Someone mentioned recently that the slowness of rustc is in large part due to llvm. I know that is probably orthogonal to the work here, but I do like the idea of building the compiler with different toolchains, and that there may be follow on effects down the line.
Depends on the workload, but yes, codegen is a huge part of the total compilation time.
That said, that doesn't mean LLVM is always where the fixes need to be. For instance, one reason rustc spends a lot of time in LLVM is that rustc feeds more code to LLVM than it should, and relies on the LLVM optimizer to improve it. Over time, we're getting better about how much code we throw at LLVM, and that's providing performance improvements.
I'm completely ignorant so forgive me if this is obvious: in the effort of the parent article - to compile rustc with gcc - will rustc still be feeding lots of code to LLVM, or would that code now be fed to gcc?
Which codegen backend the building compiler uses is independent of which codegen backend(s) the built compiler uses.
Similarly, you can build Clang using itself or using GCC. The resulting compiler should behave the same and produce the same machine code, even if its own machine code is somewhat different.
The produced binaries could still have artifacts from the original compiler in them, e.g. if "compiler built-in" libraries or standard libraries were compiled with the original compiler.
Both GCC and rustc use a multi-stage build process where the new compiler builds itself again, so you reach an idempotent state where no artifacts from the original compiler are left.
There's an experimental compiler backend [1] using cranelift [1] that's supposed to improve debug build times. I never see it mentioned often in threads about Rust's long compilation time so I'm not sure if I'm missing something.
It's slow because the borrow checker is NP complete. LLVM may or may not generate slower code than GCC would for rustc, but I doubt it's anywhere close to the primary cause of the lack of snappy.
You're wrong it's been debunked that the borrow checker is any appreciable part of the compile time - Steve Klabnik actually verified it on here somewhere.
I say “usually” because of course sometimes bugs happen and of course you can conduct degenerate stress tests. But outside of those edge cases, it’s not an issue. If it were, blog posts that talk about lowering compile times would be discussing avoiding the borrow checker to get better times, but they never do. It’s always other things.
Is there any tool for Rust that does profiling that detects what part of compilation time is caused by what? Like, a tool that reports:
- Parsing: x ms
- Type checking: y ms
- LLVM IR generation: z ms
And have there been any statistics done on that across open-source projects, like mean, median, percentiles and so on?
I am asking because it should depend a lot on each project what is costly in compile time, making it more difficult to analyse. And I am also curious about how many projects are covered by "edge cases", if it is 1%, 0.1%, 0.01%, and so on.
The post my original comment is on discusses doing this at length.
> And have there been any statistics done on that across open-source projects, like mean, median, percentiles and so on?
I am not aware of any. But in all the posts on this topic over the years, codegen always ends up being half the time. It’s why cargo check is built the way it is, and why it’s always faster than a full build. If non-codegen factors were significant with any regularity, you’d be seeing reports of check being super slow compared to build.
I have actually seen a few posts here and there of 'cargo check' being slow. I have also heard of complaints of rust-analyzer being slow, though rust-analyzer may be doing more than just 'cargo check'.
>Triage notes (AFAIUI): #132625 is merged, but the compile time is not fully clawed back as #132625 is a compromise between (full) soundness and performance in favor of a full revert (full revert would bring back more soundness problems AFAICT)
Update: They fixed it, but another issue surfaced, possibly related, as I read the comments.
I do not know if the borrow checker and Rust's type system have been formalized. There are stacked borrows and tree borrows, and other languages experimenting with features similar to borrow checking, but without formal algorithms like Algorithm W or J for Hindley-Milner, or formalizations like of Hindley-Milner and problems like typability for them, I am not sure how one can prove the complexity class of the borrow checking problem nor a specific algorithm, like https://link.springer.com/chapter/10.1007/3-540-52590-4_50 does for ML.
I also am not sure how to figure out if the complexity class of some kind of borrow checking has something to do with the exponential compile times of some practical Rust projects after they upgraded compiler version, for instance in https://github.com/rust-lang/rust/issues/75992 .
It would be good if there was a formal description of at least one borrow checking algorithm as well as the borrow checking "problem", and maybe also analysis of the complexity class of the problem.
There isn't a formal definition of how the borrow checking algorithm works, but if anyone is interested, [0] is a fairly detailed if not mathematically rigorous description of how the current non-lexical lifetime algorithm works.
The upcoming Polonius borrow checking algorithm was prototyped using Datalog, which is a logical programming language. So the source code of the prototype [1] effectively is a formal definition. However, I don't think that the version which is in the compiler now exactly matches this early prototype.
EDIT: to be clear, there is a polonius implementation in the rust compiler, but you need to use '-Zpolonius=next' flag on a nightly rust compiler to access it.
Interesting. The change to sets of loans is interesting. Datalog, related to Prolog, is not a language family I have a lot of experience with, only a smidgen. They use some kind of solving as I recall, and are great at certain types of problems and explorative programming. Analyzing the performance of them is not always easy, but they are also often used for problems that already are exponential.
>I recommend watching the video @nerditation linked. I believe Amanda mentioned somewhere that Polonius is 5000x slower than the existing borrow-checker; IIRC the plan isn't to use Polonius instead of NLL, but rather use NLL and kick off Polonius for certain failure cases.
That slowdown might be temporary, as it is optimized over time, if I had to guess, since otherwise there might then be two solvers in compilers for Rust. It would be line with some other languages if the worst-case complexity class is something exponential.
> IRC the plan isn't to use Polonius instead of NLL, but rather use NLL and kick off Polonius for certain failure cases.
Indeed. Based on the last comment on the tracking issue [0], it looks like they have not figured out whether they will be able to optimize Polonius enough before stabilization, or if they will try non-lexical lifetimes first.
Someone mentioned recently that the slowness of rustc is in large part due to llvm. I know that is probably orthogonal to the work here, but I do like the idea of building the compiler with different toolchains, and that there may be follow on effects down the line.