In short, the maximum possible speed is the same (+/- some nitpicks), but there can be significant differences in typical code, and it's hard to define what's a realistic typical example.
The big one is multi-threading. In Rust, whether you use threads or not, all globals must be thread-safe, and the borrow checker requires memory access to be shared XOR mutable. When writing single-threaded code takes 90% of effort of writing multi-threaded one, Rust programmers may as well sprinkle threads all over the place regardless whether that's a 16x improvement or 1.5x improvement. In C, the cost/benefit analysis is different. Even just spawning a thread is going to make somebody complain that they can't build the code on their platform due to C11/pthread/openmp. Risk of having to debug heisenbugs means that code typically won't be made multi-threaded unless really necessary, and even then preferably kept to simple cases or very coarse-grained splits.
To be honest, I think a lot of the justification here is just a difference in standard library and ease of use.
I wouldn't consider there to be any notable effort in making thread build on target platforms in C relative to normal effort levels in C, but it's objectively more work than `std::thread::spawn(move || { ... });`.
Despite benefits, I don't actually think the memory safety really plays a role in the usage rate of parallelism. Case in point, Go has no implicit memory safety with both races and atomicity issues being easy to make, and yet relies much heavier on concurrency (with a parallelism degree managed by the runtime) with much less consideration than Rust. After all, `go f()` is even easier.
(As a personal anecdote, I've probably run into more concurrency-related heisenbugs in Go than I ever did in C, with C heisenbugs more commonly being memory mismanagement in single-threaded code with complex object lifetimes/ownership structures...)
Yes. All `&mut` references in Rust are equivalent to C's `restrict` qualified pointers. In the past I measured a ~15% real world performance improvement in one of my projects due to this (rustc has/had a flag where you can turn this on/off; it was disabled by default for quite some time due to codegen bugs in LLVM).
I was confused by this at first since `&T` clearly allows aliasing (which is what C's `restrict` is about). But I realize that Steve meant just the optimization opportunity: you can be guaranteed that (in the absence of UB), the data behind the `&T` can be known to not change in the absence of a contained `UnsafeCell<T>`, so you don't have to reload it after mutations through other pointers.
what about generics (equivalent to templates in C++), which allow compile time optimizations all the way down which may not possible if the implementation is hidden behind a void*?
Aliasing info is gold dust to a compiler in various situations although the absence of it in the past can mean that they start smoking crack when it's provided.
Yes. Specifically since Rust's design prevents shared mutablity, if you have 2 mutable data-structures you know that they don't alias which makes auto vectorization a whole lot easier.
I'm still confused as to why linux requires linking against TBB for multithreading, thus breaking cmake configs without if(linux) for tbb. That stuff should be included by default without any effort by the developer.
Multithreading does not make code more efficient. It still takes the same amount of work and power (slightly more).
On a backend system where you already have multiple processes using various cores (databases, web servers, etc) it usually doesn’t make sense as a performance tool.
And on an embedded device you want to save power so it also rarely makes sense.
Multithreading can made an application more responsive and more performant to the end user. If multithreading causes an end user to have to wait less, the code is more performant.
I think personally the answer is "basically no", Rust, C and C++ are all the same kind of low-level languages with the same kind of compiler backends and optimizations, any performance thing you could do in one you can basically do in the other two.
However, in the spirit of the question: someone mentioned the stricter aliasing rules, that one does come to mind on Rust's side over C/C++. On the other hand, signed integer overflow being UB would count for C/C++ (in general: all the UB in C/C++ not present in Rust is there for performance reasons).
Another thing I thought of in Rust and C++s favor is generics. For instance, in C, qsort() takes a function pointer for the comparison function, in Rust and C++, the standard library sorting functions are templated on the comparison function. This means it's much easier for the compiler to specialize the sorting function, inline the comparisons and optimize around it. I don't know if C compilers specialize qsort() based on comparison function this way. They might, but it's certainly a lot more to ask of the compiler, and I would argue there are probably many cases like this where C++ and Rust can outperform C because of their much more powerful facilities for specialization.
I agree with this whole-heartedly. Rust is a LANGUAGE and C is a LANGUAGE. They are used to describe behaviours. When you COMPILE and then RUN them you can measure speed, but that's dependent on two additional bits that are not intrinsically part of the languages themselves.
Now: the languages may expose patterns that a compiler can make use of to improve optimizations. That IS interesting, but it is not a question of speed. It is a question of expressability.
No. As you've made clear, it's a question of being able to express things in a way that gives more information to a compiler, allowing it to create executables that run faster.
Saying that a language is about "expressability" is obvious. A language is nothing other than a form of expression; no more, no less.
Yes. But the speed is dependent on whether or not the compiler makes use of that information and the machine architecture the compiler is running it on.
Speed is a function of all three -- not just the language.
Optimizations for one architecture can lead to perverse behaviours on another (think cache misses and memory layout -- even PROGRAM layout can affect speed).
These things are out of scope of the language and as engineers I think we ought to aim to be a bit more precise. At a coarse level I can understand and even would agree with something like "Python is slower than C", but the same argument applies there as well.
But at some point objectivity ought to enter the playing field.
> On the other hand, signed integer overflow being UB would count for C/C++
C and C++ don't actually have an advantage here because this is only limited to signed integers unless you use compiler-specific intrinsics. Rust's standard library allows you to make overflow on any specific arithmetic operation UB on both signed and unsigned integers.
It's interesting, because it's a "cultural" thing like the author discusses, it's a very good point. Sure, you can do unsafe integer arithmetic in Rust. And you can do safe integer arithmetic with overflow in C/C++. But in both cases, do you? Probably you don't in either case.
"Culturally", C/C++ has opted for "unsafe-but-high-perf" everywhere, and Rust has "safe-but-slightly-lower-perf" everywhere, and you have to go out of your way to do it differently. Similarly with Zig and memory allocators: sure, you can do "dynamically dispatched stateful allocators that you pass to every function that allocates" in C, but do you? No, you probably don't, you probably just use malloc().
On the other hand: the author's point that the "culture of safety" and the borrow checker in Rust frees your hand to try some things in Rust which you might not in C/C++, and that leads to higher perf. I think that's very true in many cases.
Again, the answer is more or less "basically no, all these languages are as fast as each other", but the interesting nuance is in what is natural to do as an experienced programmer in them.
C++ isn't always "unsafe-but-high-perf". Move semantics are a good example. The spec goes to great lengths to ensure safety in a huge number of scenarios, at the cost of performance. Mostly shows up in two ways: one, unnecessary destructor calls on moved out objects, and two, allowing throwing exceptions in move constructors which prevents most optimizations that would be enabled by having move constructors in the first place (there was an article here recently on this topic).
Another one is std::shared_ptr. It always uses atomic operations for reference counting and there's no way to disable that behavior or any alternative to use when you don't need thread safety. On the other hand, Rust has both non-atomic Rc and atomic Arc.
Interestingly enough, Zig does not use the same terminology as C/C++/Rust do here. Zig has "illegal behavior," which is either "safety checked" or "unchecked." Unchecked illegal behavior is like undefined behavior. Compiler flags and in-source annotations can change the semantics from checked to unchecked or vice versa.
Anyway that's a long way of saying that you're right, integer overflow is illegal behavior, I just think it's interesting.
>in Rust and C++, the standard library sorting functions are templated on the comparison function. This means it's much easier for the compiler to specialize the sorting function, inline the comparisons and optimize around it.
I think this is something of a myth. Typically, a C compiler can't inline the comparison function passed to qsort because libc is dynamically linked (so the code for qsort isn't available). But if you statically link libc and have LTO, or if you just paste the implementation of qsort into your module, then a compiler can inline qsort's comparison function just as easily as a C++ compiler can inline the comparator passed to std::sort. As for type-specific optimizations, these can generally be done just as well for a (void *) that's been cast to a T as they can be for a T (though you do miss out on the possibility of passing by value).
That said, I think there is an indirect connection between a templated sort function and the ability to inline: it forces a compiler/linker architecture where the source code of the sort function is available to the compiler when it's generating code for calls to that function.
qsort is obviously just an example, this situation applies to anything that takes a callback: in C++/Rust, that's almost always generic and the compiler will monomorphize the function and optimize around it, and in C it's almost always a function pointer and a userData argument for state passed on the stack. (and, of course, it applies not just to callbacks, but more broadly to anything templated).
I'm actually very curious about how good C compilers are at specializing situations like this, I don't actually know. In the vast majority cases, the C compiler will not have access to the code (either because of dynamic linking like in this example, or because the definition is in another translation unit), but what if it does? Either with static linking and LTO, or because the function is marked "inline" in a header? Will C compilers specialize as aggressively as Rust and C++ are forced to do?
If anyone has any resources that have looked into this, I would be curious to hear about it.
In general "Is programming language X faster than Y" is a meaningless question. It mostly comes down to specific implementations - specific compilers, interpreters, etc.
The only case where one language is likely to be inherently faster than another is when the other language is so high level or abstracted away from the processors it is going to run on that an optimizing compiler is going to have a hard time bridging that gap. It may take more work for an optimizing compiler to generate good code for one language than another, for example by having to recognizing when aliasing doesn't exist, but again this is ultimately a matter of implementation not language.
One example where Rust enables better and faster abstractions is traits. C you can do this with some ugly methods like macros and such but in Rust it’s not the implementers choice it’s the callers choice whether to use dynamic dispatch (function pointer table in C) or static dispatch (direct function calls!)
In c the caller isn’t choosing typically. The author of some library or api decides this for you.
This turns out to be fairly significant in something like an embedded context where function pointers kill icache and rob cycles jumping through hoops. Say you want to bit bang a bus protocol using GPIO, in C with function pointers this adds maybe non trivial overhead and your abstraction is no longer (never was) free. Traits let the caller decide to monomorphize that code and get effectively register reads and writes inlined while still having an abstract interface to GPIO. This is excellent!
I probably enjoy ELF hacking more than most, but patching an ELF binary via LD_PRELOAD, linker hacks, or even manual or assisted relinking tricks are just tools in the bag of performant C/C++ (and probably Rust too, but I don't get paid to make that fast). If you care about perf and for whatever reason are using someone else's code, you should be intimately familiar with your linker, binary format, ABI, and OS in addition to your hardware. It's all bytes in the end, and these abstractions are pliable with standard tooling.
I'd usually rather have a nice language-level interface for customizing implementation, but ELF and Linux scripting is typically good enough. Binary patching is in a much easier to use place these days with good free tooling and plenty of (admittedly exploit-oriented) tutorials to extrapolate from as examples.
It's a tradeoff though, as I think traits makes the Rust build times grow really quickly. I don't know the exact characteristics of it, also I think they speed it up compared to how it used to be, but I do remember that you'll get noticeable build slowdowns the more you use traits, especially "complicated" ones.
Absolutely, was not trying to claim otherwise. But since we're engineers (at least I like to see myself as one), it's worth always keeping in mind that almost everything comes with tradeoffs, even traits :)
Someone down the line might be wondering why suddenly their Rust builds take 4x the time after merging something, and just maybe remembering this offhand comment will make them find the issue faster :)
A lot of C++ devs advocate for simple replacements for the STL that do not rely too much on zero-cost abstractions. That way you can have small binaries, fast compiles, and make a fast-debug kinda build where you only turn on a few optimizations.
That way you can get most of the speed of the Release version, with a fairly good chance of getting usable debug info.
A huge issue with C++ debug builds is the resulting executables are unusably slow, because the zero-cost abstractions are not zero cost in debug builds.
It's never the case that only one thing is important.
In the extreme, you surely wouldn't accept a 1 day or even 1 week build time for example? It seems like that could be possible and not hypothetical for a 1 week build since a system could fuzz over candidate compilation, and run load tests and do PGO and deliver something better. But even if runtime performance was so important that you had such a system, it's obvious you wouldn't ever have developer cycles that take a week to compile.
Build time also even does matter for release: if you have a critical bug in production and need to ship the fix, a 1 hour build time can still lose you a lot here. Release build time doesn't matter until it does.
People do have cold Rust compiles that can push up into measured in hours. Large crates often take design choices that are more compile time friendly shape.
Note that C++ also has almost as large problem with compile times with large build fanouts including on templates, and it's not always realistic for incremental builds to solve either especially time burnt on linking, e.g. I believe Chromium development often uses a mode with .dlls dynamic linking instead of what they release which is all static linked exactly to speed up incremental development. The "fast" case is C not C++.
> I believe Chromium development often uses a mode with .dlls dynamic linking instead of what they release which is all static linked exactly to speed up incremental development. The "fast" case is C not C++.
Folks have worked tirelessly to improve the speed of the Rust compiler, and it's gotten significantly faster over time. However, there are also language-level reasons why it can take longer to compile than other languages, though the initial guess of "because of the safety checks" is not one of them, those are quite fast.
> How slow are we talking here?
It really depends on a large number of factors. I think saying "roughly like C++" isn't totally unfair, though again, it really depends.
AFAIK, it's not the traits that does it but rather the generics.
Rust does make it a lot easier to use generics which is likely why using more traits appears to be the cause of longer build times. I think it's just more that the more traits you have, the more likely you are to stumble over some generic code which ultimately generates more code.
> AFAIK, it's not the traits that does it but rather the generics.
Aah, yes, that sounds more correct, the end result is the same, I failed to remember the correct mechanism that led to it. Thank you for the correction!
I almost ignored this post because I can't stand this particular war, where examples are cherry picked to prove either answer.
I'm very happy to see the nuanced take in this article, slowly deconstructing the implicit assumptions proposed by the person asking this question, to arrive at the same conclusion that I long have. I hope this post reaches the right people.
A particular language doesn't have a "speed", a particular implementation may have, and the language may have properties that make it difficult to make a fast implementation (of those specific properties/features) given the constraints of our current computer architectures. Even then, there's usually too many variables to make a generalized statement, and the question often presumes that performance is measured as total cpu time.
I like to say that there are two primary factors when we talk about how "fast" a language is:
1. What costs does the language actively inject into a program?
2. What optimizations does the language facilitate?
Most of the time, it's sufficient to just think about the first point. C and Rust are faster than Python and Javascript because the dynamic nature of the latter two requires implementations to inject runtime checks all over the place to enable that dynamism. Rust and C simply inject essentially zero active runtime checks, so membership in this club is easy to verify.
The second one is where we get bogged down, because drawing clean conclusions is complicated by the (possibly theoretical) existence of optimizing compilers that can leverage the optimizability inherent to the language, as well as the inherent fragility of such optimizations in practice. This is where we find ourselves saying things like "well Rust could have an advantage over C, since it frequently has more precise and comprehensive aliasing information to pass to the optimizer", though measuring this benefit is nontrivial and it's unclear how well LLVM is thoroughly utilizing this information at present. At the same time, the enormous observed gulf between Rust in release mode (where it's as fast as C) and Rust in debug mode (when it's as slow as Ruby) shows how important this consideration is; Rust would not have achieved C speeds if it did not carefully pick abstractions that were amenable to optimization.
It's also interesting to think about this in terms of the "zero cost abstractions"/"zero overhead abstractions" idea, which Stroustrup wrote as "What you don't use, you don't pay for. What you do use, you couldn't hand code any better". The first sentence is about 1, and the second one is about what you're able to do with 2.
Is Javascript significantly slower? It is extremely common in the real world and so a lot of effort has gone into optimizing it - v8 is very good. Yes C and Rust enable more optimizations: they will be slightly faster, but javascript has had a lot of effort put into making it run fast.
Yes, for most real-world examples JavaScript is significantly slower; JIT isn’t free and can be very sensitive to small code changes, you also have to consider the garbage collector.
Speed is also not the only metric, Rust and C enable much better control over memory usage. In general, it is easier to write a memory-efficient program in Rust or C than it is in JS.
I think there's a third question, but I don't know quite how to phrase it. Maybe "how real-world fast is the language?" or "how fast is the language in the hands of someone who isn't obsessively thinking about speed?"
That is, most of the time, most of the users aren't thinking about how to squeeze the last tenth of a percent of speed out of it. They aren't thinking about speed at all. They're thinking about writing code that works at all, and that hopefully doesn't crash too often. How fast is the language for them? Does it nudge them toward faster code, or slower? Are the default, idiomatic ways of writing things the fast way, or the slow way?
The question is what do we mean by "a fast language"? We could mean it to be how fast the fastest code that a performance expert in that language, with no resource constraints, could write. Or, we can restrict it to "idiomatic" code. Or we can say that a fast language is the one where an average programmer is most likely to produce fast code with a given budget (in which case probably none of the languages mentioned here are among the fastest).
It's compilers and compiler optimizations that make code run fast. The real question is if the Rust language and the richer memory semantics it has help the Rust compiler to provide a bit more context for optimizing that the C compiler wouldn't have do unless you hand optimize your code.
If you do hand optimize your code, all bets are off. With both languages. But I think the notion that the Rust compiler has more context for optimizing than the C compiler is maybe not as controversial as the notion that language X is better/faster than language Y. Ultimately, producing fast/optimal code in C kind of is the whole point of C. And there aren't really any hacks you can do in C that you can't do in Rust, or vice versa. So, it would be hard to make the case that Rust is slower than C or the other way around.
However, there have been a few rewrites of popular unix tools in Rust that benchmark a bit faster than their C equivalents. Could those be optimized in C. Probably; but they just haven't. But there is a case there of arguing that maybe Rust code is a bit easier to make fast than C code.
Right, but if we assume that programmers' compensation is statistically correlated with their skill, then we can drop "average" and just talk about budget.
I may be biased, but I think that if you have a budget that's reasonable in the industry for some project size and includes not only the initial development but also maintenance and evolution over the software's lifetime, especially when it's not small (say over 200KLOC), and you want to choose the language that would give you the fastest outcome, you will not get a faster program than if you chose Java. To get a faster program in any language, if possible, would require a significantly higher budget (especially for the maintenance and evolution).
I don't think so, but it may not be far behind. More importantly, though, I'm fairly confident it won't be Assembly, or C, or C++, or Rust, or Zig, but also not Python, or TS/JS. The candidates would most likely include Java, C#, and Go.
Purely by the numbers, an "average programmer" is much more likely to use Javascript, Python, or Java. The native languages have been a bit of a niche field since the late 90's (i.e. heavily slanted towards OS, embedded, and gamedev folks)
Well, if we define a reasonable yardstick for this, and try to write idiomatic code without low-level hacks and optimization, Rust has the potential to be faster because its more strict aliasing rules might enable optimizations that the C compiler wouldn't try.
However I remember reading a few years back that due to the Rust frontend not communicating these opportunities to LLVM, and LLVM not being designed to take advantage of them, the real-world gains do not always materialize.
Also sometimes people write code in Rust that does not compile under the borrow checker rules, and alleviate this issue either by cloning objects or using RefCell, both of which have a runtime cost.
The article does not mention the possible additional optimisation opportunities that arise in Rust code due to stricter aliasing rules of references. But I don’t have an example in mind. Does anyone know of an example of it happening in real code?
> When noalias annotations were first disabled in 2015 it resulted in between 0-5% increased runtime in various benchmarks.
This leaves us with a few relevant questions:
Were those benchmarks representative of real world code? (They're not linked, so we cannot know. The author is reliable, as far as I'm concerned, but we have no way to verify this off-hand comment directly, I link to it specifically because I'd take the author at their word. They do not make any claim about this, specifically.)
Those benchmarks are for Rust code with optimizations turned off and back on again, not Rust code vs C code. Does that make this a good benchmark of the question, or a bad one?
These were llvm's 'noalias' markers, which were written for `restrict` in C. Do those semantics actually take full advantage of Rust's aliasing model, or not? Could a compiler which implements these optimizations in a different way do better? (I'm actually not fully sure of the latest here, and I suspect some corners would be relying on the stacked borrows vs tree borrows stuff being finalized)
Another issue we have to consider here for the measurements taken then is that it was miscompiling, which, to me, calls into question how much we can trust that performance change.
Additionally, it was 10 years ago and LLVM has changed. It could be that LLVM does better now, or it could do worse. I would actually be interested in seeing some benchmarks with modern rustc.
Not exactly real world, but real code example demonstrating strict aliasing rule in action for C++. https://godbolt.org/z/WvMb34Kea Rust should have even more opportunities of this due to restrictions it has for writable references.
There are 2 main differences between versions with and without strict aliasing. Without strict aliasing compiler can't assume that the result accumulator doesn't change during the loop and it has to repeatedly read/write it each iteration. With strict aliasing it can just read it to register, do the looping and write the result back at the end once. Second effect is that with strict aliasing enabled compiler can vectorize the loop processing 4 floats at the same time, most likely the same uncertainty of counter prevents vecotorization without strict aliasing.
If you want something slightly simpler example you can disable vectorization by adding '-fno-tree-vectorize'. With it disabled there is still difference in handling of counter.
Using restrict pointers and multiple same type input arrays it would probably be possible to make something closer to real world example.
Many C programs are vailid C++ and are faster when compiled with a C++ compiler because of those stricter aliasing and type rules. Like you though I have no examples.
I believe this advantage is currently mostly theoretical, as the code ultimately gets compiled with LLVM which does not fully utilize all the additional optimization opportunities.
LLVM doesn't fully utilize all the power, but it does use an increasing amount every year. Flang and Rust have both given LLVM plenty of example code and a fair number of contributors who want to make LLVM work better for them.
When the optimizer knows writes can't change the reads, it can reorder and coalesce them. The main benefit of that is enabling autovectorization in more cases. Otherwise it saves a few loads here and there.
Is there a common pattern for "Is language X faster than language Y" ?
Like what is your definition of faster :
faster to developer, to start, to execute, to handle different workload with the same binaries (like JIT).
I feel like another optimization that rust code can exploit is uninhabited types.
When combined with generics and sum types these can lead to entire branches being unreachable at the type level. Like Option<!> or Result<T, !>, rust hasn't stablized !, but you can declare them other ways such as an empty enum with no variants.
struct field alignment/padding isn't part of the C spec iirc (at least not in the way mentioned in the article), but it's almost always done that way, which is important for having a stable abi
also, if performance is critical to you, profile stuff and compare outputted assembly, more often than not you'll find that llvm just outputs the same thing in both cases
See "6.7.3.2 Structure and union specifiers", paragraph 16 & 17:
> Each non-bit-field member of a structure or union object is aligned in an implementation-defined manner appropriate to its type.
> Within a structure object, the non-bit-field members and the units in which bit-fields reside have addresses that increase in the order in which they are declared.
It is indeed part of the standard. It says "Within a structure object, the non-bit-field members and the units in which bit-fields reside have
addresses that increase in the order in which they are declared"[1] which doesn't allow implementations to reorder fields, at least according to my understanding.
> struct field alignment/padding isn't part of the C spec iirc
It's part of the ABI spec. It's true that C evolved in an ad hoc way and so the formal rigor got spread around to a bunch of different stakeholders. It's not true that C is a lawless wasteland where all behavior is subject to capricious and random whims, which is an attitude I see a lot in some communities.
People write low level software to deal with memory layout and alignment every day in C, have for fourty years, and aren't stopping any time soon.
Sorry, maybe stupid question. But can't this be decided by some benchmarks, using some of the features in the article that purport to make Rust faster?
Part of what I'm getting at here is that you have to decide what is in those benchmarks in the first place. Yes, benchmarks would be an important part of answering this question, but it's not just one question: it's a bunch of related but different questions.
I think the only reasonable way to interpret this question is "is Rust written by reasonably competent Rust developer spending a reasonable amount of time faster/slower than an equally competent C developer spending the same amount of time".
I don't think a language should count as "fast" if it takes an expert or an inordinate amount of time to get good performance, because most code won't have that.
So on those grounds I would say Rust probably is faster than C, because it makes it much much easier to use multithreading and more optimised libraries. For example a lot of C code uses linked lists because they're easy to write in C, even when a vector would be faster and more appropriate. Multithreading can just be a one line change in Rust.
Depends. If it takes an assembly programmer 8 hours to implement <X>, can an equally proficient Python programmer spending 8 hours to implement <X> create a faster program?
Let's say they only need 2 hours to get the <X> to work, and can use the remaining 6 hours for optimizing. Can 6 hours of optimizing a Python program make it faster than the assembly program?
The answer isn't obvious, and certainly depends on the specific <X>. I can imagine various <X> where even unlimited time spent optimizing Python code won't produce faster results than the assembly code, unless you drop into C/C++/Zig/Rust/D and write a native Python extension (and of course, at that point you're not comparing against Python, but that native language).
Back then the C implementation of the (i.e., "one") micro benchmark beat the Rust implementation. I could squeeze out more performance by precisely controlling the loop unrolling. Nowadays, I don't really care and operate under the assumption that "Python is faster than $X and if it is not, it is still fast enough!"
As I just posted, any speed comparison needs to be based on specific implementations (compiler A vs compiler B), not languages.
When it comes to assembly, the "compiler" is the person writing the code, and while assembly gives you the maximum flexibility to potentially equal or outperform any compiler for any language, there are not too many people with the skill to do that, especially when writing large programs (which due to the effort required are rarely written in assembler). In general there is much more potential for improving the speed of programs by changing the design and using better algorithms, which is where high level languages offer a big benefit by making this easier.
This post might get the record for people responding to the title without reading the article. Jeez people, it takes five seconds to discover that it subverts expectations.
Where C application code often suffers, but by no means always, is the use of memory for data structures. A nice big chunk of static memory will make a function fast, but I’ve seen many C routines malloc memory, do a strcpy, compute a bit, and free it at the end, over and over, because there’s no convenient place to retain the state. There are no vectors, no hash maps, no crates.io and cargo to add a well-optimized data structure library.
It is for this reason I believe that Rust, and C++, have an advantage over C when it comes to writing fast code, because it’s much easier to drop in a good data structure. To a certain extent I think C++ has an advantage over Rust due to easier and better control over layout.
I'd certainly agree that malloc is the Achilles heel of any real world C. Overall though C++ was not a particularly good solution to memory efficiency since having OO available made the situation look like a fast sprint to the cake shop.
I love Betteridge's Law, and so one small thing I was trying to do here was subvert it a bit. Instead of "no," in this case, the answer is "the question is malformed."
>If we assume C is the ‘fastest language,’ whatever that means
I agree that it has no meaning. Speed(language) is undefined, therefore there is no faster language.
I get this often because python is referred to as a slow language, but since a python programmer can write more features than a C programmer in the same time, at least in my space, it causes faster programs in python, because some of those features are optimizations.
Now speed(program(language,programmer)) is defined, and you could do an experiment by having programmers of different languages write the same program and compare its execution times.
Only if it is repeatable. We have no information on what they learned in the two failed attempts - it is likely that they learned from the failures and started other architectural changes that enabled the final one to work. As such we cannot say anything about this.
Rust does have some interesting features, which restrict what you are allowed to do and thus make some things impossible but in turn make other things easier. It is highly likely that those restrictions are part of what made this possible. Given infinite resources (which you never have) a C++ implementation could be faster because it has better shared data concepts - but those same shared data concepts make it extremely hard to reason about multi-threaded code and so humanly you might not be able to make it work.
In short, the previous two attempts were done by completely different groups of different people, a few years apart. Your direct question about if direct wisdom from these two attempts was shared, either between them, or used by Stylo, isn't specifically discussed though.
> a C++ implementation could be faster because it has better shared data concepts
tl;dr: Rust officially allows you to write inline assembly so it's fast, but in C it's not officially specified as part of the language. Plus more points which do not actually indicate Rust is faster than C.
... well, that's what I get for reading an article with a silly title.
That’s not how I would summarize what I wrote, for what it’s worth. My summary would be “the question is malformed, you need to first state what the boundaries are for comparison before you can make any conclusions.” I think this is an interesting thing to discuss because many people assume that the answer to “is x faster than C?” to be “no” for all values of X.
> many people assume that the answer to “is x faster than C?” to be “no” for all values of X.
This is because C does so little for you -- bounds checking must be done explicitly for instance, like you mention in the article, so C is "faster" unless you work around rust's bounds checking. It reminds me of some West Virginia residents I know who are very proud of how low their taxes are -- the roads are falling apart, but the taxes are very low! C is this way too.
C is pretty optimally fast in the trivial case, but once you add bounds checking and error handling and memory management its edge is much much smaller (for Rust and Zig and other lowish-level languages)
In the real world the difference is rarely significant assuming great programmings implement great algorithms. However those two assumptions are rarely true.
I read the post to see how you would answer, not because I was unclear about what the answer would be, because the only possible answer here is “sometimes.” I especially like the point that Rust can be faster because it enables you to write different things. As I never tire of getting downvoted for saying, I’ve improved the speed of a program by replacing C with Python, because nobody could figure out how to write the right thing in C. If even Python can do this, it must apply to just about every pair of languages.
The article felt fairly dispassionate and even-handed to me, and I say this as someone who dislikes Klabnik very much and also dislikes the Rust community (especially its insidious, forced MIT rewrites of popular GPL software, with which they also break backwards compatibility). It is worth mentioning that there are certain things about Rust that conceivably could make it faster, e.g., const by default (theoretically facilitating certain optimizations), but in practice, thus far, do not.
Instead, I'd say that Rust & C are close enough, speed-wise, that (1) which one is faster will depend on small details of the particular use case, or (2) the speed difference will matter less than other language considerations.
The big one is multi-threading. In Rust, whether you use threads or not, all globals must be thread-safe, and the borrow checker requires memory access to be shared XOR mutable. When writing single-threaded code takes 90% of effort of writing multi-threaded one, Rust programmers may as well sprinkle threads all over the place regardless whether that's a 16x improvement or 1.5x improvement. In C, the cost/benefit analysis is different. Even just spawning a thread is going to make somebody complain that they can't build the code on their platform due to C11/pthread/openmp. Risk of having to debug heisenbugs means that code typically won't be made multi-threaded unless really necessary, and even then preferably kept to simple cases or very coarse-grained splits.
I wouldn't consider there to be any notable effort in making thread build on target platforms in C relative to normal effort levels in C, but it's objectively more work than `std::thread::spawn(move || { ... });`.
Despite benefits, I don't actually think the memory safety really plays a role in the usage rate of parallelism. Case in point, Go has no implicit memory safety with both races and atomicity issues being easy to make, and yet relies much heavier on concurrency (with a parallelism degree managed by the runtime) with much less consideration than Rust. After all, `go f()` is even easier.
(As a personal anecdote, I've probably run into more concurrency-related heisenbugs in Go than I ever did in C, with C heisenbugs more commonly being memory mismanagement in single-threaded code with complex object lifetimes/ownership structures...)
Using pthread in C, for example, TBB is not required.
Not sure about C11 threads, but I have always thought that GLIBC just uses pthread under the hood.
On a backend system where you already have multiple processes using various cores (databases, web servers, etc) it usually doesn’t make sense as a performance tool.
And on an embedded device you want to save power so it also rarely makes sense.
Are people making user facing apps in rust with GUIs?
yes
However, in the spirit of the question: someone mentioned the stricter aliasing rules, that one does come to mind on Rust's side over C/C++. On the other hand, signed integer overflow being UB would count for C/C++ (in general: all the UB in C/C++ not present in Rust is there for performance reasons).
Another thing I thought of in Rust and C++s favor is generics. For instance, in C, qsort() takes a function pointer for the comparison function, in Rust and C++, the standard library sorting functions are templated on the comparison function. This means it's much easier for the compiler to specialize the sorting function, inline the comparisons and optimize around it. I don't know if C compilers specialize qsort() based on comparison function this way. They might, but it's certainly a lot more to ask of the compiler, and I would argue there are probably many cases like this where C++ and Rust can outperform C because of their much more powerful facilities for specialization.
Now: the languages may expose patterns that a compiler can make use of to improve optimizations. That IS interesting, but it is not a question of speed. It is a question of expressability.
Saying that a language is about "expressability" is obvious. A language is nothing other than a form of expression; no more, no less.
Speed is a function of all three -- not just the language.
Optimizations for one architecture can lead to perverse behaviours on another (think cache misses and memory layout -- even PROGRAM layout can affect speed).
These things are out of scope of the language and as engineers I think we ought to aim to be a bit more precise. At a coarse level I can understand and even would agree with something like "Python is slower than C", but the same argument applies there as well.
But at some point objectivity ought to enter the playing field.
C and C++ don't actually have an advantage here because this is only limited to signed integers unless you use compiler-specific intrinsics. Rust's standard library allows you to make overflow on any specific arithmetic operation UB on both signed and unsigned integers.
"Culturally", C/C++ has opted for "unsafe-but-high-perf" everywhere, and Rust has "safe-but-slightly-lower-perf" everywhere, and you have to go out of your way to do it differently. Similarly with Zig and memory allocators: sure, you can do "dynamically dispatched stateful allocators that you pass to every function that allocates" in C, but do you? No, you probably don't, you probably just use malloc().
On the other hand: the author's point that the "culture of safety" and the borrow checker in Rust frees your hand to try some things in Rust which you might not in C/C++, and that leads to higher perf. I think that's very true in many cases.
Again, the answer is more or less "basically no, all these languages are as fast as each other", but the interesting nuance is in what is natural to do as an experienced programmer in them.
Another one is std::shared_ptr. It always uses atomic operations for reference counting and there's no way to disable that behavior or any alternative to use when you don't need thread safety. On the other hand, Rust has both non-atomic Rc and atomic Arc.
Then, I raise you to Zig which has unsigned integer overflow being UB.
Anyway that's a long way of saying that you're right, integer overflow is illegal behavior, I just think it's interesting.
https://doc.rust-lang.org/std/intrinsics/fn.unchecked_add.ht...
I think this is something of a myth. Typically, a C compiler can't inline the comparison function passed to qsort because libc is dynamically linked (so the code for qsort isn't available). But if you statically link libc and have LTO, or if you just paste the implementation of qsort into your module, then a compiler can inline qsort's comparison function just as easily as a C++ compiler can inline the comparator passed to std::sort. As for type-specific optimizations, these can generally be done just as well for a (void *) that's been cast to a T as they can be for a T (though you do miss out on the possibility of passing by value).
That said, I think there is an indirect connection between a templated sort function and the ability to inline: it forces a compiler/linker architecture where the source code of the sort function is available to the compiler when it's generating code for calls to that function.
I'm actually very curious about how good C compilers are at specializing situations like this, I don't actually know. In the vast majority cases, the C compiler will not have access to the code (either because of dynamic linking like in this example, or because the definition is in another translation unit), but what if it does? Either with static linking and LTO, or because the function is marked "inline" in a header? Will C compilers specialize as aggressively as Rust and C++ are forced to do?
If anyone has any resources that have looked into this, I would be curious to hear about it.
The only case where one language is likely to be inherently faster than another is when the other language is so high level or abstracted away from the processors it is going to run on that an optimizing compiler is going to have a hard time bridging that gap. It may take more work for an optimizing compiler to generate good code for one language than another, for example by having to recognizing when aliasing doesn't exist, but again this is ultimately a matter of implementation not language.
In c the caller isn’t choosing typically. The author of some library or api decides this for you.
This turns out to be fairly significant in something like an embedded context where function pointers kill icache and rob cycles jumping through hoops. Say you want to bit bang a bus protocol using GPIO, in C with function pointers this adds maybe non trivial overhead and your abstraction is no longer (never was) free. Traits let the caller decide to monomorphize that code and get effectively register reads and writes inlined while still having an abstract interface to GPIO. This is excellent!
I'd usually rather have a nice language-level interface for customizing implementation, but ELF and Linux scripting is typically good enough. Binary patching is in a much easier to use place these days with good free tooling and plenty of (admittedly exploit-oriented) tutorials to extrapolate from as examples.
Someone down the line might be wondering why suddenly their Rust builds take 4x the time after merging something, and just maybe remembering this offhand comment will make them find the issue faster :)
That way you can get most of the speed of the Release version, with a fairly good chance of getting usable debug info.
A huge issue with C++ debug builds is the resulting executables are unusably slow, because the zero-cost abstractions are not zero cost in debug builds.
In the extreme, you surely wouldn't accept a 1 day or even 1 week build time for example? It seems like that could be possible and not hypothetical for a 1 week build since a system could fuzz over candidate compilation, and run load tests and do PGO and deliver something better. But even if runtime performance was so important that you had such a system, it's obvious you wouldn't ever have developer cycles that take a week to compile.
Build time also even does matter for release: if you have a critical bug in production and need to ship the fix, a 1 hour build time can still lose you a lot here. Release build time doesn't matter until it does.
Note that C++ also has almost as large problem with compile times with large build fanouts including on templates, and it's not always realistic for incremental builds to solve either especially time burnt on linking, e.g. I believe Chromium development often uses a mode with .dlls dynamic linking instead of what they release which is all static linked exactly to speed up incremental development. The "fast" case is C not C++.
Bevy, a Rust ECS framework for building games (among other things), has a similar solution by offering a build/rust "feature" that enables dynamic linking (called "dynamic_linking"). https://bevy.org/learn/quick-start/getting-started/setup/#dy...
Folks have worked tirelessly to improve the speed of the Rust compiler, and it's gotten significantly faster over time. However, there are also language-level reasons why it can take longer to compile than other languages, though the initial guess of "because of the safety checks" is not one of them, those are quite fast.
> How slow are we talking here?
It really depends on a large number of factors. I think saying "roughly like C++" isn't totally unfair, though again, it really depends.
Rust does make it a lot easier to use generics which is likely why using more traits appears to be the cause of longer build times. I think it's just more that the more traits you have, the more likely you are to stumble over some generic code which ultimately generates more code.
Aah, yes, that sounds more correct, the end result is the same, I failed to remember the correct mechanism that led to it. Thank you for the correction!
Tbf this applies to Rust too. If the author writes
they have forced the caller into dynamic dispatch.Had they written
the choice would've remained open to the callerI'm very happy to see the nuanced take in this article, slowly deconstructing the implicit assumptions proposed by the person asking this question, to arrive at the same conclusion that I long have. I hope this post reaches the right people.
A particular language doesn't have a "speed", a particular implementation may have, and the language may have properties that make it difficult to make a fast implementation (of those specific properties/features) given the constraints of our current computer architectures. Even then, there's usually too many variables to make a generalized statement, and the question often presumes that performance is measured as total cpu time.
1. What costs does the language actively inject into a program?
2. What optimizations does the language facilitate?
Most of the time, it's sufficient to just think about the first point. C and Rust are faster than Python and Javascript because the dynamic nature of the latter two requires implementations to inject runtime checks all over the place to enable that dynamism. Rust and C simply inject essentially zero active runtime checks, so membership in this club is easy to verify.
The second one is where we get bogged down, because drawing clean conclusions is complicated by the (possibly theoretical) existence of optimizing compilers that can leverage the optimizability inherent to the language, as well as the inherent fragility of such optimizations in practice. This is where we find ourselves saying things like "well Rust could have an advantage over C, since it frequently has more precise and comprehensive aliasing information to pass to the optimizer", though measuring this benefit is nontrivial and it's unclear how well LLVM is thoroughly utilizing this information at present. At the same time, the enormous observed gulf between Rust in release mode (where it's as fast as C) and Rust in debug mode (when it's as slow as Ruby) shows how important this consideration is; Rust would not have achieved C speeds if it did not carefully pick abstractions that were amenable to optimization.
It's also interesting to think about this in terms of the "zero cost abstractions"/"zero overhead abstractions" idea, which Stroustrup wrote as "What you don't use, you don't pay for. What you do use, you couldn't hand code any better". The first sentence is about 1, and the second one is about what you're able to do with 2.
Speed is also not the only metric, Rust and C enable much better control over memory usage. In general, it is easier to write a memory-efficient program in Rust or C than it is in JS.
That is, most of the time, most of the users aren't thinking about how to squeeze the last tenth of a percent of speed out of it. They aren't thinking about speed at all. They're thinking about writing code that works at all, and that hopefully doesn't crash too often. How fast is the language for them? Does it nudge them toward faster code, or slower? Are the default, idiomatic ways of writing things the fast way, or the slow way?
If you do hand optimize your code, all bets are off. With both languages. But I think the notion that the Rust compiler has more context for optimizing than the C compiler is maybe not as controversial as the notion that language X is better/faster than language Y. Ultimately, producing fast/optimal code in C kind of is the whole point of C. And there aren't really any hacks you can do in C that you can't do in Rust, or vice versa. So, it would be hard to make the case that Rust is slower than C or the other way around.
However, there have been a few rewrites of popular unix tools in Rust that benchmark a bit faster than their C equivalents. Could those be optimized in C. Probably; but they just haven't. But there is a case there of arguing that maybe Rust code is a bit easier to make fast than C code.
I'd say most people use this definition, with the caveat that there's no official "average programmer", and everyone has different standards.
However I remember reading a few years back that due to the Rust frontend not communicating these opportunities to LLVM, and LLVM not being designed to take advantage of them, the real-world gains do not always materialize.
Also sometimes people write code in Rust that does not compile under the borrow checker rules, and alleviate this issue either by cloning objects or using RefCell, both of which have a runtime cost.
The first is, we do have some amount of empirical evidence here: Rust had to turn its aliasing optimizations on and off again a few times due to bugs in LLVM. A comment from 2021: https://github.com/rust-lang/rust/issues/54878#issuecomment-...
> When noalias annotations were first disabled in 2015 it resulted in between 0-5% increased runtime in various benchmarks.
This leaves us with a few relevant questions:
Were those benchmarks representative of real world code? (They're not linked, so we cannot know. The author is reliable, as far as I'm concerned, but we have no way to verify this off-hand comment directly, I link to it specifically because I'd take the author at their word. They do not make any claim about this, specifically.)
Those benchmarks are for Rust code with optimizations turned off and back on again, not Rust code vs C code. Does that make this a good benchmark of the question, or a bad one?
These were llvm's 'noalias' markers, which were written for `restrict` in C. Do those semantics actually take full advantage of Rust's aliasing model, or not? Could a compiler which implements these optimizations in a different way do better? (I'm actually not fully sure of the latest here, and I suspect some corners would be relying on the stacked borrows vs tree borrows stuff being finalized)
Additionally, it was 10 years ago and LLVM has changed. It could be that LLVM does better now, or it could do worse. I would actually be interested in seeing some benchmarks with modern rustc.
There are 2 main differences between versions with and without strict aliasing. Without strict aliasing compiler can't assume that the result accumulator doesn't change during the loop and it has to repeatedly read/write it each iteration. With strict aliasing it can just read it to register, do the looping and write the result back at the end once. Second effect is that with strict aliasing enabled compiler can vectorize the loop processing 4 floats at the same time, most likely the same uncertainty of counter prevents vecotorization without strict aliasing.
If you want something slightly simpler example you can disable vectorization by adding '-fno-tree-vectorize'. With it disabled there is still difference in handling of counter.
Using restrict pointers and multiple same type input arrays it would probably be possible to make something closer to real world example.
Also note that C++ does not have restrict, formally speaking, though it is a common compiler extension. It's a C feature only!
also, if performance is critical to you, profile stuff and compare outputted assembly, more often than not you'll find that llvm just outputs the same thing in both cases
See "6.7.3.2 Structure and union specifiers", paragraph 16 & 17:
> Each non-bit-field member of a structure or union object is aligned in an implementation-defined manner appropriate to its type.
> Within a structure object, the non-bit-field members and the units in which bit-fields reside have addresses that increase in the order in which they are declared.
[1] https://open-std.org/JTC1/SC22/WG14/www/docs/n3220.pdf section 6.7.3.2, paragraph 17.
It's part of the ABI spec. It's true that C evolved in an ad hoc way and so the formal rigor got spread around to a bunch of different stakeholders. It's not true that C is a lawless wasteland where all behavior is subject to capricious and random whims, which is an attitude I see a lot in some communities.
People write low level software to deal with memory layout and alignment every day in C, have for fourty years, and aren't stopping any time soon.
Part of what I'm getting at here is that you have to decide what is in those benchmarks in the first place. Yes, benchmarks would be an important part of answering this question, but it's not just one question: it's a bunch of related but different questions.
I don't think a language should count as "fast" if it takes an expert or an inordinate amount of time to get good performance, because most code won't have that.
So on those grounds I would say Rust probably is faster than C, because it makes it much much easier to use multithreading and more optimised libraries. For example a lot of C code uses linked lists because they're easy to write in C, even when a vector would be faster and more appropriate. Multithreading can just be a one line change in Rust.
Let's say they only need 2 hours to get the <X> to work, and can use the remaining 6 hours for optimizing. Can 6 hours of optimizing a Python program make it faster than the assembly program?
The answer isn't obvious, and certainly depends on the specific <X>. I can imagine various <X> where even unlimited time spent optimizing Python code won't produce faster results than the assembly code, unless you drop into C/C++/Zig/Rust/D and write a native Python extension (and of course, at that point you're not comparing against Python, but that native language).
[0] tldr: "I think that there are so many variables that it is difficult to draw generalized conclusions."
Back then the C implementation of the (i.e., "one") micro benchmark beat the Rust implementation. I could squeeze out more performance by precisely controlling the loop unrolling. Nowadays, I don't really care and operate under the assumption that "Python is faster than $X and if it is not, it is still fast enough!"
When it comes to assembly, the "compiler" is the person writing the code, and while assembly gives you the maximum flexibility to potentially equal or outperform any compiler for any language, there are not too many people with the skill to do that, especially when writing large programs (which due to the effort required are rarely written in assembler). In general there is much more potential for improving the speed of programs by changing the design and using better algorithms, which is where high level languages offer a big benefit by making this easier.
Where C application code often suffers, but by no means always, is the use of memory for data structures. A nice big chunk of static memory will make a function fast, but I’ve seen many C routines malloc memory, do a strcpy, compute a bit, and free it at the end, over and over, because there’s no convenient place to retain the state. There are no vectors, no hash maps, no crates.io and cargo to add a well-optimized data structure library.
It is for this reason I believe that Rust, and C++, have an advantage over C when it comes to writing fast code, because it’s much easier to drop in a good data structure. To a certain extent I think C++ has an advantage over Rust due to easier and better control over layout.
I agree that it has no meaning. Speed(language) is undefined, therefore there is no faster language.
I get this often because python is referred to as a slow language, but since a python programmer can write more features than a C programmer in the same time, at least in my space, it causes faster programs in python, because some of those features are optimizations.
Now speed(program(language,programmer)) is defined, and you could do an experiment by having programmers of different languages write the same program and compare its execution times.
That is a damn good reason to choose Rust over C++, even if the Rust implementation of the "same" thing should be a bit slower.
Rust does have some interesting features, which restrict what you are allowed to do and thus make some things impossible but in turn make other things easier. It is highly likely that those restrictions are part of what made this possible. Given infinite resources (which you never have) a C++ implementation could be faster because it has better shared data concepts - but those same shared data concepts make it extremely hard to reason about multi-threaded code and so humanly you might not be able to make it work.
In short, the previous two attempts were done by completely different groups of different people, a few years apart. Your direct question about if direct wisdom from these two attempts was shared, either between them, or used by Stylo, isn't specifically discussed though.
> a C++ implementation could be faster because it has better shared data concepts
What concepts are those?
... well, that's what I get for reading an article with a silly title.
This is because C does so little for you -- bounds checking must be done explicitly for instance, like you mention in the article, so C is "faster" unless you work around rust's bounds checking. It reminds me of some West Virginia residents I know who are very proud of how low their taxes are -- the roads are falling apart, but the taxes are very low! C is this way too.
C is pretty optimally fast in the trivial case, but once you add bounds checking and error handling and memory management its edge is much much smaller (for Rust and Zig and other lowish-level languages)
Instead, I'd say that Rust & C are close enough, speed-wise, that (1) which one is faster will depend on small details of the particular use case, or (2) the speed difference will matter less than other language considerations.