[linuxsh-dev] Re: [linux-sh:02245] Binutils

[NIIBE Y. <gn...@m1...>]

Michael Eager wrote:
 > What problem does this patch fix?

I've encountered following problem.

   Upgraded to glibc 2.2.5 and newer binutils.

   Build whole things.

   Building libgcc.so, libobjc.so, libgcj.so, the assertion in
   elf32-sh.c:sh_elf_relocate_section
	 BFD_ASSERT (off != (bfd_vma) -1);
   failed.  (And with this libobjc.so, half of testsuites failed).

This is because we have relocation of type GOT32 for __cxa_finalize in
crtbeginS.o and it's offset is 0 (ZERO).  The value ZERO is valid
here.  However, when got.offset is 0, it is wrongly overwritten to -1
in elf.c:_bfd_elf_link_hash_copy_indirect, this is cause of assertion
failure.

Note that got.refcount and got.offset is union, and placed in same
memory.  It works fine provided got.offset > 0, but when got.offset
==0, accessing as got.refcount, it looks it's not used (from view of
refcount).

Backend with can_refcount==0 used to work fine.  But it seems for me
that can_refcount==1 is the trend.

In backend with can_refcount==1, got.refcount is used untill last time.
At last time, the role will change to have offset and offset value will
be set.
--