------- Comment #3 from bkoz at gcc dot gnu dot org 2009-04-21 02:46 -------
I consider this undefined behaviour, so enhancement is the correct severity. People wanting to mess with non-standard init order should probably be taking steps to insure that libstdc++ is initialized first anyway. Not doing so is arguably a bug..... especially as it's not hard to do: ie: #include <iostream> void f1() __attribute__ ((constructor (101))); void f1() { std::ios_base::Init i; std::cout << "f1" << std::endl; } int main() { } But anyway........ libstdc++ is currently not using the first 100 slots reserved for the GNU implementation as per: http://gcc.gnu.org/onlinedocs/gcc/C_002b_002b-Attributes.html#C_002b_002b-Attributes No efforts or audit have ever been made as to enshrine a specified init order for libstd++. Locales, mutexes, and __mt_allocator would all have to be scoped and modified as code is/was changed. I consider this doable, but probably as much work in the long run as the efforts for symbol versioning. There may be an advantage to doing this for size reasons, as then <iostream>'s // For construction of filebuffers for cout, cin, cerr, clog et. al. static ios_base::Init __ioinit; could be killed. If there was some kind of real advantage, I would be more into this whole idea. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39796