[Bug gold/20512] New: Assertion failure with relocation against symbol #0

[aabel at google dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=20512

            Bug ID: 20512
           Summary: Assertion failure with relocation against symbol #0
           Product: binutils
           Version: unspecified
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: gold
          Assignee: ccoutant at gmail dot com
          Reporter: aabel at google dot com
                CC: ian at airs dot com
  Target Milestone: ---

Here is a minimal example:

bug.c:
-----------------------------------------------------------------
extern int __a;

static int __attribute__((__used__)) A (void) 
{
        return &__a>(int*)1;
}
-----------------------------------------------------------------

bug.lds:
-----------------------------------------------------------------
OUTPUT_FORMAT("elf64-x86-64")
SECTIONS
{
  __a = .;
}
----------------------------------------------------------------

Commands:
$ gcc -fPIC -O2 -c -o bug_1.o bug.c
$ ld -r -o bug_2.o -T bug.lds bug_1.o
$ objcopy bug_2.o bug_3.o
$ ld -shared -o libbug.so bug_3.o


If the program is compiled like this, the assertion "this->local_sym_index_ !=
0" in line 1144 in output.cc fails. 

If we use bug_2.o instead of bug_3.o in the last step, the problem does not
occur. One difference between bug_2.o and bug_3.o is the following:

readelf -r bug_2.o:
Relocation section '.rela.text' at offset 0xe8 contains 1 entries:
  Offset          Info           Type           Sym. Value    Sym. Name +
Addend
000000000005  000500000009 R_X86_64_GOTPCREL 0000000000000000 __a - 5

readelf -r bug_3.o:
Relocation section '.rela.text' at offset 0x570 contains 1 entries:
  Offset          Info           Type           Sym. Value    Sym. Name +
Addend
000000000005  000000000009 R_X86_64_GOTPCREL                    -5


So objcopy changes the relocation from a relocation against symbol #5 (which is
__a, and has the value 0) to a relocation against symbol #0 (which is required
to be a symbol with value 0). According to
https://sourceware.org/ml/binutils/2004-03/msg00020.html "a reloc against that
symbol is well-defined."

When gold tries to write out this relocation against symbol 0 (i.e.,
local_sym_index_==0), the assertion mentioned above fails.

-- 
You are receiving this mail because:
You are on the CC list for the bug.