[Bug libstdc++/69881] with gcc-6 of today building gcc-4.9 fails

[bernd.edlinger at hotmail dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69881

--- Comment #21 from Bernd Edlinger <bernd.edlinger at hotmail dot de> ---
(In reply to Jonathan Wakely from comment #20)
> (In reply to Bernd Edlinger from comment #19)
> > But I am sure that #define __need_offsetof can simply not work.
> 
> So don't define that. It's undefined behaviour.
> 


hmmm...

I am not sure it I tried to ride on a dead horse here...

Maybe I should just leave the c/* and c_std/* headers as they are.
I should just leave the attachment here, if some one has use for
c_std/* headers.


> > I am not sure if it can compile at all, because if I try
> > there are the same bogus symlinks to missing c/-headers.
> > 
> > But before the build reaches the point where that matters,
> > the cstdlib simply includes <stdlib.h> which fails to declare
> > std::abort.
> > 
> > Is this c/cstdlib intended for targets that have C++ aware stdlib.h?
> 
> Yes.
> 
> > When we leave it as it is, everyone will point with the finger
> > on us, because it looks like a gcc-bug.
> > 
> > Alternative would be to detect that any of the 6 known
> > __need_-defines are present when invoking cstddef and
> > give an explicit #error, but this mis-use worked previously.
> > Or maybe a #warning would be a compromise?
> > 
> > What do you think?
> 
> I think if we do anything then I prefer your original suggestion to just
> #undef them.
> 
> But I'm not yet convinced we should do anything. This already failed with
> old versions of GCC:
> 
> // get size_t
> #define __need_size_t
> #include <cstddef>
> // get everything else
> #include <cstddef>
> std::max_align_t m;

Yes, but I don't know how often I've boot-strapped gcc-4.9 already
and I never noticed anything of that undefined behaviour until now.

The problem is that header A can define __need_size_t, and include
the wrong system header file.  That leaks into header B
which includes <cstddef> and now we get a cryptic error,
while previously we just did not have the std::max_align_t,
which is also bad, if you need it but if you don't need it,
you will never know.


BTW: we have a similar problem with cstdarg:

#define __need___va_list
#include <cstdarg>

fails to compile.

and c/cstdarg has:
#undef __need___va_list
#include_next <stdarg.h>


I would like to add this to c_global/cstdarg:

 #pragma GCC system_header

+#ifdef __need___va_list
+#warning __need___va_list is not supported in cstdarg
+#undef __need___va_list
+#endif
 #include <bits/c++config.h>
 #include <stdarg.h>

that would be rather comfortable for the user.