The Universe of Disco

Suffering from "make install"
I am writing application X, which uses the nonstandard perl modules DBI, DBD::SQLite, and Template. These might not be available on the target system, so I got the idea to include them in the distribution for X and have the build process for X build and install the modules. X already carries its own custom Perl modules in X/lib anyway, so I can just install DBI and the others into X/lib and everything will Just Work. Or so I thought.

After building DBI, for example, how do you get it to install itself into X/lib instead of the default system-wide location, which only the super-user has permission to modify?

There are at least five solutions to this common problem.

Uh-oh. If solution #1 had worked, people would not have needed to invent solution #2. If solution #2 had worked, people would not have needed to invent solution #3. Since there are five solutions, there is a good chance that none of them work.

You can, I am informed:

• Set PREFIX=X when building the Makefile
• Set INSTALLDIRS=vendor and VENDORPREFIX=X when building the Makefile
  • Or maybe instead of VENDORPREFIX you need to set INSTALLVENDORLIB or something
  • Or maybe instead of setting them while building the Makefile you need to set them while running the make install target
• Set LIB=X/lib when building the Makefile
• Use PAR
• Use local::lib

Some of these items fail because they just plain fail. For example, the first thing everyone says is that you can just set PREFIX to X. No, because then the module Foo does not go into X/lib/Foo.pm. It goes into X/Foo/lib/perl5/site_perl/5.12.23/Foo.pm. Which means that if X does use lib 'X/lib'; it will not be able to find Foo.

The manual (which goes by the marvelously obvious and easily-typed name of ExtUtils::MakeMaker, by the way) is of limited help. It recommends solving the problem by travelling to Paterson, NJ, gouging your eyes out with your mom's jewelry, and then driving over the Passaic River falls. Ha ha, just kidding. That would be a big improvement on what it actually suggests, for three reasons. First, it is clear and straightforward. Second, it would feel better than the stuff it does suggest. And third, it would actually solve your problem, although obliquely.

It turns out there is a simple solution that doesn't involve travelling to New Jersey. The first thing you have to do is give up entirely on trying to use make install to install the modules. It is completely broken for this application, because even if the destination could somehow be forced to be what you wanted—and, after all, why would you expect that make install would let you configure the destination directory in a simple fashion?—it would still install not only the contents of MODULE/lib, but also the contents of MODULE/bin, MODULE/man, MODULE/share, MODULE/pus, MODULE/dork, MODULE/felch, and MODULE/scrotum, some of which you probably didn't want.

So no. But the solution is actually simple. The normal module build process (as distinct from the install process) puts all this crap under MODULE/blib. The test suite is run against the blib installation. So the test programs have the same problem that X has. If they can find the stuff under blib, so can X, by replicating the layout under blib and then doing what the test suite does.

In fact, the modules are installed into the proper subdirectories of MODULE/blib/lib. So the simple solution is just to build the module and then, instead of trying to get the installer to put the right stuff in the right place, use cp -pr MODULE/blib/lib/* X/lib. Problem solved.

For modules with a shared library, you need to copy MODULE/blib/arch/auto/* into X/lib/auto also.

I remember suffering over this at least ten years ago, when a student in a class I was teaching asked me how to do it and I let ExtUtils::MakeMaker make a monkey of me. I was amazed to find myself suffering over it once again. I am relieved to have found the right answer.

This is one of those days when I am not happy with software. It sometimes surprises me how many of those days involve make.

Dennis Ritchie once said that "make is like Pascal. Everybody likes it, so they go in and change it." I never really thought about this before, but it now occurs to me that probably Ritchie meant that they like make in about the same way that they like bladder stones. Because Dennis Ritchie probably does not like Pascal, and actually nobody else likes Pascal either. They may say they do, and they may even think they do, but if you look a little closer it always turns out that the thing they like is not actually Pascal, but some language that more or less resembles Pascal. Unfortunately, the changes people make to make tend to make it bigger and wartier, and this improves make about as much as it would improve a bladder stone.

I would like to end this article on a positive note. If you haven't already, please read Recursive make Considered Harmful and be prepared to be blinded by the Glorious Truth therein.

[Other articles in category /prog] permanent link