As a side note, there is a libffi wrapper on Rust that is exactly leveraging this code and data separation of closures in Rust: https://docs.rs/libffi
I've been using this on my vibewasm project to provide host function conversion to keep a C callable function in the front surface but doing my own custom wasm calling convention while capturing a persistent context pointer to the wasm store.
There is a side effect though: it is essentially unsound as you have to leak the object to the libffi closure which is a form of JIT -- dynamic code generation, it is, meaning Rust will have no way of knowing the lifetime of the pointer, or you have to always keep the libffi closure alive, meaning it is a permanent leak. I tried mitigate this by storing the closure in the wasm store, and we use that as the ultimate closure lifetime by designation -- any libffi function callback post store destruction is undefined behavior though.
The biggest friction I experience with respect to rust closures is their inability to be generic: I cannot implement a method that takes a closure generic over its argument(s).
So then I'm forced to define a trait for the function, define a struct (the closure) to store the references I want to close over, choose the mutability and lifetimes, instantiate it manually and pass that. Then the implementation of the method (that may only be a few lines) is not located inline so readability may suffer.
if I'm not mistaken (and I very well may be!) my primary confusion with closures comes from the fact that: the trait they implement (FnOnce / Fn / FnMut) depends entirely upon what happens inside the closure.
It will automatically implement the most general, relaxed version (FnMut I think?) and only restrict itself further to FnOnce and Fn based on what you do inside the closure.
So, it can be tricky to know what's going on, and making a code change can change the contract of the closure and therefore where and how it can be used.
(I invite rust experts to correct me if any of the above is mistaken - I always forget the order of precedence for FnOnce/Fn/FnMut and which implies which)
The three Fn* types correspond to the three ways you can refer to a value: &T, &mut T, T. Fn captures its environment by shared reference, FnMut by exclusive reference, and FnOnce by value, and everything flows from that. Calling a Fn is the same as using a reference. Calling a FnMut is the same as using a mutable reference (you can do it as many times as you want but no two uses may overlap in time). And calling a FnOnce is the same as moving a value (you can do it at most once).
> I always forget the order of precedence for FnOnce/Fn/FnMut
The way I remember the ordering is by thinking about the restrictions the various Fn traits provide from a caller's perspective:
1. FnOnce can only ever be called once and cannot be called concurrently. This is the most restrictive.
2. FnMut can be called multiple times but cannot be called concurrently. This is less restrictive than FnOnce.
3. Fn can be called multiple times and can even be called concurrently. This is the least restrictive.
So going from most to least restrictive gives you `FnMut: FnOnce` and `Fn: FnMut`.
Fn can only be called concurrently if its environment is Sync, which is often true but not necessarily.
It’s more precise to say that Fn can be called even when you only have shared access to it, which is a necessary, but not sufficient, condition for being able to be called concurrently.
`Fn`, `FnMut`, and `FnOnce` can also implement and not implement `Sync` (also `Send`, `Clone`, `Copy`, lifetime bounds, and I think `use<...>` applies to `impl Fn...` return types).
The least restrictive for the caller is Fn (you can call it whenever), then FnMut (you can call it only if you have exclusive access to it, as many times as you want), then FnOnce (you can call it only if you have exclusive owned access, and calling it once destroys it).
The least restrictive for the function itself is the opposite order: FnOnce (it can do anything to its environment, including possibly consuming things without putting them back into a consistent state), followed by FnMut (it has exclusive access to its environment, and so is allowed to mutate it, but not destroy it), followed by Fn (it has only shared access to its environment and therefore is not allowed to mutate it).
Since these orders are inverses of each other, functions that are easier to write are harder to call and vice versa. That’s why they implement the trait with the minimum amount of power possible, so that they can be called in more places.
It isn't really type inference. Each closure gets a unique type. Rather it's an automatic decision of what traits (think roughly "superclasses" I guess if you aren't familiar with traits/typeclasses) to implement for that type.
No, I don't think so, not unless there's some feature of Haskell type classes I'm completely unaware of.
If anything it's closer to SFINAE in C++ where it tries to implement methods but then doesn't consider it an error if it fails. Then infers type-classes based on the outcome of the SFINAE process. Or the macro analogy another poster made isn't bad (with the caveat that it's a type system aware macro - which at least in rust is strange).
I am not sure how Haskell works but I think what the previous poster meant is that the types get determined at compile time. Closures are akin to macros except you can't see the expanded code.
My issue is that there is no easy way to know the signature generated by the closure (unlike a function) until you read the compiler errors and even then it's some cryptic mess because closures are anonymous structs. Or maybe I missed some LSP config/extension?
This isn't the right framing IMO. Closures actually aren't complicated with GC for the same reason structs with references aren't complicated, at least as far as the programmer is concerned. You could say functional languages "hide the mess" there too, but even if you take that perspective, it's nothing to do with closures in particular. Closures are just one of the things that need memory, and memory management is tricky without GC.
Functional programming languages usually don't support linear/affine types, non-gc references and mutations.
Their closures are essentially the equivalent of Rust's `Rc<dyn Fn(...) -> ...>` with some sugar for expressing the function type and hiding all the `.clone()`s needed.
It's easy to get simplier results if you support less features and use cases.
Closures are pretty simple in relation to their captures lifetimes, but they do have a lot of complexity in how the lifetimes of their argument and return type are computed. The compiler has to basically infer them, and that can easily go very wrong. The only reason it works most of the time is because closures are immediately passed to functions whose trait bound specify the expected signature of the closure, but once you deviate a little bit from the common case things start to break down. For example if the bound is `F: SomeTrait` where `SomeTrait` is implemented for `FnOnce(&' i32) -> &i32` the inference will break. Similarly if you store the closure in a local variable before passing it to the function. This used to come up pretty often for "async" closures that were supposed to take a reference as input, since it's impossible to specify their correct trait bound using directly the `Fn*` traits. There are a bunch of related issues [1] in the rustc repo if you search for closure and higher ranked lifetimes.
I really wanted just yesterday to create a dyn AsyncFnMut, which apparently still needs async-trait to build the stable. but I was pretty much unable to figure out how to make that work with a lambda. saying this is all trivial once you understand the borrow machinery is really understating it.
The comment above isn't saying that closures are trivial. Once you understand the borrow checker, you understand that it's a miracle that closures in Rust can possibly work at all, given Rust's other dueling goals of being a GC-less language with guaranteed memory safety despite letting closures close over arbitrary references. Rust is in uncharted territory here, drawing the map as it goes.
As a side note, there is a libffi wrapper on Rust that is exactly leveraging this code and data separation of closures in Rust: https://docs.rs/libffi
I've been using this on my vibewasm project to provide host function conversion to keep a C callable function in the front surface but doing my own custom wasm calling convention while capturing a persistent context pointer to the wasm store.
There is a side effect though: it is essentially unsound as you have to leak the object to the libffi closure which is a form of JIT -- dynamic code generation, it is, meaning Rust will have no way of knowing the lifetime of the pointer, or you have to always keep the libffi closure alive, meaning it is a permanent leak. I tried mitigate this by storing the closure in the wasm store, and we use that as the ultimate closure lifetime by designation -- any libffi function callback post store destruction is undefined behavior though.
The biggest friction I experience with respect to rust closures is their inability to be generic: I cannot implement a method that takes a closure generic over its argument(s).
So then I'm forced to define a trait for the function, define a struct (the closure) to store the references I want to close over, choose the mutability and lifetimes, instantiate it manually and pass that. Then the implementation of the method (that may only be a few lines) is not located inline so readability may suffer.
You most definitely can.
Or did I not understand what you meant?Something like this isn’t possible
The workaround:Do you have an example of this? I'm not sure I follow it exactly.
if I'm not mistaken (and I very well may be!) my primary confusion with closures comes from the fact that: the trait they implement (FnOnce / Fn / FnMut) depends entirely upon what happens inside the closure.
It will automatically implement the most general, relaxed version (FnMut I think?) and only restrict itself further to FnOnce and Fn based on what you do inside the closure.
So, it can be tricky to know what's going on, and making a code change can change the contract of the closure and therefore where and how it can be used.
(I invite rust experts to correct me if any of the above is mistaken - I always forget the order of precedence for FnOnce/Fn/FnMut and which implies which)
The three Fn* types correspond to the three ways you can refer to a value: &T, &mut T, T. Fn captures its environment by shared reference, FnMut by exclusive reference, and FnOnce by value, and everything flows from that. Calling a Fn is the same as using a reference. Calling a FnMut is the same as using a mutable reference (you can do it as many times as you want but no two uses may overlap in time). And calling a FnOnce is the same as moving a value (you can do it at most once).
> I always forget the order of precedence for FnOnce/Fn/FnMut
The way I remember the ordering is by thinking about the restrictions the various Fn traits provide from a caller's perspective:
So going from most to least restrictive gives you `FnMut: FnOnce` and `Fn: FnMut`.Fn can only be called concurrently if its environment is Sync, which is often true but not necessarily.
It’s more precise to say that Fn can be called even when you only have shared access to it, which is a necessary, but not sufficient, condition for being able to be called concurrently.
I'm not sure myself, does Fn correspond to reentrancy or is there some detail I am missing?
`Fn`, `FnMut`, and `FnOnce` can also implement and not implement `Sync` (also `Send`, `Clone`, `Copy`, lifetime bounds, and I think `use<...>` applies to `impl Fn...` return types).
EDIT: https://news.ycombinator.com/item?id=46750011 also mentioned `AsyncFn`, `AsyncFnMut`, and `AsyncFnOnce`.
Easiest mnemonic to remember precedence is simply ordering by the length of their names.
FnOnce
FnMut
Fn
The least restrictive for the caller is Fn (you can call it whenever), then FnMut (you can call it only if you have exclusive access to it, as many times as you want), then FnOnce (you can call it only if you have exclusive owned access, and calling it once destroys it).
The least restrictive for the function itself is the opposite order: FnOnce (it can do anything to its environment, including possibly consuming things without putting them back into a consistent state), followed by FnMut (it has exclusive access to its environment, and so is allowed to mutate it, but not destroy it), followed by Fn (it has only shared access to its environment and therefore is not allowed to mutate it).
Since these orders are inverses of each other, functions that are easier to write are harder to call and vice versa. That’s why they implement the trait with the minimum amount of power possible, so that they can be called in more places.
This is correct. But it’s not really surprising, it’s type inference.
It isn't really type inference. Each closure gets a unique type. Rather it's an automatic decision of what traits (think roughly "superclasses" I guess if you aren't familiar with traits/typeclasses) to implement for that type.
So you're saying... it's type inference of type classes, just like in Haskell?
No, I don't think so, not unless there's some feature of Haskell type classes I'm completely unaware of.
If anything it's closer to SFINAE in C++ where it tries to implement methods but then doesn't consider it an error if it fails. Then infers type-classes based on the outcome of the SFINAE process. Or the macro analogy another poster made isn't bad (with the caveat that it's a type system aware macro - which at least in rust is strange).
I am not sure how Haskell works but I think what the previous poster meant is that the types get determined at compile time. Closures are akin to macros except you can't see the expanded code.
My issue is that there is no easy way to know the signature generated by the closure (unlike a function) until you read the compiler errors and even then it's some cryptic mess because closures are anonymous structs. Or maybe I missed some LSP config/extension?
I wish there was more customizability with regards to captures.
the move keywords captures everything. Sometimes I want a little bit more flexibility, like C++ lambdas.
This article discusses making captures more flexible: https://smallcultfollowing.com/babysteps/blog/2025/10/22/exp...
I agree it would be nice, in particular to make it easier to understand when learning the concept.
And now we have the Async version of each of those, which makes me very happy and reduces the shenanigans and lifetime issues.
Also worthy of note is that there is talk to add a syntax for explicit captures.
Closures are the bread and butter of functional programming, but Rust made closures a complicated mess.
Closures are a complicated mess. Functional programming languages hide the mess with garbage collection.
This isn't the right framing IMO. Closures actually aren't complicated with GC for the same reason structs with references aren't complicated, at least as far as the programmer is concerned. You could say functional languages "hide the mess" there too, but even if you take that perspective, it's nothing to do with closures in particular. Closures are just one of the things that need memory, and memory management is tricky without GC.
Machine code and LLVM are complicated messes. Higher-level language hide a lot, but sometimes issues pop up, even in Rust e.g. inline heuristics (https://nnethercote.github.io/perf-book/inlining.html).
Functional programming languages usually don't support linear/affine types, non-gc references and mutations.
Their closures are essentially the equivalent of Rust's `Rc<dyn Fn(...) -> ...>` with some sugar for expressing the function type and hiding all the `.clone()`s needed.
It's easy to get simplier results if you support less features and use cases.
If you understand the borrow checker, closures are just not that much on top of things.
In fact I can’t remember the last time I had to fight with them.
Closures are pretty simple in relation to their captures lifetimes, but they do have a lot of complexity in how the lifetimes of their argument and return type are computed. The compiler has to basically infer them, and that can easily go very wrong. The only reason it works most of the time is because closures are immediately passed to functions whose trait bound specify the expected signature of the closure, but once you deviate a little bit from the common case things start to break down. For example if the bound is `F: SomeTrait` where `SomeTrait` is implemented for `FnOnce(&' i32) -> &i32` the inference will break. Similarly if you store the closure in a local variable before passing it to the function. This used to come up pretty often for "async" closures that were supposed to take a reference as input, since it's impossible to specify their correct trait bound using directly the `Fn*` traits. There are a bunch of related issues [1] in the rustc repo if you search for closure and higher ranked lifetimes.
[1]: https://github.com/rust-lang/rust/issues?q=is%3Aopen%20is%3A...
I really wanted just yesterday to create a dyn AsyncFnMut, which apparently still needs async-trait to build the stable. but I was pretty much unable to figure out how to make that work with a lambda. saying this is all trivial once you understand the borrow machinery is really understating it.
> saying this is all trivial
The comment above isn't saying that closures are trivial. Once you understand the borrow checker, you understand that it's a miracle that closures in Rust can possibly work at all, given Rust's other dueling goals of being a GC-less language with guaranteed memory safety despite letting closures close over arbitrary references. Rust is in uncharted territory here, drawing the map as it goes.
Async is the stuff that messes up everything. Closures are not complicated.
Well... Rust is not a functional language, so it is not surprising that its closures are complicated.