Testcase for symbol redefinition problem

["Kevin P. Fleming" <kpfleming at linuxfromscratch dot org>]

http://linuxfromscratch.org/~kpfleming/linktest.tar

A few days ago I posted a message about a problem some of our LFS 
developers are having building UML kernels. H.J. Lu asked that I repost 
when I could supply a small testcase, so I'm doing that.

The above tarball, when unpacked and "make" run, will demonstrate the 
problem we are having. Note that the make run will always fail, because 
the link instructions are incomplete; what I'm concerned about here is 
an additional error that occurs on some systems and not others.

The error that occurs on _some_ systems only involves ld reporting that 
the "strcpy" symbol has been multiply-defined, in unmap_fin.o (which got 
if from glibc) and in strcpy.o (where it was manually defined). What's 
unusual is that we cannot find any reliable means to identify which 
systems will have this problem and which ones won't.

In the most degenerate case, we have a user with two systems both 
running Fedora Core 2 with identical packages (no updates have been 
applied, these are installs straight off the FC2 install discs). On one 
system the link fails, on the other it succeeds. The only difference we 
know of between the two systems is that one is an AMD K6-3 processor 
(which succeeds) and the other is an Intel Pentium 2 (which fails).

I also have test reports from a variety of other systems, involving 
binutils versions 2.14, 2.15, 2.15.90.0.* and 2.15.91.0.*. These reports 
also cover multiple gcc versions and multiple glibc versions. If you are 
interested, the reports are in this thread:

http://linuxfromscratch.org/pipermail/lfs-hackers/2004-July/001740.html

We have not run across this problem building anything other than UML, so 
it could very well be a problem in the UML Makefile and/or linker 
scripts. We originally were concerned that our toolchain builds were 
somehow at fault, but we now have test reports of failure on FC2, 
Gentoo, and ArchLinux.