Stefan’s musings

Writing Portable Build Systems With Faber


Faber is a project that evolved out of the need for a modern, portable build tool. Even today most existing tools are either tied to a specific platform (e.g., “GNU Make”, “nmake”), programming language (e.g. “ant”, “Rake”), or project (e.g. “”). A few projects have tried to break out of these bounds (notably “CMake” and “SCons”), but are overly complex, obscure, or inefficient.

One aspect that sets Faber apart from all of the above is its use of Python rather than a domain-specific language, to define the build logic.

In this post I will illustrate a few specific design aspects that allow the build logic (here expressed in “fabscripts”) to become platform-agnostic.

artefacts, actions, rules

Faber, just as any other build system, is concerned about building things, which we’ll call artefacts. They are updated using actions, which describe how certain “target” artefacts are built, typically from “source” artefacts. The connection from “source” to “target” is expressed using rules. Let’s consider a very simple example:

compile = action('compile', 'g++ -c -o $(<) $(>)')
link = action('link', 'g++ -o $(<) $(>)')

obj = rule(compile, 'hello.o', source='hello.cpp')
exe = rule(link, 'hello', source=obj)

This forms a very simple pipeline such that invoking faber hello will first compile ‘hello.o’ from ‘hello.cpp’, then link ‘hello’ from ‘hello.o’. (The syntax of the action commands above should be intuitively clear: “$(<)” and “$(>)” denote the target and source artefacts, respectively.)

This is the fundamental structure of all build systems, i.e. some file defines a set of artefacts and the way to update them, then the build tool can figure out the exact chain of actions that needs to be executed for the desired artefact to be updated.

While the above is extremely simple, both notationally as well as conceptually, it unfortunately isn’t very portable. Not all systems may have a compiler called “g++” installed. Calling conventions, i.e. the way to pass optional and non-optional arguments, may differ, too. And finally, the names used for the built artefacts may also be platform-dependent. While on Linux, “hello” may be a fine name for an executable, on Windows we may prefer to name it “hello.exe”.

So how can we add the required flexibility to the above script to make it portable while preserving the conceptual simplicity ?

encapsulating actions into tools

While the conceptual steps to compile the “hello” binary are the same on all platforms, the exact actions can vary. The “compile” and “link” actions are carried out by a compiler, but the exact spelling of the commands depend both on the compiler, as well as on the platform.

Faber defines tools to encapsulate concrete actions like the above. For example, it defines a “cxx” tool, with (member) actions “compile” and “link” (as well as a few others), that lets us rewrite the above build script as

from import cxx

obj = rule(cxx.compile, 'hello.o', source='hello.cpp')
exe = rule(, 'hello', source=obj)

“cxx.compile” now is an abstract action, which faber will try to match with a concrete implementation, depending on the platform and build parameters. Calling faber hello may actually perform the same action as before (if g++ is detected), or it may execute cl, if it is being executed on Windows and a MSVC toolchain is detected. Users may also specify on the command line which tool to pick (if multiple compilers are available), for example by running faber Further, tools may be configured in a config file by instantiating the appropriate compiler classes, such as::

mingwxx = gxx(command='x86_64-w64-mingw32-g++', features=target(arch='w64'))

With that, faber target.arch=w64 will find this MinGW compiler and cross-compile the code for Windows 64.

parametrizing actions with features

Our original example above used extremely simple ‘compile’ and ‘link’ commands. Real commands include different compiler flags (from header search paths to code-generation options) that need to be added.

Faber abstacts such parameters into (mostly) platform-agnostic features which will be mapped to tool-specific flags:

from import cxx, include, define

obj = rule(cxx.compile, 'hello.o', source='hello.cpp',
           featuers=(include('/search/path'), define('ANSWER=42'))
exe = rule(, 'hello', source=obj)

In this example, we add an “include” and a “define” feature specifically to the “hello.o” artefact. We can also define them globally, for example by invoking faber include=/search/path define=ANSWER=42.

As you can guess, the “include” feature adds a header search path (which maps to an -I option for g++ on Linux, and /I for cl on Windows), and the “define” feature adds a macro definition.

using implicit rules

So far we have used explicit rules to define the build pipeline from “hello.cpp” to “hello”. Typically, a project consists of many source files that all need to be compiled with the same (or at least, similar) options.

Artefacts have a type (e.g. “cxx”, “obj”, “bin”), which is either explicitly specified, or inferred from their name. Tools may define implicit rules, which map between artefact types. I.e., rather than calling a rule to update a specific target artefact (such as “hello.o”) from a given source artefact (such as “hello.cpp”), a tool calls an implicit rule which defines how to update a target artefact type (such as “obj”) from a source artefact type (such as “cxx”).

Compiler instances will declare their respective “compile” and “link” actions to build “obj” from “cxx” sources, and “bin” from “obj”, respectively, which faber can use to chain multiple implicit rules to build e.g. a “bin” from “cxx” sources.

In fact, this process is so common that faber encapsulates all this into higher-order artefacts. We can thus eliminate even more platform-specific code by rewriting the above as

from faber.artefacts.binary import binary

hello = binary('hello', 'hello.cpp')

“binary” is derived from “artefact”. Rather than being created by invoking a “rule”, it is constructed directly, using name and a source arguments. Note there is no action argument any longer. Invoking faber hello on this fabscript will still perform the expected actions, as can be seen by the output:

gxx.compile hello.o hello
...made 2 artefacts...

Can you guess what the following fabscript does ?

from faber.artefacts.library import library
from faber.artefacts.binary import binary

greet = library('greet', 'greet.cpp')
hello = binary('hello', ['hello.cpp', greet])

Hint: here is the output faber hello may produce:

gxx.compile greet.o
gxx.archive greet
gxx.compile hello.o hello
...made 4 artefacts...

In this post I have tried to give a quick overview of how to write portable build logic that works on different platforms, with different compilers.

For an in depth look please refer to Faber’s documentation.

Faber is a very young project. As such, it will lack quite a number of features, or support for more tools or platforms. Oh, and there may be one or two bugs left. ;-)

The source code is here.

I would welcome any contribution !

posted in: software