Do this: 
% readelf -wf /lib/tls/libpthread.so.0 > /dev/null 
readelf: Warning: unsupported or unknown DW_CFA_50 
readelf: Warning: unsupported or unknown DW_CFA_50 
 
This also happens on 9.3 already (I haven't tested back earlier).  This 
currently confuses valgrind 3.0 's dwarf reader on SuSE Linux, making it 
work suboptimally. 
 
Same happens when looking at /lib/tls/libc.so.6 .  The problem is in the fsync 
and close syscalls.  I created a buildsystem for STABLE, and looked at the 
.os files: 
 
% readelf -wf cc-nptl/misc/fsync.os 
The section .eh_frame contains: 
 
00000000 00000014 00000000 CIE 
  Version:               1 
  Augmentation:          "zR" 
  Code alignment factor: 1 
  Data alignment factor: -4 
  Return address column: 8 
  Augmentation data:     1b 
 
  DW_CFA_def_cfa: r4 ofs 4 
  DW_CFA_offset: r8 at cfa-4 
  DW_CFA_nop 
  DW_CFA_nop 
 
00000018 00000020 0000001c FDE cie=00000000 pc=00000000..00000065 
  DW_CFA_advance_loc: 12 to 0000000c 
  DW_CFA_register: r3 in r2 
  DW_CFA_advance_loc: 13 to 00000019 
  DW_CFA_restore: r3 
  DW_CFA_advance_loc: 31 to 00000038 
  DW_CFA_def_cfa_offset: 8 
readelf: Warning: unsupported or unknown DW_CFA_50 
 
(same for cc-nptl/io/close.os).  The problem is how NPTL syscalls with 1 
argument are constructed, for which no arch specific one exists.  E.g. 
the fsync syscall is constructed by feeding GCC with this on stdin: 
------------------------------------------ 
#include <sysdep-cancel.h> 
PSEUDO (__libc_fsync, fsync, 1) 
 ret 
PSEUDO_END(__libc_fsync) 
libc_hidden_def (__libc_fsync) 
weak_alias (__libc_fsync, fsync) 
libc_hidden_weak (fsync) 
------------------------------------------ 
 
This expands directly to assembler, parts of it are: 
 
__libc_fsync: .L__libc_fsyncSTART: cmpl $0, %gs:12 
 jne .Lpseudo_cancel 
 .type __fsync_nocancel,@function 
 .globl __fsync_nocancel 
 __fsync_nocancel: movl %ebx, %edx 
 .LSAVEBX1: movl 4(%esp), %ebx 
 movl $118, %eax 
 int $0x80 
 movl %edx, %ebx 
 .LRESTBX1: 
 cmpl $-4095, %eax 
 jae 0f 
 ret 
 .size __fsync_nocancel,.-__fsync_nocancel 
 .Lpseudo_cancel: call __libc_enable_asynccancel 
 movl %eax, %ecx 
 movl %ebx, %edx 
 .LSAVEBX2: movl 4(%esp), %ebx 
 movl $118, %eax 
 int $0x80 
 
 movl %edx, %ebx 
 .LRESTBX2: 
 pushl %eax 
 .LPUSHSTATE: movl %ecx, %eax 
 call __libc_disable_asynccancel 
 popl %eax 
 .LPOPSTATE: cmpl $-4095, %eax 
 jae 0f 
 .Lpseudo_end: 
 
.... 
 
.section .eh_frame,"a",@progbits 
.... 
 .LSTARTFDE: .long .LSTARTFDE-.LSTARTFRAME 
 .long .L__libc_fsyncSTART-. 
 .long .L__libc_fsyncEND-.L__libc_fsyncSTART 
 .uleb128 0 
 .byte 0x40+.LSAVEBX1-.L__libc_fsyncSTART 
 .byte 9 
 .uleb128 3 
 .uleb128 2 
 .byte 0x40+.LRESTBX1-.LSAVEBX1 
 .byte 0xc3 
 .byte 0x40+.LPUSHSTATE-.LRESTBX1 
 .byte 14 
 .uleb128 8 
 .byte 0x40+.LSAVEBX2-.LPUSHSTATE  <<<<<<<<<<<< here 
 .byte 9 
 .uleb128 3 
 .uleb128 2 
 .byte 0x40+.LRESTBX2-.LSAVEBX2 
 .byte 0xc3 
 .byte 0x40+.LPOPSTATE-.LRESTBX2 
 .byte 14 
 .uleb128 4 
 .align 4 
 .LENDFDE: 
 
On the marked line the difference .LSAVEBX2-.LPUSHSTATE is used, but 
.LSAVEBX2 is in front of .LPUSHSTATE, hence it is negative, and in our 
case results in 0x32 overall, i.e. our 50 readelf complains about. 
 
I looked a bit at nptl/sysdeps/unix/sysv/linux/i386/sysdep-cancel.h . 
The body of the syscalls is this: 
  L(pseudo_cancel): 
    CENABLE 
    SAVE_OLDTYPE_##args 
    PUSHCARGS_##args 
    DOCARGS_##args 
    movl $SYS_ify (syscall_name), %eax; 
    ENTER_KERNEL; 
    POPCARGS_##args; 
    POPSTATE_##args 
    cmpl $-4095, %eax; 
    jae SYSCALL_ERROR_LABEL; 
  L(pseudo_end): 
 
Now look at some more macros: 
# define SAVE_OLDTYPE_0 movl %eax, %ecx; 
# define SAVE_OLDTYPE_1 SAVE_OLDTYPE_0 
# define SAVE_OLDTYPE_2 pushl %eax; L(PUSHSTATE): 
 
# define POPSTATE_0 \ 
 pushl %eax; L(PUSHSTATE): movl %ecx, %eax; CDISABLE; popl %eax; L(POPSTATE): 
# define POPSTATE_1     POPSTATE_0 
# define POPSTATE_2     xchgl (%esp), %eax; CDISABLE; popl %eax; L(POPSTATE): 
 
Look at how with 2 or more args .LPUSHSTATE is set before the PUSHCARGS macro 
(which contains the .LSAVEBX2 label for instance, for all args).  But 
with 0 and 1 argument this label is set by POPSTATE, hence after the use 
of the PUSHCARGS macro.  For 0 arguments this doesn't matter it seems, 
as the macro setting up eh_frame for 0 args doesn't refer to .LSAVEBX2. 
 
But EH_FRAME_1 has this expression: 
    .byte 0x40+L(SAVEBX2)-L(PUSHSTATE);         /* DW_CFA_advance_loc+N */ 
hence it needs SAVEBX2 _after_ PUSHSTATE, but that's only the case for 
2 or more args. 
 
So something must be changed, I'm just not sure what.  If someone from the 
glibc maintainers could look at this, it would be nice.