Warning The GN build is experimental and best-effort. It might not work, and if you use it you're expected to feel comfortable to unbreak it if necessary. LLVM's official build system is CMake, if in doubt use that. If you add files, you're expected to update the CMake build but you don't need to update GN build files. Reviewers should not ask authors to update GN build files. Keeping the GN build files up-to-date is on the people who use the GN build.
GN is a metabuild system. It always creates ninja files, but it can create some IDE projects (MSVC, Xcode, ...) which then shell out to ninja for the actual build.
Its main features are that GN is very fast (it currently produces ninja files for LLVM's build in 35ms on the author's laptop, compared to 66s for CMake) -- a 2000x difference), and since it's so fast it doesn't aggressively cache, making it possible to switch e.g. between release and debug builds in one build directory.
The main motivation behind the GN build is that some people find it more convenient for day-to-day hacking on LLVM than CMake. Distribution, building just parts of LLVM, and embedding the LLVM GN build from other builds are a non-goal for the GN build.
This is a good overview of GN.
GN only works in the monorepo layout.
- Obtain a gn binary. If gn is not already on your PATH, run llvm/utils/gn/get.py to download a prebuilt gn binary if you're on a 64-bit X86 system running Linux, macOS, or Windows, or build gn yourself if you're on a different platform or don't want to trust prebuilt binaries.
- In the root of the monorepo, run llvm/utils/gn/gn.py gen out/gn. out/gn is the build directory, it can have any name, and you can have as many as you want, each with different build settings. (The gn.py script adds --dotfile=llvm/utils/gn/.gn --root=. and just runs regular gn; you can manually pass these parameters and not use the wrapper if you prefer.)
- Run e.g. ninja -C out/gn check-lld to build all prerequisites for and run the LLD tests.
By default, you get a release build with assertions enabled that targets the host arch. You can set various build options by editing out/gn/args.gn, for example putting is_debug = true in there gives you a debug build. Run llvm/utils/gn/gn.py args --list out/gn to see a list of all possible options. After touching out/gn/args.gn, just run ninja, it will re-invoke gn before starting the build.
GN has extensive built-in help; try e.g. gn help gen to see the help for the gen command. The full GN reference is also available online.
GN has an autoformatter: git ls-files '*.gn' '*.gni' | xargs -n 1 gn format after making GN build changes is your friend.
To not put BUILD.gn into the main tree, they are all below utils/gn/secondary. For example, the build file for llvm/lib/Support is in utils/gn/secondary/llvm/lib/Support.
Sometimes after pulling in the latest changes, the GN build doesn't work. Most of the time this is due to someone adding a file to CMakeLists.txt file. Run llvm/utils/gn/build/sync_source_lists_from_cmake.py to print a report of which files need to be added to or removed from BUILD.gn files to match the corresponding CMakeLists.txt. You have to manually read the output of the script and implement its suggestions.
If new CMakeLists.txt files have been added, you have to manually create a new corresponding BUILD.gn file below llvm/utils/gn/secondary/.
If the dependencies in a CMakeLists.txt file have been changed, you have to manually analyze and fix.
GN believes in using GN arguments to configure the build explicitly, instead of implicitly figuring out what to do based on what's available on the current system.
configure is used for three classes of feature checks:
- compiler checks. In GN, these could use exec_script to identify the host compiler at GN time. For now the build has explicit toggles for compiler features. (Maybe there could be a script that writes args.gn based on the host compiler). It's possible we'll use exec_script() for this going forward, but we'd have one exec_script call to identify compiler id and version, and then base GN arg default values of compiler id and version instead of doing one exec_script per feature check. (In theory, the config approach means a new os / compiler just needs to tweak the checks and not the code, but in practice a) new os's / compilers are rare b) they will require code changes anyhow, so the configure tradeoff seems not worth it.)
- library checks. For e.g. like zlib, GN thinks it's better to say "we require zlib, else we error at build time" than silently omitting features. People who really don't want to install zlib can explicitly set the GN arg to turn off zlib.
- header checks (does system header X exist). These are generally not needed (just keying this off the host OS works fine), but if they should become necessary in the future, they should be done at build time and the few targets that need to know if header X exists then depend on that build-time check while everything else can build parallel with it.
- LLVM-specific build toggles (assertions on/off, debug on/off, targets to build, ...). These map cleanly to GN args (which then get copied into config.h in a build step).
For the last two points, it would be nice if LLVM didn't have a single config.h header, but one header per toggle. That way, when e.g. llvm_enable_terminfo is toggled, only the 3 files caring about that setting would need to be rebuilt, instead of everything including config.h.
GN doesn't believe in users setting arbitrary cflags from an environment variable, it wants the build to be controlled by .gn files.