Undefined symbols for NETCDF 4.2

[dongli]

I used ifort to compile NETCDF:

export FC=ifort
brew install netcdf --enable-fortran

The compilation was successful, but when I built my codes, "ifort" complained:

Undefined symbols for architecture x86_64:
  "_netcdf_mp_nf90_strerror_", referenced from:
...

So I checked the symbols in /usr/local/lib/libnetcdff.5.dylib by nm, there is no _netcdf_mp_nf90_... symbols, but /usr/local/lib/libnetcdff.a has. What is wrong?

[2bits]

@Sharpie does this look like something he can easily adjust for his compiler choice, or are we suggesting people use gfortran if at all possible to build this?

@dongli can you give the us the simplest Fortran source code that will reproduce this?

[Sharpie]

This will be difficult to debug as I don't have access to ifort any longer, so I cannot test it. Generally, if you're not using our gfortran distribution to build things you may be on your own since we can't debug what we can't reproduce.

However, it would be very useful to see the output of brew instal -v netcdf --enable-fortran in a gist as I have a couple of guesses as to what may be going wrong but can't say for sure unless I see a build log.

[dongli]

@2bits Any Fortran code calling NETCDF can reproduce this:

program main

    use netcdf

    integer ncid, ierr

    ierr = nf90_open("test.nc", nf90_nowrite, ncid)

end program main

@Sharpie I can't copy the output in a gist, maybe it is too long?

When I linked with the static libs, it succeeded:

ifort -o test test.f90 /usr/local/lib/libnetcdf.a /usr/local/lib/libnetcdff.a -lcurl -lhdf5 -lhdf5_hl

[2bits]

I get an error when I try to compile that code. I named it test.f90 and ran

$ gfortran -o test test.f90 /usr/local/lib/libnetcdf.a /usr/local/lib/libnetcdff.a -lcurl -lhdf5 -lhdf5_hl
test.f90:3.14:

    use netcdf
             1
Fatal Error: Can't open module file 'netcdf.mod' for reading at (1): No such file or directory

[dongli]

@2bits You need to install netcdf, and please use ifort if possible, since I got no problem with gfortran. Thanks!

[2bits]

I did install netcdf with --enable-fortan and I don't have that file, netcdf.mod anywhere on my system. Not sure what the story is. I'll look around for ifort.

[jbeezley]

I have a bit of experience with netcdf so I might be able to help.

@2bits: It seems to be creating NETCDF.mod rather than netcdf.mod. Try to add -I/usr/local/include. Strange that the compiler seems to be looking for it in lower case. I wonder if the gfortran modules are working on case sensitive file systems.

@dongli: I don't have access to ifort on mac, but I tried on Linux, and it seems fine. Can you gist nm /usr/local/lib/libnetcdff.5.dylib? The build log would also help. I have had problems pasting into github's gist form at times (usually because the output contains non-printable characters). You might try creating an empty gist, then cloning git repo it creates and adding the content from the commandline. I have had to resort to that before.

[Sharpie]

The reason I ask for the install output is that ifort and GCC/Clang (can't tell which you are using without a build log) could have different name mangling schemes. So, if the dynamic library was assembled by GCC/Clang instead of ifort, then there may be some extra underscoring going on.

If this guess is correct, you may need to pass some extra flags to ifort in order to correctly resolve symbols in the dynamic library.

[2bits]

@jbeezley thanks for the info. This worked:

gfortran -o test test.f90 -lnetcdf -lnetcdff -lcurl -lhdf5 -lhdf5_hl -I/usr/local/include

What I don't understand is what I'm looking for. The symbol the example uses, nf90_open exists in both dylibs. I assume the idea is that when it's built with ifort, that will be missing from the dylib. I will try to reproduce that.

But the OP said this was missing: _netcdf_mp_nf90_strerror_ I don't have that in either library. This is the closest thing that I have:

$ nm /usr/local/lib/libnetcdff.a | grep nf90_strerror
0000000000000066 T ___netcdf__nf90_strerror
0000000000047fd8 S ___netcdf__nf90_strerror.eh

$ nm /usr/local/lib/libnetcdff.dylib | grep nf90_strerror
000000000000ae70 T ___netcdf__nf90_strerror

Because netcdf always has two underscores following it, netcdf__, this sounds like the extra underscoring that @Sharpie mentioned. I certainly don't have any symbols in either library that have _mp_ in the middle of their names.

$ nm /usr/local/lib/libnetcdff.a | grep _mp_
00000000000004a4 T _nf__create_mp_
00000000000009b8 S _nf__create_mp_.eh
0000000000000591 T _nf__open_mp_
00000000000009e8 S _nf__open_mp_.eh
0000000000000672 T _nf_delete_mp_
0000000000000a18 S _nf_delete_mp_.eh
                 U _nf__create_mp_
                 U _nf__open_mp_

$ nm /usr/local/lib/libnetcdff.dylib | grep _mp_
0000000000001f84 T _nf__create_mp_
0000000000002071 T _nf__open_mp_
0000000000002152 T _nf_delete_mp_

[2bits]

It looks like I have to give them my name, address, and phone number to try Intel's iFort for 30 days. I think I'll pass.

[jbeezley]

ifort and gfortran mangle names differently particularly with functions in modules. For ifort, it generally looks like mp. gfortran uses . There are also issues with how many underscores go after the name. The reason I asked for a gist of the symbol names in libnetcdff.dylib was to see what sort of convention was in this file and if the strerror function was missing entirely or just mangled differently.

In any case, I suspect this is probably an upstream issue. Splitting the fortran and c interfaces has been a difficult transition.

[dongli]

@jbeezley I have upload the output of nm /usr/local/lib/libnetcdff.5.dylib into https://gist.github.com/3016300.

It looks like that there is no F90 interface.

@Sharpie The compilation output of brew install -v netcdf --enable-fortran --default-fortran-flags into https://gist.github.com/3016337.

[jbeezley]

None of the fortran 90 subroutines are being linked into the shared library. Here is the reason why:

ifort: command line warning #10006: ignoring unknown option '-force_load,../f90/.libs/libnetcdff90.a'

They need to fix this upstream with the proper flag for ifort.

[2bits]

Nice one. I wonder if make check would have found that. I reported it upstream via email. @dongli if you can determine what the valid flag is for iFort, please let us know, and I'll forward that to the netcdf developers.

[dongli]

@jbeezley There seems a similar issue http://www.unidata.ucar.edu/support/help/MailArchives/netcdf/msg10626.html, which is related to libtool. The libtool on my Mac Pro is Apple Inc. version cctools-822.

[dongli]

If the default gfortran could be version "4.7", I will not use ifort.

[2bits]

That issue you linked implies the problem is with libtool. If you have automake and libtool brewed, then you can try to run autoreconf like this:

  def install
    system 'autoreconf', '-ivf'
    ENV.fortran if fortran?

    common_args = %W[
      --disable-dependency-tracking
      --prefix=#{prefix}
      --enable-static
      --enable-shared
    ]

After each system 'make install' I added a system 'make check' and everything seems to work when I try it.

[dongli]

@2bits I have added system 'autoreconf', '-ivf' as you mentioned, but encountered the following errors:

==> Downloading http://www.unidata.ucar.edu/downloads/netcdf/ftp/netcdf-cxx4-4.2.tar.gz
Already downloaded: /Library/Caches/Homebrew/netcdf-cxx4-4.2.tar.gz
/usr/bin/tar xf /Library/Caches/Homebrew/netcdf-cxx4-4.2.tar.gz
==> ./configure --disable-dependency-tracking --prefix=/usr/local/Cellar/netcdf/4.2 --enable-static --enable-shared
./configure --disable-dependency-tracking --prefix=/usr/local/Cellar/netcdf/4.2 --enable-static --enable-shared
./configure: line 437: sed: command not found
./configure: line 548: sed: command not found
./configure: line 887: sed: command not found
./configure: line 913: sed: command not found
./configure: line 913: sed: command not found

[2bits]

Strange. Maybe each subformula needs autoreconf also. Gist the output of brew install -vd --enable-fortran netcdf. We'll be looking to see if it finds sed for the main configure stage. Consider adding two more autoreconf statements, one in each subformula:

    NetcdfCXX.new.brew do
      system 'autoreconf', '-ivf'
      system './configure', *common_args
      system 'make install'
      system 'make check'
    end unless no_cxx?

    NetcdfFortran.new.brew do
      system 'autoreconf', '-ivf'
      system './configure', *common_args
      system 'make install'
      system 'make check'
    end if fortran?

[dongli]

@2bits I added autoreconf in each subformula as you said, and the output is in https://gist.github.com/3016337.

==> Downloading http://www.unidata.ucar.edu/downloads/netcdf/ftp/netcdf-cxx4-4.2.tar.gz
Already downloaded: /Library/Caches/Homebrew/netcdf-cxx4-4.2.tar.gz
/usr/bin/tar xf /Library/Caches/Homebrew/netcdf-cxx4-4.2.tar.gz
==> autoreconf -ivf
autoreconf -ivf
Error: #<BuildError: Failed executing: autoreconf -ivf>

[2bits]

@dongli I just got that same error. You hit a bug in our formula. We set the PATH in the formula incorrectly. Below you can see how it has an extra space.

$ echo $PATH
/usr/local/Cellar/netcdf/4.2/bin /usr/local/bin:/usr/local/sbin:/usr/local/share/python:/usr/bin:/bin:/usr/sbin:/sbin:/usr/X11/bin:/usr/local/Library/Contributions/cmds

The space instead of a : is causing these problems. The bad code is this:

    ENV.prepend 'PATH', bin

Change that to be:

    ENV.prepend 'PATH', bin, ':'

[Sharpie]

Oops, that was my bad. Should be fixed now.

[2bits]

Thanks @Sharpie. It turns out that the netcdf website has a FAQ about iFort and MacOSX:
http://www.unidata.ucar.edu/software/netcdf/docs/known_problems.html#intel-fortran-macosx

They give us the fix too. We probably only have to check if FC == 'ifort' and if so, set
lt_cv_ld_force_load=no.

[dongli]

@2bits It works now!

[2bits]

Great news. We'll add the lt_cv_ld_force_load=no to the netcdf formula if FC is set to ifort.