[Bug gold/15200] Runtime undefined reference to __exidx_start/_end

[petechou at gmail dot com]

http://sourceware.org/bugzilla/show_bug.cgi?id=15200

--- Comment #9 from pete <petechou at gmail dot com> 2013-03-13 02:03:32 UTC ---
If there is a libfoo.so, it's possible to implement part of functions in
libfoo.so and then link against libfoo.so again to produce a new library. I
think the "linker-defined" symbols should also have higher priority than the
symbols in DSO.

And I also think we should have the same behavior on defining symbol value
whether only_if_ref is set or not?

The patch will affect 4 ways to define output symbols:
1. do_define_in_output_data
2. do_define_in_output_segment
3. do_define_as_constant
4. add_undefined_symbol_from_command_line

For 1, 2, and 3, it makes sense to use the given output section, segment, and
constant to define symbol values if the "linker-defined" symbols have higher
priority.

As for 4, I think it doesn't matter.

-
Thanks,
Pete

(In reply to comment #8)
> Hi Pete,
> 
> 1. My concern of your patch is that it affects all targets, not just
> ARM.  While it may fix the problem on ARM, I cannot speak for authors
> of other backends.
> 2. I am don't have power approve a patch, you need to convince Ian
> Taylor that this is the right thing to do for all targets.
> 3. This bug only happens when DSOs incorrectly export __exidx_start &
> __exidx_end.  These DSOs are considered broken and I strongly
> recommend you moving to a newer NDK version with DSOs that have proper
> visibility.  Both ld & gold no longer export these symbols.  The
> reason for not exporting is explained here in the ld patch:
> 
> http://sourceware.org/ml/binutils/2009-11/msg00367.html
> 
> -Doug
> 
> 
> On Wed, Mar 6, 2013 at 10:43 PM, petechou at gmail dot com
> <address@hidden> wrote:
> > http://sourceware.org/bugzilla/show_bug.cgi?id=15200
> >
> > --- Comment #7 from pete <petechou at gmail dot com> 2013-03-07 06:43:08 
> > UTC ---
> > I can think visibility does matter, but only for exporting the symbol or 
> > not.
> > In the previous case, gold does define the __bss symbols but not use the
> > definition in DSO, so I think we should still define section/segment symbols
> > locally.
> >
> > -Pete
> >
> > (In reply to comment #6)
> >> The symbols do not have same visibility.  gold in general is
> >> compatible with GNU ld.  If you look at src/ld/emulparams/, you will
> >> set that the __bss symbols are exported but the __exidx symbols are
> >> defined hidden.  gold matches the behaviour of ld.  If you look at
> >> defstd.cc in gold, you will again see that some of these symbols are
> >> hidden but some are not.
> >>
> >> -Doug
> >>
> >>
> >> On Wed, Mar 6, 2013 at 9:31 PM, petechou at gmail dot com
> >> <address@hidden> wrote:
> >> > http://sourceware.org/bugzilla/show_bug.cgi?id=15200
> >> >
> >> > --- Comment #5 from pete <petechou at gmail dot com> 2013-03-07 05:31:25 
> >> > UTC ---
> >> > (In reply to comment #4)
> >> >> The symbols __exidx_end & __exidx_start were exported incorrectly in
> >> >> the past.  This is fixed in both ld & gold in the sense that these
> >> >> symbols are defined by not exported.  I think defining these symbols
> >> >> when there is a non-weak definition in a DSO is not the right thing to
> >> >> do.  This is basically multiple definition and usually is an error.
> >> >
> >> > If so, all section/segment symbols should be defined (if needed) by not
> >> > exported?
> >> >
> >> > But when there is a non-weak definition in a DSO, it's still possible to 
> >> > find
> >> > ld.gold define and export segment symbols like __bss_start (See
> >> > gold.defstd.cc:214. only_if_ref is set to false). And there is no 
> >> > multiple
> >> > definition error.
> >> >
> >> > $ readelf -Ds libc.so | grep __bss_start
> >> >   268 311: 0000d898     0 NOTYPE  GLOBAL DEFAULT ABS __bss_start__
> >> >   619 470: 0000d898     0 NOTYPE  GLOBAL DEFAULT ABS __bss_start
> >> >
> >> > $ arm-linux-androideabi-ld.gold -shared -o libplasma.so plasma.o libc.so
> >> >
> >> > $ readelf -s libplasma.so | grep __bss_start
> >> >     73: 00004148     0 NOTYPE  GLOBAL DEFAULT  ABS __bss_start
> >> >    168: 00004148     0 NOTYPE  GLOBAL DEFAULT  ABS __bss_start
> >> >
> >> >> I am not working on the Android NDK so I can't give you a good
> >> >> suggestion here.  Does using the latest NDK fixes your problem?
> >> >>
> >> >
> >> > Yes, there is no this issue with the latest NDK.
> >> >
> >> > -
> >> > Thanks,
> >> > Pete
> >> >
> >> > --
> >> > Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
> >> > ------- You are receiving this mail because: -------
> >> > You are on the CC list for the bug.
> >
> > --
> > Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
> > ------- You are receiving this mail because: -------
> > You are on the CC list for the bug.

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.