stddef: Fix test-stddef compilation error on FreeBSD 12/x86

[Bruno Haible]

On FreeBSD 12.0/x86, on the stable-202301 branch, I'm seeing this compilation
error:

../../gltests/test-stddef.c:62:1: error: static_assert failed "static assertion 
failed"
static_assert (__alignof__ (double) <= __alignof__ (max_align_t));
^              ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/usr/include/assert.h:73:23: note: expanded from macro 'static_assert'
#define static_assert   _Static_assert
                        ^
../gllib/assert.h:256:7: note: expanded from macro '_Static_assert'
      _GL_STATIC_ASSERT (__VA_ARGS__, "static assertion failed", -)
      ^                  ~~~~~~~~~~~
../gllib/assert.h:241:48: note: expanded from macro '_GL_STATIC_ASSERT'
# define _GL_STATIC_ASSERT(R, DIAGNOSTIC, ...) _Static_assert (R, DIAGNOSTIC)
                                               ^               ~
1 error generated.


What happens is that the fix for a wrong _Alignof that we have in the
stdalign / alignasof module causes max_align_t to have too small alignment.

              |  original    |  ISO C compliant  |
              |  _Alignof    |     _Alignof      |
--------------+--------------+-------------------+
long long     |     8        |       4           |
--------------+--------------+-------------------+
double        |     8        |       4           |
--------------+--------------+-------------------+
long double   |     4        |       4           |
--------------+--------------+-------------------+

max_align_t being defined through _Alignof(long long) and _Alignof(long double),
it alignment shrinks from 8 to 4, and thus
   __alignof__ (double) <= __alignof__ (max_align_t)
is no longer fulfilled.

This patch fixes it: the configure test "checking for good max_align_t..."
now reports no instead of yes, and max_align_t gets overridden.


2023-01-27  Bruno Haible  <bruno@clisp.org>

        stddef: Fix test-stddef compilation error on FreeBSD 12/x86.
        * m4/stddef_h.m4 (gl_STDDEF_H): When testing for good max_align_t, use
        an ISO C compliant definition of _Alignof.

diff --git a/m4/stddef_h.m4 b/m4/stddef_h.m4
index c0553d6f2f..a2322ebb7e 100644
--- a/m4/stddef_h.m4
+++ b/m4/stddef_h.m4
@@ -1,4 +1,4 @@
-# stddef_h.m4 serial 12
+# stddef_h.m4 serial 13
 dnl Copyright (C) 2009-2023 Free Software Foundation, Inc.
 dnl This file is free software; the Free Software Foundation
 dnl gives unlimited permission to copy and/or distribute it,
@@ -22,7 +22,14 @@ AC_DEFUN_ONCE([gl_STDDEF_H],
     [gl_cv_type_max_align_t],
     [AC_COMPILE_IFELSE(
        [AC_LANG_PROGRAM(
-          [[#include <stddef.h>
+          [[/* On FreeBSD 12.0/x86, max_align_t defined by <stddef.h> has
+               the correct alignment with the default (wrong) definition of
+               _Alignof, but a wrong alignment as soon as we activate an
+               ISO C compliant _Alignof definition.  */
+            #if ((defined __GNUC__ && 4 <= __GNUC__) || defined __clang__) && 
!defined __cplusplus
+             #define _Alignof(type) __builtin_offsetof (struct { char __a; 
type __b; }, __b)
+            #endif
+            #include <stddef.h>
             unsigned int s = sizeof (max_align_t);
             #if defined __GNUC__ || defined __clang__ || defined 
__IBM__ALIGNOF__
             int check1[2 * (__alignof__ (double) <= __alignof__ (max_align_t)) 
- 1];