http://gcc.gnu.org/ml/gcc-patches/2005-03/msg02767.html
Reproducing the failure requires a shared library used with dlopen and
dlclose from an executable linked with -E, which is beyond the scope of
the GCC testsuite. The patch has been tested by the PR submitter and
with OpenHPI, which ran into the same problems. The original problem
can be duplicated with ACE+TAO test DLL_Test, which passes with the fix
on mainline, 4.0, and 3.4.
May I check in the new patch for mainline? For 4.0? For 3.4? As
noted above, it's a regression from GCC 3.3.