Bug 7824

It would seem to be some kind of incompatibility between the compiler and
glibc (perhaps triggered by certain compiler options).  I tried reproducing
the bug with a small test case (enclosed) and I tried -Wall, -O2, and -pedantic
options and that didn't seem to do it.  Can you send a self-contained test case?
(usually the best way to do that is to take one of the source files which is
giving the error and send the output of gcc -E on that source file).

------------------ begin sep.c
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int
main (int argc, char **argv)
{
  char *p;
  char *q;
  int i;
  if (argc != 2)
    {
      printf ("Usage: sep tok1,tok2,tok3,...tokn\n");
      exit (EXIT_FAILURE);
    }
  p = argv[1];
  i = 0;
  while ((q = strsep (&p, ",")) != NULL)
    printf ("Token %d is \"%s\"\n", i++, q);
  exit (EXIT_SUCCESS);
}
------------------- end sep.c

I've now applied your advice and done "gcc -E ..." on one of the files reporting
that error. However, the results are a little large to post here, so can I email
them to you separately?

One thing though: It looks like it's something to do with gcc's optimisation, as
`gcc -Os` doesn't produce those warnings, although both `gcc -O` and `gcc -O2`
do.

For reference, here's the original compilation command for the section that I
used -E on:

 Q> gcc -O -Wall -Wconversion -DIOCTL -c -DREVERSEVIDEO -DSUNTRANSITS     \
 Q>     -DXWINDOW -I/usr/src/SatTrack-3.1.7/src/include                   \
 Q>     /usr/src/SatTrack-3.1.7/src/fixchecksum/fix checksum.c            \
 Q>     -o /usr/src/SatTrack-3.1.7/obj/fixchecksum.o

The following is the bash script I used to perform these tests:

 Q> #!/bin/bash
 Q>
 Q> function run() {
 Q>     gcc -E $1 -Wall -Wconversion -DIOCTL -c -DREVERSEVIDEO              \
 Q>         -DSUNTRANSITS -DXWINDOW  -I/usr/src/SatTrack-3.1.7/src/include  \
 Q>         /usr/src/SatTrack-3.1.7/src/fixchecksum/fixchecksum.c           \
 Q>                                                                         \
 Q>         | tr -s '\t' ' ' | sed 's/^ $//' | uniq
 Q> }
 Q>
 Q> run -O  | tee fixchecksum-O.c
 Q> run -O2 | tee fixchecksum-O2.c
 Q> run -Os | tee fixchecksum-Os.c

I can report that the output of the first two tests was identical. Here's a
directory listing of the resultant files:

-rw-r--r--   0/0   102694 Dec 23 19:44 1999 fixchecksum-O.c
-rw-r--r--   0/0   102694 Dec 23 19:44 1999 fixchecksum-O2.c
-rw-r--r--   0/0    47201 Dec 23 19:44 1999 fixchecksum-Os.c
-rw-r--r--   0/0     9358 Jan  4 21:21 1997 fixchecksum.c

As you can see, the result of the -Os test is considerably smaller than the
other two.

Created attachment 44 [details]
Original source + "gcc -E"'d versions thereof...

Never mind about the emails, I was able to attach the tarball here anyway.

Well, it is caused by -Wconversion.  The header contains the prototype:

  extern __inline char *__strpbrk_c2 (__const char *__s, char __accept1,
  char __accept2);

and that kind of prototype (with type "char") is exactly what -Wconversion
is supposed to be warning about, if you look at the documentation of
-Wconversion.

As for -O2, it isn't optimization as such, it is the fact that turning
on optimization causes the compiler to define __OPTIMIZE__ which the glibc
headers test for when deciding whether to use inlines and #define and so on.

The solution?  The only two that are springing to mind are (1) persuade the
glibc people not to do that, or (2) don't use -Wconversion.  I think I'd
suggest the latter.  I'm not thinking of anything that GCC should be
doing differently (although I suppose a case could be made for __extension__).

That makes sense, but...

If that's what the problem is, why doesn't it show with -Os as well?

The only reason I used -Wconversion is because it was used in the original, and
I can also confirm that the SAME files, when compiled under RedHat 4.2 as
distributed, and with the SAME options, did NOT produce those warnings.