Bug 17485

This change:

commit 58eb0862c3471b931bf3412d6d02bac447d5a653
Author: Roland McGrath <roland@hack.frob.com>
Date:   Fri Jun 27 09:58:51 2014 -0700

    MIPS: Consolidate NPTL/non versions of vfork

regressed MIPS builds of librt.so, depending on how the BFD linker
internals work out, which generally depends on multilib options, either
an empty entry is left in the dynamic symbol table or a build failure
is triggered like this:

.../mips-linux-gnu/bin/ld: .../nptl/libpthread.so: invalid string offset 2765592330 >= 5154 for section `.dynstr'
.../nptl/libpthread.so: could not read symbols: File format not recognized
collect2: error: ld returned 1 exit status

This is a bug in GNU LD and also GAS, but I think we have a bug too, in
the MIPS version of pt-vfork.S, that triggers this obscure corner case.

Before the change it assembled to have a symbol table like this:

Symbol table '.symtab' contains 20 entries:
   Num:    Value  Size Type    Bind   Vis      Ndx Name
     0: 00000000     0 NOTYPE  LOCAL  DEFAULT  UND
     1: 00000000     0 SECTION LOCAL  DEFAULT    1
     2: 00000000     0 SECTION LOCAL  DEFAULT    3
     3: 00000000     0 SECTION LOCAL  DEFAULT    4
     4: 00000001     0 NOTYPE  LOCAL  DEFAULT  ABS LOCALSZ
     5: 00000018     0 NOTYPE  LOCAL  DEFAULT  ABS FRAMESZ
     6: 00000014     0 NOTYPE  LOCAL  DEFAULT  ABS GPOFF
     7: 00000000     0 SECTION LOCAL  DEFAULT   11
     8: 00000000     0 SECTION LOCAL  DEFAULT   13
     9: 00000000     0 SECTION LOCAL  DEFAULT    9
    10: 00000000     0 SECTION LOCAL  DEFAULT   14
    11: 00000000     0 SECTION LOCAL  DEFAULT   16
    12: 00000000     0 SECTION LOCAL  DEFAULT    5
    13: 00000000     0 SECTION LOCAL  DEFAULT    6
    14: 00000000     0 SECTION LOCAL  DEFAULT    7
    15: 00000000     0 SECTION LOCAL  DEFAULT   18
    16: 00000000   112 FUNC    GLOBAL DEFAULT    1 __vfork
    17: 00000000     0 OBJECT  GLOBAL DEFAULT  UND _gp_disp
    18: 00000000     0 NOTYPE  GLOBAL DEFAULT  UND __syscall_error
    19: 00000000   112 FUNC    WEAK   DEFAULT    1 vfork

With the change in place the symbol table is like this instead:

Symbol table '.symtab' contains 21 entries:
   Num:    Value  Size Type    Bind   Vis      Ndx Name
     0: 00000000     0 NOTYPE  LOCAL  DEFAULT  UND
     1: 00000000     0 SECTION LOCAL  DEFAULT    1
     2: 00000000     0 SECTION LOCAL  DEFAULT    3
     3: 00000000     0 SECTION LOCAL  DEFAULT    4
     4: 00000001     0 NOTYPE  LOCAL  DEFAULT  ABS LOCALSZ
     5: 00000018     0 NOTYPE  LOCAL  DEFAULT  ABS FRAMESZ
     6: 00000014     0 NOTYPE  LOCAL  DEFAULT  ABS GPOFF
     7: 00000000     0 SECTION LOCAL  DEFAULT   11
     8: 00000000     0 SECTION LOCAL  DEFAULT   13
     9: 00000000     0 SECTION LOCAL  DEFAULT    9
    10: 00000000     0 SECTION LOCAL  DEFAULT   14
    11: 00000000     0 SECTION LOCAL  DEFAULT   16
    12: 00000000     0 SECTION LOCAL  DEFAULT    5
    13: 00000000     0 SECTION LOCAL  DEFAULT    6
    14: 00000000     0 SECTION LOCAL  DEFAULT    7
    15: 00000000     0 SECTION LOCAL  DEFAULT   18
    16: 00000000   140 FUNC    GLOBAL DEFAULT    1 __vfork
    17: 00000000     0 OBJECT  GLOBAL DEFAULT  UND _gp_disp
    18: 00000000     0 NOTYPE  GLOBAL DEFAULT  UND __syscall_error
    19: 00000000   140 FUNC    WEAK   DEFAULT    1 vfork
    20: 00000000     0 NOTYPE  GLOBAL DEFAULT  UND vfork@GLIBC_2.0

-- notice the undefined versioned entry for `vfork@GLIBC_2.0'.  This is
what confuses LD and also shouldn't have been produced by GAS in the
first place.

There's no further reference to that symbol, i.e. no relocation
anywhere across this object file, which is why LD removes it from
libpthread.so's dynsym table, however it doesn't realise early enough
it shouldn't be propagating the reference to the dynsym table in the
first place and as a result it leaves an uninitialised slot that may or
may not be empty.

The ultimate reason for this symbol's creation is a:

	.symver	__libc_vfork, vfork@GLIBC_2.0

directive however without any `__libc_vfork' definition or reference.
This is what the GAS manual has to say about this case:

"  For ELF targets, the `.symver' directive can be used like this:
     .symver NAME, NAME2@NODENAME

"  If the symbol NAME is not defined within the file being assembled,
all references to NAME will be changed to NAME2@NODENAME.  If no
reference to NAME is made, NAME2@NODENAME will be removed from the
symbol table."

so the symbol is supposed not to be emitted.  It appears though that
this bug has been present in GNU binutils since forever and our other
targets define `__libc_vfork', possibly just as an alias to `__vfork',
so I think we should simply do the same for MIPS.  The symbol does not
make it to the dynsym table so such a change will not affect our ABI.