Conf42 Rustlang 2022 - Online

Get maximum benefit from zero-cost abstractions

Video size:


This talk provides dozens of tips for improving your Rust code base. Rather than using slides, it is structured by providing “before and after” views of code examples that demonstrate a particular point.

Tips will be presented in themes:

Use idiomatic Rust

Using idioms improves maintenance by allowing everyone in the project (including people who will arrive at the project in 2 years’ time) to communicate using a shared vocabulary.

Examples drawn from:

  • naming conventions
  • API guidelines
  • clippy lints

Trust the abstractions

When coming from other languages, it’s tempting to try to improve performance by remove abstractions. In Rust, that’s usually the wrong approach. Iteration using higher-order functions results in typically results in generated code that’s more memory efficient than what you would .

Reducing cognitive load

Rust offers you lots of control. This has the downside of forcing you to make make choices when you would rather be writing your application. Some of the things that might affect learners:

  • Should you avoid .clone() in your code base?
  • What’s the difference between the Default trait and using a ::new() static method
  • How to choose between two similarly-named types, e.g. HashMap vs BTreeMap

Size matters

When you can fit more within your CPU’s memory caches, your program will run faster.

Smart pointers

What is a smart pointer? Which smart pointer should you choose?

Polymorphism patterns (aka enums are more powerful than they look)

There are a few ways to polymorphism in Rust. Generics and trait objects are well-known, but did you know that you can often get quite far with enums and the match keyword?


  • The talk's title is get maximum benefit from zero cost abstractions. The aim is to get the most from rust while keeping your runtime performance extremely high. There are three guide places to look for some really useful resources after this.
  • Programming is a team sport, and writing software is easier than reading software. What we really want to do is enable the rest of our team to follow along with our code. We do that by maintaining the conventions of the ecosystem. Where you can, implement as much as possible.
  • There are linters inside the ecosystem. Clippy rust format is available via cargo format inside your crate. You can make things harder for yourself by adding a pre commit hook. Try to remember to take a reference to the borrowed version of a type rather than borrowing an own version.
  • The first step to kind of using modeling is actually two just write a document describing how you want the program to work. A fuzzer is a program which generates random inputs for your functions. Every time it breaks your program gets stronger because you fix the problem.
  • The new type pattern is very very handy and is very confusing or alien to people who come to rust. Booleans and options should really be returned from functions. Rust focuses on the idea of a zero cost abstraction.
  • There are very few traits inside the rust standard library that have more than one method that you need to implement. To make it smoother, there is a crate calledenum dispatch which takes a lot of that pain away. It's much, much faster.
  • That's everything that I have for you today. You are very welcome to say hi in the comments. Hit me up on Twitter and let's see where we can go. Take care.


This transcript was autogenerated. To make changes, submit a PR.
Hi everybody, and welcome to this talk. We are going to be spending some time to think, but how to really improve our rust code? The talk's title is get maximum benefit from zero cost abstractions. I'm really going to try to think about the ways to get the most from rust while keeping your runtime performance extremely high. One thing to note, though, is that clone of the talk is actually completely new. There are some three guide places to look for some really useful resources after this. One is the API guidelines, another is the index for Clippy, and then is a repository idiomatic rust that you can search for. The API guidelines provide probably hundreds, at least dozens of really useful and practical tips for being able to use trust effectively. Clippy provides a very definitely provides hundreds and hundreds of programmatic checks for your code, along with explanations in all of them. So, for example, in this case, we're checking to see whether or not we are comparing for known constants that are already defined inside the standard library. By following Clippy's recommendations, we actually get much more closer to our intent with and we remove bugs from our code. Okay, so now I've got acknowledgments out of the way. I would also like to define zero cost magic doesn't actually exist. We can't get something for nothing. So zero cost, in the sense of zero cost abstractions actually means zero additional cost. You couldn't have written something better yourself. Possibly a more technically precise way to express this would be to say zero marginal cost. Lastly, I want to also point out that zero runtime zero cost also relates to runtime performance, and therefore we can trade off compile time execution if it will provide us with faster runtime code. That is, builds can take longer if the program will run faster. Okay, let's start off with a couple of very small quick tips. Some really useful literals exist in the trust language. One of them I quite like is the byte literal for being able to encode ASCII as an integer as a u eight integer, rather than as the can or char type, which takes four bytes. If we prefix a character literal with a b, we get a u eight value. Capital a is the number 65 as u eight value. There is a similar way for us to be able to keep two inject unicode literals, sorry, unicode. Inside our source code. With the slash u, we can add in any code point rather than encoding the actual source literal ourselves. Okay, so now let's touch on idiomatic trust. We had a look at clippy before, but what we really want to do is enable the rest of our team to follow along with our code. And we do that by maintaining the conventions of the ecosystem. Programming is a team sport, and writing software is easier than reading software. So by that I mean it's harder to follow along someone else's mental model or someone else's thinking when you're just reading the source code. In trust, I would say that getters and setters are kind of generally not useful. We have traits for interfaces, we don't use inheritance, and so therefore the same benefits, if there were any. I'm not confident that there were the same benefits of Java style getters and setters are sort of not available in trust. Like all rules, or most rules, at least there's an exception. And that is if you have, let's say a wrapper type that provides access to one logical thing, a get, or might be might be useful. So for example, from the standard library we have these non zero types which provide access. Two, the raw value and cell also does a similar thing. They have a get method which returns the inner type. Are there other conventions? I've touched on the API guidelines right at the start, when you are dealing with conversions of your types, you want to use the right method. By that I mean either as or to or into, depending on how you are performing the conversion, as well as sticking with the conventions of the ecosystem relating to generating iterators over some collection. You should also eagerly implement rates if your type is able to be compared for equality with other types. Implement partial equality. Implement clone where you can, or partial, or where it's appropriate, because your consumers or the people that are importing your code cannot implement those traits themselves. For your type, it's impossible for them to take a foreign trait like has and a foreign type. Let's say something that you've created and implement hash. For that foreign type. To do so, they would need to create a new type around it and it's just kind of annoying. So where you can, you should implement as much as possible. Want to talk a little bit about actually creating some other practices which will lend themselves to quality software? There are linters inside the ecosystem. We've talked about Clippy before, and to invoke it we use cargo. Clippy rust format is available via cargo format inside your crate. For extra points, you can make things harder for yourself by adding a pre commit hook. This will actually you could ask git to run cargo format on your behalf and fail if their formatting isn't correct. In fact, you can ask cargo format to update the code itself potentially, which might, may or may not be whats you want. So to set up the hook, you first create a file called pre commit or pre commit inside the hooks directory of your hidden git directory, make it executable, and then git will invoke it every time you run git commit. Here's an example. We need the hashmap syntax at the start because we didn't provide a file extension. I like to add a comment saying what this file actually does, and I've kind of got a cheat available, which is that if I want to skip any of these checks, I can just invoke git commit with cheat equals one, and then the pre commit will pass. Otherwise I ensure that the formatting is correct. I run the rust compiler under its check mode, which is a fast version of the compiler, and then I run Clippy, which is a more lengthier thing to check. At first I'm checking formatting, then I'm checking that the code compiles, and then I'm checking that the code is ergonomic. If those all pass, I'm allowed to commit my code. Something else relate, which is a little bit more complex, is when you define a function and you are taking a reference to some type, try to remember to take a reference to the borrowed version of the type rather than borrowing an own version. So for example, instead of borrowing a string, we borrow a string, or we take a string slice as an argument. It'll be easier for your callers. It turns out, for technical reasons relating to Rust's dref trait, that if you accept a string slides, you'll actually enable your callers to call your function with a reference to a string, because a reference to a string actually dereferences to an STR with a lowercase s. So what we want to do is change this string, and then all we should need to be able two do here is just change string into stir, and then we can now suddenly call is all caps with oh, that's not correct, and we're done. It's not always going to be this simple. Some other code will be slightly more complex. The same thing is, just to be very clear, we want to avoid things like taking a reference two a box of, let's say t. What we shouldn't prefer to say is just take the reference to a slice of t. The second form here is going to be easier on your callers and will still enable people who have boxed the type to be given access to it. Okay, sorry if that's confused. Some people let's reduce the cost of monomorphization. That sounds a bit crazy. Every specialized function that your code creates or your compiler creates whenever you use generics, and instantiate whenever you use generics, the compiler will generate a specialized version of a function for all of its input types. This takes up space. We can actually avoid some of that space by being a bit sneaky. Let me show what I mean. I've got a function here is all caps, which is actually doing exactly the same thing AWS before, but now we also want to accept a cow for whatever reason, just because it's the type in the standard library. That is the most fun to say. Now I've got here a crazy looking generic type which says that I'll accept a reference to any t, whether t de references to str to a string lowercase s. This will create a new version of is all caps for all of its input types. What I'm going to do instead is duplicate the function two start with and then change it to change the one that I want to call to not being a generic. And then I'm going to give it an underscore to indicate thats it's private. And then in the public method, although public because I'm not using the pub keyword, although I can I just call is all caps with the prefix underscore. So now what's triplicated? Or if I had say a cow and then a string and then a string slice. Only this smaller fraction of the code itself is actually triplicated. But there's only one version of the calling of the function that ends up being the body of the work. But there's a trick. We need to ask rust not to inline the code, otherwise things code will all get injected to all of the specific versions of the code that we want to create. So that's good. Okay, so now we have reduced the cost of monomorphization, at least the space cost. Let's talk a little bit about testing. And I say testing generally because actually what I really want to speak about is formal methods. If you have people using your software, you really don't want that software to break. And formal methods, including formal verification, are probably still in the difficult, like in the spectrum of easy to difficult. They're probably still in the difficult slides, but they are coming towards easy may be too strong. They're kind of moving towards learnable. Modeling involves spending time. So we kind of create a specification for how we want our program to run. And then the first step to kind of using modeling is actually two just write a document describing how you want the program to work. This could be called readme driven development. You kind of write the documentation before the code. This gives you time to think it through and design your API without actually having anyone dependent on your code. There are many ways that are more advanced than that, including formal verification, other languages, but writing a document actually will save you intense amount of debugging time later on. I make that as a claim without evidence, so if you disagree with that, please fire up in the comments. That's fine. I would like to make this claim a form of testing that's less common than unit tests, but it's still really valuable. Is this thing called property testing in rust, the prop test crate provides the ability for you to test a range of inputs for functions rather than just one test at a time or, sorry, a given unit test. Instead of giving it one specific input and checking its output, you can trust ask for random inputs within code range. It's actually very interesting and quite revealing sometimes about it finds the edge cases for you. And if you really want to find the edge cases, then you fuzz your inputs. A fuzzer is a program which generates random inputs for your functions. Now this sounds really complicated and difficult. It turns out though that fuzzing libraries actually make it relatively easy for you to just fuzz a specific function of your program. And you kind of write these test handlers, kind of like these baby programs thats call one function of your API and then just kind of give it random input like things thats should never ever appear in practice and see what breaks. And every time it breaks your program gets stronger because you fix the problem. You fix the problem, right? Okay, so lastly, maybe not. Lastly, there's actually plenty of slides to go. The new type pattern is very very handy and is very confusing or alien to people who come to rust. I've got this problem and that is we've got two thermometers reading the temperature and they actually are reading the same temperature. This is 20 degrees celsius. And turns out that 20 degrees celsius is also an integer in Fahrenheit, which is quite fun to know. So what we really want when we calculate our average temperature is 20 degrees celsius or 68 degrees fahrenheit, depending on if you are one of the one countries I believe that uses Fahrenheit in the world anyway, but instead we get 44. Now this is not a problem with the type system. These are all floating point values. Trust compiler dozens not care really that you have made like a logical error. What we do instead is we wrap f 32 or floating point value in our unit. Now, when we go to compare them or add them together and then divide, the program will fail to compile. Now, you may be wondering, is failing to compile really preferable to crashing or no, is failing to compile really preferable? And the answer is yes, actually, you don't want things to just silently work that are broken. Like it's better, two, not have the thing start than to things for something to fall over half the way through and even worse, fall over without you being ever told because you get a valid input, a valid output when you try to create the average temperature. One subtle way to improve this even further, even further is two, say that booleans and options should really be returned from functions. Now that's because if you assign this boolean from this call to is alive or it's a method, I suppose, and you would get a bool out, it's like true. Further later on in your code, you now need to wonder, why do I have a boolean in my code ten lines later? It kind of becomes a little bit confusing. You need like a variable that's well named, and if it's not, then you've just kind of got a crazy variable that you like. True or false? What does that actually mean? It's hard to know. It might be you're comparing a quality with something. So instead we can change from an is underscore method. So that's a method that returns a boolean. Instead we return just our own enum. And this way it doesn't matter where our variable appears. It's always going to refer to. It's always going to be sort of self documentating that it's a value that encodes the state of whether or not something is valid or alive or not in itself. Oh, this is actually a lie. It doesn't have self in there. Nearly version of this slide actually included type parameter in lifecycle. But then I thought that people might get picky and complain that I also needed to implement clone and so forth, so I should have fixed that up. I like market traits. So this is a really nice part of the type system and really focuses on this idea of a zero cost abstraction. And that's partially because I think traits are kind of one of the centerpieces of rust. Within the standard library, you get a small set of marker traits like copy, send, sync, sync, copy, send, size, synced and unpin that live within the standard marker module. And it's hard to read there, but actually at the bottom it's saying that size is so common that you actually need to opt out of it in your own types. So the compiler will always derive size for you unless you ask it not to explicitly. The idea behind these types is that they are providing information to the type system that takes up no internal representation in the final binary or final executable itself. So they take up zero space, and they are therefore zero cost in our definition. From the start of the page start of the talk, we can get further with zero cost abstractions. It turns out that using option wisely, if you are dealing with references, is super, super handy. Stupid thing to say. If you wrap a reference type that is a ampersand t or a box t in option, it takes up no more space in memory. That's because rust guarantees that option, sorry, that references which are pointers are never null. So there's a free bit pattern available that can be used for pattern matching. We can make our rust code easier to use by avoiding a couple of gotchas. We want to avoid magical typecasting with draf. So if you remember a couple of slides ago, we talked about the dereference trait and that we kind of were able to accept multiple types. So this was a reference to a string versus a string slides. And it turns out that you can base this to create something called draf polymorphism. It's a known antipattern. Now this is a more difficult code example, but let's see if I can show you what the problem is. Whats I might want to do if I was to come from an object oriented programming language is to kind of use draf to recreate inheritance. I've got here a base class with a greet method, and then I have a person class that has inside of it a member base. And the Draf implementation for person involves returning self dot base. So when the person is dereferenced, it returns this base class and the base class has an object. Now if I have a main method, I can use some sort of one of the magical things about rust thats is this dot operator. Implicitly, dereferences on demand will enable me to call a greet method from Greeter just by because person has implemented draf, we do not want to abuse this. It's like a very sharp edge feature. It will cause problems and it's going to be very very confusing. We have two ways out of this. We can either use a trait. So trait greet will enable us to say that the person can implement greet and actually rely on the default method itself. We don't even need any extra code. And now we can, we now we can call greet from person. In fact, we've simplified a bunch of other stuff because we don't need multiple types and so forth. It alternatively, if we really want our base class, if we really wanted to kind of create something which has kind of templated, we need to write the facade ourselves. So we just create a greet method and then we're given access. Two, the base greet, you see, thats the person's greet implementation is just dispatched to the internal class. And now if I have main greeter greet does the right thing. It's slightly less ergonomic than breaking the rules, but please don't break the rules. It really will cause problems. I promise to improve your rust code. You should make it impossible to create like partially constructed types. It's really tempting to add like an is valid method or validate step, but actually doing so will just create mistakes because people are lazy and will forget people are humans. We can actually encode the validation logic inside our constructors and return result rather than t. This is a form of programming called making illegal states unrepresentable. But in this case, I have sort of a building and a height. Now the building's height must be nonzero, and what I want to avoid is validation. So in order to get around this, I have some cheater code. In the new method. I actually I'll just copy and paste and I'll go back a slide. My new method becomes a we return result, and if height is nothing, we return zero height. So I needed an error type in there also, and otherwise I return. Okay, this makes it impossible for there ever to be a valid building object around that has the illegal type. Now there is a little bit of extra bureaucracy around because we need kind of a result type and we need an error type. By the way, if you have implemented debug, then you have implicitly implemented error as well, because there's an automatic, but that is, and this isn't spelled correctly, but this will provide us with significantly stronger, more robust software. I got a couple of other pieces of advice before we wrap up. You've probably heard of generics, which is static dispatch, and trait objects, which dynamic dispatch. But I'm here to tell you today that there's a third way called enum dispatch. We create can enum type, which encodes, which kind of encapsulates all the possible states or all the morphisms, all the types that our thing could be, and then we match on it inside the calling code. The downside is that it becomes slightly unwieldy to use inside the functions that make use of this kind of supertype. You need to match on every single instance of it. Now to make it smoother, there is a crate called enum dispatch which takes a lot of that pain away. It's much, much faster. And some of the benchmarks that this crate provides we're talking about sort of ten times performance gains even inside the static dispatch case, which seems kind of crazy. But I encourage you two take a look and see if your trait has more than two or three methods. There's probably an opportunity for you to refine your design. There are very few traits inside the rust standard library that have more than one method that you need to implement. When they do have multiple traits. Sorry, multiple methods inside the one trait. A lot of them are provided via default implementations. If you have a trait thats is very narrow but deep, you'll find that it's much easier for your callers to make use of, versus an API that is kind of broad but shallow is more specific. It's not general enough, and you'll find that it's really difficult for people that are using your trait to make use of. That's everything that I have for you today. I really hope that you have enjoyed the talk. Hopefully I've said a few things that you've agreed with, a few things that are new, and possibly even a few things that you disagree with. You are very welcome to say hi in the comments and let's see if we can start a discussion. Hit me up on Twitter and let's see where we can go. Take care.

Tim McNamara

Senior Software Engineer @ AWS

Tim McNamara's LinkedIn account Tim McNamara's twitter account

Awesome tech events for

Priority access to all content

Video hallway track

Community chat

Exclusive promotions and giveaways