Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

What is gup?

Gup is a general purpose, recursive, top down software build system. It doesn't care what language or framework your project uses. It has almost no syntax. It doesn't even care what language you write your build rules in, which is a fairly novel concept for a build system. It freely allows nesting of components and projects, and doesn't care where you build things from (similar to a recursive make setup, but without the downsides).

It's very similar to redo, which was designed by Daniel J. Bernstein and implemented by Avery Pennarun. So I can't take much credit for inventing it, it's really a refinement of an already-great idea.

Gup doesn't require you to specify your dependencies up-front, just an executable script which will build a target. This can be written in any language that supports shebang-style scripts (files beginning with #!/path/to-interpreter), and you can freely mix projects or components with build scripts written in different languages.

In each build script script, you declare your dependencies as you need them, simply by telling gup to make sure they're up to date. This allows you to keep "the place where you use a file" and "the place where you declare a dependency on a file" colocated, and even to create abstractions that handle these details for you. This makes it much easier to create correct dependency information compared to systems like make which rely on discipline and intimate knowledge of build internals to maintain correct dependencies.

How do I install it?


gup is packaged in nixpkgs, so you can do nix-env -iA nixpkgs.gup to install it in your profile, or add pkgs.gup to your own package's buildInputs.

Bundled with your project

For repositories where you don't want to make everyone set up gup as above, you can just commit the gup python script to your project's repository:

  • From a git checkout or released tarball:

     cp python/bin/* [your-project-workspace]/tools

(If python/bin doesn't exist, you may need to make python first).

Then, you can run it as ./tools/gup.

Windows users:

If you want Windows users (perhaps you are one) to be able to build you project, remember to include the gup.exe shim. This a portable .exe which just runs the python script named "gup" in whatever directory you place it, passing along all arguments.

Note: When run from a relative or absolute path, gup will bootstrap itself by adding the directory containing it to $PATH, so that executing gup will do the right thing from within a build script. So you may wish to place it in its own directory to avoid accidentally adding other scripts to $PATH.

Bundled builders

Gup includes some generic builder scripts (in [./builders][]), which are typically installed alongside gup. These are intended to be used as builders in a Gupfile (prefixed with !). They are provided as-is, and will not necessarily be as portable or as stable as gup itself. If you bundle gup into your project, you should also include any scripts you make use of (you can place them next to the gup binary, as gup will automatically add its directory to your $PATH when building).

Python version requirements

Gup works under both python 2 and 3, so it should work with whatever recent python is on your $PATH.

I don't routinely test under old python versions, so if you need to run gup in anything below 2.7 and you find something broken, patches welcome.

Using it via make

For convenience, you may wish to provide a Makefile that simply delegates to gup. You can find an example Makefile in the resources/ directory.

How does it build stuff?

The simplest build script is an executable which creates a single target. This is called a "direct build script", and must be named "[target-name].gup". If you've ever written a shell script to (re)generate a single file, you know the basics of writing a direct build script. Here's a simple example:

Here's a simple example script which would be used to build the file named target:


gup -u source
cat source > "$1"
echo "## Built by gup" >> "$1"

When you ask gup to update target, with:

$ gup -u target

It will run this script, (almost) as if you had written ./target.gup target.

Gup also supports indirect build scripts, which maps multiple targets (e.g. "any .o file") to a single build script (which in this case knows how to build any .o file). These mappings are specified in a file named Gupfile.

More examples

Please see the examples directory - this contains example build scripts covering gup's main features, plus a file with detailed descriptions of what's going on.

Invoking gup:

gup has a few options, but most of the time you'll just be using:

$ gup -u <target>

The -u means "update" - i.e. only build the target if it is out of date. This is how you update a target from the command line. Simply enough, it's also how you update a dependency from within a build script.

Note: Unlike many build tools, gup doesn't have the concept of a "project directory" - you can build any target from anywhere - building target from within dir/ is exactly the same as building dir/target, building ../dir/target from some-other-dir/, or building /full/path/to/dir/target from anywhere.


In each build script script, you declare your dependencies as you need them, simply by telling gup to make sure they're up to date. So just as you run gup -u <target> to build a target, each build script can itself run gup -u <input> to make sure each file that it depends on is up to date. If the file is buildable by gup, it will be built before the script continues. If not, gup will remember that this target depends on <input>.

This allows you to keep "the place where you use a file" and "the place where you declare a dependency on a file" colocated, and even to create abstractions that handle these details for you. This makes it much easier to create correct dependency information compared to systems like make which rely on discipline and intimate knowledge of build internals to maintain correct dependencies.

You don't have to specify all your dependencies in one place - you can call gup -u from any place in your script. All calls made while building your target are appended to the dependency information of the target being built. This is done by setting environment variables when invoking build scripts, so it's completely parallel-safe.

Also, gup will by default include a dependency on the build script used to generate a target. If you change the build script (or add a new build script with a higher precedence for a given target), that target will be rebuilt.

Build script execution:

gup invokes your build script as scriptname output_path target. Wherever possible, your build script should generate its result in output_path. This is an absolute path to a temporary file that, once your buildscript has completed successfully, will replace the current target file. There are a few benefits to this approach:

  1. By generating this file instead of writing to target directly, you can ensure that target is always the result of a successful build, and not a partially-created file from a build script that failed halfway through.

  2. If your script does not generate anything in output_path, gup will automatically delete target for you. This means that you won't get left with stale built files if your build script changes to no longer produce any output.

  3. When building a directory, you can be assured that output_path does not yet exist. This means you can create it via mkdir, and not have to worry about cleaning up a previous version (gup will do this for you on a successful build).

If your build script cannot create output_path, it's OK to write to target instead. Often this is necessary when shelling out to other tools where it's not easy to direct their output to a specific file. Note: If you write your results to target directly and there's a chance that the build won't touch target (because it turns out to be a no-op), you must touch target in your build script. If you don't modify the mtime of target, gup will think that your build script produced no output, and delete target (due to point #2 above).

gup in depth

The search path

When you run gup build/foo from /home/tim/proj, gup will look for direct build scripts, in this order:

  • build/foo.gup
  • build/gup/foo.gup
  • gup/build/foo.gup
  • ../gup/proj/build/foo.gup
  • ../../gup/tim/proj/build/foo.gup
  • ../../../gup/home/tim/proj/build/foo.gup

Obviously, it's getting unlikely that you would keep a root /gup folder, but gup doesn't really know where your project might end, so it just keeps searching up to the root of your filesystem for consistency (e.g you may have projects nested inside other projects, you wouldn't want it to stop at the root of the innermost project, because then it would change behaviour based on where it was called from).

Notice that this search pattern means you can choose whether to keep your build scripts co-located with targets, or whether you keep them in a "shadow" directory structure under gup/ somewhere. Having .gup files next to targets can be convenient for small projects, but larger projects will usually want to use a gup/ folder structure. Also, a gup/ structure can contain build scripts for targets in directories that don't (yet) exist, which can be very useful.

After gup has exhausted the search for direct build scripts, it will start looking for indirect build scripts, which are always named Gupfile. A Gupfile has the syntax:


That is, a non-indented line names a build script - any file path (relative to the directory containing this Gupfile). A BUILDSCRIPT prefixed with ! will search for the named executable on $PATH, rather than in the current directory.

Each following indented line is a target pattern, which may include the * wildcard to match anything except a "/", and ** to match anything including "/". A target pattern beginning with ! inverts the match.

The search for Gupfiles follows a similar pattern as the search for direct build scripts, but if it fails to find a match for "foo", it'll start to look in higher directories for a Gupfile that knows how to build "build/foo".

using pattern "foo":

- build/Gupfile
- build/gup/Gupfile
- gup/build/Gupfile
- ../gup/proj/build/Gupfile
- (...)

using pattern "build/foo":

- Gupfile
- gup/Gupfile
- ../../gup/tim/proj/foo.gup
- ../../../gup/home/tim/proj/foo.gup

using pattern "proj/build/foo":

- ../Gupfile
- ../gup/Gupfile
- ../../gup/tim/proj/foo.gup
- ../../../gup/home/tim/proj/foo.gup


This search behaviour may seem a little hard to follow, but it's fairly intuitive once you actually start using it. It's also worth noting that if you create a "more specific" build script for a target which has already been built, gup will know that the target needs to be rebuilt using the new build script, just like it rebuilds a target when the build script is modified.

Running build scripts

Once gup has found the script to build a given target, it runs it as an executable. More concretely:

It will interpret shebang lines so that you don't actually have to make your script executable, but you can also use an actual executable (with no shebang line) if you need to.

Relative script paths (starting with ./) in the shebang line are supported, and are resolved relative to the directory containing the script, rather than the current working directory.

If your interpreter is just a program name (e.g. #!python), it will be searched for in $PATH, making this a terse equivalent of #!/usr/bin/env python.

On windows, UNIX-y paths like /bin/bash may not be present, so you should either use #!bash or #!/usr/bin/env bash (gup has special support for /usr/bin/env in shebang lines, even if that path doesn't exist locally).

When a build script is run, its working directory ($PWD) is typically set to the directory containing the Gupfile (for indirect targets), or the script itself (for direct <name>.gup targets). The exception is that when a Gupfile or script lives inside a gup/ directory, it is run from the equivalent path without the gup/ directory. This means that $2 will always correspond to the matched pattern in a Gupfile.

For example:

  1. building a/b/c with a direct gupfile at a/b/c.gup will invoke a/b/c.gup from a/b with $2 set to c.

  2. For the following src/Gupfile contents:



  • gup src/tmp/foo would run src/default.gup from src/, with $2 set to tmp/foo.
  • gup src/foo.o would run src/scripts/default.o.gup from src, with $2 set to foo.o
  • gup src/foo.a would run gup-link (found on $PATH) from src, with $2 set to foo.a

But if the same files were nested under gup/ (e.g gup/src/Gupfile, gup/src/default.gup, etc, then the paths would remain the same - the gup/ path containing the actual script being executed would be accessible only via $0.

Again, these rules may seem complex to read, but the actual behaviour is fairly intuitive - rules get run from the target tree (not inside any /gup/ path), and from a directory depth consistent with the location of the Gupfile (or .gup file).


Gup exports a few environment variables which are used to propagate settings to sub-invocations of gup. If you like, you can use some of these to influence your build script too.

  • GUP_TARGET: You should not care about the value of this, but it will always be set in a gup build. By testing for the absence of this variable, you can have a build script which also does something useful when run directly (not through gup).

  • GUP_XTRACE: Set to either "0" or "1". When you pass the -x or --trace flags to gup, this variable will be set "1". This can be used to enable extra output from your build script (e.g. set -x in bash, or increasing the log verbosity in other languages).

  • GUP_VERBOSE: "0" is the default, and 1 is added for each -v flag passed to gup. May be a negative number if -q is used. You normally shouldn't need to use this, as -v is typically for debugging gup itself. But it can be useful if you need more levels than GUP_XTRACE provides.

Sample "shadow" directory structure

The ability to keep build scripts in a separate gup/ directory is useful for cleanliness, but also for defining targets whose destination doesn't actually exist yet. For example, you might have:


Where the Gupfile contains rules like:



When building build/windows/foo, it would run the script gup/build/build-foo from the directory build/ and "$2" (the target path) set to windows/foo, from which it could determine which OPTS file to use. When gup finds a build script for a path that doesn't yet exist (build/windows/ in this case), it will create this directory just prior to running the build script - so your script can always assume the base directory for the target it's building will exist.

This also allows you to keep your source and built files separate - your clean task can be as simple as rm -rf build/.

For a working example of a shadow directory, see the examples directory.

Checksum tasks

In the above example, there's a bug. If I add a new file to src/, gup doesn't know that all should be rebuilt, because it called gup -u on the files that existed at the time it was last built.

The solution for this is to use a checksum task, like so:


set -eu

find src -type f > "$1"
gup --always
gup --contents "$1"


set -eu

gup -u inputs
files="$(cat inputs)"
gup -u $files
cat $files > "$1"

The inputs task will always be re-run (at most once per gup invocation), but after it runs, it will only be considered out of date (and cause all to be rebuilt) if the contents of the file passed to gup --contents changed since the last time it was built - in this case, if the number or names of files in src/ changes.

You can also pass contents to gup --contents via stdin, for pure "stamp" tasks where the output is not actually used for anything.


Gup (as of version 0.8.0) adds explicit support for having generated builders. i.e. if you are building a target x using generate-x, then generate-x is allowed to be a target itself, and generate-x will be both a dependency of x and also a target.

Why would you want to do that? Compiled buildscripts! gup has always supported arbitrary executables as builders, but before now it didn't treat any builders as possible targets. You would have to gup -u <buildscript> explicitly in every target which was built by <buildscript>, or gup -u buildscript as a pre-step to the rest of your targets. Both of these solutions are easy to get wrong.

Big caveat: In general, gup searches quite a few paths in order to determine whether a path is a buildable target or just a plain file. In the case of plain files this is unavoidable, since gup knows nothing about them.

But to ensure implementation of this advanced feature doesn't impact the performance of the typical case, builders-of-builders have an additional limitation: they must be declared as a literal path (i.e. not a wildcard match) in the same Gupfile as the target(s) they are responsible for building.

That means you cannot have a direct foo.gup builder be buildable. Instead, you should pick a different name (to avoid confusion), say make-foo, and write a Gupfile containing:



This is a little awkward, but by requiring that the builder for make-foo exists in the same Gupfile which defines make-foo as a builder, the check for "is this builder itself buildable?" becomes essentially free.


gup can build targets in parallel - just pass in the maximum number of concurrent tasks after -j or --jobs.

Gup's python implementation uses a parallel jobserver, compatible with Make.

Due to limitations in inheriting open file descriptors, it also implements a jobserver backed by a named pipe. The named pipe is used by default because it is more robust, but the make-compatible jobserver is used when we detect we're being executed from make.

This is a bunch of complexity for a small amount of gain (sharing Make's jobserver, which is broken by many modern runtimes anyway).

In version 0.8.0, the make-compatible jobserver in the OCaml version was reworked into an actual RPC system, communicating via unix datagram sockets. The initial gup process acts as a server, and all child gup instances simply delegate commands to the root process.

This has one main benefit, in that we can more aggressively cache information needed to determine the state of a target, since all checks are done in the same process.

It also opens up the possibility for embedding the client side of this RPC protocol directly in builders, which may improve performance for large builds.

Using bash

bash is convenient as it's almost always installed on UNIX-like systems. It has a lot of gotchas, though. One benefit of gup is that it doesn't care what interpreter you use, so I encourage you to use a better language than bash for anything non-trivial.

I tend to use python, but ruby and perl are other common choices. If your project is a library or program for a particular language, it's often a good choice to write your build scripts in that same language - that way anyone contributing source code can probably understand your build code, too!

Relationship to redo

gup is based heavily on the redo build system. Most of the core ideas come from redo, since I used (and enjoyed) redo for quite some time. But as I encountered a number of bugs that were hard to fix while staying within the design of redo, I decided I could make a more reliable system by changing some of the details.

If you're familiar with redo, here are the main (intentional) differences:

  • gup will rebuild targets if the stored metadata goes missing or is incorrect, rather than silently(*) assuming they never need to be built again.

    (*) redo does print a warning, but that's too easily missed in automated build processes, and requires manual intervention to fix the problem.

  • gup uses Gupfiles to specify exactly what should be built by each generic build script, rather than segregating targets only by their file extension or resorting to a that tries to build every nonexistent file.

    This will sometimes mean more work for you, but allows gup to be completely confident about what is buildable, without ever having to guess based on the current state of your working copy.

  • gup allows you to keep your source tree clean, by using a "shadow" gup/ directory - no need to scatter build and output files amongst your source code.

As a side effect of the above differences, gup never thinks that a target is a source when it's actually a target, or vice versa. It does mean than an existing (source) file could be overwritten by an improperly written Gupfile if you accidentally tell gup it was a target (and your build script succeeds). This is true of many build systems though, and if you keep a clean separation between source and generated files (e.g by placing all generated files under build/), it's reasonably hard to do the wrong thing by accident.

Where can I discuss stuff?

If it's a straightforward bug or feature request, it can probably go straight on the github issues page.

For all other discussion, there's a mailing list.

What platforms does it work on?

The (fairly extensive) automated test suite passes on Linux, OSX and Windows. The tests are automatically run on Linux (python2 & python3) whenever I push new code, the other platforms only get tested whenever I worry that I might have broken something. If you notice any breakage, please let me know by opening a github issue (or just emailing me) and I'll do my best to fix it.

The ocaml version is more or less experimental - I use it in daily life, but I have (as yet) made no attempt to test it on things that aren't linux.

How buggy is it?

It's a relatively simple build system compared to many, and I've been using it on various projects for a few years. I think it's pretty stable, but there could still be bugs lurking. Please report any you find to the github issues page.

Similarly, if you think anything is confusing in this documentation, please suggest improvements. I already exactly how gup works, so I may be explaining it badly.

How stable is the design?

Strictly speaking, I make no guarantees. However, gup has been around a while now, and I'm pretty happy with it. Being the lazy guy that I am, it seems likely that gup won't change in a backwards-incompatible way unless there's a very good reason for it.

However, nothing is set in stone - please raise any issues you find or improvements you can think of as a github issue. If something is broken, I'd much rather know about it now than later :)

Why the name gup?

It's short, easy to type, and it contains up (as in update).

I pronounce it with a hard "g", rhyming with "cup". You can think of it as "gee-up" or "get up" or even "get thyself upwards" if you like, but you should just say "gup" if you have cause to say it out loud.

Hacking on gup:

The python code (under python/ and test/) is currently the supported version of gup.

There is an experimental ocaml implementation (under ocaml/), which I intend to be 100% interchangeable with the python version. The purpose of the ocaml version is twofold: It should be quite a bit faster than python, and I wanted to learn OCaml. The ocaml version is less easily portable than the python version though, and probably doesn't even compile on non-linux systems yet.

I don't require contributors to make changes to both codebases (i.e fixing a bug in just the python code is fine with me). It would of course be convenient if you do update the python & ocaml code simultaneously when submitting a change, but I know that's a lot to ask, so I will assume responsibility of keeping the codebases in sync as necessary.

If you do add tests or fix bugs, please include tests. These are written in python, and are run over both the python & ocaml versions. Doing this (aside from generally being a good idea) will help avoid disparity between the two versions.


If you have nix, you can just run nix-shell and then use make.

Otherwise, you'll need to get dependencies manually:

For python/:

  • PyChecker (optional; you can disable this by exporting $SKIP_PYCHECKER=1)
  • gmcs (only if you want to run on windows; provided by the mono-core package on RPM systems and mono-gmcs on debian-based systems). If you don't need Windows support, you can ignore this. The shim is tiny and changes rarely, so you should also be fine to use gup.exe from a release tarball.

For ocaml/, the best bet is to look at the before_install script in .travis.yml and follow those instructions.


The gup build process itself uses make, but you can use the included ./make or ./make.bat if you don't already have make installed.

Most operations work from the appropriate directory - e.g running make bin, make unit-test, make integration-test in either the ocaml/ or python/ directories will do the right thing.

To run the automated tests, you will also need some standard GNU utilities, including bash, find, cat, etc. On Windows, MSYS provides these.


Gup is distributed under th LGPL - see the LICENCE file.

Copyrights and are adapted from the redo project, which is LGPL and is Copyright Avery Pennarun and some of is adapted from the ZeroInstall project, which is LGPL and is copyright Thomas Leonard.

All other source code is Copyright Tim Cuthbertson, 2013-2014.


A better make, inspired by djb's redo.







No packages published