Cross-compiling perl

Build process for perl modules is unbelievably complex and awfully unsuited for cross-compiling. Perl-cross takes some shortcuts to make it work, but it has its limitations.

To add a module to your build, unpack it into cpan/ directory before running configure. Check naming scheme there: a module called Some::Module should be placed in cpan/Some-Module.

Native builds with cross-compiler perl are not supported. With rare exceptions, it is not possible to build a module on the target machine. Everything has to be cross-compiled.

The problem with modules

Building a perl module requires fully functional perl interpreter and a bunch of rather complex modules available at build time.

That's a kind of chicken-and-egg problem, perl and modules are needed to build perl and modules.

How the problem is resolved in perl-cross: miniperl is instructed to use module sources directly for non-xs modules, without building them, and xs modules are replaced with stubs from cnf/stub/ directory. The entry point for the whole thing is miniperl_top. It runs miniperl with a bunch of -I options to make it look like all the required modules are available.

Makefile rules for modules

For various reasons, module-related make rules only apply to modules found by configure. A module is always a directory located under cpan/ or ext/; check cnf/ on how exactly configure decides which directories to use, and the type of module (XS/non-XS).

The modules to be built are listed in $nonxs_ext, $static_ext, $dynamic_ext; additionally, $disabled_nonxs_ext and $disabled_dynamic_ext variables list modules that were found but won't be built by "make modules" or "make all".

Consider a module located in cpan/Some-Module; its perl name is likely Some::Module. Assuming it was correctly found by configure, the command to make it is

	make cpan/Some-Module
which will in turn call
	make cpan/Some-Module/pm_to_blib

pm_to_blib, is a real file and it's used as a flag for the whole module, which allows to avoid costly recursive make runs. As long as pm_to_blib is up-to-date, make won't attempt to rebuild the module. This system is not very stable, and it is possible get unfinished build with all pm_to_blibs in place; right now there's no good way to deal with it except for removing pm_to_blib files manually and re-running make.

Here's what make does for a target like "cpan/Some-Module/pm_to_blib":

Most of these operations require miniperl, so it will be built before attempting to make any of the module targets.

Re-building modules

Sometimes pm_to_blib file gets touched before the module is built completely. Typically this means there were built errors, but it can also happen when MakeMaker decides to Makefile needs to be re-built. As long as pm_to_blib is up-to-date, make won't be invoked for this module and the build won't be finished.

There are two possible resolution. First, if the module name is known, removing pm_to_blib manually will force rebuilt. Second,

	make modules-reset

will remove pm_to_blib from all non-disabled modules. Even the latter is relatively cheap — it will not force a complete rebuild, just "make -C cpan/Some-Module" invocations for all modules.

Cleaning up modules

Running make clean on a module requires an up-to-date Makefile for that module, which in turn depends on usable miniperl, MakeMaker and its subdependecies. Running make clean may prompt re-doing half the build. That's rather counter-intuitive, but that's how MakeMaker works.

Top-level Makefile will only invoke "make clean" for modules that have pre-built Makefile. The idea is that if there's no Makefile, the module has never been built and doesn't require any cleaning. It's not always true. To ensure all modules are really cleaned up,

	make modules-makefiles modules-clean
can be used. Note that it will try to build Makefiles for all (non-disabled) modules, potentially running per-module configure and other nasty things.

Module configuration

Some modules have configure-like tests in their Makefile.PLs, which sometimes can't handle cross-compilation very well. A notable case is Time::HiRes that depends on d_nanosleep and d_clock_* from cnf/hints/linux.

Other modules from the perl distribution seem to avoid this, but third-party modules may be a problem. There is no good solution here. Fixing Makefile.PL and/or hinting the values may help in some cases.

Some modules analyze $^O value at build time, confusing host and target platforms. At build time, $^O describes the host system, and $Config{osname} should be used for the target.