Sunday, March 1, 2020

Unity build test with Meson & LibreOffice

In a previous blog post we managed to build a notable chunk of LibreOffice with Meson. This has since been updated so you can build all top level apps (Writer, Calc, Impress, Draw). The results do not actually run, so the conversion may seem pointless. That is not the case, though, because once you have this build setup you can start doing interesting experiments on a large real world C++ code base. One of these is unity builds.

A unity build is a surprisingly simple technique to speed up builds, especially C++. The idea is that instead of compiling source files individually, you build them in larger batches by creating files like these and compiling them instead:

// etc

One of the main downsides of this technique is writing and maintaining the unity files by hand. Fortunately Meson has builtin support for creating unity files for build targets. In the next release you can even specify how many source files you want to have in a single unity source. Enabling unity builds for a single target is simple and consists of adding the following keyword argument to a target's definition:

override_options: ['unity=on']

In this experiment we used LibreOffice's Writer target, which is a single shared library consisting of almost 700 source files. The unity block size was Meson's default of 4.

The results

Compiling source code as a unity build for the first time usually leads to build failures. This is exactly what happened here as well. There are many reasons for this. The most popular are name clashes from static and anonymous namespace functions and of classes with common names that are used without namespace qualifications. LO turned out to have at least three different classes called Size. All these issues need to be fixed either by renaming or by adding namespace qualifiers. This is boring manual work, but on the other hand you may find duplicated static functions across your code base.

Once all the issues have been fixed we can do actual measurements. The test machine was an 8 thread i7-3770 with an SSD drive and the buildmode was set to debug. The full build times are these:

Regular   10m 
Unity      4m 32s

This unity build is over 50% faster than the regular one, which is a fairly typical result. The difference would have been even bigger if we had used a bigger unity block size. Incremental builds were tested by deleting one object file and rebuilding.

Regular 26s
Unity   22s

Typically incremental unity builds are slower than regular ones, but in this case it is actually faster. This is probably because unity builds produce smaller output files that have less debug data. Thus the linker has less work. Increasing the unity block size would make the incremental build time slower.

Converting the code to compile as a unity build took on the order of three hours. Ironically most of it was spent waiting for compilations to finish. Since this saves around 5 minutes per build, the time investment is recovered after the development team has done 60 full builds. The code changes have been done with as little effort as possible, so polishing this to production quality would probably take 2-3 times as long.

Do try this at home!

The code is available in the unitytest branch of this repo, it has only been tested with Meson trunk on Linux. There are two targets that use unity builds: vcl and sw. You can toggle unity builds as well as the unity block size per target with the override_options keyword argument.

1 comment:

  1. The big reason even incremental builds (rebuilding a single object file) are faster is because there are much less object files to link together and there are much fewer weak symbols to de-duplicate - think of all templates and inline functions in headers - it's the linker's job to remove all such 'weak' symbols and leave just one.

    There are a lot more details on unity builds here: