🚀 go-pugleaf

In gcc version 2.7.2 running on an Estrella Powerpc with Aix 4.1.3.0

 Gcc seems to have a problem whith compiling kernel modules which
import symbol lists.

 Here is a small example to reproduce the problem:

example.exp:
#!/unix
pincode

example.c:
int test(void)
	{
	pincode(test);
	}


 When compiling this with the following command line I get an error message:
gcc -nostdlib -Wl,-e -Wl,test -Wl,-bimport:./example.exp example.c -lsys -lcs
ys
collect2: /: library not found

 #!/unix (in example.exp) means that the .exp file contains symbols
exported by the kernel, no user mode shared library is involved. This
line is used in most kernel export files: /lib/kernex.exp,
/lib/aio.exp, /lib/pcmciacs.exp, /lib/rcm.exp, /lib/sockets.exp,
/lib/statcmd.exp, /lib/syscalls.exp, /lib/threads.exp and /lib/vme.exp
.
 Apparently gcc doesn't know this fact, and assumes that the name
after #! is always a library which exists in filespace and can be
read.

 Gcc-2.6.3 used to grok those just fine and produced working kernel
modules. The problem seems to stem from gcc's ld: If I place the old
ld from 2.6.3 in gcc-2.7.2's tree, it works.

 Here is the relevant code snippet in collect2.c where the message is
generated.

  2621        while (idx < ldh.l_istlen)
  2622          {
  2623            char *implib = impbuf + idx;
  2624            char *impmem = implib + strlen (implib) + 1;
  2625            char *soname = 0;
  2626            LDFILE *libptr = NULL;
  2627            struct prefix_list *pl;
  2628            ARCHDR ah;
  2629            idx += strlen (implib) + 1;
  2630            if (!implib[0])
  2631              continue;
  2632            idx += strlen (impmem) + 1;
  2633            if (*implib == '/')
  2634              {
  2635                if (access (soname, R_OK) == 0)
  2636                  soname = implib;
  2637              }
  2638            else
  2639              {
  2640                char *temp = alloca (libpath.max_len + strlen (implib) + 1
);
  2641                for (pl = libpath.plist; pl; pl = pl->next)
  2642                  {
  2643                    strcpy (temp, pl->prefix);
  2644                    strcat (temp, implib);
  2645                    if (access (temp, R_OK) == 0)
  2646                      {
  2647                        soname = temp;
  2648                        break;
  2649                      }
  2650                  }
  2651              }
  2652            if (!soname)
  2653              {
  2654                fatal ("%s: library not found", implib);
  2655                continue;
  2656              }


 This doesn't seem quite right to me. Indeed, initially (at line
2625), soname is set to 0. If implib begins with a slash (which is the
case here), it tests the existance of the file named soname. This
happens at line 2635. At this point, soname is still a NULL pointer!
If the file exists, soname because implib (line 2636). However, this
branch is never taken, as adress 0 probably just contains random
characters denoting a non-existant file :-) Thus soname will stay 0.
Execution then continues at line 2652 where the error message is
generated.
 The intent was probably to find out whether implib was readable, and
not soname.

 However, after changing line 2635 from 'if (access (soname, R_OK) ==
0)' to 'if (access (implib, R_OK) == 0)' the error message becomes:
  collect2: /: not an import library

 After changing line 2630 from if (!implib[0]) to if (!implib[0] ||
(implib[0]=='/' && !implib[1])), the message becomes:
 collect2: unix: library not found

 Obviously, there's no easy way to convince ld that /unix is not a
library existing in filespace but rather the kernel. At this point I
gave up and went back to 2.6.3's ld.

 Alain