Speeding Up the Rust Compiler
blog.mozilla.org
blog.mozilla.org
Admittedly part of the blame falls on me, I'm a big fan of printf-debugging and I tend to only use debuggers as a last resort.
All that said, I totally agree that there are times when faster compiles are really useful.
That being said, if I had to defend myself, I'd point out that printf debugging has a few advantages:
- It's often more lightweight and less intrusive than debuggers, which makes it less likely to encounter an "heisengbug" that disappears when you attempt to debug it. This is especially true for timing-sensitive bugs (which occur in multithreaded code as you point out, but not only).
- Debuggers are often environment-specific, if you change language, environment or even simply editor you might have to re-learn how to use your debugger. Printf will always be here for you.
- I do a lot of work in embedded environments, including low level and bare metal stuff (bootloaders, drivers etc...). While these days there's generally some debugger support available for these targets it's often more limited or much more intrusive. If I put a breakpoint in an interrupt handler I basically freeze the entire kernel when it triggers which can sometimes do more harm than good if I'm trying to figure out what's going on. And again, in these environments the debugging solutions are often proprietary and sometimes quite expensive.
It is true though that if there are bugs in the toolchain, it affects me a lot more. For my personal projects, I actually avoid windows and "fancy" GUI based tools. I also painstakingly write all my code so they either compile fast or can hot-reload.
Though for webassembly I need the generated wasm to exist and be loaded before I can really see what's happening :\
Thank you for the hard work here!
Btw, mentioned in the article are these tools: “ All of the above improvements (and most of the ones in my previous posts) I found by profiling with Cachegrind, Callgrind, DHAT, and counts, but there are plenty of other profilers out there.” Does anyone have any good resources on using these with Rust, or just in general with C or C++?
Go has made quite a few language concessions to be fast to compile and is designed with that in mind. And it really shows, using auto reloading in Go feels like using Ruby or Python which is just great.
From my usage standpoint Go has basically no compile time at all.
And awful compile times can be a huge hindrance to productivity. I used to do web dev in Scala, but waiting for the sleepy compiler is one of the reasons is switched to Go.
Scala is a nice language but the long compile times in combination with the vm/jetty cycle feels quite a bit slower than Rust.
Main idea: Essentially, compile only what has changed, re-use compiled forms of everything that hasn't.
Essentially, pre-compiled dependencies will be sourced and plugged in to a Rust project. This presents great security concerns, but also great opportunities to advance software development. A large system has a vast network of crate dependencies. Certain versions of these crates will be pre-compiled, each uniquely identifiable through a hash. Some of these crates will even be audited and potentially certified. Hundreds of black box, pre-compiled crates will each be sandboxed in a very secure fashion such that it does only as specified and no more. No sandboxed library can reach beyond its advertised behavior. Unchanged, locally-developed parts can also be compiled and used through this flow as well. WASM projects are facilitating much of this work.
I may be missing important parts from this explanation, so hopefully it is corrected by someone more knowledgeable.
That's what incremental compiling does, and it has been enabled for a while. Unfortunately, it doesn't help on fresh builds.
> Essentially, pre-compiled dependencies will be sourced and plugged in to a Rust project.
There are two aspects here. The more recent one, which you are thinking of, is building and shipping procedural macros as WASM binaries. Procedural macros are very powerful, but they take a token stream as input to allow for future changes to the language. Because of this, basically every macro uses a Rust parser crate called syn. Since it's a fully-fledged parser, syn takes a while to compile (30 seconds or so, certainly not minutes), which many people find annoying, and it gets worse if you end up using different versions of syn for different proc macro crates. The plan here is to build the proc macros (including syn) somewhere on the Rust infrastructure and ship the binaries to the users. The sandboxing story is somewhat complicated: proc macros and build scripts can legitimately read or download files, or produce non-deterministic output in other ways. A WASM macro would not be able to do this, so the whole thing would be opt-in. It also provides no benefit for crates that don't use procedural macros. See https://github.com/dtolnay/watt for more details.
The other possible avenue is MIR-only rlibs. When you compile a dependency, you get machine code (with the exception of generic code). It might be possible to compile crates to MIR (again, on the Rust infrastructure) and only do the final codegen on the user's computer. But that's still complex, and not necessarily much faster. See https://github.com/rust-lang/rust/issues/38913.
While this is true, the compiler is not yet fully incremental in my understanding, there's still a lot left to do here.
I see you've never imported the Kubernetes client library.
If the compiler catches bugs that would otherwise only be found at run time, then the additional compile time pays for itself many times over in terms of productivity.
What they have going for them is not depending on LLVM.
The alternative cranelift backend here has been making slow but steady progress, and seems to be approaching a useable state: https://github.com/bjorn3/rustc_codegen_cranelift/issues
Read, IR1, IR2, Assembly.
The existence of LLVM IR is not the problem- rather, it's how it gets used (on both sides of the API). Generating a lot of naive IR and letting the optimizer clean it up, for example, has a large cost.
And, while I'm not too up to date on the details of Jai, the last thing I heard it was still very fast to compile even in release builds that did use LLVM. That is, the Jai compiler is smarter about how it generates IR.
Well, even Turbo Pascal for Windows v1.5 (last TP version before Delphi happened) is more feature rich than Go. :)
[1] http://docwiki.embarcadero.com/RADStudio/Rio/en/LLVM-based_D...
The only thing missing being lifetimes.
F# might not win marathons, it is still faster than rustc, even when adding NGEN or Mono AOT into the pipeline.
Not really...F# has a "slow" compiler too.
Any advanced type system, whether F#, C++, Haskell, Scala, Rust, TypeScript...is inevitably going to have a significantly slower compiler than a more basic system (Go, Java, C).
That's just part of the tradeoff; though obviously you can optimize within those bounds.
It's about as universal as the runtime "rule": compiled perf > interpreted perf. Not technically inviolable, but practically so.
---
P.S. And while the F# compiler is indeed faster than rustc, it's not apples-to-apples. F# compiles to CLR; Rust compiles through LLVM all the way to native. If Rust deferred work to runtime with an LLVM interpreter (http://llvm.org/docs/CommandGuide/lli.html ?), it would improve compile-time performance, at the cost of runtime overhead.
A SASS (CSS preprocessor language) compiler can be very fast because it does very little.
Native compilers must do significantly more transformation than other compilers, e.g. F# CLR compiler. This is doubly so if you request optimized output (but that's probably not the case in this discussion).
Naturally Rust could offer an interpreter of some sort, however it still isn't there today, so we got to use what is available.
I'm not sure I understand the problem with rust build times. In Ada, at least with GNAT, a full rebuild of 200KLOC can be a bit long, depending on your use of generics, and your number of cores (thanks to AMD, build times will soon be ridiculously short). But next builds with slight modifications are quite fast, thanks to modular compilation (every module built independently, same as for C, thanks to spec/body separation if you just change the body of a module you just recompile the specific module) and incremental compilation (only rebuild what changed and their dependencies).
Is there something inherently slow in the Rust compiler that disallows those ?
I know that writing an Ada compiler that could compile units independently seemed impossible to everyone at first, until (the late, sadly) Robert Dewar worked it out with RMS : https://news.ycombinator.com/item?id=15880160
I don't know Rust enough, just that what they're doing is amazing, and I hope they're not too focused on the small scale optimizations (which are great!) and look at the high level optimizations too, and especially to what's been done elsewhere through sweat and pain.
Most people might also not be aware that Rational Software started their business by selling Ada machines, where one could enjoy an experience somehow similar to Lisp Machines, just with Ada instead.
Thanks for the link.
Agree a lot with that. I write quite a bit of code in Rust and I use the JetBrains CLion IDE with the IdeaVim and Rust plugins.
CLion is very helpful when working with Rust code. It understands the language quite well and because of that it can help you by pointing out things that aren't going to work without having to constantly recompile your code.
For students and faculty members, JetBrains give out licenses free of charge that are valid for 1 year, and which can be renewed while you are still a student or faculty member. https://www.jetbrains.com/student/
If you write a lot of Rust, I suggest that you download the 30 day evaluation version of CLion and install the Rust plugin and see how it compares.
Not necessarily true. The d compiler runs incredibly fast, for compiling the type of code you'd write in go; and it only slows down if you use a lot of complicated features like templates or CTFE.
522 results in servo. Granted, some (but not all) are in tests.
`.unwrap()` is expected to be in every code base out there, as there is rarely much point in handling errors such as tainted mutex locks.
if foo.is_some() {
let unwrapped = foo.unwrap();
}
if let Some(unwrapped) = foo {
}
pub fn new(mut streams: Vec<TokenStream>) -> TokenStream {
match streams.len() {
0 => TokenStream::empty(),
1 => streams.pop().unwrap(),
_ => TokenStream::Stream(Lrc::new(streams)),
}
}
But I agree, most of the time using `match` (or its simplified form `if let`) or one of these combinators instead of `unwrap()` is better.
The message helps a bit more, and I also want to avoid debug unwraps that were forgotten about.
Though, the more I think about it the more a macro makes sense to me, purely for code search. If we used a macro, something like `impossible!(streams.pop())` then I can code search for `.unwrap` and `.expect`.
Hmm
IIUC, panicking via unwrap() or expect() IS graceful error handling. It unwinds the stack and releases resources. We should be clearer because it's confusing to people less familiar with Rust who will confuse this with actually crashing and immediately terminating.
Personally, I think if there's nothing you can actually do to handle this error then unwrap() or expect() are the correct choices. There's no point writing more code just to bubble an error that you can't handle.
This is often poor advice, because continuing in a corrupt state can often be worse (and harder to debug) than a clean panic. E.g. it's better to crash than to overwrite a save file with corrupt data, or transfer money to the wrong account, or show one user data belonging to another user.
Let's say you have CLI tool - instead of message "XX is a directory - file expected" your tool will just fail silently.
There are much more examples. Graceful shutdown is what we have to do, even laziest of us.
.expect(format!("{} is a directory - file expected")) gives exactly the behaviour you've described for a CLI tool.
Graceful shutdown means releasing locks, deleting PID files etc when encountering an unrecoverable error. Not just any error.
When language allows you to make all errors "expected", it's really stupid to don't use it. Only reason is laziness.
Many languages don't have Result types to return so there you can't do much sometimes to prevent data corruption or something worse. But in languages where you can return more than just "false", it's not acceptable.
AFAIK, only with the default panic=unwind mode. With the panic=abort mode, it aborts the whole process instead of unwinding.
This isn't the right advice to give, and is only going to confuse beginners. Panicking via unwrap/expect is perfectly fine, whether in a library or an application, only when the panic represents a bug. For example, accessing an element of a Vec with an incorrect index because of a logic error. If a panic occurs, then it should reflect a bug that is intended to be fixed.
Stated differently, it should be impossible for an end user to cause a Rust application to panic. If they can, then it's a bug.
This advice permits use of unwrap/expect anywhere, so long as its occurrence is never expected.
I’m not sure that was the best example but I agree with your overall point.
Also, Clippy's defaults are developed to be quite opinionated. It makes many suggestions that I (and many others) disagree with. So Clippy's defaults should not be used as a bludgeon for what's idiomatic and correct.
As an example it’s quite easy to fall astray of taking your perspective to say read a packet off the network, and panic on malformed data. I know you’d agree that would be a inappropriate time to use panic, as that would crash any upstream program with trivial DOS exploits. But this is easy to do if you say have an API that takes a value, translates it to something else and panics on invalid inputs.
My advice is generally in line with the GP, pornel—use the tools at hand for avoiding panics where it’s easy to do so, such as ?, map, etc. Only panic/unwrap on bad library API usage (which I think is what you’re suggesting). Feel free to panic anywhere in main.
People often translate “ use of unwrap/expect anywhere, so long as its occurrence is never expected” in ways that can create major bugs as the software goes into more use.
> but in practice I think using unwrap or panicking in libraries is usually wrong
Very strongly disagree. I've created dozens of Rust libraries, and probably all of them have dozens of code paths that can panic. Other core libraries do the same. Just a simple slice access, e.g., `slice[i]`, is a line of code that could panic. (Since it's just a shorter way of writing `slice.get(i).unwrap()`.) I don't think we should be giving advice that runs contrary to how important libraries actually work. A lot of people learn to code by reading others' code, and when we give advice like "don't use unwrap/expect," they get confused when they see that virtually every widely used piece of Rust code violates it.
The key here is really that one shouldn't panic unless the panic itself is indicative of a bug somewhere. Blanket advice like "don't use unwrap/panic in libraries" is bad because---as I argued in the Clippy issue I posted in a sibling comment---it's effectively a prohibition against runtime invariants themselves. As a programmer, you insert a panic when you've failed (for any number of reasons) to capture the invariant in the type system. Blanket advice saying that one shouldn't use unwrap/expect in these circumstances leads one toward a path of much more complex APIs with error types that are never constructed unless there is a bug in the code. (It's likely that actually adding all of those error types will be so annoying that it's plausible the programmer will give up on Rust.)
> As an example it’s quite easy to fall astray of taking your perspective to say read a packet off the network, and panic on malformed data. I know you’d agree that would be a inappropriate time to use panic, as that would crash any upstream program with trivial DOS exploits. But this is easy to do if you say have an API that takes a value, translates it to something else and panics on invalid inputs.
I think this is a reflection of error handling being hard. It's just as easy to code a program with the mistaken assumption that a particular file path will always point to a valid and readable file. It takes a bit of learning to understand which things you can rely on never happening and which you can't. In the mean time, I don't think we should be giving advice that is very easy to misinterpret into a conclusion that just isn't tenable. Personally, I think it's much easier to talk about end user behavior. If there's a panic, then it's a bug that ought to be fixed. If you follow that, it's hard to go wrong.
You could rephrase it with, "If there's a reachable code path that panics, then the code path should be removed." You could then talk about what "reachable" means, i.e., code paths that are determine based on data that the program doesn't control vs code paths that are enforced for all inputs to the program.
You’ve said don’t panic for code paths that the program doesn’t control. That is nuanced, and often is not always obvious and easy to reason about, but the language makes those exceptional cases obvious, so why not use that to your advantage?
edit: and by the way, I didn't say that on indexes, I totally agree with you. Though, sometimes it is valuable to go to the extreme and prevent malicious index values from untrusted sources crashing your program...
I think we are disagreeing mostly over pedagogy.
> That is nuanced, and often is not always obvious and easy to reason about
Right. I think this is what I meant by "error handling is hard." But this really boils down to understanding what a runtime invariant is, and that takes time to learn. It's hard to do good error handling without internalizing that.
> but the language makes those exceptional cases obvious, so why not use that to your advantage?
I think you are, specifically by using unwrap/expect. In many such cases, the unwrap/expect has a comment explaining why it's impossible to panic.
> edit: and by the way, I didn't say that on indexes, I totally agree with you. Though, sometimes it is valuable to go to the extreme and prevent malicious index values from untrusted sources crashing your program...
Well, `&slice[i]` and `slice.get(i).unwrap()` are equivalent. So if we give blanket advice like "don't use unwrap/expect in libraries" then the latter gets caught up in that advice while the former doesn't, even though they are exactly equivalent. To me, this reveals the problem in that pedagogy because it focuses too much on one particularly common manifestation rather than the thing that actually matters: a panic visible to an end user is always a bug.
To be clear, the thing I am disagreeing with is the advice to "not use unwrap/expect; use case analysis instead." I can appreciate a pedagogy that simplifies things upfront with the cost of getting some corner cases wrong. But unwrap/expect are used too much in too many valid cases IMO for that type of strategy to be effective.
Because the problem that I'm seeing in Swift code (which also has a "crash on logic errors" approach) is that bugs can bring your whole system down, which is especially bad on server-side apps with multiple threads. Yes, you can use supervisord and/or load balancers, but you still lose in-flight requests.
By contrast, in a language with a runtime like Java or Ruby, you can catch almost everything at the top level, so you could just have some logic that generates a "whoops something went wrong" response in case of a serious error.
The point here is resiliency. I agree that you want to catch bugs early, and you should immediately abort execution once bad things happen; I also fundamentally agree that you don't want to litter your code with error types or other constructs that are supposed to never actually be used. But faults do happen in practice, because we write buggy code, and in such a case, it's good if you can isolate the fault and recover at a higher level. Languages like Erlang take this idea to an extreme.
This is currently a real problem for us with Swift, so I was wondering whether Rust has a better solution here.
Yes: https://doc.rust-lang.org/std/panic/fn.catch_unwind.html
However, this comes with a critical caveat: unwinding can be disabled when compiling an application, which means one cannot guarantee that catch_unwind will actually work. (If unwinding is disabled, then panics turn into unrecoverable aborts.)
For this reason, and because the standard error handling mechanism is done through return values, panicking is not an acceptable way to do robust error handling in Rust. Recovering from panics is useful in niche scenarios, like keeping a web server running even if a request causes a panic or in tests for ensuring that all tests run even if one panics. Which basically matches your key concern here.
(To be clear, I think this is mostly orthogonal to my original comment in this thread. :-))
But in general, exception handling is hard. There is value in having locally-unrecoverable, but globally-recoverable faults, stuff like that is probably where languages like Erlang shine.
For off by 1 errors, in Rust it’s often better to turn to iterators where appropriate than to say using indexes (it’s also more efficient in most cases). Clippy can also catch those issues.
> For off by 1 errors, in Rust it’s often better to turn to iterators where appropriate than to say using indexes (it’s also more efficient in most cases).
The off-by-one errors I made weren't related to indexing collections. I don't remember anymore what it was exactly, it got caught by the very first tests I wrote before even trying to use anything. I do use iterators whenever they make sense and I write new iterators whenever it makes sense. Still, good note from you.
I’ve found that it is helpful in avoiding those and additional directing you to more performant choices. YMMV.
So basically your code is not even a little bug-free, it just probably doesn't have a certain small class of bug in it.
It won't likely crash, beyond that: all bets are off.
I've now encountered several pieces of rust software that have no error / exceptional case handling and just panic at things like unexpected command-line arguments.
I wouldn't go so far as to argue that it is at this point. But I think rust advocates probably ought to stop arguing as though it's axiomatically true. Maybe Mozilla could be talked into funding an academic study comparing defect rates in rust vs other languages used for systems programming.
Beyond substantiating the rust-improves-reliability trope, it could also identify areas for improvement where it doesn't.
Of course, once you add / change logic, I agree with your assessment that the statement is hopelessly naïve in most practical use cases.
I've translated pure-logic business code (with fairly thorough, albeit static, tests that all passed) from Python to Rust, and the compiler's complained about subtle errors that would have taken down a business using the code in days had they been put into production. Turns out there was a flaw in the spec.
Leveraging Rust's type system to logically separate the meanings of different integers, and only permitting arithmetic operations where meaningful (e.g. no distance+time, distance/time=speed), also uncovered a couple of even more subtle bugs – though luckily assertions would've caught these in the event of them actually making it through to production.
If it compiles, it probably works. But, more importantly, Rust forces you to write code that works in the first place. You have to think through the ramifications of what you're doing, sometimes to hold the entire function in your head at once, and that means you have to understand what you're doing. If you don't, it complains, and the chances are you won't be able to get it to compile until you do understand it.
And if it does compile when it's wrong, it's obviously wrong and your tests will fail. Most of the time.
Isn't programming fun?
Basically, if you've programmed in unsafe languages such as C, C++, Assembly or any language that has `null`, you know what Rust brings to the table. The Rust compiler does prevent very large classes of bugs all the while being non garbage collected. That's quite a feat!
And I'm far from being a Rust fanboy. I work with PHP, Javascript and devops tools in my day job. I'm not doing PR here. Just sharing my experience. Feel free to share yours.
Does this imply that existing benchmarks weren’t catching the problem here because they were avoiding making use of a feature because it was too slow? That seems like a strange way to write a benchmark, especially if tokens were actually in common use in the compiler itself.
The bit about benchmarks using tokens sounds to me like he's talking about harness code, rather than the test subject code. You can speed up running the benchmarks without necessarily affecting the benchmark results themselves.
Benchmark harnesses are real programs too.
a debug compile takes about 5 minutes from clean. Release takes 18 minutes. (1.38.0)
There is a compiler flag you can use to list where time is being spent and it showed all my time was spent dealing with those trait bounds.
There is an open issue to make that not necessary.
If your time is spent linking, you can swap the lld linker in
Performance improvements in this post are from november'19 to december'19 - IMHO the november version _already_ had impressive optimizations.
Maybe they are working on it in parallel.
Rust was partly created because most of the transistors on the computers are heavily underutilized, and multiprocessing with C++ correctly is extremely hard.
Rust compiler is written in Rust, so it would be a perfect showcase of taking advantage of the multi-processing safety of the language.
I know that there are global variables in the compiler that the compiler team is getting rid of, but at this point that should be the main focus, as I see that most of the easy huge gains of single-threaded improvements are over.
https://www.youtube.com/watch?v=Wh20eXfMOSk&t=8s
I found the meeting notes: