Tup Backend

The Tup build backend is an alternative to the default Make backend. The Tup build system is designed for fast and correct incremental builds. A top-level no-op build should be under 2 seconds, and clobbers should rarely be required. It is currently only available for Linux Desktop builds – other platforms like Windows or OSX are planned for the future.

As part of the mozbuild architecture, the Tup backend shares a significant portion of frontend (developer-facing) code in the build system. When using the Tup backend, ./mach build is still the entry point to run the build system, and moz.build files are still used for the build description. Familiar parts of the build system like configure and generating the build files (the Reticulating splines... step) are virtually identical in both backends. The difference is that mach invokes Tup instead of Make under the hood to do the actual work of determining what needs to be rebuilt. Tup is able to perform this work more efficiently by loading only the parts of the DAG that are required for an incremental build. Additionally, Tup instruments processes to see what files are read from and written to in order to verify that dependencies are correct.

For more detailed information on the rationale behind Tup, see the Build System Rules and Algorithms paper.

Installation

You’ll need to install the Tup executable, as well as the nightly rust/cargo toolchain (Note: Replace $topsrcdir with the path to your mozilla-central source tree):

cd ~/.mozbuild && $topsrcdir/mach artifact toolchain --from-build linux64-tup
rustup install nightly
rustup default nightly

Configuration

Your mozconfig needs to describe how to find the executable if it’s not in your PATH, and enable the Tup backend:

export TUP=~/.mozbuild/tup/tup
ac_add_options --enable-build-backends=Tup

Configuring Parallel Jobs

To override the default number of jobs run in parallel, set MOZ_PARALLEL_BUILD in your mozconfig:

mk_add_options MOZ_PARALLEL_BUILD=8

What Works

You should expect a Linux desktop build to generate a working Firefox binary from a ./mach build, and be able to run test suites against it (eg: mochitests, xpcshell, gtest). Top-level incremental builds should be fast enough to use them during a regular compile/edit/test cycle. If you wish to stop compilation partway through the build to more quickly iterate on a particular file, you can expect ./mach build objdir/path/to/file.o to correctly produce all inputs required to build file.o before compiling it. For example, you don’t have to run the build system in various subdirectories to get generated headers built in the right order.

Currently Unsupported / Future Work

There are a number of features that you may use in the Make backend that are currently unsupported for the Tup backend. We plan to add support for these in the future according to developer demand and build team availability.

  • sccache - This is currently under active development to support icecream-like functionality, which likely impacts the same parts that would affect Tup’s dependency checking mechanisms. Note that icecream itself should work with Tup.
  • Incremental Rust compilation - see bug 1468527
  • Watchman integration - This will allow Tup to skip the initial Scanning filesystem... step, saving 1-2 seconds of startup overhead.
  • More platform support (Windows, OSX, etc.)
  • Packaging in automation - This is still quite intertwined with Makefiles
  • Tests in automation - Requires packaging

Multiple object directories and object directories outside of the source tree

Common workflows involving multiple object directories which may be outside of the source directory are supported by Tup, however there are some things to consider when using these configurations.

  • Using multiple object directories works as expected, however you may find pulling and attempting to build will fail due to tup attempting to parse a Tupfile in an object directory other than the active object directory. As a workaround, activate the object directory containing the failing file and run configure before proceeding.
  • If the currently active object directory is located outside of the source directory, the tup backend will prompt the user to run tup init --no-sync in a common ancestor directory of the source directory and object directory. If this ancestor contains too many files (it’s the home directory, for instance), this will slow down tup’s initial file scan. Anecdotally multiple object directories will only incur marginal scanning overhead, but care should be exercised when choosing a directory layout.

Tup builds in automation

Tup builds run on integration branches as Btup in treeherder. There are some aspects of the Tup builds that are currently implemented outside of the make build system, and divergences may cause the Btup job to fail until the build is completely integrated. There are two known situations this has come up.

  • Changes to the xpidl/webidl/ipdl build - these are given special treatment by the make backend that is re-created in the tup backend. An update to a Makefile implementing one of these parts of the build may require a corresponding change to python/mozbuild/mozbuild/backend/tup.py
  • Build scripts generating new outputs - due to the implementation of Tup, outputs from build.rs that run during the build must be known to Tup before the build starts. The current outputs are enumerated in python/mozbuild/mozbuild/backend/cargo_build_defs.py. Modifying the set of outputs from a build script will require a corresponding update to this file.

Partial tree builds in tup

Partial tree builds are possible in tup by invoking ./mach build <relative srcdir path>, however the behavior differs from the make backend. A partial tree build will result in running the commands in Tupfiles in the specified subtree, building .o files and other outputs, but unlike make it will take changed dependencies into account and build everything necessary to update those outputs as well. Also unlike make it will not attempt to find downstream commands that depend on these files, i.e. programs and libraries from other parts of the tree will not be linked. A top level incremental build should be performed to run all commands that depend on changes to the local tree.

How to Contribute

At the moment we’re looking for early adopters who are developing on the Linux desktop to try out the Tup backend, and share your experiences with the build team (see Contact).

  • Are there particular issues or missing features that prevent you from using the Tup backend at this time?
  • Do you find that top-level incremental builds are fast enough to use for every build invocation?
  • Have you needed to perform a clobber build to fix an issue?

Contact

If you have any issues, feel free to file a bug blocking buildtup, or contact mshal or chmanchester in #build on IRC.