Bug#1109256: libkpathsea-dev: redeclares Standard C functions with incompatible prototype, causing FTBFS elsewhere
Package: libkpathsea-dev
Version: 2024.20240313.70630+ds-6
Severity: important
Tags: ftbfs upstream forky sid
Control: block 1096597 by -1
Control: affects -1 + src:evince
In #1096597, Matthias Klose reported that evince FTBFS with gcc-15 as
default (it is not the default for trixie, but will become the default
in forky). This seems to be caused by libkpathsea-dev's headers,
combined with gcc-15 defaulting to C23.
As an example of the class of issue I'm reporting, libkpathsea-dev's
c-std.h has this:
>#ifdef HAVE_STDLIB_H
>#include <stdlib.h>
>/* Include <stdlib.h> before <stddef.h>, to help avoid NULL
> redefinitions on some systems. (We don't include <stddef.h>
> ourselves any more, but FYI.) */
>#else
>/* It's impossible to say for sure what the system will deign to put in
> <stdlib.h>, but let's hope it's at least this. */
>extern char *getenv ();
>#endif /* not HAVE_STDLIB_H */
and evince includes <stdlib.h> followed by <kpathsea/c-std.h>.
The code in c-std.h assumes that the build system of every dependent
package will check for stdlib.h and define HAVE_STDLIB_H if found, but
there is a tendency to stop doing that, and instead assume that all
platforms comply with a 1989 C standard by now. In particular Evince
does not check for <stdlib.h>, and includes it unconditionally (I
assume).
In C11 or older, `extern char *getenv ()` declares getenv() as a
function taking unspecified parameters and returning a string, but in
C23 it declares getenv() as a function taking *no* parameters and
returning a string, hence the compiler sees this as a declaration that
is incompatible with the one it already saw in <stdlib.h>, and errors
out:
>/usr/include/kpathsea/c-std.h:48:14: error: conflicting types for ‘getenv’; have ‘char *(void)’
> 48 | extern char *getenv ();
> | ^~~~~~
An equivalent failure mode is seen in the evince build log for other
Standard C functions like strtok() and strstr(). There might be others,
I didn't inspect all of the kpathsea headers.
evince could presumably work around this by conditionally or
unconditionally defining HAVE_STDLIB_H, etc., before it includes
kpathsea headers; but it shouldn't have to, because each library should
be self-contained.
I think <kpathsea/c-std.h> should do something more like this
(pseudo-patch, untested) so that it treats __STDC__ as implying that
every Standard C header is present:
-#ifdef HAVE_STDLIB_H
+#if defined(__STDC__) || defined(HAVE_STDLIB_H)
#include <stdlib.h>
...
(or it could check __STDC_VERSION__ if preferred)
... or it could just assume that every platform has Standard C headers
by now, since C89 is several decades old.
Thanks,
smcv
Reply to: