id,summary,reporter,owner,description,type,status,milestone,component,version,severity,resolution,keywords,cc
10850,"gcc-x86-implementation of ""atomic_exchange_and_add"" triggers intel's ""unitialized variable"" runtime check",fabian.hachenberg@…,Peter Dimov,"The current implementation of atomic_exchange_and_add for gcc x86 is (http://svn.boost.org/svn/boost/trunk/boost/smart_ptr/detail/sp_counted_base_gcc_x86.hpp) as following:

inline int atomic_exchange_and_add( int * pw, int dv )
{
    // int r = *pw;
    // *pw += dv;
    // return r;

    int r;

    __asm__ __volatile__
    (
        ""lock\n\t""
        ""xadd %1, %0"":
        ""=m""( *pw ), ""=r""( r ): // outputs (%0, %1)
        ""m""( *pw ), ""1""( dv ): // inputs (%2, %3 == %1)
        ""memory"", ""cc"" // clobbers
    );

    return r;
}

This unfortunately triggers the ""unitialized variable"" runtime check of the Intel c++ compiler. Since r is actually superfluous, a simple patch can fix the problem:

inline int atomic_exchange_and_add( int * pw, int dv )
{
    // int r = *pw;
    // *pw += dv;
    // return r;
 
    __asm__ __volatile__
    (
        ""lock\n\t""
        ""xadd %1, %0"":
        ""+m""( *pw ), ""+r""( dv ): // input/output (%0, %1)
                :
        ""memory"", ""cc"" // clobbers
    );
 
    return dv;
}
 
My  colleague, who wrote the patch, checked that the generated assembler code is almost identical to the original one. ""Almost"" in the sense, that the removal of r changes some offsets.",Patches,new,To Be Determined,smart_ptr,Boost Development Trunk,Problem,,,