This is the mail archive of the binutils@sourceware.org mailing list for the binutils project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

mips64-elf .rel.dyn sections


The mips64-elf toolchain is creating .rel.dyn sections, even though it does not support dynamic linking, and does not create dynamic sections.  I attached a testcase.

Both old and new binutils call mips_elf_rel_dyn_section to create the .rel.dyn section when a R_MIPS_32 reloc is seen, in case it is needed.  In old binutils, this gets removed as a zero-size section in _bfd_mips_elf_size_dynamic_sections.  In new binutils, we call allocate_dynrelocs first, which creates dummy relocs in the .rel.dyn section, and now it can't be removed as it is no longer a zero size section.

The bug seems to be in allocate_dynrelocs in elfxx-mips.c, which does
  if (! info->relocatable
      && hmips->possibly_dynamic_relocs != 0
      && (h->root.type == bfd_link_hash_defweak
          || !h->def_regular
          || info->shared))

However, we are looking at a reference here, not a definition, so none of the h->def_* bits are set, though we do have h->ref_* bits set.  

Breakpoint 2, allocate_dynrelocs (h=0x817c8ec, inf=0x81692e0) at ../../new-combined/bfd/elfxx-mips.c:8551
(gdb) print *h
$1 = {root = {root = {next = 0x817c574, string = 0x8198b86 "i", hash = 13803625\
}, type = bfd_link_hash_defined, non_ir_ref = 0, u = {undef = {next = 0x0, abfd\
 = 0x8193a54}, def = {next = 0x0, section = 0x8193a54, value = 0}, i = {next = \
0x0, link = 0x8193a54, warning = 0x0}, c = {next = 0x0, p = 0x8193a54, size = 0\
}}}, indx = -1, dynindx = -1, got = {refcount = -1, offset = 184467440737095516\
15, glist = 0xffffffff, plist = 0xffffffff}, plt = {refcount = -1, offset = 184\
46744073709551615, glist = 0xffffffff, plist = 0xffffffff}, size = 4, type = 1,\
 other = 0, target_internal = 0, ref_regular = 1, def_regular = 0, ref_dynamic \
= 0, def_dynamic = 0, ref_regular_nonweak = 1, dynamic_adjusted = 0, needs_copy\
 = 0, needs_plt = 0, non_elf = 0, hidden = 0, forced_local = 0, dynamic = 0, ma\
rk = 0, non_got_ref = 0, dynamic_def = 0, dynamic_weak = 0, pointer_equality_ne\
eded = 0, unique_global = 0, dynstr_index = 0, u = {weakdef = 0x0, elf_hash_val\
ue = 0}, verinfo = {verdef = 0x0, vertree = 0x0}, vtable = 0x0}
(gdb)

This suggests that maybe the line
    || !h->def_regular
Should instead be
    || (!h->def_regular && !h->ref_regular)
I am not sure if the h->ref_regular_nonweak bit needs to be tested also or instead of ref_regular.

I don't have any mips-linux system using PLTs and copy-relocs yet, so I don't have an easy way to test that.  I also don't have a binutils assignment yet, so I can't formally contribute a patch for this.  I can open a bugzilla bug report if necessary.

Jim

Attachment: mips64-elf-rel-dyn.txt
Description: mips64-elf-rel-dyn.txt


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]