[llvm-dev] Unwind info in .debug_frame incorrect for addresses beyond function epilog

Sandeep Raju via llvm-dev llvm-dev at lists.llvm.org
Tue Aug 18 14:31:42 PDT 2015


Hi All,

We discovered an issue with the unwind info generated by LLVM in the
.debug_frame section. It appears that the cfa_offset for instructions
beyond the function epilog is incorrect.

The problem is best depicted with the following testcase:

$ cat t.c
int foo(int i, int j, int k, int l, int m, int n, int o, int p, int q)
{
   return 0;
}

int main()
{
   return foo (1,2,3,4,5,6,7,8,9);
}

Compiled thusly:

$ clang --target=i686-pc-linux -fomit-frame-pointer -m32 -march=i686
-fPIC -g -o t t.c

Running it in a debugger (look for annotations "<====" ):

(gdb) file t
Reading symbols from t...done.
(gdb) disassemble foo
Dump of assembler code for function foo:
   0x08048380 <+0>:     push   %ebp
   0x08048381 <+1>:     push   %ebx
   0x08048382 <+2>:     push   %edi
   0x08048383 <+3>:     push   %esi
   0x08048384 <+4>:     sub    $0x34,%esp
   0x08048387 <+7>:     mov    0x68(%esp),%eax
   0x0804838b <+11>:    mov    0x64(%esp),%ecx
   0x0804838f <+15>:    mov    0x60(%esp),%edx
   0x08048393 <+19>:    mov    0x5c(%esp),%esi
   0x08048397 <+23>:    mov    0x58(%esp),%edi
   0x0804839b <+27>:    mov    0x54(%esp),%ebx
   0x0804839f <+31>:    mov    0x50(%esp),%ebp
   0x080483a3 <+35>:    mov    %eax,0xc(%esp)
   0x080483a7 <+39>:    mov    0x4c(%esp),%eax
   0x080483ab <+43>:    mov    %eax,0x8(%esp)
   0x080483af <+47>:    mov    0x48(%esp),%eax
   0x080483b3 <+51>:    mov    %eax,0x4(%esp)
   0x080483b7 <+55>:    xor    %eax,%eax
   0x080483b9 <+57>:    mov    %eax,(%esp)
   0x080483bc <+60>:    mov    0x4(%esp),%eax
   0x080483c0 <+64>:    mov    %eax,0x30(%esp)
   0x080483c4 <+68>:    mov    0x8(%esp),%eax
   0x080483c8 <+72>:    mov    %eax,0x2c(%esp)
   0x080483cc <+76>:    mov    %ebp,0x28(%esp)
   0x080483d0 <+80>:    mov    %ebx,0x24(%esp)
   0x080483d4 <+84>:    mov    %edi,0x20(%esp)
   0x080483d8 <+88>:    mov    %esi,0x1c(%esp)
   0x080483dc <+92>:    mov    %edx,0x18(%esp)
   0x080483e0 <+96>:    mov    %ecx,0x14(%esp)
   0x080483e4 <+100>:   mov    0xc(%esp),%ecx
   0x080483e8 <+104>:   mov    %ecx,0x10(%esp)
   0x080483ec <+108>:   mov    (%esp),%eax
   0x080483ef <+111>:   add    $0x34,%esp <==== epilog
   0x080483f2 <+114>:   pop    %esi
   0x080483f3 <+115>:   pop    %edi
   0x080483f4 <+116>:   pop    %ebx
   0x080483f5 <+117>:   pop    %ebp
   0x080483f6 <+118>:   ret
End of assembler dump.
(gdb) break *0x080483ef  <==== break just before the epilog to
illustrate the issue
Breakpoint 1 at 0x80483ef: file t.c, line 3.
(gdb) run
Starting program: /auto/wssanraju-sjc/unw/issue1_3/t
warning: Unable to find dynamic linker breakpoint function.
GDB will be unable to debug shared library initializers
and track explicitly loaded dynamic code.
warning: Could not load shared library symbols for 2 libraries, e.g.
/lib/libc.so.6.
Use the "info sharedlibrary" command to see the complete listing.
Do you need "set solib-search-path" or "set sysroot"?

Breakpoint 1, 0x080483ef in foo (i=1, j=2, k=3, l=4, m=5, n=6, o=7,
p=8, q=9) at t.c:3
3          return 0;
(gdb) bt <==== debugger able to generate correct backtrace
#0  0x080483ef in foo (i=1, j=2, k=3, l=4, m=5, n=6, o=7, p=8, q=9) at t.c:3
#1  0x080484c2 in main () at t.c:8
(gdb) stepi <==== execute one instruction, now we're in the epilog
0x080483f2 in foo (i=8, j=7, k=6, l=5, m=4, n=3, o=2, p=1, q=134513858) at t.c:3
3          return 0;
(gdb) bt <==== debugger not able to generate correct backtrace and
arguments are displayed incorrectly
#0  0x080483f2 in foo (i=8, j=7, k=6, l=5, m=4, n=3, o=2, p=1,
q=134513858) at t.c:3
#1  0x00000006 in ?? ()
#2  0x00000004 in ?? ()
#3  0x00000003 in ?? ()
#4  0x00000002 in ?? ()
#5  0x00000001 in ?? ()
#6  0x00000005 in ?? ()
#7  0x00000009 in ?? ()
#8  0x00000008 in ?? ()
#9  0x00000007 in ?? ()
#10 0x080496f4 in ?? ()
Backtrace stopped: previous frame inner to this frame (corrupt stack?)
(gdb)

Now, we look at the .debug_frame section in the object file for the
same testcase:

$ clang --target=i686-pc-linux -fomit-frame-pointer -m32 -march=i686
-fPIC -c -o t.o -g t.c
$ readelf --debug-dump=frames t.o
The section .debug_frame contains:

00000000 00000010 ffffffff CIE
  Version:               4
  Augmentation:          ""
  Code alignment factor: 4
  Data alignment factor: 0
  Return address column: 1

  DW_CFA_advance_loc: 240 to 000000f0
  DW_CFA_same_value: r12
  DW_CFA_advance_loc4: 2416320528 to 90062100

00000014 00000024 00000000 FDE cie=00000000 pc=00000000..00000077
  DW_CFA_advance_loc: 4 to 00000004
  DW_CFA_def_cfa_offset: 8
  DW_CFA_advance_loc: 4 to 00000008
  DW_CFA_def_cfa_offset: 12
  DW_CFA_advance_loc: 4 to 0000000c
  DW_CFA_def_cfa_offset: 16
  DW_CFA_advance_loc: 4 to 00000010
  DW_CFA_def_cfa_offset: 20
  DW_CFA_advance_loc: 12 to 0000001c <=== for all addresses after
0x1c, the offset is 72 !
  DW_CFA_def_cfa_offset: 72
  DW_CFA_offset: r6 at cfa+0
  DW_CFA_offset: r7 at cfa+0
  DW_CFA_offset: r3 at cfa+0
  DW_CFA_offset: r5 at cfa+0
  DW_CFA_nop
...
...

Disassembly of foo, with relative addresses:

$ objdump --disassemble t.o

t.o:     file format elf32-i386

Disassembly of section .text:


00000000 <foo>:
   0:   55                      push   %ebp
   1:   53                      push   %ebx
   2:   57                      push   %edi
   3:   56                      push   %esi
   4:   83 ec 34                sub    $0x34,%esp
   7:   8b 44 24 68             mov    0x68(%esp),%eax
   b:   8b 4c 24 64             mov    0x64(%esp),%ecx
   f:   8b 54 24 60             mov    0x60(%esp),%edx
  13:   8b 74 24 5c             mov    0x5c(%esp),%esi
  17:   8b 7c 24 58             mov    0x58(%esp),%edi
  1b:   8b 5c 24 54             mov    0x54(%esp),%ebx
  1f:   8b 6c 24 50             mov    0x50(%esp),%ebp
  23:   89 44 24 0c             mov    %eax,0xc(%esp)
  27:   8b 44 24 4c             mov    0x4c(%esp),%eax
  2b:   89 44 24 08             mov    %eax,0x8(%esp)
  2f:   8b 44 24 48             mov    0x48(%esp),%eax
  33:   89 44 24 04             mov    %eax,0x4(%esp)
  37:   31 c0                   xor    %eax,%eax
  39:   89 04 24                mov    %eax,(%esp)
  3c:   8b 44 24 04             mov    0x4(%esp),%eax
  40:   89 44 24 30             mov    %eax,0x30(%esp)
  44:   8b 44 24 08             mov    0x8(%esp),%eax
  48:   89 44 24 2c             mov    %eax,0x2c(%esp)
  4c:   89 6c 24 28             mov    %ebp,0x28(%esp)
  50:   89 5c 24 24             mov    %ebx,0x24(%esp)
  54:   89 7c 24 20             mov    %edi,0x20(%esp)
  58:   89 74 24 1c             mov    %esi,0x1c(%esp)
  5c:   89 54 24 18             mov    %edx,0x18(%esp)
  60:   89 4c 24 14             mov    %ecx,0x14(%esp)
  64:   8b 4c 24 0c             mov    0xc(%esp),%ecx
  68:   89 4c 24 10             mov    %ecx,0x10(%esp)
  6c:   8b 04 24                mov    (%esp),%eax
  6f:   83 c4 34                add    $0x34,%esp
  72:   5e                      pop    %esi
  73:   5f                      pop    %edi
  74:   5b                      pop    %ebx
  75:   5d                      pop    %ebp
  76:   c3                      ret
  77:   90                      nop


Since at the function epilog, the stack pointer is adjusted, it needs
to generate a new offset for instructions after esp?

The use case for this is, when the execution is interrupted in the
function epilog by a signal handler, it will not be able to unwind the
backtrace correctly.

Any pointers on how to fix this issue?

Thanks,
Sandeep


More information about the llvm-dev mailing list