addr2line ends up off by 1 line numbers when -O0 is used

[Britton Kerin]

I'm not sure whether to blame libc (backtrace() function) or gcc or
addr2line, but if I dump the pointers from backtrace() to
address_file then use

  addr2line -e my_exe -f -i <address_file >call_locations

to recover the line numbers, the line numbers are off by 1 statement
for some stack frames when -O0 has been used, but not if -O2 has been
used at compile time.

The documentation for backtrace() suggests that things should work
better with -O0 but I seem
to get the opposite.

I'm on gcc 4.8.1, addr2lines binutils 2.26 and libc 2.11.3-4 debian package:
$ ldd --version
ldd (Debian EGLIBC 2.11.3-4) 2.11.3

I thought this might be related to -fomit-frame-pointer.  I thought it might be
since -O>0 apparently activates this feature, making stack unwinding more
complex.  But adding -fomit-frame-pointer neither fixes the problem when
-O0 is used, nor causes it when -O2 is used.