[Bug binutils/13406] New: version script: please allow anonymous tag to be combined with other tags

[lionel at mamane dot lu]

http://sourceware.org/bugzilla/show_bug.cgi?id=13406

             Bug #: 13406
           Summary: version script: please allow anonymous tag to be
                    combined with other tags
           Product: binutils
           Version: 2.21
            Status: NEW
          Severity: enhancement
          Priority: P2
         Component: binutils
        AssignedTo: address@hidden
        ReportedBy: address@hidden
    Classification: Unclassified


Created attachment 6054
  --> http://sourceware.org/bugzilla/attachment.cgi?id=6054
patch to allow anonymous version tag combined with non-anonymous version tag

In a version script, currently one is not allowed to use an anonymous version
tag along with another tag. I'm not sure why that is,
http://cygwin.com/ml/binutils/2001-12/msg00182.html says "per Ulrich's
request".

I would like that to be allowed.

My use case is within LibreOffice, which is structured as an executable that
(lazily, as the need my arise) dlopen()s various libraries, which themselves
lazily dlopen() the libraries they need, ... These libraries use version link
scripts along the lines of:


UDK_3_0_0 {
    global:
        _ZTI*; _ZTS*; # weak RTTI symbols for C++ exceptions
        component_writeInfo;
        component_getFactory;
    local:
        *;
};

In particular, they hide (not export) all symbols by default. This plays out
badly with an optimisation within libstdc++, which is that when an empty
std::string is created, it does not get memory allocated from the heap, but the
internal data pointer points to a unique static address allocated by the linker
via a unique symbol (_S_empty_rep_storage). The std::string destructor
special-cases that address to avoid calling heap deallocation on it.

The situation is that a library using such a version script (an internal
library of LibreOffice, mysqlc.uno.so; that is the "native" driver to MySQL for
LibreOffice) ldopen()s a library that does *not* use such a version script (a
library that is external to LibreOffice, libmysqlcppconn.so; that is the
official C++ API to MySQL: http://www.mysql.com/downloads/connector/cpp/).

The core of the problem is that because of the "local: *;" in the version
script of mysqlc.uno.so, mysqlc.uno.so and libmysqlcppconn.so have a
*different* _S_empty_rep_storage symbol; the linker does *not* merge them. So
when code in mysqlc.uno.so calls a function in libmysqlcppconn.so, gets an
empty std::string and that std::string goes out of scope:

1) The empty std::string is constructed within libmysqlcppconn.so, so its
internal data pointer points to libmysqlcppconn's _S_empty_rep_storage.

2) That std::string is destructed in mysqlc.uno.so, so (code equivalent to)
this is executed within mysqlc.uno.so:

if (the_string.pData != &_S_empty_rep_storage)
   free(the_string.pData)

As it is compared to the address of mysqlc.uno's _S_empty_rep_storage, the test
evaluates to false, and heap deallocation is called on libmysqlcppconn.so's
_S_empty_rep_storage, which was not heap allocated, and all hell breaks loose
;-) In particular, in libstdc++ debug mode (compile with -D_GLIBCXX_DEBUG),
this immediately aborts the whole program.

To avoid that, I would like to add to mysqlc.uno.so's version link script:

{
    global:
        _ZNSs4_Rep20_S_empty_rep_storageE;
};

Patch is attached. Tested, works for me, solves the described problem within
LibreOffice.

This would also solve bug 12548 as a side-effect :)

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.