Why are std:: symbols forcibly exported in shared libraries?

Calmarius_58805 0
Viewed 445

If your shared library uses C++ inside, then all the std:: templates and types referenced by it will be exported as weak symbols. Even if you use -fvisiblity=hidden and -Wl,--exclude-libs,ALL. The only way to hide those symbols is using a version script.

What's the purpose of these forced exports?

Is there any harm in hiding them with version scripts?

shared-libraries linker gcc linux c++
1 Answers

So I found two (rejected) GCC-bugreports discussing this issue:

The reasoning seems to be:

  • Easier programming along library borders when using stl
  • Easier testing of stl
  • It's difficult overall to hide template implementation details consistently

Check out this comment by Benjamin Kosnik especially. Cited for preservation purposes:

This isn't a bug, but rather part of a deliberate linkage strategy.

For C++, types that are to be used across shared libraries have to be visible. See:

http://gcc.gnu.org/onlinedocs/gcc/Code-Gen-Options.html#Code-Gen-Options

And in particular:

Note that '-fvisibility' does affect C++ vague linkage entities. This means that, for instance, an exception class that will be thrown between DSOs must be explicitly marked with default visibility so that the 'type_info' nodes will be unified between the DSOs.

Thus, the rationale for libstdc++ having namespace std have visibility "default." If you were to hack in support for allowing namespace std to have hidden visibility, and run the testsuite with -fvisibility=hidden (see attached patch) you would notice the breakdown in testresults, with mass failures. Thus, it is provided for information purposes only.

In the libstdc++ source files, anonymous namespaces are used for specific entities that have both local/hidden linkage. This use of anonymous namespaces is considered superior to attribute hidden as it uses ISO C++ and is thus more portable than vendor extensions (pragmas or attributes).

However, in libstdc++ header files, attribute hidden has been difficult to use. One might think that perhaps all the implementation details could be moved to say std::__detail, and then marked with attribute hidden. Then, many of these helper functions would be marked as hidden in your example below.

There are some pitfalls with this approach:

  1. all these implementation base types that are used by default derived classes would have to be default
  2. same with implementation details that use static locals
  3. same with virtual functions, etc etc.
  4. thus you end up with a pretty limited set of hidden things

In addition, this may exacerbate the dlopen + RTLD_LOCAL + weak symbol issue for C++.

user_71360130
Related