How to compile distributable Fortran binaries on Mac OS X Mountain Lion?

Question

Since Apple have stopped distributing gfortran with Xcode, how should I compile architecture independent Fortran code? I have Mac OS X Mountain Lion (10.8), and XCode 4.4 installed, with the Command Line Tools package installed.

Apple's Native Compilers

As far as I can tell, the Xcode C / C++ / ObjC compilers use a fork of the GNU compiler collection, with llvm as a backend; the latter I figure enables compiling and optimising "universal" binaries, for both Intel and PPC architectures.

3rd party binary Fortran compilers

HPC

I've only found a single website that distributes a binary version of gfortran specifically for Mountain Lion: the HPC website. However, I failed to get this to compile SciPy, and later saw in SciPy's README that it is "known to generate buggy scipy binaries".

CRAN/R

SciPy's recommended (free) Fortran compiler is the one on CRAN's R server, but this has not been updated for Mountain Lion yet. They provide instructions and a script for Building a Universal Compiler, but, again, this hasn't been updated for Mountain Lion yet..

G95

The G95 project hasn't had an update since 2010, so I didn't try it.. Anyone tried this on Mountain Lion?

MacPorts

I guess this will be the easiest way to get gfortran installed, but port search gfortran comes up with nothing, and I've not had any joy with MacPorts in the past (no offence to MacPorts; it's looks like a very active project, but I've been spoilt with Linux package managers, my favourite manager being aptitude) so on Mac OS X I've compiled software and libraries from source code in the past. Never been a problem 'til now...

Building a Fortran compiler

Having dug around on the internet a lot in the last couple of days, I've found other Fortran compilers, but I've failed to get any to cross-compile universal binaries, or to compile SciPy.

GCC - The Gnu Compiler Collection

I compiled the entire GCC collection (v4.6.3), including autotools, automake, libtool and m4 - like the GCC wiki and this blog describe - but the resulting compilers didn't compile universal binaries, probably because LLVM wasn't used as a backend.

DragonEgg

DragonEgg is a "gcc plugin that replaces GCC's optimisers and code-generators ... with LLVM". This looks interesting, but I don't know how I could use it to compile 'llvm-gfortran-4.x'. Can this be done?

Compatibility

Libraries

The compiler that comes with Xcode is (a fork of?) GCC v4.2. But GCC's current release and development branches are versions 4.6 and 4.7, respectively. Apparently, a GNU license change, or something, stopped Apple from updating to more modern versions of GCC. So, if I was to build dynamic libraries made with GCC's gfortran v4.6, could they then be linked with C code compiled by Xcode's native compiler? At a minimum, I figure resulting Mach-O binaries need both x86_64 and i386 code paths. Do GCC provide backwards compatibility with Apple's (forks of?) GCC? I know gfortran has the -ff2c flag, but is this stable across versions?

Compile flags

The GCC Fortran compiler I built from source didn't support the use of the -arch compile flag. I had been including the flags -arch x86_64 -arch i386 in both CFLAGS and FFLAGS environment variables on earlier OSX versions (Snow Leopard to Lion). Python's distutils, and probably other OSX compilers, expect these flags to work, when configured to build apps or frameworks, using Xcode's universal SDK.

In case you're wondering what compile flags I use, I've uploaded the script I use to pastebin, which I source before I compile anything, using: source ~/.bash_devenv.

The Ideal OSX Fortran Compiler

• Create ppc and intel (32 and 64bit) universal binaries, specified by using the -arch flags.
• Makes binaries compatible with XCode's linker.
• Compiles SciPy, giving no errors (compatible with numpy's distutils and f2py).

I don't use Xcode so much, but integration with it would surely benefit other users. Even Intel are still having problems integrating ifort into Xcode 4.4, so this is not something I expect to work..

If you read all the above, then thank you! You can probably tell that I'm not averse to building my own Fortran compiler from source, but is it even possible? Have I missed something? A configure flag maybe? And if such a compiler is not available yet, then why not?!

(Update:) Apple's GCC

Apple provide the source code for their patched version of GCC, at opensource.apple.com. This actually includes the source code for gfortran, but what do you know - it doesn't compile (easily). I'm in the process of writing a build script to get this to work. Unfortunately, I've had to apply a couple of patches, and learn about "the Apple way" of building GNU software. This is the way to go I think. Any reasons why it shouldn't be? I'll update with an answer if I get it to work...

Answer 1

I know you said you don't like MacPorts, but if you install the gcc48 port, it does in fact include gfortran (although you'll also have to do sudo port select --set gcc mp-gcc48 to get it to set up the symlink named gfortran).

Also, FWIW, the MacPorts option is not necessarily a binary - MacPorts can actually build it from source, which is why it sometimes takes a while. On the other hand, it also sometimes seems to get archived binaries from somewhere, but I think it depends on what the original author of the portfile uploaded.

Answer 2

You could use brew (or Homebrew) to install gfortran.

$ brew install gfortran

Answer 3

In case you didn't already notice this: In newer versions of Xcode you have to explicitly install command line tools in the following way:

Preferences -> Downloads -> Components

And then click the "install" button for command line tools. This includes gfortran:

> gfortran -v
Using built-in specs.
Target: i686-apple-darwin10
Thread model: posix
gcc version 4.2.1 (Apple Inc. build 5664)

Admittedly, this does not solve all of my fortran needs (in some cases "./configure" scripts will complain that they cannot "compile a simple fotran program").

Answer 4

I managed to compile after installing gfortran from http://r.research.att.com/tools/gcc-42-5666.3-darwin11.pkg, as explained here. I had to try to open the package a couple of times, though. First time it said that only apps from App Store can be installed. After installing gfortran, python setup.py build and python setup.py install worked fine. The unit tests of scipy though give a fairly high number of fails, not sure it's normal.

Ran 5481 tests in 82.079s

FAILED (KNOWNFAIL=13, SKIP=42, errors=11, failures=72)
<nose.result.TextTestResult run=5481 errors=11 failures=72>

Answer 5

I ended up compiling gfortran the source code provided at Apple's developer tools source code page. This seems to be working okay now - I've successfully compiled x86-64 and i386/i686 LAPACK, ATLAS and BLAS fortran libraries - but there are some ranlib tests which fail, when running make -k test in the build directories. (I could provide more info on that pastebin or somewhere, if someone wants...)

Build process

After asking the question, I downloaded Apple's llvmgcc42 source code tar archive, which includes the source code for llvm/gcc C, C++, ObjC and fortran compilers, and spent some time trying to compile a universal build of gfortran. The build takes about 30-60 minutes on my quad-core 2.8GHz Mac Pro, and became quite an involved process, so I wrote a set of build scripts for it, which I've shared at github.com.

....

I'll keep a tar archive of my build here for the time being, if anyone would like a copy. (Updated 26-Sep-2012) It'll only work if installed with a prefix of /usr/local/ though, unless you run install_name_tool on the executables and dylibs, to change the prefix from /usr/local to wherever you want to put it. You can test install_name's with otool -L filename (more info on the reasons for this is here).

The final build I'm now using also includes updates to the gcc/fortran and libgfortran directories, which I got from GNU GCC 4.2.4. These sources I got from my local GCC's mirror. There were only minor changes between 4.2.1 and 4.2.4, and the build scripts include the patches needed to upgrade the code.

The build-gfortran.sh script I wrote downloads missing dependencies (mpfr and gmp), compiles and cross-compiles them, patches differing headers with architecture-dependent preprocessor macros, and runs lipo to create universal binaries and libraries, eventually supporting both i386 and x86_64 architectures. The process is similar for llvmCore, and then GCC. I mostly copied code from the build_llvm and build_gcc bash scripts provided with Apple's llvmgcc42, but had some of it had to be modified, including a few lipo and install_name_tool commands.

The official way to compile Apple's gcc, using Xcode's gnumake just didn't work for me. I thought this should work just byadding "fortran" to the LANGUAGES variable in build_gcc.

With regards to compiling Scipy, still can't get that building perfectly. I've had to use clang and clang++ as C/C++ compilers, or else I get EXC_BAD_ACCESS malloc errors. Haven't tried the gcc/g++ compilers I built, just used the system ones. This is as reported for Lion, on the Scipy install page. I'm down to 11 errors and 1 failure, which are all raised from the same 3 function calls (_fitpack._bspleval, numeric.asarray, testing.utils.chk_same_position). Think that's pretty good, but I'd like every test to pass...