[Trilinos-Users] Teuchos_MPIReduction

[Marco Morandini]

Bill Spotz wrote:
> Marco,
> 
> So I guess my hypothesis from before was just a brick wall.
> 

In the meantime, I've found that the last problem described in

http://software.sandia.gov/bugzilla/show_bug.cgi?id=1777

reproduces my error.
However, the submitter solves the problem
using mpic++ intead of mpicc.

> My working theory now is that we are just seeing porting issues.  The 
> python wrappers are tested every night on a Linux box, but not a SUSE 
> Linux box.

Question: is the automatic test configured --with-gnumake?
If yes, what happens if you try to build without that configure option?
Perhaps my link problems were due to the fact that I was not
using that configure option. See below.

> 
> Why you need to patch the swig interface files with "#include"s I don't 
> know, but I can add them to the repository/distribution without problem.
> 
> As for the unresolved symbols at run time, and the fix of reordering the 
> libraries, this seems to be a common porting issue.  Let me know what 
> symbols are not being found and what the order of libraries is on the 
> link line, for both the "bad" build and the "good" one.  I can then 
> patch the setup.py script so that it does the right thing in a portable 
> way.

I'm not currently able to reproduce the problem.
I can try, but it will take time, and I don't know if it
worth the effort.

I think that now everything is fine because now I'm configuring with
--with-gnumake (thanks for the suggestion!)
and (but I've not checked that this is actually the case)
duplicate libraries are already deleted, in the correct order,
by the perl script strip_dup_libs.pl.
For this reason, I think that the problem occurs when trilinos
is not configured --with-gnumake
This is what I used to do before starting this thread.
And, at that time, I had to reduce the number of packages;
this is because my swig segfaults with a too long list of arguments and,
whithout gnumake, strip_dup_incl_paths.pl is not run, and
a too long command line is passed to swig.

Now, configuring --with-gnumake I've added new packages :),
but I've lost the unique combination of packages
that I was able to find in order to build trilinos --without-gnumake :(

> 
> By the way, it wasn't entirely clear from your message.  Are you 
> commenting out "uniquifyList(options)" or are you replacing it with 
> "uniquifyList(options,True)"?

I was commenting out "uniquifyList(options)".
Looking at MakefileVariables.py I think that a
cleaner solution could be to use "uniquifyList(options,True)",
but I've not tried it.

If you usually configure --with-gnumake, and you can confirm
that something is broken without that option, what about
checking for gnumake at configure time and aborting
if it is not found?

Marco