Re: Keep config.h idempotent

[Paul Eggert]

On 2023-01-25 06:05, Bruno Haible wrote:
> What happens, is that this compilation unit #includes <config.h> five times,
> and somewhere between the first and the fifth inclusion, poke's code does
>    #define returns_nonnull
>
> This is their particular way of not emitting this particular GCC/clang
> attribute, and is OK since 'returns_nonnull' is not a keyword and not a
> token in the reserved namespace.

Is this diagnosis correct? It appears to me that poke 2.90.1's "#define 
returns_nonnull" is not a problem, because the Gnulib code never uses 
that identifier. Instead, the Gnulib code does this:

#if _GL_HAS_ATTRIBUTE (returns_nonnull)

which uses this macro:

# define _GL_HAS_ATTRIBUTE(attr) __has_attribute (__##attr##__)

and so has the same effect as this:

#if __has_attribute (__returns_nonnull__)

regardless of what returns_nonnull was defined to, because C doesn't 
macro-expand immediate operands of ##.

If I understand things correctly the problem with poke 2.90.1 wasn't its 
"#define returns_nonnull"; it was its "#define __returns_nonnull__", 
which indeed defines a token in the reserved namespace.

> Hence this sequence
>
>   #include <config.h>
>   #define returns_nonnull
>   #include <config.h>
>
> is supposed to work fine.

Isn't there still a problem with something like the following, even with 
the latest Gnulib?

   #define __returns_nonnull__
   #include <config.h>

That is, it seems to me that the real problem here is a collision with 
poke's "#define __returns_nonnull__" before <config.h>, not with 
including <config.h> twice.

Obviously we can't make Gnulib immune to arbitrary #defines of reserved 
names. This is not to say that we shouldn't accommodate GNU poke, just 
that it's not clear that the cure is right here, as a general rule.

More generally, I don't think Gnulib should imply to users that it's OK 
to include <config.h> twice. That's an uncommon practice and is a recipe 
for trouble, as too many programs do fun stuff in their configure.ac 
that Gnulib has no control over. Instead, we should clarify Gnulib's 
documentation to say that compilation units should include <config.h> 
first, before including anything else, and should not include <config.h> 
later.