On Fri, May 12, 2006 at 05:52:46PM -0700, H. J. Lu wrote:
> On Fri, May 12, 2006 at 01:54:13PM -0700, H. J. Lu wrote:
> > On Fri, May 12, 2006 at 10:05:12AM -0700, H. J. Lu wrote:
> > > Hi Richard,
> > > 
> > > Your patch:
> > > 
> > > http://sourceware.org/ml/binutils/2004-11/msg00226.html
> > > 
> > > assumes that CIE/FDE are aligned at the pointer size. But it isn't
> > > necessarily true. See
> > > 
> > > http://sources.redhat.com/bugzilla/show_bug.cgi?id=2657
> > > 
> > 
> > This patch fixes 2 PRs 2655/2657. PR 2655 is a gcc bug
> > 
> > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27576
> > 
> > PR 2657 is we don't properly shrink CIE/FDE.
> > 
> 
> It turns out that _bfd_elf_discard_section_eh_frame will always pad
> CIE/FDE record to the pointer size boundary since it calls
> size_of_output_cie_fde to set the CIE/FDE record size. Even if the next
> .eh_frame section is marked for 4 byte alignment, it always will be
> aligned at the pointer size. Fixing PR 2657 will also fix PR 2655.
> I am enclosing an upated patch and 2 testcases for x86-64.
> 
> 

We shouldn't pad the .eh_frame section to its section alignment. We
only need to make sure that CIE/FDE records are aligned at pointer
size. Otherwise, we may generate bad .eh_frame section if the input
section alignment isn't pointer size. I am enclosing a new patch. I
also added a new testcase.


H.J.