Cross-linking woes: what am I doing wrong?

[Bernd Jendrissek <berndj at prism dot co dot za>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hello, me again

I used to think my problem was with the mingw GCC, it being patched and
all.  But now I think I'll start blaming ld for my woes.  Let's see...

I have two files, foo.c and bar.c:

foo.c:
void shiftit(unsigned char *buf, int n)
{
	int i;

	for (i = 0; i < n; i++) {
		buf[i] >>= 1;
	}
}

bar.c:
void shiftit(unsigned char *buf, int n);

char name[] = "foobar";

int main()
{
	shiftit(name, 5);
}
/usr/cross/bin/i386-mingw32-objdump -d -r bar.o:

bar.o:     file format pe-i386

Disassembly of section .text:

0000000000000000 <_main>:
   0:	55                   	push   %ebp
   1:	31 c0                	xor    %eax,%eax
   3:	89 e5                	mov    %esp,%ebp
   5:	83 ec 08             	sub    $0x8,%esp
   8:	83 e4 f0             	and    $0xfffffff0,%esp
   b:	e8 fc ff ff ff       	call   c <_main+0xc>
			c: DISP32	__alloca
  10:	e8 fc ff ff ff       	call   11 <_main+0x11>
			11: DISP32	___main
  15:	c7 04 24 00 00 00 00 	movl   $0x0,(%esp,1)
			18: dir32	.data
  1c:	b8 05 00 00 00       	mov    $0x5,%eax
  21:	89 44 24 04          	mov    %eax,0x4(%esp,1)
  25:	e8 fc ff ff ff       	call   26 <_main+0x26>
			26: DISP32	_shiftit
  2a:	89 ec                	mov    %ebp,%esp
  2c:	5d                   	pop    %ebp
  2d:	c3                   	ret    
	...
Disassembly of section .text:

0000000000000000 <.text>:

	[same stuff again]

(Why are there multiple copies of .text in bar.o???)

So in bar.o, we can all see the DISP32 reloc to shiftit; presumably the
diff from the reloc offset (26) to its value (shiftit) has to get added
to the data already there (-4), to yield the PC relative offset to
shiftit from the *next* instruction.  Am I right so far?  (I know I'm
right about how i386 jumps and calls work; I mean am I right about how
the relocation is supposed to work?)

But now, when I link foo.o and bar.o into an executable:

/usr/cross/bin/i386-mingw32-ld -Bdynamic -o foo.exe /usr/cross/lib/gcc-lib/i386-mingw32/3.3.3/../../../../i386-mingw32/lib/crt2.o /usr/cross/lib/gcc-lib/i386-mingw32/3.3.3/crtbegin.o -L/usr/cross/lib/gcc-lib/i386-mingw32/3.3.3 -L/usr/cross/lib/gcc-lib/i386-mingw32/3.3.3/../../../../i386-mingw32/lib foo.o bar.o -lmingw32 -lgcc -lmoldname -lmingwex -lmsvcrt -luser32 -lkernel32 -ladvapi32 -lshell32 -lmingw32 -lgcc -lmoldname -lmingwex -lmsvcrt /usr/cross/lib/gcc-lib/i386-mingw32/3.3.3/crtend.o

the relocation seems to get screwed up:

/usr/cross/bin/i386-mingw32-objdump -d -r foo.exe:

foo.exe:     file format efi-app-ia32

Disassembly of section .text:

	...

0000000000401ac0 <_shiftit>:
  401ac0:	55                   	push   %ebp
	...

0000000000401b40 <_main>:
  401b40:	55                   	push   %ebp
  401b41:	31 c0                	xor    %eax,%eax
  401b43:	89 e5                	mov    %esp,%ebp
  401b45:	83 ec 08             	sub    $0x8,%esp
  401b48:	83 e4 f0             	and    $0xfffffff0,%esp
  401b4b:	e8 cc 04 00 00       	call   40201c <LBE7+0xb>
  401b50:	e8 b7 01 00 00       	call   401d0c <LBE3+0x1>
  401b55:	c7 04 24 00 30 40 00 	movl   $0x403000,(%esp,1)
  401b5c:	b8 05 00 00 00       	mov    $0x5,%eax
  401b61:	89 44 24 04          	mov    %eax,0x4(%esp,1)
  401b65:	e8 52 ff ff ff       	call   401abc <___do_sjlj_init+0x3c>

BANG!  The call misses its intended target by 4 bytes, as if the
pre-reloc instruction
  25:	e8 fc ff ff ff       	call   26 <_main+0x26>
in bar.o should have been
  25:	e8 00 00 00 00       	call   26 <_main+0x26>

This off-by-four breakage is everywhere, of course; including the
mingw-runtime, so my executables never even get a chance to start up.
(They crash in CrtDllMainStartup or whatever it's called.)

What the hell is going on here?  And more to the (my) point, what do I
have to do to be able to cross-compile to Windows?  (Everyone else using
MinGW seems to be doing just fine; what makes me so special???)

- -- 
http://voyager.abite.co.za/~berndj/ (up again for now - yay!)
"There are few or no amazing magicians in the world. Precious few
people are skilled enough to walk in on the blazing ruins of an
arbitrary sinking project, patch the holes and make it into a graceful
ocean liner. People that have this skill ought to be designing
projects correctly from the ground up, not dousing the flames of
someone else's train wreck."

   - Lewin Edwards (comp.arch.embedded)
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.0.4 (GNU/Linux)
Comment: For info see http://www.gnupg.org

iD8DBQFAe/Ew/FmLrNfLpjMRAhnIAKCCxngA7irDZA5brKhBgAZGBbo6NgCfaAju
WAMRJ/0/DkpFI526EOj2fNA=
=akg8
-----END PGP SIGNATURE-----