linux and bfd_gas

linux and bfd_gas

[Rick Sladkey]

Recently using bfd for gas on linux was made preferred in the
configuration.  This doesn't work for a.out because the relocations
for PC-relative function calls (e.g.) don't come out right.

For example:
$ cat main.s
	.file	"main.c"
gcc2_compiled.:
___gnu_compiled_c:
.text
	.align 4
.globl _main
_main:
	pushl %ebp
	movl %esp,%ebp
	call ___main
	call _sync
L1:
	movl %ebp,%esp
	popl %ebp
	ret
$ as-940906 main.s -o main.o                             # The stock 0906 gas.
$ objdump --disassemble main.o

main.o:     file format a.out-i386-linux

Disassembly of section .text:
00000000 <_main> pushl  %ebp
00000001 <_main+1> movl   %esp,%ebp
00000003 <_main+3> call   00000004 <_main+4>             # Here is the problem.
00000008 <_main+8> call   00000009 <_main+9>             # Here is the problem.
0000000d <_main+d> movl   %ebp,%esp
0000000f <_main+f> popl   %ebp
00000010 <_main+10> ret    
...
Disassembly of section .data:

$ as main.s -o main.o                                    # My fixed version.
$ objdump --disassemble main.o

test/as/main.o.good:     file format a.out-i386-linux

Disassembly of section .text:
00000000 <_main> pushl  %ebp
00000001 <_main+1> movl   %esp,%ebp
00000003 <_main+3> call   00000000 <_main>               # Here it is correct.
00000008 <_main+8> call   00000000 <_main>               # Here it is correct.
0000000d <_main+d> movl   %ebp,%esp
0000000f <_main+f> popl   %ebp
...
Disassembly of section .data:

$

The problem is that DISP32 relocations on i386-aout are relative to
the instruction and not to the section.

Here is a patch that fixes the problem for me but I am not positive
that it is correct for all targets.
-----
Wed Sep  7 03:59:18 1994  Rick Sladkey  (jrs@world.std.com)

	* write.c (write_relocs): Subtract the address from the relocation
	addend for pc_relative relocations that aren't pcrel_offset.

*** gas-940906/gas/write.c-dist	Fri Aug 26 17:52:45 1994
--- gas-940906/gas/write.c	Wed Sep  7 03:37:49 1994
***************
*** 702,707 ****
--- 702,715 ----
  	  /* bfd_perform_relocation screws this up */
  	  reloc->addend += reloc->address;
  	}
+ 
+       if (reloc->howto->pcrel_offset == false
+ 	  && reloc->howto->pc_relative == true)
+ 	{
+ 	  /* Targets like i386-aout use the addend instead of pcrel_offset.  */
+ 	  reloc->addend -= reloc->address;
+ 	}
+ 
        /* Pass bogus address so that when bfd_perform_relocation adds
  	 `reloc->address' back in, it'll come up with `data', which is where
  	 we want it to operate.  We can't just do it by fudging reloc->address,

Re: linux and bfd_gas

[Ian Lance Taylor]

   From: jrs@world.std.com (Rick Sladkey)
   Date: Wed, 7 Sep 1994 04:05:09 EDT

   Recently using bfd for gas on linux was made preferred in the
   configuration.  This doesn't work for a.out because the relocations
   for PC-relative function calls (e.g.) don't come out right.

Actually, this just broke on September 1 due to a change to make the
linux targeted gas default to defining BFD_ASSEMBLER.  When
BFD_ASSEMBLER is defined, gas reloc handling is basically broken, in
that making any particular target work requires a lot of senseless
twiddling.  This twiddling has not been done for the i386-aout target.

You can make it work by using the --disable-bfd-assembler argument
when you configure gas (this won't work for a cross compiler without a
patch I just made to configure, but it should work for native).

If you want to fix up the i386-aout target for BFD_ASSEMBLER, the
places to change are md_apply_fix and tc_gen_reloc in tc-i386.c.  Your
proposed patch to write.c will break other a.out targets.

Ian

Re: linux and bfd_gas

[Ken Raeburn]

I will switch the linux a.out configuration back, so that it doesn't
set bfd_gas.  I had done it because the non-bfd version silently
ignored "@PLT" and the like in the input file, where the bfd version
would complain about not being able to generate the relocations.

Generating correct results for correct input files is obviously more
important than warning about code the assembler shouldn't be getting
fed anyways.