I've spent the last two days trying to recompile gcc on my G4 iBook. It hasn't been pleasant and it hasn't worked. And that might be because I didn't do enough research.
In the process of doing Project Euler problems, I reached one where I needed to solve a system of linear equations. For those of you who don't remember what this looks like, it's something like:
$4x+3y = 10$
$7x-2y = 3$
While this is a reasonably simple example and can be solved by hand, they can get significantly more difficult. The standard way for solving problems in this fashion involves using matrices. (If you took a linear algebra class, this should look familiar.)
Coding a naive way to do this is not difficult. Cramer's Rule provides a way to find each value by using the determinant of matrices. (Apparently LU decomposition is faster but it looks harder to implement. I'll go with naive first and then refactor later if I need to.) The determinant of arbitrary-size square matrices can be expressed through the Laplace expansion until eventually 2x2 (or, I suppose, 3x3) matrices are found that can have their determinant determined mathematically.
In consulting Skiena's The Algorithm Design Manual, I found references to LAPACK, a linear algebra package for FORTRAN. This is not useful to me since I'm using Ruby. But if such a package exists for FORTRAN, surely one exists for Ruby.
Linalg is a linear algebra package for Ruby. Or, more precisely, it's a Ruby wrapper around LAPACK. As a result, it relies on embedding a FORTRAN compiler into the package. And this is where the pain starts.
OS X comes with LAPACK already. However, it does not come with FORTRAN support. In order to get FORTRAN support, you need to install software or recompile gcc. Failing to do a proper search and not finding this resource, I decided to compile a new gcc.
gcc 4.3.3 relies on gmp and mpfr. gmp 4.2.1, which, if you don't have it installed, gcc tells you to get from ftp://gcc.gnu.org/pub/gcc/infrastructure/, does not work on OS X. It builds and installs fine and then yields an obscure linker bug when you try to build mpfr or any other software that tries to include gmp. After an hour of fighting with this, I found out that gmp 4.2.1 was not the newest version but that 4.2.4 was. I had no issues with gmp 4.2.4.
Then there has been little love compiling gcc. At some point, gcc creates a spec file which is then used to pass the option -macosx-version-min which causes the compiler to fail. Editing the spec file and removing that lets it continue on before the next pass with a recreated spec file. Fixing that seems to correct the issue again. And then, eventually, there's a point in the build where it just fails. However, this may have been left over from something else so I tore everything down and restarted the process.
It looks like the issue with the spec file could also be of my own making. If I understand this post correctly, the issue might be that I don't have a new enough version of the developer tools installed. This may be possible since I think the version I have is the one that came with OS X 10.4 (and is, therefore, at least three years old).
With the information I've found this morning, I wonder if I could restart the process and have it run smoother. (Although the installable gfortran may be the smoothest option.) I just wish I knew days ago what I know now.
