[libcxx] type_info/bad_cast observations relating to r191397

[Chandler Carruth]

On Wed, Oct 2, 2013 at 9:10 AM, G M <gmisocpp at gmail.com> wrote:

> Hi Everyone
> Change r191397 affected the libcxx test suite results on my
> windows machine for g++ and mingw (and possibly eventually VS) builds.
>

Very sorry for the trouble, and sorry that I missed your original email.
Sending stuff to the list as well helps ensure that my filters get me to
look at it.


>
> The comment is:
> "Make the guard for external ABI libraries include the guard for
>  libsupc++ in typeinfo.cpp, bringing it into agreement with
> exception.cpp. This fixes link errors due to duplicate symbols from
> this translation unit."
>
> I don't know that r191397 is wrong but after applying that change I
> noticed some additional test suite failures. One example:
> c:\libcxx\test\language.support\support.rtti\bad.cast\bad_cast.pass.cpp
>
> which now results in a linker error where it previously compiled. The
> linker error is:
> "undefined reference to std::bad_cast::bad_cast()"
>
> Another test case may be similarly effected is:
> /libcxx/test/language.support/support.rtti/bad.typeid/bad_typeid.pass.cpp
>
> Looking at the change, it's because libcxx/src/type_info.cpp was changed
> from this:
> #if !(defined(_LIBCPPABI_VERSION) || defined(LIBCXXRT))
>
> to this:
> #if !defined(LIBCXXRT) && !defined(_LIBCPPABI_VERSION) &&
> !defined(__GLIBCXX__)
>
> and on my builds this now blocks out the definitions for bad_cast and
> type_info constructors etc. This would account for the linker error, as on
> my machine LIBCXXRT is not defined && LIBCPPABI_VERSION is not defined
> either.
>
Yes, but the critical question is, why was __GLIBCXX__ defined? If it were
left undefined, you would still have gotten the symbols.

That macro is (AFAIK) supposed to indicate that libc++ is being lined with
libsupc++, the ABI portion of libstdc++, and thus should not provide these
symbols.

> A grep of _LIBCPPABI_VERSION in the libcxx code base shows a few
> references to other code that may be of interest here.
> My machine has cxxabi.h which appears possibly also of interest.
> Does anyone know what should be linked in to the above mentioned test
> cases to restore the bad_cast and type_info definitions for g++/mingw/vs if
> they are not to be provided by libcxx, so as to make the libcxx test cases
> compile again?
>
> Alternatively if r191397 is incorrect / insufficient, what should be
> changed in libcxx or the changed guard mentioned above, that will re-enable
> libcxx to again provide these definitions if it should?
>

I feel like your build is misconfigured.

Something needs to provide these pieces of the C++ ABI standard library.
Currently, the options are:

1) libc++ (what you've been using thus far I suspect)
2) libsupc++ (from libstdc++, indicated by defining __GLIBCXX__)
3) libcxxrt (from PathScale, indicated by defining LIBCXXRT)
4) libc++abi (a sibling project to libc++, indicated by defining
_LIBCPPABI_VERSION)

I don't know that #1 is really intended to work long term. I know that
Howard indicated to me some users are still relying on this though. I'm
personally interested in #2 not breaking (which motivated my change) and in
actively supporting #4 as it was developed in close conjunction with
libc++. What is your configuration for these pieces of the runtime?

>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20131002/26104ac5/attachment.html>