[Tinycc-devel] wrong PLT generated by tcc 0.9.16

[Basile STARYNKEVITCH]

Dear All,

It seems to me that tcc generate a wrong Procedure Linkage Table.

I am on Linux Debian Sid (x86=Athlon) - libc6 2.3.1-8, gcc 2.3.2,
binutils 2.13.90.0.16-1, kernel 2.4.20

I have the following two test files

################################################################
// file esstccdl.c
// emacs Time-stamp: <2002 De 31 Tue 14h15:24 cet {esstccdl.c} >
// prcsid $Id: esstccdl.c 1.1 Tue, 31 Dec 2002 16:31:38 +0100 basile $

#include <stdlib.h>
#include <stdio.h>
#include <dlfcn.h>

int
main (int argc, char **argv)
{
  char linbuf[300];
  char *funam = "fun";
  char *sonam = "./esstccso.so";
  void *h = 0;
  FILE *f = 0;
  void (*fptr) ();
  if (argc > 2)
    funam = argv[1];
  if (argc > 3)
    sonam = argv[2];
  puts (__FILE__ " compiled at " __DATE__ " on " __TIME__ "\n");
  h = dlopen (sonam, RTLD_NOW);
  if (!h)    {
    fprintf (stderr, "dlopen %s failed : %s\n", sonam, dlerror ());
    exit (1);
  };
  printf ("dlopen %s => %p\n", sonam, h);
  fptr = dlsym (h, funam);
  if (!fptr) {
    fprintf (stderr, "dlsym %s failed : %s\n", funam, dlerror ());
    exit (1);
  };
  printf ("dlsym %s => %p\n", funam, (void *) fptr);
  f = fopen("/proc/self/maps", "r");
  printf("\n*** our (pid %d) memory map:\n", (int)getpid());
  while (!feof(f)) {
    memset(linbuf, 0, sizeof(linbuf));
    fgets(linbuf, sizeof(linbuf)-1, f);
    fputs(linbuf, stdout);
  };
  fclose(f);
  puts("*** end of map\n");
  printf("before call fptr=%p\n", (void*)fptr);
  (*fptr) ();
  printf ("end test\n");
  return 0;
} // end of main

// eof esstccdl.c
################################################################
// file esstccso.c
// emacs Time-stamp: <2002 De 31 Tue 12h40:05 cet {esstccso.c} >
// prcsid $Id: esstccso.c 1.1 Tue, 31 Dec 2002 16:31:38 +0100 basile $

#include <stdio.h>
#include <unistd.h>

void fun(void) {
  puts (__FILE__ " compiled at " __DATE__ " on " __TIME__ "\n");
  printf("pid %d\n", (int)getpid());
}


// eof esstccso.c
################################################################


The test file esstccdl.c is compiled with
   tcc -o esstccdl esstccdl.c -ldl
(adding the -g flag dont help and dont harm)
esstccdl.c dlopen the esstccso.so file and indirectly calls the fun
function there using dlsym.

The test file esstccso.c should be compiled into a shared object. With
tcc -o esstccso.so -shared esstccso.c
the resulting file has a bad PLT entry at file offset 0xff0; when
running ./esstccdl which dlopens it, a SIGSEGV occurs there (in
process adress 0x40013ff0 on my system, where the esstccso.so mapping
starts at 0x40013000).

If I compile esstccso.c with gcc using
gcc -o esstccso.so -shared esstccso.c
the resulting file has a good PLT and the ./esstccdl program runs ok.

using objdump and readelf

The gcc generated esstccso.so has 

       
Contents of section .plt:
 0558 ffb30400 0000ffa3 08000000 00000000  ................
 0568 ffa30c00 00006800 000000e9 e0ffffff  ......h.........
 0578 ffa31000 00006808 000000e9 d0ffffff  ......h.........

Relocation section '.rel.plt' at offset 0x530 contains 2 entries:
 Offset     Info    Type            Sym.Value  Sym. Name
0000180c  00001e07 R_386_JUMP_SLOT   00000000   __cxa_finalize
00001810  00002207 R_386_JUMP_SLOT   00000000   _Jv_RegisterClasses



The tcc generated esstccso.so has

Hex dump of section '.plt':
  0x00000fe0 00000000 00002128 25ff0000 212435ff .5$!...%(!......
  0x00000ff0 ffffffe0 e9000000 00680000 212025ff .% !..h.........
  0x00001000 ffffffd0 e9000000 08680000 212025ff .% !..h.........
  0x00001010 ffffffc0 e9000000 10680000 212025ff .% !..h.........
  0x00001020 080724a0 00000018 00000091 00000048 H............$..
  0x00001030 00000001 00000009 0000003c 00000020  ...<...........
  0x00001040 00000000 00000001 00000000 00000006 ................
  0x00001050 00000000 00000f85 00000f85 00000018 ................
  0x00001060 00000000 00000000 080755b8 00000000 .....U..........

it has no relocation of .plt (ie no .rel.plt section)

I have no idea of how to correct tcc for this.

-- 

Basile STARYNKEVITCH         http://starynkevitch.net/Basile/ 
email: basile<at>starynkevitch<dot>net 
alias: basile<at>tunes<dot>org 
8, rue de la Faïencerie, 92340 Bourg La Reine, France