ld ignores used libraries (especially libc) with --as-needed

[forwarded from http://bugs.debian.org/712081]

Compiling a trivial int a() {return 17;} file with
g++ -v -Wl,--as-needed t.c --shared -o libt${n}.so

With binutils 2.23.52.20130612-1:
$ readelf -d /tmp/libt3.so | grep NEEDED
$ objdump -T libt3.so | grep _cxa_fin
0000000000000000 w D *UND* 0000000000000000 __cxa_finalize

While with binutils 2.22-8 and 2.23.2:
$ readelf -d libt2.so | grep NEEDED
 0x0000000000000001 (NEEDED) Shared library: [libc.so.6]
$ objdump -T libt2.so | grep _cxa_fin
0000000000000000 w DF *UND* 0000000000000000 GLIBC_2.2.5 __cxa_finalize

i.e. the dependency on libc.so.6 is missing and what is worse:
the libc6 symbol is not properly versioned, which might mean that that
library might cause programs segfault with future libc versions.

g++ says it does:
COLLECT_GCC_OPTIONS='-v' '-shared' '-o' 'libt3.so' '-shared-libgcc'
'-mtune=generic' '-march=x86-64'
 /usr/lib/gcc/x86_64-linux-gnu/4.7/collect2 --sysroot=/ --build-id
--eh-frame-hdr -m elf_x86_64 --hash-style=gnu -shared -o libt3.so
/usr/lib/gcc/x86_64-linux-gnu/4.7/../../../x86_64-linux-gnu/crti.o
/usr/lib/gcc/x86_64-linux-gnu/4.7/crtbeginS.o
-L/usr/lib/gcc/x86_64-linux-gnu/4.7
-L/usr/lib/gcc/x86_64-linux-gnu/4.7/../../../x86_64-linux-gnu
-L/usr/lib/gcc/x86_64-linux-gnu/4.7/../../../../lib
-L/lib/x86_64-linux-gnu -L/lib/../lib -L/usr/lib/x86_64-linux-gnu
-L/usr/lib/../lib -L/usr/lib/gcc/x86_64-linux-gnu/4.7/../../..
--as-needed /tmp/cc8CjUQJ.o -lstdc++ -lm -lgcc_s -lc -lgcc_s
/usr/lib/gcc/x86_64-linux-gnu/4.7/crtendS.o
/usr/lib/gcc/x86_64-linux-gnu/4.7/../../../x86_64-linux-gnu/crtn.o

with the need for symbol __cxa_finalize coming from crtbeginS.o:

$ objdump -t /usr/lib/gcc/x86_64-linux-gnu/4.7/crtbeginS.o | grep _cxa_fi
0000000000000000 w *UND* 0000000000000000 __cxa_finalize

gcc unlike g++ is not affected because it puts a --no-as-needed before -lc

Zack Weinberg commented in the Debian report:

I think the linker is doing exactly what it's supposed to in this case
-- weak symbols *shouldn't* provoke NEEDED entries all by themselves.
A standard use case, for instance, is to take weak references to
pthread_mutex_(un)lock and then call them only if the function
addresses are non-NULL at runtime -- if that by itself dragged in
libpthread it would vitiate the optimization.

If you do the exact same test with a shared object that explicitly
calls something in libc (say, 'puts') that's a strong reference and
you will get a NEEDED entry.

As for the lack of version tagging on the weak reference, I can't say
whether this is a problem in general, but in the case of
__cxa_finalize, I'm pretty sure it isn't: a GLIBC_2.2.5 version
indicates a function that has not changed since the introduction of
symbol versioning in libc.so.6, and the spec governing
__cxa_finalize's behavior hasn't changed since 2003ish.