Streamlining LOOT development

Oct 17, 2018 3 minute read

LOOT has made its core sorting and metadata access functionality available to other applications through a C++ library since 2015. It used to be something of an afterthought, a drain on time that could have been spent improving LOOT itself, but now it sits at the heart of everything LOOT does, and that’s made it a much better library all around.

The library is currently known as the LOOT API, but let’s call it libloot here, because “LOOT API” is a terrible name and I should probably change it. Before early 2017, libloot and the LOOT application (which I’ll just call “LOOT” from now on) lived in the same repository and shared some code in a core library that they’d both statically link to.

However, there was a lot of duplication outside of that core library, because libloot and LOOT did similar things in different ways. Typically, because LOOT offered much more flexibility, it would interact with the core library in more detail, while libloot would generally stick to simpler interfaces. This meant that there was a lot of overhead when making changes, and because libloot wasn’t the core “product”, there wasn’t much incentive to keep its development up to date with the application’s. I wouldn’t leave it broken, but if a new feature was added to the application, I wouldn’t necessarily expose it through the API too.

This wasn’t a great position to be in, architecturally or as a project, so I decided to rebuild LOOT to go through libloot instead of just using the core library directly. As the issue tracker shows, there was a lot that needed changing! The whole process revealed a lot that needed to be improved in the API, not just about exposing new data access, but logging, caching and providing functions that would allow efficient data access patterns.

Once LOOT was happily running on top of libloot, I was able to split the repository in two, so LOOT now lives in loot/loot and libloot lives in loot/loot-api. I preserved the full history in both repositories, as it was both simpler and more representative to do that. An unforseen side effect of that is that LOOT’s credits page counts my commits until that split twice, making me look much busier than I am!

With the repositories split, LOOT could depend on libloot as an “external project” in its CMake config, and download and build it from source. However, having the clear separation between libloot and LOOT showed that libloot was about twice as slow to build, and that most changes to LOOT didn’t need any libloot changes, so libloot was a huge bottleneck. To resolve this, I switched to building libloot DLLs in AppVeyor and publishing them to Bintray, so that LOOT could just download a prebuilt DLL and dynamically link to it.

Some of the benefits of all this work were:

libloot development has very little overhead and naturally stays in sync with LOOT, getting new features when LOOT gets them, if it makes sense for them to be a library-level feature.
Better separation of concerns, libloot and LOOT can focus on backend and GUI functionality respectively.
Much faster CI build times for LOOT, as it can now link to a prebuilt libloot DLL, avoiding having to recompile the backend whenever the GUI changes.
Because libloot bottlenecks LOOT builds less frequently, it can now afford to provide x64 builds in addition to x86 builds.
Because LOOT is now just another libloot client, there’s more incentive in making libloot actually pleasant to use, so it’s seen a lot more refinement.
The LOOT metadata validator was reimplemented on top of libloot, and as it must now use the same interfaces as LOOT, there’s more assurance that it’ll catch issues that LOOT would see.
libloot and LOOT can have separate release cycles, and by being careful about ABI compatibility in libloot, it’s often possible to just replace the DLL in an existing LOOT build. This means it’s now much simpler to test backend changes in isolation.

Now if only it were possible call C++ from other languages without having to write wrappers!