changed the TSS code recently, so that must be the problem. This implies that there is a race between the thread_specific_ptr constructor or destructor and the wave code that uses it. Such a race is not a new problem, though it obviously has new consequences. I'll look into it. Anthony