This is the mail archive of the
binutils@sources.redhat.com
mailing list for the binutils project.
Re: PATCH: Re: EH frame optimization bug
On Thu, Oct 02, 2003 at 06:05:03PM -0700, H. J. Lu wrote:
> On Thu, Oct 02, 2003 at 02:43:45PM -0400, Daniel Jacobowitz wrote:
> > On Tue, Sep 30, 2003 at 03:05:59PM -0400, Daniel Jacobowitz wrote:
> > > Testcase derived from glibc. Take the attached files and link them:
> > > ld -o bin connect.o send.o dumb.o
> > >
> > > readelf -wf bin -> segfault.
> > >
> > > The problem is that we've got GAS-produced .eh_frame sections that look like
> > > this:
> > > [ 5] .eh_frame PROGBITS 00000000 000098 000040 00 A 0 0 4
> > >
> > > 00000000 00000012 00000000 CIE
> > >
> > > 00000016 00000026 0000001a FDE cie=00000000 pc=0000001e..00000074
> > >
> > > Then we have two almost identical object files (built from the same source
> > > file, actually). We optimize them to share a CIE.
> > >
> > > 00000000 00000012 00000000 CIE
> > > 00000016 00000026 0000001a FDE cie=00000000 pc=08048080..080480d6
> > > 00000040 00000026 00000044 FDE cie=00000000 pc=080480e0..08048136
> > >
> > > That second FDE now ends at 0x6a. The next .eh_frame section is align-4.
> > > Two bytes padding get added, and everything blows up.
> > >
> > > Something similar came up a few months ago but I can't find the reference
> > > now. What's the right thing to do here? My instinct says, grow the last
> > > FDE before the padding, but I have no idea how to do that. Is the alignment
> > > of a .eh_frame section mandated?
> >
>
> How about this patch?
It seems reasonable, but I'd have thought that it would be safer/more
useful to ignore old_fill.
>
>
> H.J.
> ---
> 2003-10-02 H.J. Lu <hongjiu.lu@intel.com>
>
> * elf-eh-frame.c (_bfd_elf_write_section_eh_frame): Pad the
> last FDE if needed.
>
> --- bfd/elf-eh-frame.c.pad 2003-08-07 09:04:30.000000000 -0700
> +++ bfd/elf-eh-frame.c 2003-10-02 18:03:43.000000000 -0700
> @@ -1026,6 +1026,38 @@ _bfd_elf_write_section_eh_frame (bfd *ab
> and 3xDW_CFA_nop as pad */
> p += 16;
> }
> + else
> + {
> + int old_fill = (sec->_raw_size
> + & ((1 << sec->alignment_power) - 1));
> + int new_fill = (sec->_cooked_size
> + & ((1 << sec->alignment_power) - 1));
> + int needed = old_fill - new_fill;
> +
> + if (needed)
> + {
> + unsigned long length;
> +
> + if (sec_info->entry[sec_info->count - 1].cie)
> + abort ();
> +
> + needed = (((1 << sec->alignment_power) + needed)
> + & ((1 << sec->alignment_power) - 1));
> +
> + buf = contents + sec_info->entry[sec_info->count - 1].new_offset;
> +
> + /* Update length. */
> + length = bfd_get_32 (abfd, buf) + needed;
> + bfd_put_32 (abfd, length, buf);
> +
> + /* Pad it with DW_CFA_nop */
> + memset (p, 0, needed);
> + p += needed;
> +
> + sec_info->entry[sec_info->count - 1].size += needed;
> + sec->_cooked_size += needed;
> + }
> + }
>
> BFD_ASSERT ((bfd_size_type) (p - contents) == sec->_cooked_size);
>
>
--
Daniel Jacobowitz
MontaVista Software Debian GNU/Linux Developer