[C++-sig] Re: Boost.Python and GCC 3.4.2

[Jonathan Brandmeyer]

On Thu, 2004-09-30 at 08:26, Patrick Hartling wrote:
> On Wed, 29 Sep 2004 13:05:57 -0400, David Abrahams
> <dave at boost-consulting.com> wrote:
> > Patrick Hartling <patrick.hartling at gmail.com> writes:
> > 
> > > The code that GCC is complaining about is the type cast in the last
> > > term of the last statement in the following:
> > >
> > > template <class Data>
> > > struct additional_instance_size
> > > {
> > >     typedef instance<Data> instance_data;
> > >     typedef instance<char> instance_char;
> > >     static const std::size_t value = sizeof(instance_data) -
> > > ((size_t)(&((instance_char *)0)->storage));
> > >
> > >
> > > };
> > >
> > > I haven't run into this problem before, and I am not sure how to go
> > > about fixing it (and submitting a patch if appropriate).  I did find a
> > > semi-related discussion on another mailing list with a reference to
> > > the relevant part of the C++ Standard:
> > >
> > >    http://www.cygwin.com/ml/ecos-discuss/2003-02/msg00214.html
> > >
> > > I have seen reports from other people on this list saying that they
> > > are using GCC 3.4 to compile Boost.Python without problems, so I may
> > > be doing something wrong.  I have attached some simple test code that
> > > reproduces the compiler error.  Is there a way to work around this, or
> > > is it a GCC bug?
> > 
> > It is a GCC bug, but not where you think.  This expression:
> > 
> >    (&((instance_char *)0)->storage)
> > 
> > is generated by GCC's C library sizeof() macro.
> 
> Do you mean offsetof() here?  I should have looked at the expansion of
> BOOST_PYTHON_OFFSETOF before going straight to the full output from
> the preprocessor.
> 
> > sizeof is required to
> > return an integral constant, by whatever means.  So the GCC people
> > need to stick in some special compiler magic to make it happy.
> 
> After digging into this some more, I think that this is a FreeBSD bug.
>  GCC 3.4.2 works just fine on Fedora Core 2 (though I haven't tried
> other Linux distributions).  On FreeBSD, stddef.h expands offsetof()
> to __offsetof() (from sys/cdefs.h).  __offsetof() is what generates
> the above code that GCC 3.4.2 does not like.  I'll let the FreeBSD
> folks know about this in hopes of seeing it fixed before FreeBSD 5.3
> gets released.
> 
> Thanks for pointing me in the right direction on this.
> 
>  -Patrick

I ran into a similar bug with a prerelease of GCC 3.4.  The bug was
ultimately corrected by the GCC team using a new __offsetof__() built-in
function and a macro that calls it in the GCC-supplied stddef.h.

HTH,
-Jonathan