Compiler · Sep 26, 2020

What's new in ReScript 8.3 (Part 2)

Hongbo Zhang
Compiler & Build System


ReScript is a soundly typed language with an optimizing compiler focused on the JS platform. It's focused on type safety, performance and JS interop. It used to be called BuckleScript.

ReScript@8.3 is now available for testing, you can try it via

npm i bs-platform@8.3.1

Following the previous post, in this post we will go through the enhancement over the build system.

Performance enhancement

The underlying build engine for bsb is ninja, it is famous for being fast to build large C++ repos.

In the last releases, we did lots of work for vertical integration into the bsb build chain. For example, we replaced the dynamic dependency parser with a minimal specialized one for bsb. we also removed the static dependencies parser which is only used for parsing C++ compiler output.

Thanks to various other low-level improvements, the final outcome is quite impressive. For example, The binary size for Mac platform is 270 KB while our vendored version is only 136KB. This is a non-trivial gain given that ninja is minimalist and already optimized by top-level C++ experts.

Note such vertical integration not only brings better performance, smaller sizes, it also brings new features

Build system enhancements for editor diagnostics

When people are coding in their favorite editors, they expect to see syntax and type errors in real-time. There are multiple ways to achieve this. The most reliable way is to always invoke the build system whenever the user saves a file. Due to not having an in-memory cache, our build system is very reliable. However we didn't yet optimize the build system for live feedback in editors. what syntax errors, type checking errors do they have when editing? There are multiple ways to achieve this, the most easy and reliable way is to always invoke the build system whenever the user saves the files, since it's the same build system without any in-memory cache, the reliability is very high, however, there's several challenges to use the build system output as editor diagnostics.

The build system/compiler has to be fast to deliver real-time feedback

Our build system is fast enough to deliver feedback for reasonable sized projects in less than 100ms. thanks to our previous hard work. We continue improving Pushing the limits of performance in the build system allows us to provide real-time feedback in editors.

The warnings for each file should not be flushed during a rebuild

For a typical file based build system, if the file A is compiled successfully with some warnings, the rebuild will not build A anymore. This is problematic if we use the build system output for editor diagnostics. Since the second build will not capture those warnings, we could use some caching mechanism to cache previous build output. But… stateful systems are not reliable and come with a whole range of different problems.

To solve such a challenging problem, we did some innovations to co-ordinate the compiler and build system. When the file A is compiled with warnings, the compiler will produce some marks to the build system, the build system will keep building but such marks are encoded in the build rules so that the second build will do the rebuild.

The benefit is two fold:

  • Rebuild will re-capture those warnings

  • Rebuild will be fast since only those files with warnings get rebuilt, it will not trigger unnecessary builds since our build is content-based build system.

The integration between compiler and build system is encoded in a specialized protocol. This makes it almost cost-free.

Notifying clients through .compiler.log

To get the build output, instead of communicating through IPC, we adopted a simple protocol. Whenever a build is done, we write the output to a file called .compiler.log. This makes the editor integration build-system agnostic, it does not need talk to the build system directly.

It also makes our build tool work with other watchers including Facebook's watchman. Watchman is a more scalable watcher tool for some specific platforms and less memory hungry, however, we still need a watchman-client to get the output of triggered job. We write the output to .compiler.log per each build, allowing clients to read compiler diagnostics when they want.

A better algorithm for removing stale outputs

Whenever we rename a file, e.g. a.res to b.res, will lead to the output of a.res being stale. Thanks to the deeper integration of the build system and compiler, we employ a more advanced strategy to remove stale outputs in this release. Pruning stale outputs is done in the beginning of each build.

There are two ways of removing staled artifacts, the second one is introduced in this release:

  • Based on live analysis and prebuilt-in knowledge

We scan lib/bs directory and check some dangling cm{i,t,j,ti} files, if it does not exist in the current build set, it is considered stale artifacts. If it is cmt file, it would trigger some hooks of genType, notably -cmt-rm.

  • Based on previous build logs We store previous compilation stats. If a file is in the previous compiler output, but no longer in the output of the new build, it is considered stale and can be removed. it is considered stale output which can be removed.

In general, strategy two is more reliable and efficient.

  • We don't need to recompute the path since it is already done by the build system.

  • When we change the in-source build to out-source build, it will still do the pruning properly

However, strategy one is easier for tooling like genType. Not every tool has knowledge of the build system.

Sometimes a combination of both strategies is needed.

  • When removing .cm* files, we use the first strategy.

  • When removing generated javascript, we use strategy two,

Happy Hacking!

Want to read more?
Back to Overview