Re: max_align_t fails with clang 6.1 on OS X 10.10

[Bruno Haible]

Werner LEMBERG wrote:
> > Could you try incorporating the std-gnu11 module into your package,
> > and see whether that fixes things?  If so, perhaps we could fix this
> > problem for everybody by having stddef depend on std-gnu11.
> 
> This approach worked

Nevertheless, I don't agree that this is the solution.

What you are trying to do is to run the configure tests with a C compiler
and then use the resulting config.h with C code [1] and C++ code [2].

[1] http://repo.or.cz/ttfautohint.git/tree/HEAD:/lib
[2] http://repo.or.cz/ttfautohint.git/tree/HEAD:/frontend

This is not supported by Autoconf, and this is not supported by Gnulib -
simply because _any_ test result (not just the one for max_align_t) may
be different in C++ than it is in C.

There is one real fix for this problem: You create separate configure.ac
files for the lib/ and for the frontend directory, and each of them uses
a separate gnulib-tool invocation. (So, you must use the options --lib
and --source-base of gnulib-tool, to make sure the two gnulib-tool results
don't interfere.) The configure.ac for frontend/ must use AC_LANG_PUSH([C++])
before gl_INIT.

This fix is future-proof: It will work even after
  - you request additional gnulib modules,
  - the C standard features change,
  - the C++ standard features change.

The other approaches are brittle: If we/you change something regarding
the max_align_t problem, you will continue to see issues
  - on other platforms,
  - when you request additional gnulib modules,
  - when the C standard features change,
  - when the C++ standard features change.

Among these other approaches are the following:

  * You could add gnulib module 'std-gnu11' to the list of gnulib modules
    you request. This has the chance of fixing other problems than the
    max_align_t one, because C++ is closer to C11 than it is to C99.
    But still, newer C or C++ standard can bring new problems.

  * Gnulib could add a dependency from 'stddef' to 'std-gnu11'. I reject
    this idea because this is a big change that affects all users, for a
    situation that is not supported.

  * Gnulib could test something like
      HAVE_MAX_ALIGN_T || (defined(__APPLE__) && defined(__MACH__) && 
defined(__cplusplus))
    instead of
      HAVE_MAX_ALIGN_T
    This is not such a big change, but it is brittle code, may only fix
    the max_align_t problem, and is for a situation that is not supported.

In summary, I don't think Gnulib should change here, and you still the choice
among a real fix and a quick-fix.

Bruno