genuine question: how much can the frontend really impact compile times at the end of the day? I would guess most of the time spent compiling is on the backend, but IANA compiler engineer
One obvious example of this would be C++, where a smarter frontend that doesn't do textual inclusion of a million lines would significantly improve compile times.
But that's low-hanging fruit "optimization", no? Once you get around it, and this has been forever, the bottleneck is in the backend generation. So if Rust has already solved this, the area where they can improve most is the backend generation?
Most C++ builds I have worked with, and all of them being MMLoC codebases, were actually bottlenecked by the high memory pressure. In part due to heavy templates and symbol visibility.
I think the history of C++ implementations shows that it's not low hanging fruit, it's a huge effort to implement and the payoffs aren't game changing.
- Serial took 10.2 seconds in the frontend and 6.2 seconds in the backend.
- Parallel tool 5.9 seconds (-42%) and 5.3 seconds (-14.5%).
So to answer your question, parallelising the frontend can have a substantial impact.
You are right that frontend is only a subset of the total compilation time - backend and linking time matter too. Happily, that's something that's being worked on too!
These graphs only include stable releases, not nightly ones. Since the multi-threaded frontend hasn't landed on stable yet, we can't see it's effect on these graphs.
The Scala.js compiler toolchain has been using similar techniques since 2015 by default. It features parallel and incremental whole-program optimizations, so it is also used for release builds. In fact it's so fast we run most advanced optimizations in development builds as well.
For those interested in more details, we published a paper about it in OOPSLA'16 [1].
This [1] is a (somewhat outdated!) design doc for how we're approaching incremental compilation in the Isograph compiler! If anyone if interested in helping build an incremental compiler, the project is being actively developed and open to new contributors! See also my recent talk at GraphQL conf [2]
In particular, we're actively implementing a simpler Salsa [3] and transform the currently-batch-mode compiler into an incremental, pull compiler.
I’m interested to see how things progress! rustc itself is still reliant on llvm and ld, but there’s some intriguing alternatives coming in cranelift and wild both written in rust themselves, and I think done in a way that would allow incremental compilation and linking to be done in a way that better suits rust.
23 comments
[ 2.7 ms ] story [ 77.5 ms ] threadBuild systems.
Now compiler passes.
Container layers.
Also in ML systems, eg dvc.
Where else is it missing?
(Although as the language complexity continues to grow maybe that question is missing the opposite force of compiler slowdowns from added work?)
Most C++ builds I have worked with, and all of them being MMLoC codebases, were actually bottlenecked by the high memory pressure. In part due to heavy templates and symbol visibility.
Your question is addressed in this blog post - Faster compilation with the parallel front-end in nightly (https://blog.rust-lang.org/2023/11/09/parallel-rustc.html). One the example shown there:
- Serial took 10.2 seconds in the frontend and 6.2 seconds in the backend.
- Parallel tool 5.9 seconds (-42%) and 5.3 seconds (-14.5%).
So to answer your question, parallelising the frontend can have a substantial impact.
You are right that frontend is only a subset of the total compilation time - backend and linking time matter too. Happily, that's something that's being worked on too!
There is an alternate backend called cranelift (https://bjorn3.github.io/2024/11/14/progress-report-nov-2024...) that compiles faster than LLVM, at the cost of producing slower binaries. That's a good trade-off for debug builds.
And there's work on changing the default linker to lld (https://blog.rust-lang.org/2024/05/17/enabling-rust-lld-on-l...).
We may see all 3 efforts land in the next year :)
- https://internals.rust-lang.org/t/slow-incremental-compilati...
- https://internals.rust-lang.org/t/feedback-request-performan... (there is more context to that being published today in a blog post)
There are several more ideas that are in even earlier stages of discussion.
For those interested in more details, we published a paper about it in OOPSLA'16 [1].
[1] https://dl.acm.org/doi/10.1145/2983990.2984013
We don't share the query infrastructure in rust-analyzer, but rust-analyzer is a query-driven system. We use Salsa (https://github.com/salsa-rs/salsa/), which is strongly inspired by rustc's query system, down to the underlying red/green algorithm (https://rustc-dev-guide.rust-lang.org/queries/incremental-co...), hence the name: Salsa.
In particular, we're actively implementing a simpler Salsa [3] and transform the currently-batch-mode compiler into an incremental, pull compiler.
[1] https://isograph.dev/docs/design-docs/incremental-compilatio...
[2] https://www.youtube.com/watch?v=sf8ac2NtwPY
[3] https://www.youtube.com/watch?v=_muY4HjSqVw