"Bradley D. LaRonde" <firstname.lastname@example.org> writes:
>> An object should never use stubs if takes the address of the function.
>> It should only use a stub for some symbol foo if every use of foo is
>> for a direct call.
> OK. So in a case where an object does take a pointer, does that mean that
> ld.so must fix the GOT entry for that symbol before handing control to the
> app (i.e. no lazy binding for that symbol)?
Right. Such objects won't use a stub, they'll just have a normal
reference to an undefined symbol. The dynamic loader must resolve
it in the usual way.
> I notice that the debian mipsel libpthread.so.0 in
> http://ftp.debian.org/pool/main/g/glibc/libc6_2.2.5-11.5_mipsel.deb has
> st_value == 0 for every UND FUNC, just like my x86 debian libraries. This
> is very different than the uClibc libpthread.so where every UND FUNC has
> st_value != 0. Interestingly if I link glibc's libpthread with uClibc's
> libc.so I see that most UND FUNCs then have st_value != 0.
You said in your original message that you'd recently upgraded to binutils
2.15-based tools. Was your libpthread.so built with them? If so, that would
explain it. I think earlier versions of ld were overly pessimistic about when
stubs could be used.
FWIW, I have a glibc-based sysroot built with gcc 3.4 and binutils 2.15.
Its libpthread uses plenty of stubs.