Packaging Guide

This guide is intended for developers or administrators who want to package software so that Spack can install it. It assumes that you have at least some familiarity with Python, and that you've read the basic usage guide, especially the part about specs.

There are two key parts of Spack:

  1. Specs: expressions for describing builds of software, and

  2. Packages: Python modules that describe how to build software according to a spec.

Specs allow a user to describe a particular build in a way that a package author can understand. Packages allow the packager to encapsulate the build logic for different versions, compilers, options, platforms, and dependency combinations in one place. Essentially, a package translates a spec into build logic.

Packages in Spack are written in pure Python, so you can do anything in Spack that you can do in Python. Python was chosen as the implementation language for two reasons. First, Python is becoming ubiquitous in the scientific software community. Second, it's a modern language and has many powerful features to help make package writing easy.

Creating & editing packages

spack create

The spack create command creates a directory with the package name and generates a file with a boilerplate package template. If given a URL pointing to a tarball or other software archive, spack create is smart enough to determine basic information about the package, including its name and build system. In most cases, spack create plus a few modifications is all you need to get a package working.

Here's an example:

$ spack create

Spack examines the tarball URL and tries to figure out the name of the package to be created. If the name contains uppercase letters, these are automatically converted to lowercase. If the name contains underscores or periods, these are automatically converted to dashes.

Spack also searches for additional versions located in the same directory of the website. Spack prompts you to tell you how many versions it found and asks you how many you would like to download and checksum:

$ spack create
==> This looks like a URL for gmp
==> Found 16 versions of gmp:


How many would you like to checksum? (default is 1, q to abort)

Spack will automatically download the number of tarballs you specify (starting with the most recent) and checksum each of them.

You do not have to download all of the versions up front. You can always choose to download just one tarball initially, and run spack checksum later if you need more versions.

Let's say you download 3 tarballs:

How many would you like to checksum? (default is 1, q to abort) 3
==> Downloading...
==> Fetching
######################################################################## 100.0%
==> Fetching
######################################################################## 100.0%
==> Fetching
######################################################################## 100.0%
==> Checksummed 3 versions of gmp:
==> This package looks like it uses the autotools build system
==> Created template for gmp package
==> Created package file: /Users/Adam/spack/var/spack/repos/builtin/packages/gmp/

Spack automatically creates a directory in the appropriate repository, generates a boilerplate template for your package, and opens up the new in your favorite $EDITOR:

# This is a template package file for Spack.  We've put "FIXME"
# next to all the things you'll want to change. Once you've handled
# them, you can save this file and test your package like this:
#     spack install gmp
# You can edit this file again by typing:
#     spack edit gmp
# See the Spack documentation for more information on packaging.
# If you submit this package back to Spack as a pull request,
# please first remove this boilerplate and all FIXME comments.
from spack import *

class Gmp(AutotoolsPackage):
    """FIXME: Put a proper description of your package here."""

    # FIXME: Add a proper url for your package's homepage here.
    homepage = ""
    url      = ""

    version('6.1.2', '8ddbb26dc3bd4e2302984debba1406a5')
    version('6.1.1', '4c175f86e11eb32d8bf9872ca3a8e11d')
    version('6.1.0', '86ee6e54ebfc4a90b643a65e402c4048')

    # FIXME: Add dependencies if required.
    # depends_on('foo')

    def configure_args(self):
        # FIXME: Add arguments other than --prefix
        # FIXME: If not needed delete the function
        args = []
        return args

The tedious stuff (creating the class, checksumming archives) has been done for you. You'll notice that spack create correctly detected that gmp uses the Autotools build system. It created a new Gmp package that subclasses the AutotoolsPackage base class. This base class provides basic installation methods common to all Autotools packages:

./configure --prefix=/path/to/installation/directory

make check
make install

For most Autotools packages, this is sufficient. If you need to add additional arguments to the ./configure call, add them via the configure_args function.

In the generated package, the download url attribute is already set. All the things you still need to change are marked with FIXME labels. You can delete the commented instructions between the license and the first import statement after reading them. The rest of the tasks you need to do are as follows:

  1. Add a description.

    Immediately inside the package class is a docstring in triple-quotes ("""). It is used to generate the description shown when users run spack info.

  2. Change the homepage to a useful URL.

    The homepage is displayed when users run spack info so that they can learn more about your package.

  3. Add depends_on() calls for the package's dependencies.

    depends_on tells Spack that other packages need to be built and installed before this one. See Dependencies.

  4. Get the installation working.

    Your new package may require specific flags during configure. These can be added via configure_args. Specifics will differ depending on the package and its build system. Implementing the install method is covered in detail later.

Passing a URL to spack create is a convenient and easy way to get a basic package template, but what if your software is licensed and cannot be downloaded from a URL? You can still create a boilerplate by telling spack create what name you want to use:

$ spack create --name intel

This will create a simple intel package with an install() method that you can craft to install your package.

What if spack create <url> guessed the wrong name or build system? For example, if your package uses the Autotools build system but does not come with a configure script, Spack won't realize it uses Autotools. You can overwrite the old package with --force and specify a name with --name or a build system template to use with --template:

$ spack create --name gmp
$ spack create --force --template autotools


If you are creating a package that uses the Autotools build system but does not come with a configure script, you'll need to add an autoreconf method to your package that explains how to generate the configure script. You may also need the following dependencies:

depends_on('autoconf', type='build')
depends_on('automake', type='build')
depends_on('libtool',  type='build')
depends_on('m4',       type='build')

A complete list of available build system templates can be found by running spack create --help.

spack edit

One of the easiest ways to learn how to write packages is to look at existing ones. You can edit a package file by name with the spack edit command:

$ spack edit gmp

So, if you used spack create to create a package, then saved and closed the resulting file, you can get back to it with spack edit. The gmp package actually lives in $SPACK_ROOT/var/spack/repos/builtin/packages/gmp/, but spack edit provides a much simpler shortcut and saves you the trouble of typing the full path.

Naming & directory structure

This section describes how packages need to be named, and where they live in Spack's directory structure. In general, spack create handles creating package files for you, so you can skip most of the details here.


A Spack installation directory is structured like a standard UNIX install prefix (bin, lib, include, var, opt, etc.). Most of the code for Spack lives in $SPACK_ROOT/lib/spack. Packages themselves live in $SPACK_ROOT/var/spack/repos/builtin/packages.

If you cd to that directory, you will see directories for each package:

$ cd $SPACK_ROOT/var/spack/repos/builtin/packages && ls

Each directory contains a file called, which is where all the python code for the package goes. For example, the libelf package lives in:


Alongside the file, a package may contain extra directories or files (like patches) that it needs to build.

Package Names

Packages are named after the directory containing So, libelf's lives in a directory called libelf. The file defines a class called Libelf, which extends Spack's Package class. For example, here is $SPACK_ROOT/var/spack/repos/builtin/packages/libelf/

from spack import *

class Libelf(Package):
    """ ... description ... """
    homepage = ...
    url = ...

    def install():

The directory name (libelf) determines the package name that users should provide on the command line. e.g., if you type any of these:

$ spack info libelf
$ spack versions libelf
$ spack install libelf@0.8.13

Spack sees the package name in the spec and looks for libelf/ in var/spack/repos/builtin/packages. Likewise, if you run spack install py-numpy, Spack looks for py-numpy/

Spack uses the directory name as the package name in order to give packagers more freedom in naming their packages. Package names can contain letters, numbers, and dashes. Using a Python identifier (e.g., a class name or a module name) would make it difficult to support these options. So, you can name a package 3proxy or foo-bar and Spack won't care. It just needs to see that name in the packages directory.

Package class names

Spack loads files dynamically, and it needs to find a special class name in the file for the load to succeed. The class name (Libelf in our example) is formed by converting words separated by - in the file name to CamelCase. If the name starts with a number, we prefix the class name with _. Here are some examples:

Module Name

Class Name





In general, you won't have to remember this naming convention because spack create and spack edit handle the details for you.

Trusted Downloads

Spack verifies that the source code it downloads is not corrupted or compromised; or at least, that it is the same version the author of the Spack package saw when the package was created. If Spack uses a download method it can verify, we say the download method is trusted. Trust is important for all downloads: Spack has no control over the security of the various sites from which it downloads source code, and can never assume that any particular site hasn't been compromised.

Trust is established in different ways for different download methods. For the most common download method --- a single-file tarball --- the tarball is checksummed. Git downloads using commit= are trusted implicitly, as long as a hash is specified.

Spack also provides untrusted download methods: tarball URLs may be supplied without a checksum, or Git downloads may specify a branch or tag instead of a hash. If the user does not control or trust the source of an untrusted download, it is a security risk. Unless otherwise specified by the user for special cases, Spack should by default use only trusted download methods.

Unfortunately, Spack does not currently provide that guarantee. It does provide the following mechanisms for safety:

  1. By default, Spack will only install a tarball package if it has a checksum and that checksum matches. You can override this with spack install --no-checksum.

  2. Numeric versions are almost always tarball downloads, whereas non-numeric versions not named develop frequently download untrusted branches or tags from a version control system. As long as a package has at least one numeric version, and no non-numeric version named develop, Spack will prefer it over any non-numeric versions.


For tarball downloads, Spack can currently support checksums using the MD5, SHA-1, SHA-224, SHA-256, SHA-384, and SHA-512 algorithms. It determines the algorithm to use based on the hash length.

Versions and fetching

The most straightforward way to add new versions to your package is to add a line like this in the package class:

class Foo(Package):

    url = ""

    version('8.2.1', '4136d7b4c04df68b686570afa26988ac')
    version('8.2.0', '1c9f62f0778697a09d36121ead88e08e')
    version('8.1.2', 'd47dd09ed7ae6e7fd6f9a816d7f5fdf6')

Versions should be listed in descending order, from newest to oldest.

Date Versions

If you wish to use dates as versions, it is best to use the format @yyyy-mm-dd. This will ensure they sort in the correct order.

Alternately, you might use a hybrid release-version / date scheme. For example, @1.3_2016-08-31 would mean the version from the 1.3 branch, as of August 31, 2016.

Version URLs

By default, each version's URL is extrapolated from the url field in the package. For example, Spack is smart enough to download version 8.2.1 of the Foo package above from

If the URL is particularly complicated or changes based on the release, you can override the default URL generation algorithm by defining your own url_for_version() function. For example, the download URL for OpenMPI contains the major.minor version in one spot and the major.minor.patch version in another:

In order to handle this, you can define a url_for_version() function like so:

    def url_for_version(self, version):
        url = "{0}/downloads/openmpi-{1}.tar.bz2"
        return url.format(version.up_to(2), version)

With the use of this url_for_version(), Spack knows to download OpenMPI 2.1.1 from but download OpenMPI 1.10.7 from

You'll notice that OpenMPI's url_for_version() function makes use of a special Version function called up_to(). When you call version.up_to(2) on a version like 1.10.0, it returns 1.10. version.up_to(1) would return 1. This can be very useful for packages that place all X.Y.* versions in a single directory and then places all X.Y.Z versions in a sub-directory.

There are a few Version properties you should be aware of. We generally prefer numeric versions to be separated by dots for uniformity, but not all tarballs are named that way. For example, icu4c separates its major and minor versions with underscores, like icu4c-57_1-src.tgz. The value 57_1 can be obtained with the use of the version.underscored property. Note that Python properties don't need parentheses. There are other separator properties as well:












Python properties don't need parentheses. version.dashed is correct. version.dashed() is incorrect.

In addition, these version properties can be combined with up_to(). For example:

>>> version = Version('1.2.3')
>>> version.up_to(2).dashed
>>> version.underscored.up_to(2)

As you can see, order is not important. Just keep in mind that up_to() and the other version properties return Version objects, not strings.

If a URL cannot be derived systematically, or there is a special URL for one of its versions, you can add an explicit URL for a particular version:

version('8.2.1', '4136d7b4c04df68b686570afa26988ac',

When you supply a custom URL for a version, Spack uses that URL verbatim and does not perform extrapolation. The order of precedence of these methods is:

  1. package-level url

  2. url_for_version()

  3. version-specific url

so if your package contains a url_for_version(), it can be overridden by a version-specific url.

If your package does not contain a package-level url or url_for_version(), Spack can determine which URL to download from even if only some of the versions specify their own url. Spack will use the nearest URL before the requested version. This is useful for packages that have an easy to extrapolate URL, but keep changing their URL format every few releases. With this method, you only need to specify the url when the URL changes.

Skipping the expand step

Spack normally expands archives (e.g. *.tar.gz and *.zip) automatically after downloading them. If you want to skip this step (e.g., for self-extracting executables and other custom archive types), you can add expand=False to a version directive.

version('8.2.1', '4136d7b4c04df68b686570afa26988ac',
        url='', expand=False)

When expand is set to False, Spack sets the current working directory to the directory containing the downloaded archive before it calls your install method. Within install, the path to the downloaded archive is available as self.stage.archive_file.

Here is an example snippet for packages distributed as self-extracting archives. The example sets permissions on the downloaded file to make it executable, then runs it with some arguments.

def install(self, spec, prefix):
    installer = Executable(self.stage.archive_file)
    installer('--prefix=%s' % prefix, 'arg1', 'arg2', 'etc.')

Download caching

Spack maintains a cache (described here) which saves files retrieved during package installations to avoid re-downloading in the case that a package is installed with a different specification (but the same version) or reinstalled on account of a change in the hashing scheme.

Version comparison

Most Spack versions are numeric, a tuple of integers; for example, apex@0.1, ferret@6.96 or py-netcdf@ Spack knows how to compare and sort numeric versions.

Some Spack versions involve slight extensions of numeric syntax; for example, py-sphinx-rtd-theme@0.1.10a0. In this case, numbers are always considered to be "newer" than letters. This is for consistency with RPM.

Spack versions may also be arbitrary non-numeric strings; any string here will suffice; for example, @develop, @master, @local. Versions are compared as follows. First, a version string is split into multiple fields based on delimiters such as ., - etc. Then matching fields are compared using the rules below:

  1. The following develop-like strings are greater (newer) than all numbers and are ordered as develop > master > head > trunk.

  2. Numbers are all less than the chosen develop-like strings above, and are sorted numerically.

  3. All other non-numeric versions are less than numeric versions, and are sorted alphabetically.

The logic behind this sort order is two-fold:

  1. Non-numeric versions are usually used for special cases while developing or debugging a piece of software. Keeping most of them less than numeric versions ensures that Spack chooses numeric versions by default whenever possible.

  2. The most-recent development version of a package will usually be newer than any released numeric versions. This allows the @develop version to satisfy dependencies like depends_on(abc, when="@x.y.z:")

Version selection

When concretizing, many versions might match a user-supplied spec. For example, the spec python matches all available versions of the package python. Similarly, python@3: matches all versions of Python3. Given a set of versions that match a spec, Spack concretization uses the following priorities to decide which one to use:

  1. If the user provided a list of versions in packages.yaml, the first matching version in that list will be used.

  2. If one or more versions is specified as preferred=True, in either packages.yaml or, the largest matching version will be used. ("Latest" is defined by the sort order above).

  3. If no preferences in particular are specified in the package or in packages.yaml, then the largest matching non-develop version will be used. By avoiding @develop, this prevents users from accidentally installing a @develop version.

  4. If all else fails and @develop is the only matching version, it will be used.

spack checksum

If you want to add new versions to a package you've already created, this is automated with the spack checksum command. Here's an example for libelf:

$ spack checksum libelf
==> Found 16 versions of libelf.

How many would you like to checksum? (default is 1, q to abort)

This does the same thing that spack create does, but it allows you to go back and add new versions easily as you need them (e.g., as they're released). It fetches the tarballs you ask for and prints out a list of version commands ready to copy/paste into your package file:

==> Checksummed new versions of libelf:
    version('0.8.13', '4136d7b4c04df68b686570afa26988ac')
    version('0.8.12', 'e21f8273d9f5f6d43a59878dc274fec7')
    version('0.8.11', 'e931910b6d100f6caa32239849947fbf')
    version('0.8.10', '9db4d36c283d9790d8fa7df1f4d7b4d9')

By default, Spack will search for new tarball downloads by scraping the parent directory of the tarball you gave it. So, if your tarball is at, Spack will look in for links to additional versions. If you need to search another path for download links, you can supply some extra attributes that control how your package finds new versions. See the documentation on list_url and list_depth.


  • This command assumes that Spack can extrapolate new URLs from an existing URL in the package, and that Spack can find similar URLs on a webpage. If that's not possible, e.g. if the package's developers don't name their tarballs consistently, you'll need to manually add version calls yourself.

  • For spack checksum to work, Spack needs to be able to import your package in Python. That means it can't have any syntax errors, or the import will fail. Use this once you've got your package in working order.

Finding new versions

You've already seen the homepage and url package attributes:

from spack import *

class Mpich(Package):
   """MPICH is a high performance and widely portable implementation of
      the Message Passing Interface (MPI) standard."""
   homepage = ""
   url      = ""

These are class-level attributes used by Spack to show users information about the package, and to determine where to download its source code.

Spack uses the tarball URL to extrapolate where to find other tarballs of the same package (e.g. in spack checksum, but this does not always work. This section covers ways you can tell Spack to find tarballs elsewhere.


When spack tries to find available versions of packages (e.g. with spack checksum), it spiders the parent directory of the tarball in the url attribute. For example, for libelf, the url is:

url = ""

Here, Spack spiders to find similar tarball links and ultimately to make a list of available versions of libelf.

For many packages, the tarball's parent directory may be unlistable, or it may not contain any links to source code archives. In fact, many times additional package downloads aren't even available in the same directory as the download URL.

For these, you can specify a separate list_url indicating the page to search for tarballs. For example, libdwarf has the homepage as the list_url, because that is where links to old versions are:

class Libdwarf(Package):
    homepage = ""
    url      = ""
    list_url = homepage


libdwarf and many other packages have a listing of available versions on a single webpage, but not all do. For example, mpich has a tarball URL that looks like this:

url = ""

But its downloads are in many different subdirectories of So, we need to add a list_url and a list_depth attribute:

class Mpich(Package):
    homepage   = ""
    url        = ""
    list_url   = ""
    list_depth = 1

By default, Spack only looks at the top-level page available at list_url. list_depth = 1 tells it to follow up to 1 level of links from the top-level page. Note that here, this implies 1 level of subdirectories, as the mpich website is structured much like a filesystem. But list_depth really refers to link depth when spidering the page.

Fetching from code repositories

For some packages, source code is provided in a Version Control System (VCS) repository rather than in a tarball. Spack can fetch packages from VCS repositories. Currently, Spack supports fetching with Git, Mercurial (hg), Subversion (svn), and Go.

To fetch a package from a source repository, Spack needs to know which VCS to use and where to download from. Much like with url, package authors can specify a class-level git, hg, svn, or go attribute containing the correct download location.

Many packages developed with Git have both a Git repository as well as release tarballs available for download. Packages can define both a class-level tarball URL and VCS. For example:

class Trilinos(CMakePackage):

    homepage = ""
    url      = ""
    git      = ""

    version('develop', branch='develop')
    version('master',  branch='master')
    version('12.12.1', 'ecd4606fa332212433c98bf950a69cc7')
    version('12.10.1', '667333dbd7c0f031d47d7c5511fd0810')
    version('12.8.1',  '9f37f683ee2b427b5540db8a20ed6b15')

If a package contains both a url and git class-level attribute, Spack decides which to use based on the arguments to the version() directive. Versions containing a specific branch, tag, or revision are assumed to be for VCS download methods, while versions containing a checksum are assumed to be for URL download methods.

Like url, if a specific version downloads from a different repository than the default repo, it can be overridden with a version-specific argument.


In order to reduce ambiguity, each package can only have a single VCS top-level attribute in addition to url. In the rare case that a package uses multiple VCS, a fetch strategy can be specified for each version. For example, the rockstar package contains:

class Rockstar(MakefilePackage):

    homepage = ""

    version('develop', git='')
    version('yt', hg='')


Git fetching supports the following parameters to version:

  • git: URL of the git repository, if different than the class-level git.

  • branch: Name of a branch to fetch.

  • tag: Name of a tag to fetch.

  • commit: SHA hash (or prefix) of a commit to fetch.

  • submodules: Also fetch submodules recursively when checking out this repository.

Only one of tag, branch, or commit can be used at a time.

Default branch

To fetch a repository's default branch:

class Example(Package):

    git = ""


This download method is untrusted, and is not recommended. Aside from HTTPS, there is no way to verify that the repository has not been compromised, and the commit you get when you install the package likely won't be the same commit that was used when the package was first written. Additionally, the default branch may change. It is best to at least specify a branch name.


To fetch a particular branch, use the branch parameter:

version('experimental', branch='experimental')

This download method is untrusted, and is not recommended. Branches are moving targets, so the commit you get when you install the package likely won't be the same commit that was used when the package was first written.


To fetch from a particular tag, use tag instead:

version('1.0.1', tag='v1.0.1')

This download method is untrusted, and is not recommended. Although tags are generally more stable than branches, Git allows tags to be moved. Many developers use tags to denote rolling releases, and may move the tag when a bug is patched.


Finally, to fetch a particular commit, use commit:

version('2014-10-08', commit='9d38cd4e2c94c3cea97d0e2924814acc')

This doesn't have to be a full hash; you can abbreviate it as you'd expect with git:

version('2014-10-08', commit='9d38cd')

This download method is trusted. It is the recommended way to securely download from a Git repository.

It may be useful to provide a saner version for commits like this, e.g. you might use the date as the version, as done above. Or, if you know the commit at which a release was cut, you can use the release version. It's up to the package author to decide what makes the most sense. Although you can use the commit hash as the version number, this is not recommended, as it won't sort properly.


You can supply submodules=True to cause Spack to fetch submodules recursively along with the repository at fetch time. For more information about git submodules see the manpage of git: man git-submodule.

version('1.0.1', tag='v1.0.1', submodules=True)


If a project is hosted on GitHub, any valid Git branch, tag, or hash may be downloaded as a tarball. This is accomplished simply by constructing an appropriate URL. Spack can checksum any package downloaded this way, thereby producing a trusted download. For example, the following downloads a particular hash, and then applies a checksum.

version('', 'd035e4bc704d136db79b43ab371b27d2',


Fetching with Mercurial works much like Git, but you use the hg parameter.

Default branch

Add the hg attribute with no revision passed to version:

class Example(Package):

    hg = ""


This download method is untrusted, and is not recommended. As with Git's default fetching strategy, there is no way to verify the integrity of the download.


To fetch a particular revision, use the revision parameter:

version('1.0', revision='v1.0')

Unlike git, which has special parameters for different types of revisions, you can use revision for branches, tags, and commits when you fetch with Mercurial. Like Git, fetching specific branches or tags is an untrusted download method, and is not recommended. The recommended fetch strategy is to specify a particular commit hash as the revision.


To fetch with subversion, use the svn and revision parameters.

Fetching the head

Simply add an svn parameter to the package:

class Example(Package):

    svn = ""


This download method is untrusted, and is not recommended for the same reasons as mentioned above.

Fetching a revision

To fetch a particular revision, add a revision argument to the version directive:

version('develop', revision=128)

This download method is untrusted, and is not recommended.

Unfortunately, Subversion has no commit hashing scheme like Git and Mercurial do, so there is no way to guarantee that the download you get is the same as the download used when the package was created. Use at your own risk.

Subversion branches are handled as part of the directory structure, so you can check out a branch or tag by changing the URL. If you want to package multiple branches, simply add a svn argument to each version directive.


Go isn't a VCS, it is a programming language with a builtin command, go get, that fetches packages and their dependencies automatically. It can clone a Git repository, or download from another source location. For example:

class ThePlatinumSearcher(Package):

    homepage = ""
    go       = ""


Go cannot be used to fetch a particular commit or branch, it always downloads the head of the repository. This download method is untrusted, and is not recommended. Use another fetch strategy whenever possible.

and is not recommended. Use another fetch strategy whenever possible.


Many software packages can be configured to enable optional features, which often come at the expense of additional dependencies or longer build-times. To be flexible enough and support a wide variety of use cases, Spack permits to expose to the end-user the ability to choose which features should be activated in a package at the time it is installed. The mechanism to be employed is the spack.directives.variant() directive.

Boolean variants

In their simplest form variants are boolean options specified at the package level:

class Hdf5(AutotoolsPackage):
        'shared', default=True, description='Builds a shared version of the library'

with a default value and a description of their meaning / use in the package. Variants can be tested in any context where a spec constraint is expected. In the example above the shared variant is tied to the build of shared dynamic libraries. To pass the right option at configure time we can branch depending on its value:

def configure_args(self):
    if '+shared' in self.spec:

As explained in Variants the constraint +shared means that the boolean variant is set to True, while ~shared means it is set to False. Another common example is the optional activation of an extra dependency which requires to use the variant in the when argument of spack.directives.depends_on():

class Hdf5(AutotoolsPackage):
    variant('szip', default=False, description='Enable szip support')
    depends_on('szip', when='+szip')

as shown in the snippet above where szip is modeled to be an optional dependency of hdf5.

Multi-valued variants

If need be, Spack can go beyond Boolean variants and permit an arbitrary number of allowed values. This might be useful when modeling options that are tightly related to each other. The values in this case are passed to the spack.directives.variant() directive as a tuple:

class Blis(Package):
        'threads', default='none', description='Multithreading support',
        values=('pthreads', 'openmp', 'none'), multi=False

In the example above the argument multi is set to False to indicate that only one among all the variant values can be active at any time. This constraint is enforced by the parser and an error is emitted if a user specifies two or more values at the same time:

$ spack spec blis threads=openmp,pthreads
Input spec
blis threads=openmp,pthreads

==> Error: multiple values are not allowed for variant "threads"

Another useful note is that Python's None is not allowed as a default value and therefore it should not be used to denote that no feature was selected. Users should instead select another value, like 'none', and handle it explicitly within the package recipe if need be:

if self.spec.variants['threads'].value == 'none':

In cases where multiple values can be selected at the same time multi should be set to True:

class Gcc(AutotoolsPackage):
        'languages', default='c,c++,fortran',
        values=('ada', 'brig', 'c', 'c++', 'fortran',
                'go', 'java', 'jit', 'lto', 'objc', 'obj-c++'),
        description='Compilers and runtime libraries to build'

Within a package recipe a multi-valued variant is tested using a key=value syntax:

if 'languages=jit' in spec:

Complex validation logic for variant values

To cover complex use cases, the spack.directives.variant() directive could accept as the values argument a full-fledged object which has default and other arguments of the directive embedded as attributes.

An example, already implemented in Spack's core, is spack.variant.DisjointSetsOfValues. This class is used to implement a few convenience functions, like spack.variant.any_combination_of():

class Adios(AutotoolsPackage):
        values=any_combination_of('flexpath', 'dataspaces'),
        description='Enable dataspaces and/or flexpath staging transports'

that allows any combination of the specified values, and also allows the user to specify 'none' (as a string) to choose none of them. The objects returned by these functions can be modified at will by chaining method calls to change the default value, customize the error message or other similar operations:

class Mvapich2(AutotoolsPackage):
        description='List of the process managers to activate',
            ('auto',), ('slurm',), ('hydra', 'gforker', 'remshell')
            "'slurm' or 'auto' cannot be activated along with "
            "other process managers"

Resources (expanding extra tarballs)

Some packages (most notably compilers) provide optional features if additional resources are expanded within their source tree before building. In Spack it is possible to describe such a need with the resource directive :


Based on the keywords present among the arguments the appropriate FetchStrategy will be used for the resource. The keyword destination is relative to the source root of the package and should point to where the resource is to be expanded.

Licensed software

In order to install licensed software, Spack needs to know a few more details about a package. The following class attributes should be defined.


Boolean. If set to True, this software requires a license. If set to False, all of the following attributes will be ignored. Defaults to False.


String. Contains the symbol used by the license manager to denote a comment. Defaults to #.


List of strings. These are files that the software searches for when looking for a license. All file paths must be relative to the installation directory. More complex packages like Intel may require multiple licenses for individual components. Defaults to the empty list.


List of strings. Environment variables that can be set to tell the software where to look for a license if it is not in the usual location. Defaults to the empty list.


String. A URL pointing to license setup instructions for the software. Defaults to the empty string.

For example, let's take a look at the package for the PGI compilers.

# Licensing
license_required = True
license_comment  = '#'
license_files    = ['license.dat']
license_vars     = ['PGROUPD_LICENSE_FILE', 'LM_LICENSE_FILE']
license_url      = ''

As you can see, PGI requires a license. Its license manager, FlexNet, uses the # symbol to denote a comment. It expects the license file to be named license.dat and to be located directly in the installation prefix. If you would like the installation file to be located elsewhere, simply set PGROUPD_LICENSE_FILE or LM_LICENSE_FILE after installation. For further instructions on installation and licensing, see the URL provided.

Let's walk through a sample PGI installation to see exactly what Spack is and isn't capable of. Since PGI does not provide a download URL, it must be downloaded manually. It can either be added to a mirror or located in the current directory when spack install pgi is run. See ミラー for instructions on setting up a mirror.

After running spack install pgi, the first thing that will happen is Spack will create a global license file located at $SPACK_ROOT/etc/spack/licenses/pgi/license.dat. It will then open up the file using the editor set in $EDITOR, or vi if unset. It will look like this:

# A license is required to use pgi.
# The recommended solution is to store your license key in this global
# license file. After installation, the following symlink(s) will be
# added to point to this file (relative to the installation prefix):
#   license.dat
# Alternatively, use one of the following environment variable(s):
# If you choose to store your license in a non-standard location, you may
# set one of these variable(s) to the full pathname to the license file, or
# port@host if you store your license keys on a dedicated license server.
# You will likely want to set this variable in a module file so that it
# gets loaded every time someone tries to use pgi.
# For further information on how to acquire a license, please refer to:
# You may enter your license below.

You can add your license directly to this file, or tell FlexNet to use a license stored on a separate license server. Here is an example that points to a license server called licman1:

SERVER 00163eb7fba5 27200

If your package requires the license to install, you can reference the location of this global license using self.global_license_file. After installation, symlinks for all of the files given in license_files will be created, pointing to this global license. If you install a different version or variant of the package, Spack will automatically detect and reuse the already existing global license.

If the software you are trying to package doesn't rely on license files, Spack will print a warning message, letting the user know that they need to set an environment variable or pointing them to installation documentation.


Depending on the host architecture, package version, known bugs, or other issues, you may need to patch your software to get it to build correctly. Like many other package systems, spack allows you to store patches alongside your package files and apply them to source code after it's downloaded.


You can specify patches in your package file with the patch() directive. patch looks like this:

class Mvapich2(Package):
    patch('ad_lustre_rwcontig_open_source.patch', when='@1.9:')

The first argument can be either a URL or a filename. It specifies a patch file that should be applied to your source. If the patch you supply is a filename, then the patch needs to live within the spack source tree. For example, the patch above lives in a directory structure like this:


If you supply a URL instead of a filename, you need to supply a sha256 checksum, like this:


Spack includes the hashes of patches in its versioning information, so that the same package with different patches applied will have different hash identifiers. To ensure that the hashing scheme is consistent, you must use a sha256 checksum for the patch. Patches will be fetched from their URLs, checked, and applied to your source code. You can use the spack sha256 command to generate a checksum for a patch file or URL.

Spack can also handle compressed patches. If you use these, Spack needs a little more help. Specifically, it needs two checksums: the sha256 of the patch and archive_sha256 for the compressed archive. archive_sha256 helps Spack ensure that the downloaded file is not corrupted or malicious, before running it through a tool like tar or zip. The sha256 of the patch is still required so that it can be included in specs. Providing it in the package file ensures that Spack won't have to download and decompress patches it won't end up using at install time. Both the archive and patch checksum are checked when patch archives are downloaded.


patch keyword arguments are described below.

sha256, archive_sha256

Hashes of downloaded patch and compressed archive, respectively. Only needed for patches fetched from URLs.


If supplied, this is a spec that tells spack when to apply the patch. If the installed package spec matches this spec, the patch will be applied. In our example above, the patch is applied when mvapich is at version 1.9 or higher.


This tells spack how to run the patch command. By default, the level is 1 and spack runs patch -p 1. If level is 2, spack will run patch -p 2, and so on.

A lot of people are confused by level, so here's a primer. If you look in your patch file, you may see something like this:

--- a/src/mpi/romio/adio/ad_lustre/ad_lustre_rwcontig.c 2013-12-10 12:05:44.806417000 -0800
+++ b/src/mpi/romio/adio/ad_lustre/ad_lustre_rwcontig.c 2013-12-10 11:53:03.295622000 -0800
@@ -8,7 +8,7 @@
  *   Copyright (C) 2008 Sun Microsystems, Lustre group

-#define _XOPEN_SOURCE 600
+//#define _XOPEN_SOURCE 600
 #include <stdlib.h>
 #include <malloc.h>
 #include "ad_lustre.h"

Lines 1-2 show paths with synthetic a/ and b/ prefixes. These are placeholders for the two mvapich2 source directories that diff compared when it created the patch file. This is git's default behavior when creating patch files, but other programs may behave differently.

-p1 strips off the first level of the prefix in both paths, allowing the patch to be applied from the root of an expanded mvapich2 archive. If you set level to 2, it would strip off src, and so on.

It's generally easier to just structure your patch file so that it applies cleanly with -p1, but if you're using a patch you didn't create yourself, level can be handy.


This tells spack where to run the patch command. By default, the working directory is the source path of the stage (.). However, sometimes patches are made with respect to a subdirectory and this is where the working directory comes in handy. Internally, the working directory is given to patch via the -d option. Let's take the example patch from above and assume for some reason, it can only be downloaded in the following form:

--- a/romio/adio/ad_lustre/ad_lustre_rwcontig.c 2013-12-10 12:05:44.806417000 -0800
+++ b/romio/adio/ad_lustre/ad_lustre_rwcontig.c 2013-12-10 11:53:03.295622000 -0800
@@ -8,7 +8,7 @@
  *   Copyright (C) 2008 Sun Microsystems, Lustre group

-#define _XOPEN_SOURCE 600
+//#define _XOPEN_SOURCE 600
 #include <stdlib.h>
 #include <malloc.h>
 #include "ad_lustre.h"

Hence, the patch needs to applied in the src/mpi subdirectory, and the working_dir='src/mpi' option would exactly do that.

Patch functions

In addition to supplying patch files, you can write a custom function to patch a package's source. For example, the py-pyside package contains some custom code for tweaking the way the PySide build handles RPATH:

    def patch(self):
        """Undo PySide RPATH handling and add Spack RPATH."""
        # Figure out the special RPATH
        pypkg = self.spec['python'].package
        rpath = self.rpath
            self.prefix, pypkg.site_packages_dir, 'PySide'))

        # Add Spack's standard CMake args to the sub-builds.
        # They're called BY so we have to patch it.
            r'OPTION_CMAKE, ' + (
                '"-DCMAKE_INSTALL_RPATH=%s",' % ':'.join(rpath)),

        # PySide tries to patch ELF files to remove RPATHs
        # Disable this and go with the one we set.
        if self.spec.satisfies('@1.2.4:'):
            rpath_file = ''
            rpath_file = ''

        filter_file(r'(^\s*)(rpath_cmd\(.*\))', r'\1#\2', rpath_file)

        # TODO: rpath handling for PySide 1.2.4 still doesn't work.
        # PySide can't find the Shiboken library, even though it comes
        # bundled with it and is installed in the same directory.

        # PySide does not provide official support for
        # Python 3.5, but it should work fine
        filter_file("'Programming Language :: Python :: 3.4'",
                    "'Programming Language :: Python :: 3.4',\r\n        "
                    "'Programming Language :: Python :: 3.5'",

A patch function, if present, will be run after patch files are applied and before install() is run.

You could put this logic in install(), but putting it in a patch function gives you some benefits. First, spack ensures that the patch() function is run once per code checkout. That means that if you run install, hit ctrl-C, and run install again, the code in the patch function is only run once. Also, you can tell Spack to run only the patching part of the build using the spack patch command.

Dependency patching

So far we've covered how the patch directive can be used by a package to patch its own source code. Packages can also specify patches to be applied to their dependencies, if they require special modifications. As with all packages in Spack, a patched dependency library can coexist with other versions of that library. See the section on depends_on for more details.

Inspecting patches

If you want to better understand the patches that Spack applies to your packages, you can do that using spack spec, spack find, and other query commands. Let's look at m4. If you run spack spec m4, you can see the patches that would be applied to m4:

$ spack spec m4
Input spec

m4@1.4.18%clang@9.0.0-apple patches=3877ab548f88597ab2327a2230ee048d2d07ace1062efe81fc92e91b7f39cd00,c0a408fbffb7255fcc75e26bd8edab116fc81d216bfd18b473668b7739a4158e,fc9b61654a3ba1a8d6cd78ce087e7c96366c290bc8d2c299f09828d793b853c8 +sigsegv arch=darwin-highsierra-x86_64
    ^libsigsegv@2.11%clang@9.0.0-apple arch=darwin-highsierra-x86_64

You can also see patches that have been applied to installed packages with spack find -v:

$ spack find -v m4
==> 1 installed package
-- darwin-highsierra-x86_64 / clang@9.0.0-apple -----------------
m4@1.4.18 patches=3877ab548f88597ab2327a2230ee048d2d07ace1062efe81fc92e91b7f39cd00,c0a408fbffb7255fcc75e26bd8edab116fc81d216bfd18b473668b7739a4158e,fc9b61654a3ba1a8d6cd78ce087e7c96366c290bc8d2c299f09828d793b853c8 +sigsegv

In both cases above, you can see that the patches' sha256 hashes are stored on the spec as a variant. As mentioned above, this means that you can have multiple, differently-patched versions of a package installed at once.

You can look up a patch by its sha256 hash (or a short version of it) using the spack resource show command:

$ spack resource show 3877ab54
    path:       /home/spackuser/src/spack/var/spack/repos/builtin/packages/m4/gnulib-pgi.patch
    applies to: builtin.m4

spack resource show looks up downloadable resources from package files by hash and prints out information about them. Above, we see that the 3877ab54 patch applies to the m4 package. The output also tells us where to find the patch.

Things get more interesting if you want to know about dependency patches. For example, when dealii is built with boost@1.68.0, it has to patch boost to work correctly. If you didn't know this, you might wonder where the extra boost patches are coming from:

$ spack spec dealii ^boost@1.68.0 ^hdf5+fortran | grep '\^boost'
        ^boost@1.68.0%clang@9.0.0-apple+atomic+chrono~clanglibcpp cxxstd=default +date_time~debug+exception+filesystem+graph~icu+iostreams+locale+log+math~mpi+multithreaded~numpy patches=2ab6c72d03dec6a4ae20220a9dfd5c8c572c5294252155b85c6874d97c323199,b37164268f34f7133cbc9a4066ae98fda08adf51e1172223f6a969909216870f ~pic+program_options~python+random+regex+serialization+shared+signals~singlethreaded+system~taggedlayout+test+thread+timer~versionedlayout+wave arch=darwin-highsierra-x86_64
$ spack resource show b37164268
    path:       /home/spackuser/src/spack/var/spack/repos/builtin/packages/dealii/boost_1.68.0.patch
    applies to: builtin.boost
    patched by: builtin.dealii

Here you can see that the patch is applied to boost by dealii, and that it lives in dealii's directory in Spack's builtin package repository.

Handling RPATHs

Spack installs each package in a way that ensures that all of its dependencies are found when it runs. It does this using RPATHs. An RPATH is a search path, stored in a binary (an executable or library), that tells the dynamic loader where to find its dependencies at runtime. You may be familiar with LD_LIBRARY_PATH on Linux or DYLD_LIBRARY_PATH on Mac OS X. RPATH is similar to these paths, in that it tells the loader where to find libraries. Unlike them, it is embedded in the binary and not set in each user's environment.

RPATHs in Spack are handled in one of three ways:

  1. For most packages, RPATHs are handled automatically using Spack's compiler wrappers. These wrappers are set in standard variables like CC, CXX, F77, and FC, so most build systems (autotools and many gmake systems) pick them up and use them.

  2. CMake also respects Spack's compiler wrappers, but many CMake builds have logic to overwrite RPATHs when binaries are installed. Spack provides the std_cmake_args variable, which includes parameters necessary for CMake build use the right installation RPATH. It can be used like this when cmake is invoked:

    class MyPackage(Package):
        def install(self, spec, prefix):
            cmake('..', *std_cmake_args)
  3. If you need to modify the build to add your own RPATHs, you can use the self.rpath property of your package, which will return a list of all the RPATHs that Spack will use when it links. You can see this how this is used in the PySide example above.

Parallel builds

By default, Spack will invoke make() with a -j <njobs> argument, so that builds run in parallel. It figures out how many jobs to run by determining how many cores are on the host machine. Specifically, it uses the number of CPUs reported by Python's multiprocessing.cpu_count().

If a package does not build properly in parallel, you can override this setting by adding parallel = False to your package. For example, OpenSSL's build does not work in parallel, so its package looks like this:

class Openssl(Package):
    homepage = ""
    url      = ""

    version('1.0.1h', '8d6d684a9430d5cc98a62a5d8fbda8cf')

    parallel = False

Similarly, you can disable parallel builds only for specific make commands, as libdwarf does:

class Libelf(Package):

    def install(self, spec, prefix):
        configure("--prefix=" + prefix,

        # The mkdir commands in libelf's install can fail in parallel
        make("install", parallel=False)

The first make will run in parallel here, but the second will not. If you set parallel to False at the package level, then each call to make() will be sequential by default, but packagers can call make(parallel=True) to override it.


We've covered how to build a simple package, but what if one package relies on another package to build? How do you express that in a package file? And how do you refer to the other package in the build script for your own package?

Spack makes this relatively easy. Let's take a look at the libdwarf package to see how it's done:

class Libdwarf(Package):
    homepage = ""
    url      = ""
    list_url = homepage

    version('20130729', '4cc5e48693f7b93b7aa0261e63c0e21d')


    def install(self, spec, prefix):


The highlighted depends_on('libelf') call tells Spack that it needs to build and install the libelf package before it builds libdwarf. This means that in your install() method, you are guaranteed that libelf has been built and installed successfully, so you can rely on it for your libdwarf build.

Dependency specs

depends_on doesn't just take the name of another package. It can take a full spec as well. This means that you can restrict the versions or other configuration options of libelf that libdwarf will build with. For example, suppose that in the libdwarf package you write:


Now libdwarf will require libelf at exactly version 0.8. You can also specify a requirement for a particular variant or for specific compiler flags:

depends_on('libelf debug=True')
depends_on('libelf cppflags="-fPIC"')

Both users and package authors can use the same spec syntax to refer to different package configurations. Users use the spec syntax on the command line to find installed packages or to install packages with particular constraints, and package authors can use specs to describe relationships between packages.

Version ranges

Although some packages require a specific version for their dependencies, most can be built with a range of version. For example, if you are writing a package for a legacy Python module that only works with Python 2.4 through 2.6, this would look like:


Version ranges in Spack are inclusive, so 2.4:2.6 means any version greater than or equal to 2.4 and up to and including 2.6. If you want to specify that a package works with any version of Python 3, this would look like:


Here we leave out the upper bound. If you want to say that a package requires Python 2, you can similarly leave out the lower bound:


Notice that we didn't use @:3. Version ranges are inclusive, so @:3 means "up to and including 3".

What if a package can only be built with Python 2.6? You might be inclined to use:


However, this would be wrong. Spack assumes that all version constraints are absolute, so it would try to install Python at exactly 2.6. The correct way to specify this would be:


A spec can contain multiple version ranges separated by commas. For example, if you need Boost 1.59.0 or newer, but there are known issues with 1.64.0, 1.65.0, and 1.66.0, you can say:


Dependency types

Not all dependencies are created equal, and Spack allows you to specify exactly what kind of a dependency you need. For example:

depends_on('cmake', type='build')
depends_on('py-numpy', type=('build', 'run'))
depends_on('libelf', type=('build', 'link'))

The following dependency types are available:

  • "build": made available during the project's build. The package will be added to PATH, the compiler include paths, and PYTHONPATH. Other projects which depend on this one will not have these modified (building project X doesn't need project Y's build dependencies).

  • "link": the project is linked to by the project. The package will be added to the current package's rpath.

  • "run": the project is used by the project at runtime. The package will be added to PATH and PYTHONPATH.

One of the advantages of the build dependency type is that although the dependency needs to be installed in order for the package to be built, it can be uninstalled without concern afterwards. link and run disallow this because uninstalling the dependency would break the package.

If the dependency type is not specified, Spack uses a default of ('build', 'link'). This is the common case for compiler languages. Non-compiled packages like Python modules commonly use ('build', 'run'). This means that the compiler wrappers don't need to inject the dependency's prefix/lib directory, but the package needs to be in PATH and PYTHONPATH during the build process and later when a user wants to run the package.

Dependency patching

Some packages maintain special patches on their dependencies, either to add new features or to fix bugs. This typically makes a package harder to maintain, and we encourage developers to upstream (contribute back) their changes rather than maintaining patches. However, in some cases it's not possible to upstream. Maybe the dependency's developers don't accept changes, or maybe they just haven't had time to integrate them.

For times like these, Spack's depends_on directive can optionally take a patch or list of patches:

class SpecialTool(Package):
    depends_on('binutils', patches='special-binutils-feature.patch')

Here, the special-tool package requires a special feature in binutils, so it provides an extra patches=<filename> keyword argument. This is similar to the patch directive, with one small difference. Here, special-tool is responsible for the patch, so it should live in special-tool's directory in the package repository, not the binutils directory.

If you need something more sophisticated than this, you can simply nest a patch() directive inside of depends_on:

class SpecialTool(Package):
                      when='@:1.3'),   # condition on binutils
        when='@2.0:')                  # condition on special-tool

Note that there are two optional when conditions here -- one on the patch directive and the other on depends_on. The condition in the patch directive applies to binutils (the package being patched), while the condition in depends_on applies to special-tool. See patch directive for details on all the arguments the patch directive can take.

Finally, if you need multiple patches on a dependency, you can provide a list for patches, e.g.:

class SpecialTool(Package):

As with patch directives, patches are applied in the order they appear in the package file (or in this case, in the list).


You may wonder whether dependency patching will interfere with other packages that depend on binutils. It won't.

As described in patching, Patching a package adds the sha256 of the patch to the package's spec, which means it will have a different unique hash than other versions without the patch. The patched version coexists with unpatched versions, and Spack's support for handling_rpaths guarantees that each installation finds the right version. If two packages depend on binutils patched the same way, they can both use a single installation of binutils.


Spack provides a mechanism for dependencies to provide variables that can be used in their dependents' build. Any package can declare a setup_dependent_environment() function, and this function will be called before the install() method of any dependent packages. This allows dependencies to set up environment variables and other properties to be used by dependents.

The function declaration should look like this:

    def setup_dependent_environment(self, spack_env, run_env, dependent_spec):
        spack_env.set('QTDIR', self.prefix)

Here, the Qt package sets the QTDIR environment variable so that packages that depend on a particular Qt installation will find it.

The arguments to this function are:

  • spack_env: List of environment modifications to be applied when the dependent package is built within Spack.

  • run_env: List of environment modifications to be applied when the dependent package is run outside of Spack. These are added to the resulting module file.

  • dependent_spec: The spec of the dependent package about to be built. This allows the extendee (self) to query the dependent's state. Note that this package's spec is available as self.spec.

A good example of using these is in the Python package:

    def setup_dependent_environment(self, spack_env, run_env, dependent_spec):
        """Set PYTHONPATH to include the site-packages directory for the
        extension and any other python extensions it depends on."""

        # If we set PYTHONHOME, we must also ensure that the corresponding
        # python is found in the build environment. This to prevent cases
        # where a system provided python is run against the standard libraries
        # of a Spack built python. See issue #7128
        spack_env.set('PYTHONHOME', self.home)

        path = os.path.dirname(self.command.path)
        if not is_system_path(path):
            spack_env.prepend_path('PATH', path)

        python_paths = []
        for d in dependent_spec.traverse(
                deptype=('build', 'run', 'test')):
            if d.package.extends(self.spec):

        pythonpath = ':'.join(python_paths)
        spack_env.set('PYTHONPATH', pythonpath)

        # For run time environment set only the path for
        # dependent_spec and prepend it to PYTHONPATH
        if dependent_spec.package.extends(self.spec):
            run_env.prepend_path('PYTHONPATH', join_path(
                dependent_spec.prefix, self.site_packages_dir))

The first thing that happens here is that the python command is inserted into module scope of the dependent. This allows most python packages to have a very simple install method, like this:

def install(self, spec, prefix):
    python('', 'install', '--prefix={0}'.format(prefix))

Python's setup_dependent_environment method also sets up some other variables, creates a directory, and sets up the PYTHONPATH so that dependent packages can find their dependencies at build time.


Sometimes packages have known bugs, or limitations, that would prevent them to build e.g. against other dependencies or with certain compilers. Spack makes it possible to express such constraints with the conflicts directive.

Adding the following to a package:

conflicts('%intel', when='@1.2')

we express the fact that the current package cannot be built with the Intel compiler when we are trying to install version "1.2". The when argument can be omitted, in which case the conflict will always be active. Conflicts are always evaluated after the concretization step has been performed, and if any match is found a detailed error message is shown to the user.


Spack's support for package extensions is documented extensively in spack module tcl loads. This section documents how to make your own extendable packages and extensions.

To support extensions, a package needs to set its extendable property to True, e.g.:

class Python(Package):
    extendable = True

To make a package into an extension, simply add simply add an extends call in the package definition, and pass it the name of an extendable package:

class PyNumpy(Package):

Now, the py-numpy package can be used as an argument to spack activate. When it is activated, all the files in its prefix will be symbolically linked into the prefix of the python package.

Some packages produce a Python extension, but are only compatible with Python 3, or with Python 2. In those cases, a depends_on() declaration should be made in addition to the extends() declaration:

class Icebin(Package):
    extends('python', when='+python')
    depends_on('python@3:', when='+python')

Many packages produce Python extensions for some variants, but not others: they should extend python only if the appropriate variant(s) are selected. This may be accomplished with conditional extends() declarations:

class FooLib(Package):
    variant('python', default=True, description= \
        'Build the Python extension Module')
    extends('python', when='+python')

Sometimes, certain files in one package will conflict with those in another, which means they cannot both be activated (symlinked) at the same time. In this case, you can tell Spack to ignore those files when it does the activation:

class PySncosmo(Package):
    # py-sncosmo binaries are duplicates of those from py-astropy
    extends('python', ignore=r'bin/.*')

The code above will prevent everything in the $prefix/bin/ directory from being linked in at activation time.


You can call either depends_on or extends on any one package, but not both. For example you cannot both depends_on('python') and extends(python) in the same package. extends implies depends_on.


As covered in Filesystem Views, the spack view command can be used to symlink a number of packages into a merged prefix. The methods of PackageViewMixin can be overridden to customize how packages are added to views. Generally this can be used to create copies of specific files rather than symlinking them when symlinking does not work. For example, Python overrides add_files_to_view in order to create a copy of the python binary since the real path of the Python executable is used to detect extensions; as a consequence python extension packages (those inheriting from PythonPackage) likewise override add_files_to_view in order to rewrite shebang lines which point to the Python interpreter.

Activation & deactivation

Adding an extension to a view is referred to as an activation. If the view is maintained in the Spack installation prefix of the extendee this is called a global activation. Activations may involve updating some centralized state that is maintained by the extendee package, so there can be additional work for adding extensions compared with non-extension packages.

Spack's Package class has default activate and deactivate implementations that handle symbolically linking extensions' prefixes into a specified view. Extendable packages can override these methods to add custom activate/deactivate logic of their own. For example, the activate and deactivate methods in the Python class handle symbolic linking of extensions, but they also handle details surrounding Python's .pth files, and other aspects of Python packaging.

Spack's extensions mechanism is designed to be extensible, so that other packages (like Ruby, R, Perl, etc.) can provide their own custom extension management logic, as they may not handle modules the same way that Python does.

Let's look at Python's activate function:

    def activate(self, ext_pkg, view, **args):
        ignore = self.python_ignore(ext_pkg, args)

        super(Python, self).activate(ext_pkg, view, **args)

        extensions_layout = view.extensions_layout
        exts = extensions_layout.extension_map(self.spec)
        exts[] = ext_pkg.spec

        self.write_easy_install_pth(exts, prefix=view.get_projection_for_spec(

This function is called on the extendee (Python). It first calls activate in the superclass, which handles symlinking the extension package's prefix into the specified view. It then does some special handling of the easy-install.pth file, part of Python's setuptools.

Deactivate behaves similarly to activate, but it unlinks files:

    def deactivate(self, ext_pkg, view, **args):
        args.update(ignore=self.python_ignore(ext_pkg, args))

        super(Python, self).deactivate(ext_pkg, view, **args)

        extensions_layout = view.extensions_layout
        exts = extensions_layout.extension_map(self.spec)
        # Make deactivate idempotent
        if in exts:
            del exts[]

Both of these methods call some custom functions in the Python package. See the source for Spack's Python package for details.

Activation arguments

You may have noticed that the activate function defined above takes keyword arguments. These are the keyword arguments from extends(), and they are passed to both activate and deactivate.

This capability allows an extension to customize its own activation by passing arguments to the extendee. Extendees can likewise implement custom activate() and deactivate() functions to suit their needs.

The only keyword argument supported by default is the ignore argument, which can take a regex, list of regexes, or a predicate to determine which files not to symlink during activation.

Virtual dependencies

In some cases, more than one package can satisfy another package's dependency. One way this can happen is if a package depends on a particular interface, but there are multiple implementations of the interface, and the package could be built with any of them. A very common interface in HPC is the Message Passing Interface (MPI), which is used in many large-scale parallel applications.

MPI has several different implementations (e.g., MPICH, OpenMPI, and MVAPICH) and scientific applications can be built with any one of them. Complicating matters, MPI does not have a standardized ABI, so a package built with one implementation cannot simply be relinked with another implementation. Many package managers handle interfaces like this by requiring many similar package files, e.g., foo, foo-mvapich, foo-mpich, but Spack avoids this explosion of package files by providing support for virtual dependencies.


In Spack, mpi is handled as a virtual package. A package like mpileaks can depend on it just like any other package, by supplying a depends_on call in the package definition. For example:

class Mpileaks(Package):
    homepage = ""
    url = ""

    version('1.0', '8838c574b39202a57d7c2d68692718aa')


Here, callpath and adept-utils are concrete packages, but there is no actual package file for mpi, so we say it is a virtual package. The syntax of depends_on, is the same for both. If we look inside the package file of an MPI implementation, say MPICH, we'll see something like this:

class Mpich(Package):

The provides("mpi") call tells Spack that the mpich package can be used to satisfy the dependency of any package that depends_on('mpi').

Versioned Interfaces

Just as you can pass a spec to depends_on, so can you pass a spec to provides to add constraints. This allows Spack to support the notion of versioned interfaces. The MPI standard has gone through many revisions, each with new functions added, and each revision of the standard has a version number. Some packages may require a recent implementation that supports MPI-3 functions, but some MPI versions may only provide up to MPI-2. Others may need MPI 2.1 or higher. You can indicate this by adding a version constraint to the spec passed to provides:


Suppose that the above provides call is in the mpich2 package. This says that mpich2 provides MPI support up to version 2, but if a package depends_on("mpi@3"), then Spack will not build that package with mpich2.

provides when

The same package may provide different versions of an interface depending on its version. Above, we simplified the provides call in mpich to make the explanation easier. In reality, this is how mpich calls provides:

provides('mpi@:3', when='@3:')
provides('mpi@:1', when='@1:')

The when argument to provides allows you to specify optional constraints on the providing package, or the provider. The provider only provides the declared virtual spec when it matches the constraints in the when clause. Here, when mpich is at version 3 or higher, it provides MPI up to version 3. When mpich is at version 1 or higher, it provides the MPI virtual package at version 1.

The when qualifier ensures that Spack selects a suitably high version of mpich to satisfy some other package that depends_on a particular version of MPI. It will also prevent a user from building with too low a version of mpich. For example, suppose the package foo declares this:

class Foo(Package):

Suppose a user invokes spack install like this:

$ spack install foo ^mpich@1.0

Spack will fail with a constraint violation, because the version of MPICH requested is too low for the mpi requirement in foo.

Abstract & concrete specs

Now that we've seen how spec constraints can be specified on the command line and within package definitions, we can talk about how Spack puts all of this information together. When you run this:

$ spack install mpileaks ^callpath@1.0+debug ^libelf@0.8.11

Spack parses the command line and builds a spec from the description. The spec says that mpileaks should be built with the callpath library at 1.0 and with the debug option enabled, and with libelf version 0.8.11. Spack will also look at the depends_on calls in all of these packages, and it will build a spec from that. The specs from the command line and the specs built from package descriptions are then combined, and the constraints are checked against each other to make sure they're satisfiable.

What we have after this is done is called an abstract spec. An abstract spec is partially specified. In other words, it could describe more than one build of a package. Spack does this to make things easier on the user: they should only have to specify as much of the package spec as they care about. Here's an example partial spec DAG, based on the constraints above:


digraph { mpileaks -> mpi mpileaks -> "callpath@1.0+debug" -> mpi "callpath@1.0+debug" -> dyninst dyninst -> libdwarf -> "libelf@0.8.11" dyninst -> "libelf@0.8.11" }

This diagram shows a spec DAG output as a tree, where successive levels of indentation represent a depends-on relationship. In the above DAG, we can see some packages annotated with their constraints, and some packages with no annotations at all. When there are no annotations, it means the user doesn't care what configuration of that package is built, just so long as it works.


An abstract spec is useful for the user, but you can't install an abstract spec. Spack has to take the abstract spec and "fill in" the remaining unspecified parts in order to install. This process is called concretization. Concretization happens in between the time the user runs spack install and the time the install() method is called. The concretized version of the spec above might look like this:

mpileaks@2.3%gcc@4.7.3 arch=linux-debian7-x86_64
    ^callpath@1.0%gcc@4.7.3+debug arch=linux-debian7-x86_64
        ^dyninst@8.1.2%gcc@4.7.3 arch=linux-debian7-x86_64
            ^libdwarf@20130729%gcc@4.7.3 arch=linux-debian7-x86_64
                ^libelf@0.8.11%gcc@4.7.3 arch=linux-debian7-x86_64
        ^mpich@3.0.4%gcc@4.7.3 arch=linux-debian7-x86_64

digraph { "mpileaks@2.3\n%gcc@4.7.3\n arch=linux-debian7-x86_64" -> "mpich@3.0.4\n%gcc@4.7.3\n arch=linux-debian7-x86_64" "mpileaks@2.3\n%gcc@4.7.3\n arch=linux-debian7-x86_64" -> "callpath@1.0\n%gcc@4.7.3+debug\n arch=linux-debian7-x86_64" -> "mpich@3.0.4\n%gcc@4.7.3\n arch=linux-debian7-x86_64" "callpath@1.0\n%gcc@4.7.3+debug\n arch=linux-debian7-x86_64" -> "dyninst@8.1.2\n%gcc@4.7.3\n arch=linux-debian7-x86_64" "dyninst@8.1.2\n%gcc@4.7.3\n arch=linux-debian7-x86_64" -> "libdwarf@20130729\n%gcc@4.7.3\n arch=linux-debian7-x86_64" -> "libelf@0.8.11\n%gcc@4.7.3\n arch=linux-debian7-x86_64" "dyninst@8.1.2\n%gcc@4.7.3\n arch=linux-debian7-x86_64" -> "libelf@0.8.11\n%gcc@4.7.3\n arch=linux-debian7-x86_64" }

Here, all versions, compilers, and platforms are filled in, and there is a single version (no version ranges) for each package. All decisions about configuration have been made, and only after this point will Spack call the install() method for your package.

Concretization in Spack is based on certain selection policies that tell Spack how to select, e.g., a version, when one is not specified explicitly. Concretization policies are discussed in more detail in 設定ファイル. Sites using Spack can customize them to match the preferences of their own users.

spack spec

For an arbitrary spec, you can see the result of concretization by running spack spec. For example:

$ spack spec dyninst@8.0.1

dyninst@8.0.1%gcc@4.7.3 arch=linux-debian7-x86_64
    ^libdwarf@20130729%gcc@4.7.3 arch=linux-debian7-x86_64
        ^libelf@0.8.13%gcc@4.7.3 arch=linux-debian7-x86_64

This is useful when you want to know exactly what Spack will do when you ask for a particular spec.

Concretization Policies

A user may have certain preferences for how packages should be concretized on their system. For example, one user may prefer packages built with OpenMPI and the Intel compiler. Another user may prefer packages be built with MVAPICH and GCC.

See the Concretization Preferences section for more details.

Conflicting Specs

Suppose a user needs to install package C, which depends on packages A and B. Package A builds a library with a Python2 extension, and package B builds a library with a Python3 extension. Packages A and B cannot be loaded together in the same Python runtime:

class A(Package):
    variant('python', default=True, 'enable python bindings')
    depends_on('python@2.7', when='+python')
    def install(self, spec, prefix):
        # do whatever is necessary to enable/disable python
        # bindings according to variant

class B(Package):
    variant('python', default=True, 'enable python bindings')
    depends_on('python@3.2:', when='+python')
    def install(self, spec, prefix):
        # do whatever is necessary to enable/disable python
        # bindings according to variant

Package C needs to use the libraries from packages A and B, but does not need either of the Python extensions. In this case, package C should simply depend on the ~python variant of A and B:

class C(Package):

This may require that A or B be built twice, if the user wishes to use the Python extensions provided by them: once for +python and once for ~python. Other than using a little extra disk space, that solution has no serious problems.

Implementing the installation procedure

The last element of a package is its installation procedure. This is where the real work of installation happens, and it's the main part of the package you'll need to customize for each piece of software.

Defining an installation procedure means overriding a set of methods or attributes that will be called at some point during the installation of the package. The package base class, usually specialized for a given build system, determines the actual set of entities available for overriding. The classes that are currently provided by Spack are:

Base Class



General base class not specialized for any build system


Specialized class for packages built invoking hand-written Makefiles


Specialized class for packages built using GNU Autotools


Specialized class for packages built using CMake


A helper class for packages that use CUDA. It is intended to be used in combination with others


Specialized class for packages build using QMake


Specialized class for packages built using SCons


Specialized class for packages built using Waf


Specialized class for R extensions


Specialized class for Octave packages


Specialized class for Python extensions


Specialized class for Perl extensions


Specialized class for licensed Intel software


Choice of the appropriate base class for a package

In most cases packagers don't have to worry about the selection of the right base class for a package, as spack create will make the appropriate choice on their behalf. In those rare cases where manual intervention is needed we need to stress that a package base class depends on the build system being used, not the language of the package. For example, a Python extension installed with CMake would extends('python') and subclass from CMakePackage.

Installation pipeline

When a user runs spack install, Spack:

  1. Fetches an archive for the correct version of the software.

  2. Expands the archive.

  3. Sets the current working directory to the root directory of the expanded archive.

Then, depending on the base class of the package under consideration, it will execute a certain number of phases that reflect the way a package of that type is usually built. The name and order in which the phases will be executed can be obtained either reading the API docs at build_systems, or using the spack info command:

$ spack info m4
AutotoolsPackage:    m4

Safe versions:

    Name       Default   Description

    sigsegv    on        Build the libsigsegv dependency

Installation Phases:
    autoreconf    configure    build    install

Build Dependencies:


Typically, phases have default implementations that fit most of the common cases:

    def configure(self, spec, prefix):
        """Runs configure with the arguments specified in
        and an appropriately set prefix.
        options = getattr(self, 'configure_flag_args', [])
        options += ['--prefix={0}'.format(prefix)]
        options += self.configure_args()

        with working_dir(self.build_directory, create=True):

It is thus just sufficient for a packager to override a few build system specific helper methods or attributes to provide, for instance, configure arguments:

    def configure_args(self):
        spec = self.spec
        args = ['--enable-c++']

        if spec.satisfies('%clang') and not spec.satisfies('platform=darwin'):

        if spec.satisfies('%arm') and not spec.satisfies('platform=darwin'):

        if spec.satisfies('%intel'):

        if '+sigsegv' in spec:

        arch = spec.architecture
        if (arch.platform == 'darwin' and arch.os == 'sierra' and
            '%gcc' in spec):

        return args


Each specific build system has a list of attributes that can be overridden to fine-tune the installation of a package without overriding an entire phase. To have more information on them the place to go is the API docs of the build_systems module.

Overriding an entire phase

In extreme cases it may be necessary to override an entire phase. Regardless of the build system, the signature is the same. For example, the signature for the install phase is:

class Foo(Package):
    def install(self, spec, prefix):

For those not used to Python instance methods, this is the package itself. In this case it's an instance of Foo, which extends Package. For API docs on Package objects, see Package.


This is the concrete spec object created by Spack from an abstract spec supplied by the user. It describes what should be installed. It will be of type Spec.


This is the path that your install method should copy build targets into. It acts like a string, but it's actually its own special type, Prefix.

The arguments spec and prefix are passed only for convenience, as they always correspond to self.spec and self.spec.prefix respectively.

As mentioned in The build environment, you will usually not need to refer to dependencies explicitly in your package file, as the compiler wrappers take care of most of the heavy lifting here. There will be times, though, when you need to refer to the install locations of dependencies, or when you need to do something different depending on the version, compiler, dependencies, etc. that your package is built with. These parameters give you access to this type of information.

The build environment

In general, you should not have to do much differently in your install method than you would when installing a package on the command line. In fact, you may need to do less than you would on the command line.

Spack tries to set environment variables and modify compiler calls so that it appears to the build system that you're building with a standard system install of everything. Obviously that's not going to cover all build systems, but it should make it easy to port packages to Spack if they use a standard build system. Usually with autotools or cmake, building and installing is easy. With builds that use custom Makefiles, you may need to add logic to modify the makefiles.

The remainder of the section covers the way Spack's build environment works.

Forking install()

To give packagers free reign over their install environment, Spack forks a new process each time it invokes a package's install() method. This allows packages to have a sandboxed build environment, without impacting the environments ofother jobs that the main Spack process runs. Packages are free to change the environment or to modify Spack internals, because each install() call has its own dedicated process.

Environment variables

Spack sets a number of standard environment variables that serve two purposes:

  1. Make build systems use Spack's compiler wrappers for their builds.

  2. Allow build systems to find dependencies more easily

The Compiler environment variables that Spack sets are:




C compiler


C++ compiler


Fortran 77 compiler


Fortran 90 and above compiler

Spack sets these variables so that they point to compiler wrappers. These are covered in their own section below.

All of these are standard variables respected by most build systems. If your project uses Autotools or CMake, then it should pick them up automatically when you run configure or cmake in the install() function. Many traditional builds using GNU Make and BSD make also respect these variables, so they may work with these systems.

If your build system does not automatically pick these variables up from the environment, then you can simply pass them on the command line or use a patch as part of your build process to get the correct compilers into the project's build system. There are also some file editing commands you can use -- these are described later in the section on file manipulation.

In addition to the compiler variables, these variables are set before entering install() so that packages can locate dependencies easily:


Set to point to /bin directories of dependencies


Path to dependency prefixes for CMake


Path to any pkgconfig directories for dependencies


Path to site-packages dir of any python dependencies

PATH is set up to point to dependencies /bin directories so that you can use tools installed by dependency packages at build time. For example, $MPICH_ROOT/bin/mpicc is frequently used by dependencies of mpich.

CMAKE_PREFIX_PATH contains a colon-separated list of prefixes where cmake will search for dependency libraries and headers. This causes all standard CMake find commands to look in the paths of your dependencies, so you do not have to manually specify arguments like -DDEPENDENCY_DIR=/path/to/dependency to cmake. More on this is in the CMake documentation.

PKG_CONFIG_PATH is for packages that attempt to discover dependencies using the GNU pkg-config tool. It is similar to CMAKE_PREFIX_PATH in that it allows a build to automatically discover its dependencies.

If you want to see the environment that a package will build with, or if you want to run commands in that environment to test them out, you can use the spack env command, documented below.

Failing the build

Sometimes you don't want a package to successfully install unless some condition is true. You can explicitly cause the build to fail from install() by raising an InstallError, for example:

if spec.architecture.startswith('darwin'):
    raise InstallError('This package does not build on Mac OS X!')

Shell command functions

Recall the install method from libelf:

    def install(self, spec, prefix):
        make('install', parallel=False)

Normally in Python, you'd have to write something like this in order to execute shell commands:

import subprocess
subprocess.check_call('configure', '--prefix={0}'.format(prefix))

We've tried to make this a bit easier by providing callable wrapper objects for some shell commands. By default, configure, cmake, and make wrappers are are provided, so you can call them more naturally in your package files.

If you need other commands, you can use which to get them:

sed = which('sed')
sed('s/foo/bar/', filename)

The which function will search the PATH for the application.

Callable wrappers also allow spack to provide some special features. For example, in Spack, make is parallel by default, and Spack figures out the number of cores on your machine and passes an appropriate value for -j<numjobs> when it calls make (see the parallel package attribute <attribute_parallel>). In a package file, you can supply a keyword argument, parallel=False, to the make wrapper to disable parallel make. In the libelf package, this allows us to avoid race conditions in the library's build system.

Compiler flags

Compiler flags set by the user through the Spec object can be passed to the build in one of three ways. By default, the build environment injects these flags directly into the compiler commands using Spack's compiler wrappers. In cases where the build system requires knowledge of the compiler flags, they can be registered with the build system by alternatively passing them through environment variables or as build system arguments. The flag_handler method can be used to change this behavior.

Packages can override the flag_handler method with one of three built-in flag_handlers. The built-in flag_handlers are named inject_flags, env_flags, and build_system_flags. The inject_flags method is the default. The env_flags method puts all of the flags into the environment variables that make uses as implicit variables ('CFLAGS', 'CXXFLAGS', etc.). The build_system_flags method adds the flags as arguments to the invocation of configure or cmake, respectively.


Passing compiler flags using build system arguments is only supported for CMake and Autotools packages. Individual packages may also differ in whether they properly respect these arguments.

Individual packages may also define their own flag_handler methods. The flag_handler method takes the package instance (self), the name of the flag, and a list of the values of the flag. It will be called on each of the six compiler flags supported in Spack. It should return a triple of (injf, envf, bsf) where injf is a list of flags to inject via the Spack compiler wrappers, envf is a list of flags to set in the appropriate environment variables, and bsf is a list of flags to pass to the build system as arguments.


Passing a non-empty list of flags to bsf for a build system that does not support build system arguments will result in an error.

Here are the definitions of the three built-in flag handlers:

def inject_flags(pkg, name, flags):
    return (flags, None, None)

def env_flags(pkg, name, flags):
    return (None, flags, None)

def build_system_flags(pkg, name, flags):
    return (None, None, flags)


Returning [] and None are equivalent in a flag_handler method.

Packages can override the default behavior either by specifying one of the built-in flag handlers,

flag_handler = env_flags

or by implementing the flag_handler method. Suppose for a package Foo we need to pass cflags, cxxflags, and cppflags through the environment, the rest of the flags through compiler wrapper injection, and we need to add -lbar to ldlibs. The following flag handler method accomplishes that.

def flag_handler(self, name, flags):
    if name in ['cflags', 'cxxflags', 'cppflags']:
        return (None, flags, None)
    elif name == 'ldlibs':
    return (flags, None, None)

Because these methods can pass values through environment variables, it is important not to override these variables unnecessarily (E.g. setting env['CFLAGS']) in other package methods when using non-default flag handlers. In the setup_environment and setup_dependent_environment methods, use the append_flags method of the EnvironmentModifications class to append values to a list of flags whenever the flag handler is env_flags. If the package passes flags through the environment or the build system manually (in the install method, for example), we recommend using the default flag handler, or removing manual references and implementing a custom flag handler method that adds the desired flags to export as environment variables or pass to the build system. Manual flag passing is likely to interfere with the env_flags and build_system_flags methods.

In rare circumstances such as compiling and running small unit tests, a package developer may need to know what are the appropriate compiler flags to enable features like OpenMP, c++11, c++14 and alike. To that end the compiler classes in spack implement the following properties: openmp_flag, cxx98_flag, cxx11_flag, cxx14_flag, and cxx17_flag, which can be accessed in a package by self.compiler.cxx11_flag and alike. Note that the implementation is such that if a given compiler version does not support this feature, an error will be produced. Therefore package developers can also use these properties to assert that a compiler supports the requested feature. This is handy when a package supports additional variants like

variant('openmp', default=True, description="Enable OpenMP support.")

Blas, Lapack and ScaLapack libraries

Multiple packages provide implementations of Blas, Lapack and ScaLapack routines. The names of the resulting static and/or shared libraries differ from package to package. In order to make the install() method independent of the choice of Blas implementation, each package which provides it implements @property def blas_libs(self): to return an object of LibraryList type which simplifies usage of a set of libraries. The same applies to packages which provide Lapack and ScaLapack. Package developers are requested to use this interface. Common usage cases are:

  1. Space separated list of full paths

lapack_blas = spec['lapack'].libs + spec['blas'].libs
  1. Names of libraries and directories which contain them

blas = spec['blas'].libs
  1. Search and link flags

math_libs = spec['scalapack'].libs + spec['lapack'].libs + spec['blas'].libs

For more information, see documentation of LibraryList class.

Prefix objects

Spack passes the prefix parameter to the install method so that you can pass it to configure, cmake, or some other installer, e.g.:


For the most part, prefix objects behave exactly like strings. For packages that do not have their own install target, or for those that implement it poorly (like libdwarf), you may need to manually copy things into particular directories under the prefix. For this, you can refer to standard subdirectories without having to construct paths yourself, e.g.:

def install(self, spec, prefix):
    install('foo-tool', prefix.bin)

    install('foo.h', prefix.include)

    install('libfoo.a', prefix.lib)

Attributes of this object are created on the fly when you request them, so any of the following will work:

Prefix Attribute








Of course, this only works if your file or directory is a valid Python variable name. If your file or directory contains dashes or dots, use join instead:


Spec objects

When install is called, most parts of the build process are set up for you. The correct version's tarball has been downloaded and expanded. Environment variables like CC and CXX are set to point to the correct compiler and version. An install prefix has already been selected and passed in as prefix. In most cases this is all you need to get configure, cmake, or another install working correctly.

There will be times when you need to know more about the build configuration. For example, some software requires that you pass special parameters to configure, like --with-libelf=/path/to/libelf or --with-mpich. You might also need to supply special compiler flags depending on the compiler. All of this information is available in the spec.

Testing spec constraints

You can test whether your spec is configured a certain way by using the satisfies method. For example, if you want to check whether the package's version is in a particular range, you can use specs to do that, e.g.:

configure_args = [

if spec.satisfies('@1.2:1.4'):


This works for compilers, too:

if spec.satisfies('%gcc'):
    configure_args.append('CXXFLAGS="-g3 -O3"')
if spec.satisfies('%intel'):
    configure_args.append('CXXFLAGS="-xSSE2 -fast"')

Or for combinations of spec constraints:

if spec.satisfies('@1.2%intel'):
    tty.error("Version 1.2 breaks when using Intel compiler!")

You can also do similar satisfaction tests for dependencies:

if spec.satisfies('^dyninst@8.0'):

This could allow you to easily work around a bug in a particular dependency version.

You can use satisfies() to test for particular dependencies, e.g. foo.satisfies('^openmpi@1.2') or foo.satisfies('^mpich'), or you can use Python's built-in in operator:

if 'libelf' in spec:
    print "this package depends on libelf"

This is useful for virtual dependencies, as you can easily see what implementation was selected for this build:

if 'openmpi' in spec:
elif 'mpich' in spec:
elif 'mvapich' in spec:

It's also a bit more concise than satisfies. The difference between the two functions is that satisfies() tests whether spec constraints overlap at all, while in tests whether a spec or any of its dependencies satisfy the provided spec.

Accessing Dependencies

You may need to get at some file or binary that's in the installation prefix of one of your dependencies. You can do that by sub-scripting the spec:


The value in the brackets needs to be some package name, and spec needs to depend on that package, or the operation will fail. For example, the above code will fail if the spec doesn't depend on mpi. The value returned is itself just another Spec object, so you can do all the same things you would do with the package's own spec:


Multimethods and @when

Spack allows you to make multiple versions of instance functions in packages, based on whether the package's spec satisfies particular criteria.

The @when annotation lets packages declare multiple versions of methods like install() that depend on the package's spec. For example:

class SomePackage(Package):

    def install(self, prefix):
        # Do default install

    def install(self, prefix):
        # This will be executed instead of the default install if
        # the package's sys_type() is chaos_5_x86_64_ib.

    def install(self, prefix):
        # This will be executed if the package's sys_type() is
        # linux-debian7-x86_64.

In the above code there are three versions of install(), two of which are specialized for particular platforms. The version that is called depends on the architecture of the package spec.

Note that this works for methods other than install, as well. So, if you only have part of the install that is platform specific, you could do something more like this:

class SomePackage(Package):
    # virtual dependence on MPI.
    # could resolve to mpich, mpich2, OpenMPI

    def setup(self):
        # do nothing in the default case

    def setup(self):
        # do something special when this is built with OpenMPI for
        # its MPI implementations.

    def install(self, prefix):
        # Do common install stuff
        # Do more common install stuff

You can write multiple @when specs that satisfy the package's spec, for example:

class SomePackage(Package):

    def setup_mpi(self):
        # the default, called when no @when specs match

    def setup_mpi(self):
        # this will be called when mpi is version 3 or higher

    def setup_mpi(self):
        # this will be called when mpi is version 2 or higher

    def setup_mpi(self):
        # this will be called when mpi is version 1 or higher

In situations like this, the first matching spec, in declaration order will be called. As before, if no @when spec matches, the default method (the one without the @when decorator) will be called.


The default version of decorated methods must always come first. Otherwise it will override all of the platform-specific versions. There's not much we can do to get around this because of the way decorators work.

Compiler wrappers

As mentioned, CC, CXX, F77, and FC are set to point to Spack's compiler wrappers. These are simply called cc, c++, f77, and f90, and they live in $SPACK_ROOT/lib/spack/env.

$SPACK_ROOT/lib/spack/env is added first in the PATH environment variable when install() runs so that system compilers are not picked up instead.

All of these compiler wrappers point to a single compiler wrapper script that figures out which real compiler it should be building with. This comes either from spec concretization or from a user explicitly asking for a particular compiler using, e.g., %intel on the command line.

In addition to invoking the right compiler, the compiler wrappers add flags to the compile line so that dependencies can be easily found. These flags are added for each dependency, if they exist:

Compile-time library search paths * -L$dep_prefix/lib * -L$dep_prefix/lib64

Runtime library search paths (RPATHs) * $rpath_flag$dep_prefix/lib * $rpath_flag$dep_prefix/lib64

Include search paths * -I$dep_prefix/include

An example of this would be the libdwarf build, which has one dependency: libelf. Every call to cc in the libdwarf build will have -I$LIBELF_PREFIX/include, -L$LIBELF_PREFIX/lib, and $rpath_flag$LIBELF_PREFIX/lib inserted on the command line. This is done transparently to the project's build system, which will just think it's using a system where libelf is readily available. Because of this, you do not have to insert extra -I, -L, etc. on the command line.

Another useful consequence of this is that you often do not have to add extra parameters on the configure line to get autotools to find dependencies. The libdwarf install method just calls configure like this:

configure("--prefix=" + prefix)

Because of the -L and -I arguments, configure will successfully find libdwarf.h and, without the packager having to provide --with-libdwarf=/path/to/libdwarf on the command line.


For most compilers, $rpath_flag is -Wl,-rpath,. However, NAG passes its flags to GCC instead of passing them directly to the linker. Therefore, its $rpath_flag is doubly wrapped: -Wl,-Wl,,-rpath,. $rpath_flag can be overriden on a compiler specific basis in lib/spack/spack/compilers/$

The compiler wrappers also pass the compiler flags specified by the user from the command line (cflags, cxxflags, fflags, cppflags, ldflags, and/or ldlibs). They do not override the canonical autotools flags with the same names (but in ALL-CAPS) that may be passed into the build by particularly challenging package scripts.

MPI support in Spack

It is common for high performance computing software/packages to use the Message Passing Interface ( MPI). As a result of conretization, a given package can be built using different implementations of MPI such as Openmpi, MPICH or IntelMPI. That is, when your package declares that it depends_on('mpi'), it can be built with any of these mpi implementations. In some scenarios, to configure a package, one has to provide it with appropriate MPI compiler wrappers such as mpicc, mpic++. However different implementations of MPI may have different names for those wrappers.

Spack provides an idiomatic way to use MPI compilers in your package. To use MPI wrappers to compile your whole build, do this in your install() method:

env['CC'] = spec['mpi'].mpicc
env['CXX'] = spec['mpi'].mpicxx
env['F77'] = spec['mpi'].mpif77
env['FC'] = spec['mpi'].mpifc

That's all. A longer explanation of why this works is below.

We don't try to force any particular build method on packagers. The decision to use MPI wrappers depends on the way the package is written, on common practice, and on "what works". Loosely, There are three types of MPI builds:

  1. Some build systems work well without the wrappers and can treat MPI as an external library, where the person doing the build has to supply includes/libs/etc. This is fairly uncommon.

  2. Others really want the wrappers and assume you're using an MPI "compiler" – i.e., they have no mechanism to add MPI includes/libraries/etc.

  3. CMake's FindMPI needs the compiler wrappers, but it uses them to extract –I / -L / -D arguments, then treats MPI like a regular library.

Note that some CMake builds fall into case 2 because they either don't know about or don't like CMake's FindMPI support – they just assume an MPI compiler. Also, some autotools builds fall into case 3 (e.g. here is an autotools version of CMake's FindMPI).

Given all of this, we leave the use of the wrappers up to the packager. Spack will support all three ways of building MPI packages.

Packaging Conventions

As mentioned above, in the install() method, CC, CXX, F77, and FC point to Spack's wrappers around the chosen compiler. Spack's wrappers are not the MPI compiler wrappers, though they do automatically add –I, –L, and –Wl,-rpath args for dependencies in a similar way. The MPI wrappers are a bit different in that they also add -l arguments for the MPI libraries, and some add special -D arguments to trigger build options in MPI programs.

For case 1 above, you generally don't need to do more than patch your Makefile or add configure args as you normally would.

For case 3, you don't need to do much of anything, as Spack puts the MPI compiler wrappers in the PATH, and the build will find them and interrogate them.

For case 2, things are a bit more complicated, as you'll need to tell the build to use the MPI compiler wrappers instead of Spack's compiler wrappers. All it takes some lines like this:

env['CC'] = spec['mpi'].mpicc
env['CXX'] = spec['mpi'].mpicxx
env['F77'] = spec['mpi'].mpif77
env['FC'] = spec['mpi'].mpifc

Or, if you pass CC, CXX, etc. directly to your build with, e.g., --with-cc=<path>, you'll want to substitute spec['mpi'].mpicc in there instead, e.g.:

configure('—prefix=%s' % prefix,
          '—with-cc=%s' % spec['mpi'].mpicc)

Now, you may think that doing this will lose the includes, library paths, and RPATHs that Spack's compiler wrapper get you, but we've actually set things up so that the MPI compiler wrappers use Spack's compiler wrappers when run from within Spack. So using the MPI wrappers should really be as simple as the code above.


Ok, so how does all this work?

If your package has a virtual dependency like mpi, then referring to spec['mpi'] within install() will get you the concrete mpi implementation in your dependency DAG. That is a spec object just like the one passed to install, only the MPI implementations all set some additional properties on it to help you out. E.g., in mvapich2, you'll find this:

    def setup_dependent_package(self, module, dependent_spec):
        self.spec.mpicc  = join_path(self.prefix.bin, 'mpicc')
        self.spec.mpicxx = join_path(self.prefix.bin, 'mpicxx')
        self.spec.mpifc  = join_path(self.prefix.bin, 'mpif90')
        self.spec.mpif77 = join_path(self.prefix.bin, 'mpif77')
        self.spec.mpicxx_shared_libs = [
            join_path(self.prefix.lib, 'libmpicxx.{0}'.format(dso_suffix)),
            join_path(self.prefix.lib, 'libmpi.{0}'.format(dso_suffix))

That code allows the mvapich2 package to associate an mpicc property with the mvapich2 node in the DAG, so that dependents can access it. openmpi and mpich do similar things. So, no matter what MPI you're using, spec['mpi'].mpicc gets you the location of the MPI compilers. This allows us to have a fairly simple polymorphic interface for information about virtual dependencies like MPI.

Wrapping wrappers

Spack likes to use its own compiler wrappers to make it easy to add RPATHs to builds, and to try hard to ensure that your builds use the right dependencies. This doesn't play nicely by default with MPI, so we have to do a couple tricks.

  1. If we build MPI with Spack's wrappers, mpicc and friends will be installed with hard-coded paths to Spack's wrappers, and using them from outside of Spack will fail because they only work within Spack. To fix this, we patch mpicc and friends to use the regular compilers. Look at the filter_compilers method in mpich, openmpi, or mvapich2 for details.

  2. We still want to use the Spack compiler wrappers when Spack is calling mpicc. Luckily, wrappers in all mainstream MPI implementations provide environment variables that allow us to dynamically set the compiler to be used by mpicc, mpicxx, etc. Denis pasted some code from this below – Spack's build environment sets MPICC, MPICXX, etc. for mpich derivatives and OMPI_CC, OMPI_CXX, etc. for OpenMPI. This makes the MPI compiler wrappers use the Spack compiler wrappers so that your dependencies still get proper RPATHs even if you use the MPI wrappers.

MPI on Cray machines

The Cray programming environment notably uses ITS OWN compiler wrappers, which function like MPI wrappers. On Cray systems, the CC, cc, and ftn wrappers ARE the MPI compiler wrappers, and it's assumed that you'll use them for all of your builds. So on Cray we don't bother with mpicc, mpicxx, etc, Spack MPI implementations set spec['mpi'].mpicc to point to Spack's wrappers, which wrap the Cray wrappers, which wrap the regular compilers and include MPI flags. That may seem complicated, but for packagers, that means the same code for using MPI wrappers will work, even on even on a Cray:

env['CC'] = spec['mpi'].mpicc

This is because on Cray, spec['mpi'].mpicc is just spack_cc.

Checking an installation

By default, Spack assumes that a build has failed if nothing is written to the install prefix, and that it has succeeded if anything (a file, a directory, etc.) is written to the install prefix after install() completes.

Consider a simple autotools build like this:

def install(self, spec, prefix):

If you are using using standard autotools or CMake, configure and make will not write anything to the install prefix. Only make install writes the files, and only once the build is already complete.

sanity_check_is_file and sanity_check_is_dir

Unfortunately, many builds of scientific software modify the install prefix before make install. Builds like this can falsely report that they were successfully installed if an error occurs before the install is complete but after files have been written to the prefix.

You can optionally specify sanity checks to deal with this problem. Add properties like this to your package:

class MyPackage(Package):

    sanity_check_is_file = ['include/libelf.h']
    sanity_check_is_dir  = [lib]

    def install(self, spec, prefix):
        configure("--prefix=" + prefix)

Now, after install() runs, Spack will check whether $prefix/include/libelf.h exists and is a file, and whether $prefix/lib exists and is a directory. If the checks fail, then the build will fail and the install prefix will be removed. If they succeed, Spack considers the build successful and keeps the prefix in place.

Build-time tests

Sometimes packages finish to build "correctly" and issues with their run-time behavior are discovered only at a later stage, maybe after a full software stack relying on them has already been built. To avoid situations of that kind it's possible to write build-time tests that will be executed only if the option --run-tests of spack install has been activated.

The proper way to write these tests is relying on two decorators that come with any base class listed in Implementing the installation procedure.

def check_build(self):
     # Custom implementation goes here

The first decorator run_after('build') schedules this function to be invoked after the build phase has been executed, while the second one makes the invocation conditional on the fact that self.run_tests == True. It is also possible to schedule a function to be invoked before a given phase using the run_before decorator.


Default implementations for build-time tests

Packages that are built using specific build systems may already have a default implementation for build-time tests. For instance AutotoolsPackage based packages will try to invoke make test and make check if Spack is asked to run tests. More information on each class is available in the the build_systems documentation.


The API for adding tests is not yet considered stable and may change drastically in future releases.

File manipulation functions

Many builds are not perfect. If a build lacks an install target, or if it does not use systems like CMake or autotools, which have standard ways of setting compilers and options, you may need to edit files or install some files yourself to get them working with Spack.

You can do this with standard Python code, and Python has rich libraries with functions for file manipulation and filtering. Spack also provides a number of convenience functions of its own to make your life even easier. These functions are described in this section.

All of the functions in this section can be included by simply running:

from spack import *

This is already part of the boilerplate for packages created with spack create.

Filtering functions

filter_file(regex, repl, *filenames, **kwargs)

Works like sed but with Python regular expression syntax. Takes a regular expression, a replacement, and a set of files. repl can be a raw string or a callable function. If it is a raw string, it can contain \1, \2, etc. to refer to capture groups in the regular expression. If it is a callable, it is passed the Python MatchObject and should return a suitable replacement string for the particular match.


  1. Filtering a Makefile to force it to use Spack's compiler wrappers:

    filter_file(r'^CC\s*=.*',  spack_cc,  'Makefile')
    filter_file(r'^CXX\s*=.*', spack_cxx, 'Makefile')
    filter_file(r'^F77\s*=.*', spack_f77, 'Makefile')
    filter_file(r'^FC\s*=.*',  spack_fc,  'Makefile')
  2. Replacing #!/usr/bin/perl with #!/usr/bin/env perl in bib2xhtml:

                '#!/usr/bin/env perl', prefix.bin.bib2xhtml)
  3. Switching the compilers used by mpich's MPI wrapper scripts from cc, etc. to the compilers used by the Spack build:

    filter_file('CC="cc"', 'CC="%s"' %,
    filter_file('CXX="c++"', 'CXX="%s"' % self.compiler.cxx,
change_sed_delimiter(old_delim, new_delim, *filenames)

Some packages, like TAU, have a build system that can't install into directories with, e.g. '@' in the name, because they use hard-coded sed commands in their build.

change_sed_delimiter finds all sed search/replace commands and change the delimiter. e.g., if the file contains commands that look like s///, you can use this to change them to s@@@.

Example of changing s/// to s@@@ in TAU:

change_sed_delimiter('@', ';', 'configure')
change_sed_delimiter('@', ';', 'utils/FixMakefile')
change_sed_delimiter('@', ';', 'utils/FixMakefile.sed.default')

File functions

ancestor(dir, n=1)

Get the nth ancestor of the directory dir.


True if we can read and write to the file at path. Same as native python os.access(file_name, os.R_OK|os.W_OK).

install(src, dest)

Install a file to a particular location. For example, install a header into the include directory under the install prefix:

install('my-header.h', prefix.include)

An alias for os.path.join. This joins paths using the OS path separator.


Create each of the directories in paths, creating any parent directories if they do not exist.

working_dir(dirname, kwargs)

This is a Python Context Manager that makes it easier to work with subdirectories in builds. You use this with the Python with statement to change into a working directory, and when the with block is done, you change back to the original directory. Think of it as a safe pushd / popd combination, where popd is guaranteed to be called at the end, even if exceptions are thrown.

Example usage:

  1. The libdwarf build first runs configure and make in a subdirectory called libdwarf. It then implements the installation code itself. This is natural with working_dir:

    with working_dir('libdwarf'):
        configure("--prefix=" + prefix, "--enable-shared")
        install('libdwarf.a',  prefix.lib)
  2. Many CMake builds require that you build "out of source", that is, in a subdirectory. You can handle creating and cd'ing to the subdirectory like the LLVM package does:

    with working_dir('spack-build', create=True):

    The create=True keyword argument causes the command to create the directory if it does not exist.


Create an empty file at path.

Style guidelines for packages

The following guidelines are provided, in the interests of making Spack packages work in a consistent manner:

Variant Names

Spack packages with variants similar to already-existing Spack packages should use the same name for their variants. Standard variant names are:






Build shared libraries



Build static libraries






Build Python extension

If specified in this table, the corresponding default should be used when declaring a variant.

Version Lists

Spack packages should list supported versions with the newest first.

Packaging workflow commands

When you are building packages, you will likely not get things completely right the first time.

The spack install command performs a number of tasks before it finally installs each package. It downloads an archive, expands it in a temporary directory, and only then gives control to the package's install() method. If the build doesn't go as planned, you may want to clean up the temporary directory, or if the package isn't downloading properly, you might want to run only the fetch stage of the build.

A typical package workflow might look like this:

$ spack edit mypackage
$ spack install mypackage
... build breaks! ...
$ spack clean mypackage
$ spack edit mypackage
$ spack install mypackage
... repeat clean/install until install works ...

Below are some commands that will allow you some finer-grained control over the install process.

spack fetch

The first step of spack install. Takes a spec and determines the correct download URL to use for the requested package version, then downloads the archive, checks it against an MD5 checksum, and stores it in a staging directory if the check was successful. The staging directory will be located under $SPACK_HOME/var/spack.

When run after the archive has already been downloaded, spack fetch is idempotent and will not download the archive again.

spack stage

The second step in spack install after spack fetch. Expands the downloaded archive in its temporary directory, where it will be built by spack install. Similar to fetch, if the archive has already been expanded, stage is idempotent.

spack patch

After staging, Spack applies patches to downloaded packages, if any have been specified in the package file. This command will run the install process through the fetch, stage, and patch phases. Spack keeps track of whether patches have already been applied and skips this step if they have been. If Spack discovers that patches didn't apply cleanly on some previous run, then it will restage the entire package before patching.

spack restage

Restores the source code to pristine state, as it was before building.

Does this in one of two ways:

  1. If the source was fetched as a tarball, deletes the entire build directory and re-expands the tarball.

  2. If the source was checked out from a repository, this deletes the build directory and checks it out again.

spack clean

Cleans up all of Spack's temporary and cached files. This can be used to recover disk space if temporary files from interrupted or failed installs accumulate in the staging area.

When called with --stage or without arguments this removes all staged files.

When called with --downloads this will clear all resources cached during installs.

When called with --user-cache this will remove caches in the user home directory, including cached virtual indices.

To remove all of the above, the command can be called with --all.

When called with positional arguments, cleans up temporary files only for a particular package. If fetch, stage, or install are run again after this, Spack's build process will start from scratch.

Keeping the stage directory on success

By default, spack install will delete the staging area once a package has been successfully built and installed. Use --keep-stage to leave the build directory intact:

$ spack install --keep-stage <spec>

This allows you to inspect the build directory and potentially debug the build. You can use clean later to get rid of the unwanted temporary files.

Keeping the install prefix on failure

By default, spack install will delete any partially constructed install prefix if anything fails during install(). If you want to keep the prefix anyway (e.g. to diagnose a bug), you can use --keep-prefix:

$ spack install --keep-prefix <spec>

Note that this may confuse Spack into thinking that the package has been installed properly, so you may need to use spack uninstall --force to get rid of the install prefix before you build again:

$ spack uninstall --force <spec>

Graphing dependencies

spack graph

Spack provides the spack graph command for graphing dependencies. The command by default generates an ASCII rendering of a spec's dependency graph. For example:

$ spack graph hdf5
o  hdf5
| o  openmpi
| |\
| | o  hwloc
| |/| 
| | |\
| | | |\
| | | o |  libxml2
| |_|/| | 
|/| |/| | 
| | | |\ \
o | | | | |  zlib
 / / / / /
| | o | |  xz
| |  / /
| | | o  libpciaccess
| | |/| 
| |/| | 
| | | |\
| | | o |  util-macros
| | |  /
o | | |  numactl
|\ \ \ \
| |\ \ \ \
| | |_|_|/
| |/| | | 
| | |\ \ \
| | o | | |  automake
| | |\| | | 
| | | o | |  autoconf
| |_|/| | | 
|/| |/ / /
| | o | |  perl
| | o | |  gdbm
| | o | |  readline
| | o | |  ncurses
| | |/ /
| | o |  pkgconf
| |  /
| o |  libtool
|/ /
o |  m4
o |  libsigsegv
o  libiconv

At the top is the root package in the DAG, with dependency edges emerging from it. On a color terminal, the edges are colored by which dependency they lead to.

$ spack graph --deptype=link hdf5
o  hdf5
| o  openmpi
| |\
| | o  hwloc
| |/| 
| | |\
| | o |  libxml2
| |/| | 
|/| | | 
| | |\ \
o | | | |  zlib
 / / / /
| o | |  xz
|  / /
o | |  numactl
 / /
| o  libpciaccess
o  libiconv

The deptype argument tells Spack what types of dependencies to graph. By default it includes link and run dependencies but not build dependencies. Supplying --deptype=link will show only link dependencies. The default is --deptype=all, which is equivalent to --deptype=build,link,run,test. Options for deptype include:

  • Any combination of build, link, run, and test separated by commas.

  • all for all types of dependencies.

You can also use spack graph to generate graphs in the widely used Dot format. For example:

$ spack graph --dot hdf5
digraph G {
  labelloc = "b"
  rankdir = "TB"
  ranksep = "5"

  "guz5buq7nzkcixynql243u4ucye4mtyv" [label="hdf5/guz5buq"]
  "7flhg2ud3smax7gsphdhob4ucimjztlr" [label="openmpi/7flhg2u"]
  "pbcn3p2f6zqypyceohqyb2iloqcu57bk" [label="hwloc/pbcn3p2"]
  "cmn6yii6vkzjrapeabvbcna3qqqx6web" [label="libpciaccess/cmn6yii"]
  "jdxbjftheiotj6solpomva7dowrhlerl" [label="libtool/jdxbjft"]
  "olv5tj5h5wttoqa52zduhuj46apmakbg" [label="m4/olv5tj5"]
  "ywasfr4aapg3seh3af667xcbaac7cx5n" [label="libsigsegv/ywasfr4"]
  "azscwdopqzv4picsufqwfbvr6mwejkim" [label="pkgconf/azscwdo"]
  "gs6ag7ktdoiirb62t7bcagjw62szrrg2" [label="util-macros/gs6ag7k"]
  "rbuffvam2e6rkhx2hm2jfphxtcbarbtv" [label="libxml2/rbuffva"]
  "viytb2ptz5qx7fax3dajccmwod6bwxzm" [label="libiconv/viytb2p"]
  "rrennauh5rwpd7mw4dz36lcw44xtkuaa" [label="xz/rrennau"]
  "smoyzzo2qhzpn6mg6rd3l2p7b23enshg" [label="zlib/smoyzzo"]
  "yvxocdy2kcxvvgatrrst3sqzyt7yxhgn" [label="numactl/yvxocdy"]
  "fbgfsafbxmi5rirzfksqccqohfojonk6" [label="autoconf/fbgfsaf"]
  "ioh2jcxomxtumyup3jtsaaitigeovy6c" [label="perl/ioh2jcx"]
  "cyx4faxjefkoy4jaome42keqsmyjvy3z" [label="gdbm/cyx4fax"]
  "motgnr22wv3zbo37flh5eimd36ifunpv" [label="readline/motgnr2"]
  "3cvxm5csh43lgnjkzjge7slh3de5f43i" [label="ncurses/3cvxm5c"]
  "h5u4mzozx3yipvr3ror3hprc7m4yhlxu" [label="automake/h5u4mzo"]

  "7flhg2ud3smax7gsphdhob4ucimjztlr" -> "yvxocdy2kcxvvgatrrst3sqzyt7yxhgn"
  "ioh2jcxomxtumyup3jtsaaitigeovy6c" -> "cyx4faxjefkoy4jaome42keqsmyjvy3z"
  "rbuffvam2e6rkhx2hm2jfphxtcbarbtv" -> "smoyzzo2qhzpn6mg6rd3l2p7b23enshg"
  "7flhg2ud3smax7gsphdhob4ucimjztlr" -> "pbcn3p2f6zqypyceohqyb2iloqcu57bk"
  "yvxocdy2kcxvvgatrrst3sqzyt7yxhgn" -> "olv5tj5h5wttoqa52zduhuj46apmakbg"
  "fbgfsafbxmi5rirzfksqccqohfojonk6" -> "ioh2jcxomxtumyup3jtsaaitigeovy6c"
  "cmn6yii6vkzjrapeabvbcna3qqqx6web" -> "gs6ag7ktdoiirb62t7bcagjw62szrrg2"
  "cyx4faxjefkoy4jaome42keqsmyjvy3z" -> "motgnr22wv3zbo37flh5eimd36ifunpv"
  "pbcn3p2f6zqypyceohqyb2iloqcu57bk" -> "cmn6yii6vkzjrapeabvbcna3qqqx6web"
  "3cvxm5csh43lgnjkzjge7slh3de5f43i" -> "azscwdopqzv4picsufqwfbvr6mwejkim"
  "pbcn3p2f6zqypyceohqyb2iloqcu57bk" -> "azscwdopqzv4picsufqwfbvr6mwejkim"
  "rbuffvam2e6rkhx2hm2jfphxtcbarbtv" -> "viytb2ptz5qx7fax3dajccmwod6bwxzm"
  "olv5tj5h5wttoqa52zduhuj46apmakbg" -> "ywasfr4aapg3seh3af667xcbaac7cx5n"
  "yvxocdy2kcxvvgatrrst3sqzyt7yxhgn" -> "h5u4mzozx3yipvr3ror3hprc7m4yhlxu"
  "rbuffvam2e6rkhx2hm2jfphxtcbarbtv" -> "rrennauh5rwpd7mw4dz36lcw44xtkuaa"
  "pbcn3p2f6zqypyceohqyb2iloqcu57bk" -> "rbuffvam2e6rkhx2hm2jfphxtcbarbtv"
  "jdxbjftheiotj6solpomva7dowrhlerl" -> "olv5tj5h5wttoqa52zduhuj46apmakbg"
  "motgnr22wv3zbo37flh5eimd36ifunpv" -> "3cvxm5csh43lgnjkzjge7slh3de5f43i"
  "guz5buq7nzkcixynql243u4ucye4mtyv" -> "7flhg2ud3smax7gsphdhob4ucimjztlr"
  "cmn6yii6vkzjrapeabvbcna3qqqx6web" -> "azscwdopqzv4picsufqwfbvr6mwejkim"
  "h5u4mzozx3yipvr3ror3hprc7m4yhlxu" -> "ioh2jcxomxtumyup3jtsaaitigeovy6c"
  "yvxocdy2kcxvvgatrrst3sqzyt7yxhgn" -> "jdxbjftheiotj6solpomva7dowrhlerl"
  "cmn6yii6vkzjrapeabvbcna3qqqx6web" -> "jdxbjftheiotj6solpomva7dowrhlerl"
  "guz5buq7nzkcixynql243u4ucye4mtyv" -> "smoyzzo2qhzpn6mg6rd3l2p7b23enshg"
  "fbgfsafbxmi5rirzfksqccqohfojonk6" -> "olv5tj5h5wttoqa52zduhuj46apmakbg"
  "h5u4mzozx3yipvr3ror3hprc7m4yhlxu" -> "fbgfsafbxmi5rirzfksqccqohfojonk6"
  "rbuffvam2e6rkhx2hm2jfphxtcbarbtv" -> "azscwdopqzv4picsufqwfbvr6mwejkim"
  "pbcn3p2f6zqypyceohqyb2iloqcu57bk" -> "yvxocdy2kcxvvgatrrst3sqzyt7yxhgn"
  "7flhg2ud3smax7gsphdhob4ucimjztlr" -> "smoyzzo2qhzpn6mg6rd3l2p7b23enshg"
  "yvxocdy2kcxvvgatrrst3sqzyt7yxhgn" -> "fbgfsafbxmi5rirzfksqccqohfojonk6"

This graph can be provided as input to other graphing tools, such as those in Graphviz. If you have graphviz installed, you can write straight to PDF like this:

$ spack graph --dot hdf5 | dot -Tpdf > hdf5.pdf

Interactive shell support

Spack provides some limited shell support to make life easier for packagers. You can enable these commands by sourcing a setup file in the share/spack directory. For bash or ksh, run:

export SPACK_ROOT=/path/to/spack
. $SPACK_ROOT/share/spack/

For csh and tcsh run:

setenv SPACK_ROOT /path/to/spack
source $SPACK_ROOT/share/spack/setup-env.csh

spack cd will then be available.

spack cd

spack cd allows you to quickly cd to pertinent directories in Spack. Suppose you've staged a package but you want to modify it before you build it:

$ spack stage libelf
==> Trying to fetch from
######################################################################## 100.0%
==> Staging archive: ~/spack/var/spack/stage/libelf@0.8.13%gcc@4.8.3 arch=linux-debian7-x86_64/libelf-0.8.13.tar.gz
==> Created stage in ~/spack/var/spack/stage/libelf@0.8.13%gcc@4.8.3 arch=linux-debian7-x86_64.
$ spack cd libelf
$ pwd
~/spack/var/spack/stage/libelf@0.8.13%gcc@4.8.3 arch=linux-debian7-x86_64/libelf-0.8.13

spack cd here changed the current working directory to the directory containing the expanded libelf source code. There are a number of other places you can cd to in the spack directory hierarchy:

$ spack cd --help
usage: spack cd [-h] [-m | -r | -i | -p | -P | -s | -S | -b | -e ENV] ...

cd to spack directories in the shell

positional arguments:
  spec               spec of package to fetch directory for

optional arguments:
  -h, --help         show this help message and exit
  -m, --module-dir   spack python module directory
  -r, --spack-root   spack installation root
  -i, --install-dir  install prefix for spec (spec need not be installed)
  -p, --package-dir  directory enclosing a spec's file
  -P, --packages     top-level packages directory for Spack
  -s, --stage-dir    stage directory for a spec
  -S, --stages       top level stage directory
  -b, --build-dir    checked out or expanded source directory for a spec (requires it to be staged first)
  -e ENV, --env ENV  location of an environment managed by spack

Some of these change directory into package-specific locations (stage directory, install directory, package directory) and others change to core spack locations. For example, spack cd --module-dir will take you to the main python source directory of your spack install.

spack env

spack env functions much like the standard unix env command, but it takes a spec as an argument. You can use it to see the environment variables that will be set when a particular build runs, for example:

$ spack env mpileaks@1.1%intel

This will display the entire environment that will be set when the mpileaks@1.1%intel build runs.

To run commands in a package's build environment, you can simply provide them after the spec argument to spack env:

$ spack cd mpileaks@1.1%intel
$ spack env mpileaks@1.1%intel ./configure

This will cd to the build directory and then run configure in the package's build environment.

spack location

spack location is the same as spack cd but it does not require shell support. It simply prints out the path you ask for, rather than cd'ing to it. In bash, this:

$ cd $(spack location --build-dir <spec>)

is the same as:

$ spack cd --build-dir <spec>

spack location is intended for use in scripts or makefiles that need to know where packages are installed. e.g., in a makefile you might write:

DWARF_PREFIX = $(spack location --install-dir libdwarf)

Build System Configuration Support

Imagine a developer creating a CMake or Autotools-based project in a local directory, which depends on libraries A-Z. Once Spack has installed those dependencies, one would like to run cmake with appropriate command line and environment so CMake can find them. The spack setup command does this conveniently, producing a CMake configuration that is essentially the same as how Spack would have configured the project. This can be demonstrated with a usage example:

$ cd myproject
$ spack setup myproject@local
$ mkdir build; cd build
$ ../ ..
$ make
$ make install


  • Spack must have myproject/ in its repository for this to work.

  • spack setup produces the executable script in the local directory, and also creates the module file for the package. is normally run from the user's out-of-source build directory.

  • The version number given to spack setup is arbitrary, just like spack diy. myproject/ does not need to have any valid downloadable versions listed (typical when a project is new).

  • produces a CMake configuration that does not use the Spack wrappers. Any resulting binaries will not use RPATH, unless the user has enabled it. This is recommended for development purposes, not production.

  • is human readable, and can serve as a developer reference of what dependencies are being used.

  • make install installs the package into the Spack repository, where it may be used by other Spack packages.

  • CMake-generated makefiles re-run CMake in some circumstances. Use of breaks this behavior, requiring the developer to manually re-run when a CMakeLists.txt file has changed.


In order to enable spack setup functionality, the author of myproject/ must subclass from CMakePackage instead of the standard Package superclass. Because CMake is standardized, the packager does not need to tell Spack how to run cmake; make; make install. Instead the packager only needs to create (optional) methods configure_args() and configure_env(), which provide the arguments (as a list) and extra environment variables (as a dict) to provide to the cmake command. Usually, these will translate variant flags into CMake definitions. For example:

def configure_args(self):
    spec = self.spec
    return [
        '-DUSE_EVERYTRACE=%s' % ('YES' if '+everytrace' in spec else 'NO'),
        '-DBUILD_PYTHON=%s' % ('YES' if '+python' in spec else 'NO'),
        '-DBUILD_GRIDGEN=%s' % ('YES' if '+gridgen' in spec else 'NO'),
        '-DBUILD_COUPLER=%s' % ('YES' if '+coupler' in spec else 'NO'),
        '-DUSE_PISM=%s' % ('YES' if '+pism' in spec else 'NO')

If needed, a packager may also override methods defined in StagedPackage (see below).


CMakePackage is implemented by subclassing the StagedPackage superclass, which breaks down the standard Package.install() method into several sub-stages: setup, configure, build and install. Details:

  • Instead of implementing the standard install() method, package authors implement the methods for the sub-stages install_setup(), install_configure(), install_build(), and install_install().

  • The spack install command runs the sub-stages configure, build and install in order. (The setup stage is not run by default; see below).

  • The spack setup command runs the sub-stages setup and a dummy install (to create the module file).

  • The sub-stage install methods take no arguments (other than self). The arguments spec and prefix to the standard install() method may be accessed via self.spec and self.prefix.

GNU Autotools

The setup functionality is currently only available for CMake-based packages. Extending this functionality to GNU Autotools-based packages would be easy (and should be done by a developer who actively uses Autotools). Packages that use non-standard build systems can gain setup functionality by subclassing StagedPackage directly.