zlib+shlib issue with the official EL4 R binary

On Thu, 29 Mar 2007, Prof Brian Ripley wrote:

I presume you are talking about RPMs built with --with-system-zlib?

That's not recommended in the R-admin manual, and perhaps you can 
persuade
Martyn Plummer not to do it.

I've checked, and he no longer does so, sorry.  So there is a zlib 
inside R, and its entry points will be visible to a packages' .so when R 
is built as a binary, but not when it is built as a shared library.  In 
that case it depends on the fine details (including of what you did) 
whether
mypackage.so resolved against R.bin or libz.so in the first two lines of 
your table.  But it is entirely possible that it links against R.bin in 
line 2 and libz.so in line 3.

If your program is that sensitive to the version of libz I think you 
should statically link against your own copy (which is R's solution for 
its own purposes).

On Thu, 29 Mar 2007, Hin-Tak Leung wrote:

Hi,

I have got a curious issue with an R package which uses zlib, against
the official binary here:
http://cran.r-project.org/bin/linux/redhat/el4/i386/R-2.4.1-1.rh4AS.i386.rpm 

on a Redhat EL4 i686 system.

The problem is that at the end of reading a gzip'ed file within my
C code, gzgetc() returns -1 (no more to read or error) but
gzeof() doesn't return true. Now one can probably jump to conclusion
that this is a zlib problem (or mine), but it is not. Here is
the summary table (my package binary is the same one on the four
2.4.1 cases below - built against the official EL4 binary on EL4)

Rver    shlib   buildhost   runhost    status of zlib code in mypackage
2.3.1   no      EL4         EL4        ok
2.4.1   no      EL4         EL4        ok
2.4.1   yes     EL4         EL4        broken
2.4.1   yes     EL4         FC6        ok
2.4.1   yes     FC6         FC6        ok

The most curious part is entry 3 vs 4 - if I just transplant
the official R binary for EL4 and run it on FC6 (x86_64 and
loading the same mypackage i686 binary), then it works.
The dynamic linker on FC6 is very different from all the earlier
redhat releases [and supposedly a lot faster...].
(http://docs.fedoraproject.org/release-notes/fc6/en_US/sn-Devel.html#id2956225) 


so it looks like an issue due to a bad interaction for the
combination of both the shlib compile-time option and the EL4
runtime environment. (e.g. R 2.4.1 is shipped and bundled with
zlib 1.2.3 but EL4 is only equiped with zlib 1.2.1.2...).

For the moment I'll just recommend anybody who needs to run my
package (snpMatrix in http://www-gene.cimr.cam.ac.uk/clayton/software/)
on Redhat EL4 to just build and install R from source, since shlib=no
is the default for compiling from source.

Anybody has any idea how this strange brokenness of zlib might happen?

Hin-Tak Leung