assert-h: Make static_assert work on Solaris 11.4

assert-h: Make static_assert work on Solaris 11.4

[Bruno Haible]

The GNU sed prerelease testing [1] and the GNU texinfo prerelease testing [2]
have revealed compilation errors on Solaris 11.4, due to the use of
'static_assert'. The <assert.h> of Solaris 11.4 defines it in such a way
that it can only be invoked with 2 arguments. However, since 2022-09-13,
many Gnulib source files use 'static_assert' with 1 argument only.

[1] https://lists.gnu.org/archive/html/sed-devel/2022-10/msg00004.html
[2] https://lists.gnu.org/archive/html/bug-texinfo/2022-10/msg00188.html

This patch fixes it.


2022-10-23  Bruno Haible  <bruno@clisp.org>

	assert-h: Make static_assert work on Solaris 11.4.
	* m4/assert_h.m4 (gl_ASSERT_H): After including <assert.h>, on Solaris,
	redefine static_assert.

diff --git a/m4/assert_h.m4 b/m4/assert_h.m4
index c1306daef4..e892ea2f01 100644
--- a/m4/assert_h.m4
+++ b/m4/assert_h.m4
@@ -57,5 +57,11 @@ AC_DEFUN([gl_ASSERT_H],
              && __GNUG__ < 6 && __clang_major__ < 6)))
  #include <assert.h>
  #undef/**/assert
+ /* Solaris 11.4 <assert.h> defines static_assert as a macro with 2 arguments.
+    We need it also to be invocable with a single argument.  */
+ #if defined __sun && (__STDC_VERSION__ - 0 >= 201112L) && !defined __cplusplus
+  #undef static_assert
+  #define static_assert _Static_assert
+ #endif
 #endif])
 ])

Re: assert-h: Make static_assert work on Solaris 11.4

[Paul Eggert]

On 2022-10-23 07:47, Bruno Haible wrote:
>    #include <assert.h>
>    #undef/**/assert
> + /* Solaris 11.4 <assert.h> defines static_assert as a macro with 2 arguments.
> +    We need it also to be invocable with a single argument.  */
> + #if defined __sun && (__STDC_VERSION__ - 0 >= 201112L) && !defined __cplusplus
> +  #undef static_assert
> +  #define static_assert _Static_assert
> + #endif
>   #endif])

Will this approach work if code does something like the following? I 
worry that the later <assert.h> includes would collide with config.h's 
definition of static_assert.

   #include <config.h>

   #define NDEBUG 1
   #include <assert.h>

   #define NDEBUG 0
   #include <assert.h>

   static_assert (true);

Come to think of it, the latest C23 draft is a little squirrelly here, 
as its section 7.2 says that <assert.h> defines a static_assert macro. 
This must be a typo because it never goes no to say anything about what 
the macro does, and static_assert is a keyword in C23.

Also, while we're on the topic, why does the latest C23 draft require 
that when NDEBUG is defined, the assert macro is defined via "#define 
assert(...) ((void)0)" rather than as "#define assert(ignore) 
((void)0)"? What's the point of requiring the ellipsis?

Re: assert-h: Make static_assert work on Solaris 11.4

[Jeffrey Walton]

On Tue, Oct 25, 2022 at 2:13 PM Paul Eggert <eggert@cs.ucla.edu> wrote:
>
> On 2022-10-23 07:47, Bruno Haible wrote:
> >    #include <assert.h>
> >    #undef/**/assert
> > + /* Solaris 11.4 <assert.h> defines static_assert as a macro with 2 arguments.
> > +    We need it also to be invocable with a single argument.  */
> > + #if defined __sun && (__STDC_VERSION__ - 0 >= 201112L) && !defined __cplusplus
> > +  #undef static_assert
> > +  #define static_assert _Static_assert
> > + #endif
> >   #endif])
>
> Will this approach work if code does something like the following? I
> worry that the later <assert.h> includes would collide with config.h's
> definition of static_assert.
>
>    #include <config.h>
>
>    #define NDEBUG 1
>    #include <assert.h>
>
>    #define NDEBUG 0
>    #include <assert.h>
>
>    static_assert (true);
>
> Come to think of it, the latest C23 draft is a little squirrelly here,
> as its section 7.2 says that <assert.h> defines a static_assert macro.
> This must be a typo because it never goes no to say anything about what
> the macro does, and static_assert is a keyword in C23.
>
> Also, while we're on the topic, why does the latest C23 draft require
> that when NDEBUG is defined, the assert macro is defined via "#define
> assert(...) ((void)0)" rather than as "#define assert(ignore)
> ((void)0)"? What's the point of requiring the ellipsis?

Re: the ellipses. From
https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p2264r3.html,
Section 8:

    These changes are relative to N2573. If those changes are
    applied the minimal change for C++ standard proposed above
    could be used.

    In section 7.2 (Diagnostics <assert.h>) change the
    definition of the assert() macro to use elipsis instead of
    a single macro parameter:
    ...

Jeff

Re: assert-h: Make static_assert work on Solaris 11.4

[Paul Eggert]

On 2022-10-25 11:18, Jeffrey Walton wrote:
> Re: the ellipses. From
> https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p2264r3.html,
> Section 8:

Thanks. Although that sort of explains things, it doesn't explain why 
the latest draft uses ellipses only when NDEBUG is defined. Shouldn't 
ellipses also be used when NDEBUG is not defined?

Sounds like another typo (or at least lack of clarity) in the current draft.

Re: assert-h: Make static_assert work on Solaris 11.4

[Bruno Haible]

Paul Eggert asked:
> >    #include <assert.h>
> >    #undef/**/assert
> > + /* Solaris 11.4 <assert.h> defines static_assert as a macro with 2 arguments.
> > +    We need it also to be invocable with a single argument.  */
> > + #if defined __sun && (__STDC_VERSION__ - 0 >= 201112L) && !defined __cplusplus
> > +  #undef static_assert
> > +  #define static_assert _Static_assert
> > + #endif
> >   #endif])
> 
> Will this approach work if code does something like the following? I 
> worry that the later <assert.h> includes would collide with config.h's 
> definition of static_assert.
> 
>    #include <config.h>
> 
>    #define NDEBUG 1
>    #include <assert.h>
> 
>    #define NDEBUG 0
>    #include <assert.h>
> 
>    static_assert (true);

It will work. The reason is that <assert.h> looks like this:

   #ifndef _ASSERT_H
   #define _ASSERT_H
   /* Part 1: definitions that don't depend on NDEBUG */
   #endif
   /* Part 2: definitions that depend on NDEBUG */

and the definition of 'static_assert', on Solaris 11.4, happens to be in
part 1.

> Come to think of it, the latest C23 draft is a little squirrelly here, 
> as its section 7.2 says that <assert.h> defines a static_assert macro. 
> This must be a typo because it never goes no to say anything about what 
> the macro does, and static_assert is a keyword in C23.

I agree. Still, 'static_assert' is _allowed_ to be a macro, by § 6.10.9.(2).

> Also, while we're on the topic, why does the latest C23 draft require 
> that when NDEBUG is defined, the assert macro is defined via "#define 
> assert(...) ((void)0)" rather than as "#define assert(ignore) 
> ((void)0)"? What's the point of requiring the ellipsis?

I would guess that it's for compatibility for C++. ISO C++ 17 does not
require anything about the argument when NDEBUG is defined. Thus, a
program can contain

  #undef NDEBUG
  assert (new HashMap<A,B>(3).count == 0);

and the compiler should effectively ignore this. But when the types HashMap,
A, and B are not defined, the only way to ignore this line is to allow
multiple arguments to 'assert' at the preprocessor level. (Because of the
comma.)

Most <assert.h> implementations get this wrong. Effectively requiring the
programmer to insert extra parentheses:

  #undef NDEBUG
  assert ((new HashMap<A,B>(3).count == 0));

instead.

Bruno

Re: assert-h: Make static_assert work on Solaris 11.4

[Bruno Haible]

Paul Eggert wrote:
> > Re: the ellipses. From
> > https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p2264r3.html,
> > Section 8:
> 
> Thanks. Although that sort of explains things, it doesn't explain why 
> the latest draft uses ellipses only when NDEBUG is defined. Shouldn't 
> ellipses also be used when NDEBUG is not defined?

That's exactly what the standard says in § 7.2.(2):

  The assert macro shall be implemented as a macro with an ellipsis parameter,
  not as an actual function.

Bruno

Re: assert-h: Make static_assert work on Solaris 11.4

[Paul Eggert]

On 10/25/22 13:11, Bruno Haible wrote:
>    The assert macro shall be implemented as a macro with an ellipsis parameter,
>    not as an actual function.

Thanks for clarifying it for me.

I suppose the Gnulib assert-h module could be enhanced to support this 
C23 extension. Low priority for me as I don't deal with Gnulib + C++ and 
the problem comes up more rarely in C than in C++.

Re: assert-h: Make static_assert work on Solaris 11.4

[Bruno Haible]

Paul Eggert wrote:
> I suppose the Gnulib assert-h module could be enhanced to support this 
> C23 extension. Low priority for me as I don't deal with Gnulib + C++ and 
> the problem comes up more rarely in C than in C++.

I agree. Especially since I know of no system (so far) whose <assert.h>
does it right.

Bruno