Introducing Fbuild
posted on January 30, 2009 - 01:39 AM PST
by Erick Tryzelaar
filed under:
FBuild
I'd like to introduce yet another build system, this one called Fbuild. It's
a different way to build things. As opposed to pretty much every other build
system, Fbuild is designed as just a caching library for Python
3.0. It takes advantage of Python's evaluation
scheme to implicitly describe the build dependency tree. In order to evaluate
the build, we simply evaluate python functions and cache their results. This
turns out to be an elegant and easy way to describe and program build systems.
On top of this alternative design, Fbuild comes out of the box with some
pretty advanced features:
- Linux and Apple support (Windows support will come soon)
- C, C++, OCaml, Bison, and Felix builders
- Multilevel namespaces for builders
- Simple creation of new builders
- Extensive configuration system for c90, c99, most of posix, and other libraries
- On-demand configuration
- Simultaneous building
- Detects file changes via digests verses timestamps
- Pretty output
- Very speedy (though it could be much faster)
Here's a quick example, in a file called fbuildroot.py:
import fbuild.builders.c
def build():
static = fbuild.builders.c.guess_static()
static.build_exe('exe', ['a.c', 'b.c', 'c.c'])
When run with as fbuild, assuming all those files exist, produces:
determining platform : {'bsd', 'darwin', 'macosx', 'posix'}
looking for program gcc : ok /usr/bin/gcc
checking /usr/bin/gcc : ok
checking /usr/bin/gcc with -g : ok
checking /usr/bin/gcc with -O2 : ok
checking /usr/bin/gcc with -c : ok
looking for program ar : ok /usr/bin/ar
looking for program ranlib : ok /usr/bin/ranlib
checking if /usr/bin/gcc -c can make objects: ok
checking if /usr/bin/gcc -c can make libraries: ok
checking if /usr/bin/gcc -c can make exes: ok
checking if /usr/bin/gcc -c can link lib to exe: ok
* /usr/bin/gcc -MM : c.c
* /usr/bin/gcc -MM : b.c
* /usr/bin/gcc -MM : a.c
* /usr/bin/gcc : c.c -> build/c.o
* /usr/bin/gcc : b.c -> build/b.o
* /usr/bin/gcc : a.c -> build/a.o
* /usr/bin/gcc : build/a.o build/b.o build/c.o -> build/exe
You may notice a couple interesting things about the output:
Fbuildconfigures the c static builder for you and tests that it works.- Much of the output is hidden.
Fbuilddoes this because the output can get overwhelming for large projects. It can be very easy to miss warnings and messages.Fbuildhides this in a log file, normally calledbuild/fbuild.log. - The generated files are stored in a
builddirectory. This is a user-customizable directory that can be used to support different build targets, such as adebugand areleasebuild. Fbuildautomatically discovers and tracks and sorts the dependencies between each c file.
To develop the last idea further, Fbuild was designed to work with globbed
paths if you don't want to bother writing out each filename. I'll show this by
rewriting the last example using the fbuild.path.Path class. This is a
convenience class that implements many useful path functions, as well as
supports replacing the Unix-style '/' separator with the native path separator.
Lets combine that with showing off what happens if we tweak what we want built:
import fbuild.builders.c
import fbuild.path
def build():
static = fbuild.builders.c.guess_static()
lib = static.build_lib('lib', fbuild.path.Path.glob('*.c'))
static.build_exe('exe', [], libs=[lib])}}}
This will produce:
* /usr/bin/ar : build/a.o build/b.o build/c.o -> build/liblib.a
* /usr/bin/ranlib : build/liblib.a
* /usr/bin/gcc : -> build/exe
As you can see, most of the work was skipped. Fbuild detected that none of
the source files changed, so it didn't need to recompile them. Since the glob
passed all the source files to static.build_lib they were all linked together
in a library. Finally, since the arguments changed for static.build_exe,
build/exe was recompiled.
Now that you've got a taste for Fbuild, lets go through some of the reasoning
behind its design. As I mentioned previously, every build system I've found is
based off of a multi-phase declarative tree evaluation. Here's the list of the
system I looked at:
By "declarative tree evaluation", I mean that the order of evaluation is
explicitly encoded in a dependency tree. Each node has a function that takes
one or more inputs and returns one or more outputs. By "multi-phased", I
consider building the tree as one phase, walking the dependencies as another,
and evaluating the nodes as yet another. This contrasts with Fbuild, which I
consider to be a single-phase procedural build system. As python evaluates
each cached function, the dependency implicitly gets evaluated. This means
that it's possible to dynamically modify the build on demand:
Here's a stupid example of something that's difficult with Make:
import random
import fbuild.builders.c
def build():
static = fbuild.builders.c.guess_static()
static.build('exe1', ['exe1.c'])
# Exit the build early if we roll a '6'
if random.randint(1, 6) == 6: return
static.build('exe2', ['exe2.c'])
Because we don't have to build up the tree first before we evaluate it, we don't have to shoehorn weird nodes into the graph to perform an action between one node and another.
For a more realistic example of how this is useful, consider a common problem
with compilers. The compiler is normally written in one language, then the
compiler is used to compile the standard library. This essentially requires
dynamically creating new nodes for the tree, and can result in some pretty
scary build scripts. Fbuild makes it easy though:
import fbuild.builders.c
def build():
static = fbuild.builders.c.guess_static()
exe = static.build_exe('exe', [...])
my_static = fbuild.builders.c.guess_static(exe)
my_static.build_lib('lib', [...])
First, we configure the system C builder. Next, we reconfigure another C
builder with the one we just created. Finally, we use that builder to compile
our standard library. As far as I can tell, most other build systems require
some painful contortions to support this, and I suspect most projects cheat by
ignoring the dependency between the standard library and the compiler.
Fbuild inherently supports this style since we don't have to work around the
graph.
Some systems allow for dynamically specifying dependency nodes in their script,
such as in SCons. As far as I can tell, this is the simplest node you can
create:
def action(target, source, env):
pass
env = Environment(BUILDERS={'A': Builder(action=Action(action))})
env.A('foo.out', 'foo.in')
First, you define a simple function that converts a source file into a target
file. Next, you create an environment where you update the builders with your
custom builder. Finally, you use the environment to perform your action. It's
not actually that bad. Here's the equivalent in Fbuild:
import fbuild.db
@fbuild.db.caches
def action(target:fbuild.db.DST, source:fbuild.dst.SRC):
pass
def build():
action('foo.out', 'foo.in')
It's actually a little longer. Part of that is that SCons imports all of the
common code into the global namespace. This can save some lines, but it also
forces a lot of SCons into the same namespace. This forces each node to be
uniquely named, and is more likely to run into conflicts. Anyway, for Fbuild
we need to do a couple extra things to get the to cache. First, we mark that
the function is cached using the fbuild.db.caches decorator. There's also
fbuild.db.cachemethod for decorating methods, and
fbuild.db.PersistentObject parent class for caching object creation. We also
need to tell Fbuild if the arguments are source or destination files. We do
this using python 3.0's annotation support. This tells the database to check to
see if source files were modified or if the destination file was removed. If
either of these occurred, the function is rerun.
As the builders get nontrivial and more complex, Fbuild becomes a much
simpler. For instance, consider SCons's Yacc builder.
There are three main things it needs to do. First, SCons has to set up the
environment variables for yacc. Next, they register yacc with the C
builder. Finally, they have a function that uses the values from the
environment and actually evaluates yacc.
Fbuild's current bison builder is much more straight forward:
import fbuild
import fbuild.builders
import fbuild.db
from fbuild.path import Path
class Bison(fbuild.db.PersistentObject):
def __init__(self, exe=None, flags=[], *, suffix='.c'):
self.exe = fbuild.builders.find_program([exe or 'bison'])
self.flags = flags
self.suffix = suffix
def __call__(self, src:fbuild.db.SRC, dst=None, *,
suffix=None,
verbose=False,
name_prefix=None,
defines=False,
flags=[],
buildroot=None) -> fbuild.db.DST:
buildroot = buildroot or fbuild.buildroot
suffix = suffix or self.suffix
dst = Path.addroot(dst or src, buildroot).replaceext(suffix)
dst.parent.makedirs()
cmd = [self.exe]
if verbose:
cmd.append('-v')
if name_prefix is not None:
cmd.extend(('-p', name_prefix))
if defines:
cmd.append('-d')
cmd.extend(self.flags)
cmd.extend(flags)
cmd.extend(('-o', dst))
cmd.append(src)
fbuild.execute(cmd, self.exe, '%s -> %s' % (src, dst), color='yellow')
return dst
We do some simple configuration to set up the executable and flags. Then we define a function to evaluate the executable. Since we're able to wrap a callable object, we don't have to be aware of anything else happening in the tree. We also don't need to pollute the environment with extra values, when they're only needed in this builder.
It's easy to extend this paradigm to configuration. Say you want to check if a header exists. It's simple to write this:
import fbuild.builders.c
import fbuild.db
@fbuild.db.caches
def check_stdlib_h(builder):
return builder.header_exists('stdlib_h')
def build():
static = fbuild.builders.c.guess_static()
if check_stdlib_h(builder):
...
else:
...
It's even easier using fbuild.config though. That package replicates much of
the functionality in autoconf to provide a convenient lazy configuration
system:
import fbuild.builders.c
import fbuild.config.c.c90
def build():
static = fbuild.buildesr.c.guess_static()
stdlib_h = fbuild.config.c.c90.stdlib_h(static)
if stdlib_h.header:
...
else:
...
if stdlib_h.atof.return_type == 'float':
...
else:
...
There are already hundreds of tests defined in fbuild.config. It's a new
system, but it's already proven to be pretty useful.
So, that's Fbuild. Hopefully you'll find it interesting. It's still a very
young project, and the interfaces aren't stable. But if you still want to
check it out, you can find the source
here. If you want to try
it out, you'll need python 3.0 and git installed. You can check it out with:
git clone http://git.felix-lang.org/r/fbuild.git
There are a handful of examples in the examples directory. These can be run
either with examples/examples.py, or run directly as in:
cd examples/c
../../fbuild-light
Where fbuild-light is just a simple wrapper around the fbuild executable.
It's used so that you can run fbuild without it being installed.
If you have any questions, feel free to ask on the felix mailing list, or reply to the comments.