Easily build binary dependencies for Julia packages
Since there seems to be a lot of confusion surrounding the package systems and the role of this package, before we get started looking at the actual package, I want to answer a few common questions:
BinDeps is a package that provides a collection of tools to build binary dependencies for Julia packages.
Do I need to use this package if I want to build binary dependencies for my Julia package?
Absolutely not! The system is designed to give the maximum amount
of freedom to the package author in order to be able to address any
situation that one may encounter in the real world. This is achieved
by simply evaluating a file called
deps/build.jl (if it exists) in
a package whenever it is installed or updated. Thus the following
might perhaps be the simplest possible useful
one can imagine:
I want to use BinDeps, but it is missing some functionality I need (e.g. a package manager)
Since BinDeps is written in Julia it is extensible with the same ease as the rest of Julia. In particular, defining new behavior, e.g. for adding a new package manager, consists of little more than adding a type and implementing a couple of methods (see the section on Interfaces) or the WinRPM package for an example implementation.
I like the runtime features that BinDeps provides, but I don't really want to use its build time capabilities. What do you recommend?
The easiest way to do this is probably just to declare a
BuildProcess for all your declared dependencies. This way, your
custom build process will be called whenever there is an unsatisfied
library dependency and you may still use the BinDeps runtime
Is there anything I should keep in mind when extending BinDeps or writing my own build process?
BinDeps uses a fairly standard set of directories by default and if possible, using the same directory structure is advised. Currently the specified directory structure is:
deps/ build.jl # This is your build file downloads/ # Store any binary/source downloads here builds/ dep1/ # out-of-tree build for dep1, is possible dep2/ # out-of-tree build for dep2, is possible ... src/ dep1/ # Source code for dep1 dep2/ # Source code for dep2 ... usr/ # "prefix", install your binaries here lib/ # Dynamic libraries (yes even on Windows) bin/ # Excecutables include/ # Headers ...
To get a feel for the high level interface provided by BinDeps, have a look at real-world examples. The build script from the GSL pakage illustrates the simple case where only one library is needed. On the other hand, the build script from the Cairo package uses almost all the features that BinDeps currently provides and offers a complete overview. Let's take it apart, to see exactly what's going on.
As you can see Cairo depends on a lot of libraries that all need to be managed by this build script.
Every one of these library dependencies is introduced by the
library_dependency function. The only required argument
is the name of the library, so the following would be an entirely valid call:
foo = library_dependency("libfoo")
However, you'll most likely quickly run into the issue that this library is named differently on different systems.
(If this happens, you will receive an error such as
Provider Binaries failed to satisfy dependency libfoo.)
why BinDeps provides the handy
aliases keyword argument. So suppose our library is sometimes known as
other times as
libfoo-1.0.0.dylib or even
libbar.dll on windows, because the authors of the library
decided to punish windows users. In either case, we can easily declare all these in our library dependency:
foo = library_dependency("libfoo", aliases = ["libfoo", "libfoo-1", "libfoo-1.0.0", "libbar"])
So far so good! There are a couple of other keyword arguments that are currently implemented:
os = OS_NAME
Limits this dependency to certain operating systems. The same could be achieved by using the OS-specific macro, but
this setting applies to all uses of this dependency and avoids having to wrap all uses of this dependency in macros.
Note that the
os parameter must match the value of
Base.OS_NAME on the target platform with the special exception that
:Unix matches all Unix-like platforms (e.g.
Mac OS X,
As an example, consider this line from the Cairo build script:
gettext = library_dependency("gettext", aliases = ["libgettext", "libgettextlib"], os = :Unix)
depends = [dep1, dep2]
Currently unused, but in the future will be used to keep track of the dependency graph between binary dependencies to allow parallel builds. E.g.:
cairo = library_dependency("cairo", aliases = ["libcairo-2", "libcairo"], depends = [gobject, fontconfig, libpng])
Whether or not to consider this a runtime dependency. If false, its absence
will not trigger an error at runtime (and it will not be loaded), but if it
cannot be found at buildtime it will be installed. This is useful for build-time
dependencies of other binary dependencies.
You may pass a function to validate whether or not a certain library is usable,
e.g. whether or not has the correct version. To do so, pass a function that takes
(name,handle) as an argument and returns
true if the library is usable and
it not. The
name argument is either an absolute path or the library name if it is a
global system library, while the handle is a handle that may be passed to
check library symbols or the return value of a function.
Should the validation return false for a library that was installed by a provider, the
provider will be instructed to force a rebuild.
function validate_cairo_version(name,handle) f = Libdl.dlsym_e(handle, "cairo_version") f == C_NULL && return false v = ccall(f, Int32,()) return v > 10800 end ... cairo = library_dependency("cairo", aliases = ["libcairo-2", "libcairo"], validate = validate_cairo_version)
Other keyword arguments will most likely be added as necessary.
Alright, now that we have declared all the dependencies that we need let's tell BinDeps how to build them. One of the easiest ways to do so is to use the system package manager. So suppose we have defined the following dependencies:
foo = library_dependency("libfoo") baz = library_dependency("libbaz")
Let's suppose that these libraries are available in the
in apt-get and that both libraries are installed by the
baz or the
baz1 yum package, and the
baz pacman package. We may
declare this as follows:
provides(AptGet, Dict("libfoo-dev" => foo, "libbaz-dev" => baz)) provides(Yum, ["baz", "baz1"], [foo, baz]) provides(Pacman, "baz", [foo, baz])
One may remember the
provides function by thinking
provides the dependencies
The basic signature of the provides function is
provides(Provider, data, dependency, options...)
data is provider-specific (e.g. a string in all of the package manager
dependency is the return value from
library_dependency. As you saw
above multiple definitions may be combined into one function call as such:
provides(Provider, Dict(data1=>dep1, data2=>dep2), options...)
which is equivalent to (and in fact will be internally dispatched) to:
provides(Provider, data1, dep1, options...) provides(Provider, data2, dep2, options...)
If one provide satisfied multiple dependencies simultaneously,
also be an array of dependencies (as in the
Pacman cases above).
There are also several builtin options. Some of them are:
os = OS_NAME # e.g. :Linux, :Windows, :Darwin
This provider can only satisfy the library dependency on the specified
This argument takes has the same syntax as the
os keyword argument to
installed_libpath = "path"
If the provider installs a library dependency to someplace other than the standard search paths, that location can be specified here.
SHA = "sha"
Provides a SHA-256 checksum to validate a downloaded source or binary file against.
We have already seen the
Yum providers, which all take a string naming the package as
their data argument. The other build-in providers are:
URI object as its data argument and declared that the sources may be
downloaded from the provided URI. This dependency is special, because it's
success does not automatically mark the build as succeeded (in BinDeps
terminology, it's a "helper"). By default this provider expects the unpacked
directory name to be that of the archive downloaded. If that is not the case,
you may use the :unpacked_dir option to specify the name of the unpacked directory,
provides(Sources,URI("http://libvirt.org/sources/libvirt-1.1.1-rc2.tar.gz"), libvirt, unpacked_dir = "libvirt-1.1.1")
If given a
URI object as its data argument, indicates that the binaries may be
downloaded from the provided URI. It is assumed that the binaries unpack the
usr/lib. If given a
String as its data argument, provides
a custom search path for the binaries. A typical use might be to allow the
user to provide a custom path by using an environmental variable.
Common super class of various kind of build processes. The exact behavior depends on the
data argument. Some of the currently supported build processes are
A subclass of BuildProcess that that downloads the sources (as declared by the "Sources" provider) and attempts to install using Autotools. There is a plethora of options to change the behavior of this command. See the appropriate section of the manual (or even better, read the code) for more details on the available options.
A subclass of BuildProcess that takes any object that's part of the low-level interface and
could be passed to
run and simply executes that command.
To load dependencies without a runtime dependence on BinDeps, place code like the following near the start of the Package's primary file. Don't forget to change the error message to reflect the name of the package.
const depsfile = joinpath(dirname(@__FILE__), "..", "deps", "deps.jl") if isfile(depsfile) include(depsfile) else error("HDF5 not properly installed. Please run Pkg.build(\"HDF5\") then restart Julia.") end
This will make all your libraries available as variables named by the names you gave the dependency. E.g. if you declared a dependency as
libfoo variable will now contain a reference to that library that may be passed
ccall or similar functions.
The low level interface provides a number of utilities to write cross platform build scripts. It looks something like this (from the Cairo build script):
@build_steps begin GetSources(libpng) CreateDirectory(pngbuilddir) @build_steps begin ChangeDirectory(pngbuilddir) FileRule(joinpath(prefix,"lib","libpng15.dll"),@build_steps begin `cmake -DCMAKE_INSTALL_PREFIX="$prefix" -G"MSYS Makefiles" $pngsrcdir` `make` `cp libpng*.dll $prefix/lib` `cp libpng*.a $prefix/lib` `cp libpng*.pc $prefix/lib/pkgconfig` `cp pnglibconf.h $prefix/include` `cp $pngsrcdir/png.h $prefix/include` `cp $pngsrcdir/pngconf.h $prefix/include` end) end end
All the steps are executed synchronously. The result of the
may be passed to run to execute it directly, thought this is not recommended other
than for debugging purposes. Instead, please use the high level interface to tie
the build process to dependencies.
Some of the builtin build steps are:
Download a file from
remote_file create it as
Unpack the file `local_file` into the folder `folder`
Invoke autotools. Use of this build step is not recommended. Use the high level interface instead
Create the directory
cd into the directory
dir and try to remain there for this build block. Must
be the first command in a
@build_steps block and will remain active for the entire block
make with the given arguments in the given directory with the given environment.
dir does not exist invoke step and validate that the directory was created
Like Directory rule, but validates the existence of any of the files in the
Get the declared sources from the dependency dep and put them in the default
A simple way to see what libraries are required by a package, and to detect missing dependencies,
is to use BinDeps.debug("PackageName")
julia> using BinDeps
julia> BinDeps.debug("Cairo") INFO: Reading build script... The package declares 1 dependencies.
about 1 month ago