Re: Sort relocs output by ld -r

[Andreas Krebbel <krebbel at linux dot vnet dot ibm dot com>]

On Thu, Dec 04, 2014 at 05:52:10PM +1030, Alan Modra wrote:
> Applying to master.
>
> bfd/
>       PR 17666
>       * elflink.c: Include bfd_stdint.h.
>       (cmp_ext32l_r_offset, cmp_ext32b_r_offset,
>       cmp_ext64l_r_offset, cmp_ext64b_r_offset): New functions.
>       (elf_link_adjust_relocs): Sort relocs.  Free reloc hashes after
>       sorting invalidates.
> ld/testsuite/
>       * ld-powerpc/vxworks-relax.rd: Update for reloc sorting.
>       * ld-powerpc/vxworks-relax-2.rd: Likewise.
>       * ld-sh/sh64/reldl32.rd: Likewise.
>       * ld-sh/sh64/reldl64.rd: Likewise.

This seems to be break the linker TLS optimizations on S/390 :(

When downgrading from GD to IE model we rewrite the call to
__tls_get_offset to a 64 bit load instruction. This relies on the fact
that the additional relocation for the call target has already been
executed when doing the rewrite.

   f1018:       e3 20 d0 00 00 04       lg      %r2,0(%r13)
   f101e:       c0 e5 00 00 00 00       brasl   %r14,f101e <__res_init+0x1e>
                        f101e: R_390_TLS_GDCALL __libc_resp
                        f1020: R_390_PLT32DBL   __tls_get_offset+0x2

0000000f1020  39f6c00000014 R_390_PLT32DBL    0000000000000000 __tls_get_offset + 2
0000000f101e  3afb700000026 R_390_TLS_GDCALL  0000000000000008 __libc_resp + 0

Due to the reloc sorting the order changed an the PLT32DBL reloc is
executed after the rewrite and overwrites part of the load instruction
with garbage.

0000000f101e  3afb700000026 R_390_TLS_GDCALL  0000000000000008 __libc_resp + 0
0000000f1020  39f6c00000014 R_390_PLT32DBL    0000000000000000 __tls_get_offset + 2

The attached patch mimics what you did for Mips and disables the
sorting for relocations against code section.  Does that look like a
reasonable fix for this case as well?

Alternatively we might also try to get rid of this dependency.

Bye,

-Andreas-

bfd/
2015-03-13  Andreas Krebbel  <krebbel@linux.vnet.ibm.com>

	* elf-s390-common.c (elf_s390_elf_sort_relocs_p): Don't sort
	relocs against code sections.
	* elf32-s390.c: Define elf_backend_sort_relocs_p.
	* elf64-s390.c: Likewise.


diff --git a/bfd/elf-s390-common.c b/bfd/elf-s390-common.c
index 6fd1027..462da16 100644
--- a/bfd/elf-s390-common.c
+++ b/bfd/elf-s390-common.c
@@ -242,3 +242,15 @@ elf_s390_add_symbol_hook (bfd *abfd,

   return TRUE;
 }
+
+/* Whether to sort relocs output by ld -r or ld --emit-relocs, by
+   r_offset.  Don't do so for code sections.  We want to keep ordering
+   of GDCALL / PLT32DBL for TLS optimizations as is.  On the other
+   hand, elf-eh-frame.c processing requires .eh_frame relocs to be
+   sorted.  */
+
+static bfd_boolean
+elf_s390_elf_sort_relocs_p (asection *sec)
+{
+  return (sec->flags & SEC_CODE) == 0;
+}
diff --git a/bfd/elf32-s390.c b/bfd/elf32-s390.c
index f5403e2..0127eab 100644
--- a/bfd/elf32-s390.c
+++ b/bfd/elf32-s390.c
@@ -4028,6 +4028,7 @@ elf32_s390_merge_private_bfd_data (bfd *ibfd, bfd *obfd)
 #define elf_backend_grok_prstatus	      elf_s390_grok_prstatus
 #define elf_backend_plt_sym_val		      elf_s390_plt_sym_val
 #define elf_backend_add_symbol_hook           elf_s390_add_symbol_hook
+#define elf_backend_sort_relocs_p             elf_s390_elf_sort_relocs_p

 #define bfd_elf32_mkobject		elf_s390_mkobject
 #define elf_backend_object_p		elf_s390_object_p
diff --git a/bfd/elf64-s390.c b/bfd/elf64-s390.c
index e002f06..f06b58a 100644
--- a/bfd/elf64-s390.c
+++ b/bfd/elf64-s390.c
@@ -3840,6 +3840,7 @@ const struct elf_size_info s390_elf64_size_info =
 #define elf_backend_init_index_section	      _bfd_elf_init_1_index_section
 #define elf_backend_plt_sym_val		      elf_s390_plt_sym_val
 #define elf_backend_add_symbol_hook           elf_s390_add_symbol_hook
+#define elf_backend_sort_relocs_p             elf_s390_elf_sort_relocs_p

 #define bfd_elf64_mkobject		elf_s390_mkobject
 #define elf_backend_object_p		elf_s390_object_p