Incremental Compilation

Sept. 8, 2016 · Michael Woerister

I remember when, during the 1.0 anniversary presentation at the Bay Area Meetup, Aaron Turon talked about Dropbox so far having been pretty happy with using Rust in production there. The core team has been in touch with them regularly, he said, asking them, you know, what do you need? And their answer is always: faster compiles ... To the seasoned Rust user it is no surprise that this solicited a knowing chuckle from the audience. Improving compile times has actually been a major development focus after Rust reached 1.0 -- although, up to this point, much of the work towards this goal has gone into laying architectural foundations within the compiler and we are only slowly beginning to see actual results.

One of the projects that is building on these foundations, and that should help improve compile times a lot for typical workflows, is incremental compilation. Incremental compilation avoids redoing work when you recompile a crate, which will ultimately lead to a much faster edit-compile-debug cycle.

Today we are announcing an alpha version of incremental compilation, which marks an important milestone in the development of the feature: For the first time since implementation started towards the end of last year, all of the basic components are in place, the bulk of the groundwork has been done. You can give it a try in the nightly version of the compiler:

$ rustc -Zincremental=<path> ./main.rs

This will start the compiler in incremental mode, using whatever <path> you've provided as the incremental compilation cache directory. We are also working on a cargo subcommand to make this smoother, letting you just write cargo incremental build. Once things are working reliably, of course, incremental compilation will be available through the default cargo build command.

With that being said, incremental compilation is not production-ready yet: You might see crashes, you might see cases where there is no actual reduction in compile times and, most importantly, we still have to write extensive regression tests that make sure that incrementally compiled programs are always correct — so don't use it anywhere yet where it really matters. Over the next few weeks and months, however, our focus will be on making the implementation rock-solid from a correctness point of view and you will see continuous, gradual improvements in the feature's efficiency, up to a point where it will be transformative to your development experience.

This blog post will go through why and when incremental compilation is useful to begin with, how our implementation of it works, what its current development status is, and finally what's planned for the future and how you can contribute, should you be so inclined.

Why Incremental Compilation in the First Place?

Much of a programmer's time is spent in an edit-compile-debug workflow:

  • you make a small change (often in a single module or even function),
  • you let the compiler translate the code into a binary, and finally
  • you run the program or a bunch of unit tests in order to see results of the change.

After that it's back to step one, making the next small change informed by the knowledge gained in the previous iteration. This essential feedback loop is at the core of our daily work. We want the time being stalled while waiting for the compiler to produce an executable program to be as short as possible — ideally short enough as not to warrant a time-consuming and stress-inducing mental context switch: You want to be able to keep working, stay in the zone. After all, there is only so much regressive fun to be had while rustc bootstraps.

Incremental compilation is a way of exploiting the fact that little changes between compiles during the regular programming workflow: Many, if not most, of the changes done in between two compilation sessions only have local impact on the machine code in the output binary, while the rest of the program, same as at the source level, will end up exactly the same, bit for bit. Incremental compilation aims at retaining as much of those unchanged parts as possible while redoing only that amount of work that actually must be redone.

How Do You Make Something "Incremental"?

We have already heard that computing something incrementally means updating only those parts of the computation's output that need to be adapted in response to a given change in the computation's inputs. One basic strategy we can employ to achieve this is to view one big computation (like compiling a program) as a composite of many smaller, interrelated computations that build up on each other. Each of those smaller computations will yield an intermediate result that can be cached and hopefully re-used in a later iteration, sparing us the need to re-compute that particular intermediate result again.

Let's take a look at a simple example from algebra to make things more concrete. Let's see what it means to evaluate an expression of the form a+b×c incrementally. This will involve evaluating the expression once with one set of values for a, b, and c, and then evaluating it a second time with a different value for a. For the first time around, a will be 1, b will be 2, and c will be 3:

Initial Computation of a+b×c

Assume that we "saved" the intermediate results at each step, that is, we remember somewhere that b×c is 6 and a+b×c is 7. Now, in the second round, we want to know what a+b×c is if we change the value of a to 4. When we recompute the value of the expression, however, we see that we already know that b×c = 6, so we don't have to perform that computation again, and can rather skip directly to (a = 4) + (b×c = 6). We thus have computed the value of our expression in just one step instead of two, sparing us an entire, tedious multiplication.

Updating the Computation

Let's see how this scheme translates to the compiler.

An Incremental Compiler

The way we chose to implement incrementality in the Rust compiler is straightforward: An incremental compilation session follows exactly the same steps in the same order as a batch compilation session. However, when control flow reaches a point where it is about to compute some non-trivial intermediate result, it will try to load that result from the incremental compilation cache on disk instead. If there is a valid entry in the cache, the compiler can just skip computing that particular piece of data. Let's take a look at a (simplified) overview of the different compilation phases and the intermediate results they produce:

Compiler Phases and their By-Products

First the compiler will parse the source code into an abstract syntax tree (AST). The AST then goes through the analysis phase which produces type information and the MIR for each function. After that, if analysis did not find any errors, the codegen phase will transform the MIR version of the program into its machine code version, producing one object file per source-level module. In the last step all the object files get linked together into the final output binary which may be a library or an executable.

Comparing that with our algebra example from above, the pieces of AST correspond to a, b, and c, that is, they are the inputs to our incremental computation and they determine what needs to be updated as we make our way through the compilation process. The pieces of type information and MIR and the object files, on the other hand, are our intermediate results, that is, they correspond to the incremental computation cache entries we stored for b×c and a+b×c. Where a cache entry looks like b×c = 6 in the algebra example, it would look something like translate_to_obj_file(mir1, mir2, mir3) = <bits of the obj-file> in the case of the compiler.

So, this seems pretty simple so far: Instead of computing something a second time, just load the value from the cache. Things get tricky though when we need to find out if it's actually valid to use a value from the cache or if we have to re-compute it because of some changed input.

Dependency Graphs

There is a formal method that can be used to model a computation's intermediate results and their individual "up-to-dateness" in a straightforward way: dependency graphs. It looks like this: Each input and each intermediate result is represented as a node in a directed graph. The edges in the graph, on the other hand, represent which intermediate result or input can have an influence on some other intermediate result.

Let's go back to our algebra example to see what this looks like in practice:

Dependency Graph of a+b×c

As you can see, we have nodes for the inputs a, b, and c, and nodes for the intermediate results b×c and a+b×c. The edges should come as no surprise: There is one edge from b×c to b and one to c because those are the values we need to read when computing b×c. For a+b×c it's exactly the same. Note, by the way, that the above graph is a tree just because the computation it models has the form of a tree. In general, dependency graphs are directed acyclic graphs, as would be the case if we would add another intermediate result b×c+c to our computation:

Example of a non-tree Dependency Graph

What makes this data structure really useful is that we can ask it questions of the form "if X has changed, is Y still up-to-date?". We just take the node representing Y and collect all the inputs Y depends on by transitively following all edges originating from Y. If any of those inputs has changed, the value we have cached for Y cannot be relied on anymore.

Dependency Tracking in the Compiler

When compiling in incremental mode, we always build the dependency graph of the produced data: every time, some piece of data is written (like an object file), we record which other pieces of data we are accessing while doing so.

The emphasis is on recording here. At any point in time the compiler keeps track of which piece of data it is currently working on (it does so in the background in thread-local memory). This is the currently active node of the dependency graph. Conversely, the data that needs to be read to compute the value of the active node is also tracked: it usually already resides in some kind container (e.g. a hash table) that requires invoking a lookup method to access a specific entry. We make good use of this fact by making these lookup methods transparently create edges in the dependency graph: whenever an entry is accessed, we know that it is being read and we know what it is being read for (the currently active node). This gives us both ends of the dependency edge and we can simply add it to the graph. At the end of the compilation sessions we have all our data nicely linked up, mostly automatically:

Dependency Graph of Compilation Data

This dependency graph is then stored in the incremental compilation cache directory along with the cache entries it describes.

At the beginning of a subsequent compilation session, we detect which inputs (=AST nodes) have changed by comparing them to the previous version. Given the graph and the set of changed inputs, we can easily find all cache entries that are not up-to-date anymore and just remove them from the cache:

Using the Dependency Graph to Validate the Incremental Compilation Cache

Anything that has survived this cache validation phase can safely be re-used during the current compilation session.

There are a few benefits to the automated dependency tracking approach we are employing. Since it is built into the compiler's internal APIs, it will stay up-to-date with changes to the compiler, and it is hard to accidentally forget about. And if one still forgets using it correctly (e.g. by not declaring the correct active node in some place) then the result is an overly conservative, but still "correct" dependency graph: It will negatively impact the re-use ratio but it will not lead to incorrectly re-using some outdated piece of data.

Another aspect is that the system does not try to predict or compute what the dependency graph is going to look like, it just keeps track. A large part of our (yet to be written) regression tests, on the other hand, will give a description of what the dependency graph for a given program ought to look like. This makes sure that the actual graph and the reference graph are arrived at via different methods, reducing the risk that both the compiler and the test case agree on the same, yet wrong, value.

"Faster! Up to 15% or More."*

Let's take a look at some of the implications of what we've learned so far:

  • The dependency graph reflects the actual dependencies between parts of the source code and parts of the output binary.
  • If there is some input node that is reachable from many intermediate results, e.g. a central data type that is used in almost every function, then changing the definition of that data type will mean that everything has to be compiled from scratch, there's no way around it.

In other words, the effectiveness of incremental compilation is very sensitive to the structure of the program being compiled and the change being made. Changing a single character in the source code might very well invalidate the whole incremental compilation cache. Usually though, this kind of change is a rare case and most of the time only a small portion of the program has to be recompiled.

The Current Status of the Implementation

For the first spike implementation of incremental compilation, what we call the alpha version now, we chose to focus on caching object files. Why did we do that? Let's take a look at the compilation phases again and especially at how much time is spent in each one on average:

Relative Cost of Compilation Phases

As you can see, the Rust compiler spends most of its time in the optimization and codegen passes. Consequently, if this phase can be skipped at least for part of a code base, this is where the biggest impact on compile times can be achieved.

With that in mind, we can also give an upper bound on how much time this initial version of incremental compilation can save: If the compiler spends X seconds optimizing when compiling your crate, then incremental compilation will reduce compile times at most by those X seconds.

Another area that has a large influence on the actual effectiveness of the alpha version is dependency tracking granularity: It's up to us how fine-grained we make our dependency graphs, and the current implementation makes it rather coarse in places. For example, the dependency graph only knows a single node for all methods in an impl. As a consequence, the compiler will consider all methods of that impl as changed if just one of them is changed. This of course will mean that more code will be re-compiled than is strictly necessary.

Performance Numbers

Here are some numbers of how the current implementation fares in various situations. First let's take a look at the best case scenario where a 100% of a crate's object files can be re-used. This might occur when changing one crate in a multi-crate project and downstream crates need to be rebuilt but are not really affected.

Normalized Incremental Compilation Build Times

As you can see, compiling a crate for the first time in incremental mode can be slower than compiling it in non-incremental mode. This is because the dependency tracking incurs some additional cost when activated. Note that compiling incrementally can also be faster (as in the case of the regex crate). This is because incremental compilation splits the code into smaller optimization units than a regular compilation session, resulting in less time optimizing, but also in less efficient runtime code.

The last column shows the amount of time a rebuild of the crate takes when nothing has actually changed. For crates where the compiler spends a lot of time optimizing, like syntex-syntax or regex, the gain can be substantial: The incremental rebuild only takes about 22% of the time a full rebuild would need for syntex-syntax, only 16% for regex, and less than 10% for the all.rs test case of the futures-rs crate.

On the other hand, for a crate like the futures-rs library that results in little machine code when being compiled, the current version of incremental compilation makes little difference: It's only 3% faster than compiling from scratch.

The next graph shows which impact various changes made to the regex crate have on incremental rebuild times:

Build Times After Specific Changes

The numbers show that the effectiveness of incremental compilation indeed depends a lot on the type of change applied to the code. For changes with very local effect we can get close to optimal re-use (as in the case of Error::cause(), or dfa::write_vari32()). If we change something that has an impact on many places, like Compiler::new(), we might not see a noticeable reduction in compile times. But again, be aware that these measurements are from the current state of the implementation. They do not reflect the full potential of the feature.

Future Plans

The alpha version represents a minimal end-to-end implementation of incremental compilation for the Rust compiler, so there is lots of room for improvement. The section on the current status already laid out the two major axes along which we will pursue increased efficiency:

  • Cache more intermediate results, like MIR and type information, which will allow the compiler to skip more and more steps.

  • Make dependency tracking more precise, so that the compiler encounters fewer false positives during cache invalidation.

Improvements in both of these directions will make incremental compilation more effective as the implementation matures.

In terms of correctness, we tried to err on the side of caution from the get-go, rather making the compiler recompute something if we were not sure if our dependency tracking did the right thing, but there is still more that can be done.

  • We want to have many more auto-tests that make sure that various basic components of the system don't regress. This is an area where interested people can start contributing with relative ease, since one only needs to understand the Rust language and the test framework, but not the more complicated innards of the compiler's implementation. If you are interested in jumping in, head on over to the tracking issue on GitHub and leave a comment!

  • We are working on the cargo incremental tool (implemented as a Cargo subcommand for hassle-free installation and usage) that will walk a projects git history, compiling successive versions of the source code and collecting data on the efficiency and correctness of incremental versus regular compilation. If you're interested in helping out, consider yourself invited to either hack on the tool itself or downloading and running it on a project of yours. The #rustc channel on IRC is currently the best place to get further information regarding this.

Thanks to everyone in the community who has contributed directly or indirectly to incremental compilation so far! If you are interested in tackling a specific implementation problem, look for issues tagged with A-incr-comp or ask around in the #rustc channel on IRC.