With g++ 4.5.0 static initializers can be executed more than once. Here's how. Let a.cc contain the following code: #include <stdio.h> struct A { A() { static int x = printf("x initialized at @%p\n", &x); } }; extern "C" void inita(); void inita() { A x; A y; } And let b.cc be the copy of a.cc where 'inita' has been renamed to 'initb'. Then make a shared library a.so from a.cc and b.so from b.cc. Finally, make a program that dynamically loads a.so and b.so and calls inita and initb functions. Since 'x' ends up as a unique global symbol, there will be only one object in the process, but the initialization code will be run twice (the address printed will be the same).
Created attachment 20619 [details] The test case Just unpack and run 'make; make run'. The output should be something like: x initialized at @0x40021a28 x initialized at @0x40021a28 (note that the addresses are identical).
It works for me. You need to build shared libraries with -fPIC.
Or rather as you use dlopen you need to use RTLD_GLOBAL.
Re: #c2, that's just because you are using too old binutils I guess. The bug seems to be in: .type _ZGVZN1AC1EvE1x, @object .type _ZZN1AC1EvE1x, @gnu_unique_object I believe the guard needs to be @gnu_unique_object as well.
Indeed, using RTLD_GLOBAL works around the problem. However, I don't think one should just resort to using this flag with dlopen everywhere. The problem was originally found with Python modules written in C, and while one can change the dlopen flags used for Python imports, I'm not sure that that won't break something else. As Jakub has pointed out, the guard needs to be unique as well.
I guess the problem is in the !DECL_ARTIFICIAL (DECL) test in ASM_DECLARE_OBJECT_NAME macro - the guard is artificial. Not sure why that has been added. /* For template static data member instantiations or \ inline fn local statics, use gnu_unique_object so that \ they will be combined even under RTLD_LOCAL. */ \ if (USE_GNU_UNIQUE_OBJECT \ && !DECL_ARTIFICIAL (DECL) && DECL_ONE_ONLY (DECL)) \ ASM_OUTPUT_TYPE_DIRECTIVE (FILE, NAME, "gnu_unique_object"); \ else \ ASM_OUTPUT_TYPE_DIRECTIVE (FILE, NAME, "object"); \
Reopening.
Subject: Bug 44059 Author: jakub Date: Tue May 11 18:14:19 2010 New Revision: 159287 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=159287 Log: PR c++/44059 * config/elfos.h (ASM_DECLARE_OBJECT_NAME): Use qnu_unique_object even for DECL_ONE_ONLY DECL_ARTIFICIAL !TREE_READONLY decls. * config/alpha/elf.h (ASM_DECLARE_OBJECT_NAME): Likewise. * dwarf2asm.c (dw2_output_indirect_constant_1): Set TREE_READONLY on DW.ref.* decls. Modified: trunk/gcc/ChangeLog trunk/gcc/config/alpha/elf.h trunk/gcc/config/elfos.h trunk/gcc/dwarf2asm.c
Fixed on the trunk so far, will backport to 4.5 after a while if there aren't any issues with it on the trunk.
Subject: Bug 44059 Author: jakub Date: Wed Jun 30 15:16:54 2010 New Revision: 161607 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=161607 Log: Backport from mainline 2010-05-11 Jakub Jelinek <jakub@redhat.com> PR c++/44059 * config/elfos.h (ASM_DECLARE_OBJECT_NAME): Use qnu_unique_object even for DECL_ONE_ONLY DECL_ARTIFICIAL !TREE_READONLY decls. * config/alpha/elf.h (ASM_DECLARE_OBJECT_NAME): Likewise. * dwarf2asm.c (dw2_output_indirect_constant_1): Set TREE_READONLY on DW.ref.* decls. Modified: branches/gcc-4_5-branch/gcc/ChangeLog branches/gcc-4_5-branch/gcc/config/alpha/elf.h branches/gcc-4_5-branch/gcc/config/elfos.h branches/gcc-4_5-branch/gcc/dwarf2asm.c branches/gcc-4_5-branch/gcc/testsuite/g++.dg/warn/Wconversion-null-2.C (props changed) Propchange: branches/gcc-4_5-branch/gcc/testsuite/g++.dg/warn/Wconversion-null-2.C ('svn:mergeinfo' removed)
Fixed.
*** Bug 44815 has been marked as a duplicate of this bug. ***