RFC: expose-fn-type

text/3476-expose-fn-type.md

@@ -0,0 +1,269 @@
+- Feature Name: `expose-fn-type`
+- Start Date: 2023-08-20
+- RFC PR: [rust-lang/rfcs#3476](https://github.com/rust-lang/rfcs/pull/3476)
+- Rust Issue: N/A
+
+# Summary
+[summary]: #summary
+
+This exposes the function type of a function item to the user.
+
+# Motivation
+[motivation]: #motivation
+
+### DasLixou
+
+I was trying to make something similar to bevy's system functions. And for safety reasons, they check for conflicts between SystemParams, so that a function requiring `Res<A>` and `ResMut<A>` [panic](https://github.com/bevyengine/bevy/blob/main/crates/bevy_ecs/src/system/system_param.rs#L421).
+
+Then after I heard about axum's [`#[debug_handler]`](https://docs.rs/axum/latest/axum/attr.debug_handler.html) I wanted to do something similar to my copy of bevy systems, so that I get compile time errors when there is a conflict. I wanted even more, I wanted to force the user to mark the function with a specific proc attribute macro in order to make it possible to pass it into my code and call itself a system.
+
+For that, I would need to mark the type behind the function item, for example, with a trait.
+
+### madsmtm
+
+In Swift, some functions have an associated selector that you can access with [`#selector`](https://developer.apple.com/documentation/swift/using-objective-c-runtime-features-in-swift).
+
+In my crate `objc2`, it would be immensely beautiful (and useful) to be able to do something similar, e.g. access a function's selector using something like `MyClass::my_function::Selector` or `selector(MyClass::my_function)`, instead of having to know the selector name (which might be something completely different than the function name).
+
+# Terminology
+
+I'll may shorten `function` to `fn` sometimes.
+
+- **function pointer**: pointer type with the type syntax `fn(?) -> ?` directly pointing at a function, not the type implementing the `Fn[Once/Mut](?) -> ?` traits.
+- **function item** (or just function): a declared function in code. free-standing or associated to a type.
+- **function group**: many non-specific functions with the same signature (params, return type, etc.)
+- **function trait(s)**: the `Fn[Once/Mut](?) -> ?` traits
+- **function type**: the type behind a function, which also implements the function traits.
+- **fixed type**: directly named type, no generic / `impl Trait`.
+- **describe the function type**: write `fn(..) -> ? name` instead of just `fn name`.
+
+# Guide-level explanation
+[guide-level-explanation]: #guide-level-explanation
+
+As we all know, you can refer to a struct by its name and for example implement a trait
+```rust
+struct Timmy;
+impl Person for Timmy {
+    fn greet() {
+        println!("Hey it's me, Timmy!");
+    }
+}
+```
+When we want to target a specific function for a trait implementation, we somehow need to get to the type behind it. 
+Refering to the hidden type is achieved via the following syntax
+```rust
+fn is_positive(a: i32) -> bool { /* ... */ }
+impl MyTrait for fn(i32) -> bool is_positive {
+    /* ... */
+}
+```
+For function signatures, where every parameter/return-type is a fixed type and can be known just by refering to the function (so no generics or `impl Trait` parameters/return type), we can drop the redundant information:
+```rust
+fn is_positive(a: i32) -> bool { /* ... */ }
+impl MyTrait for fn is_positive {
+    /* ... */
+}
+```
+
+> 💡 NOTE: Even when we need to describe the function type but the return type is `()`, we can (just as for function pointers and function traits) drop the `-> ()` from the type. (This should also be added as a lint).
+
+---
+A function with a more complex signature, like a function that specifies `const`, `unsafe` or `extern "ABI"`, we just ignore that when naming the type:
+```rust
+const fn my_fn(a: i32) -> (i16, i16) { .. }
+impl MyTrait for fn my_fn {}
+// or with explicit declaration
+impl MyTrait for fn(i32) -> (i16, i16) my_fn { .. }
+```
+
+When having an async function, we in theory have a `impl Future<Output = ..>` as a return type, which should force us to explicitly declare the function type like so
+```rust
+async fn request_name(id: PersonID) -> String { .. }
+
+impl<F: Future<Output = String>> Requestable for fn(PersonID) -> F request_name {
+    /* ... */
+}
+```
+We can take a shortcut and use the `async` keyword, as long as the `Output` assoc type in the Future is still fixed
+```rust
+async fn request_name(id: PersonID) -> String { .. }
+
+impl Requestable for async fn request_name {
+    /* ... */
+}
+```
+
+# Reference-level explanation
+[reference-level-explanation]: #reference-level-explanation
+
+As described in the **Guide-level explanation**, with the syntax `[async] fn[(..) -> ?] <fn_path>`, we can reference the type behind the named function.
+
+When the function is for example in a different mod, it should be referenced by its path
+```rust
+mod sub {
+    fn sub_mod_fn() { .. }
+}
+trait MyTrait {}
+impl MyTrait for fn sub::sub_mod_fn {
+    /* ... */
+}
+```
+
+> ⚠️ NOTE: The same rules apply here as for normal types. Either the function item or the trait to implement mustn't be foreign for the impl. [Same as E0210](https://github.com/rust-lang/rust/blob/master/compiler/rustc_error_codes/src/error_codes/E0210.md)
+
+---
+
+It should be also possible to get the type of associated functions:
+
+```rust
+struct MyStruct;
+impl MyStruct {
+    fn new() -> Self { Self }
+}
+impl fn MyStruct::new {
+    /* ... */
+}
+```
+
+When the associated function comes from a trait, the same rules as for associated types apply here ([Ambiguous Associated Type, E0223](https://github.com/rust-lang/rust/blob/master/compiler/rustc_error_codes/src/error_codes/E0223.md)):
+
+```rust
+struct MyStruct;
+type MyTrait {
+    fn ambiguous();
+}
+impl MyTrait for MyStruct {
+    fn ambiguous() {  }
+}
+impl fn MyStruct::ambiguous { } // ERROR: ambiguous associated function
+// instead:
+impl fn <MyStruct as MyTrait>::ambiguous { } // OK
+```
+
+When the type of the associated function has generics, they will be handles as follows
+
+```rust
+struct MyStruct<T>(T);
+impl<T> MyStruct<T> {
+    fn get() -> T { .. }
+}
+
+impl<T> fn MyStruct<T>::get { }
+// or fully described:
+impl<T> fn() -> T MyStruct<T>::get { }
+```
+
+---
+
+When a function has generics, the function type is forced to be described, and the generic should be placed at it's desired position:
+```rust
+fn send<T: Send>(val: T, postal_code: u32) {}
+impl<T: Send> ParcelStation for fn(T, u32) send {
+    /* ... */
+}
+```
+
+When we have an implicit generic, the same rule applies
+```rust
+fn implicit_generic(val: impl Clone) -> impl ToString {}
+impl<T: Clone, U: ToString> for fn(T) -> U implicit_generic {
+    /* ... */
+}
+```
+
+---
+
+When functions have lifetimes, they have to be included in the types
+```rust
+fn log(text: &str) { .. }
+impl<'a> Logger for fn(&'a str) log {
+    /* ... */
+}
+```
+
+When the lifetime is explicitly defined on the function signature and there's no other rule forcing us to describe the function type, we can take a shortcut as follows
+```rust
+fn log<'a>(text: &'a str) { .. } // explicit lifetime 'a
+impl<'a> Logger for fn<'a> log {
+    /* ... */
+}
+```
+
+---
+
+Just as structs and enums have the possibility to derive traits to automatically generate code, function type have similar ways via attribute macros:
+
+```rust
+#[debug_signature]
+fn signature_test(val: i32) -> bool {
+    /* ... */
+}
+
+// Expands to
+
+fn signature_test(val: i32) -> bool {
+    /* ... */
+}
+impl DbgSignature for fn signature_test {
+    fn dbg_signature() -> &'static str {
+        "fn signature_test(val: i32) -> bool"
+    }
+}
+```
+
+Other than that, it should behave like every other type does.
+
+# Additional ToDo's
+
+## Change the fn type syntax for consistency
+
+When we try to compile the current [code snippet](https://play.rust-lang.org/?version=stable&mode=debug&edition=2021&gist=cba2a1391c3e431a7499c6bf427b350d) 
+```rust
+fn cool<'a, T: Clone>(val: &'a T) -> (i32, bool) {
+    todo!()
+}
+
+fn main() {
+    let _a: () = cool;
+}
+```
+we get following error:
+```
+error[E0308]: mismatched types
+ --> src/main.rs:6:18
+  |
+6 |     let _a: () = cool;
+  |             --   ^^^^ expected `()`, found fn item
+  |             |
+  |             expected due to this
+  |
+  = note: expected unit type `()`
+               found fn item `for<'a> fn(&'a _) -> (i32, bool) {cool::<_>}`
+
+For more information about this error, try `rustc --explain E0308`.
+```
+For consistency, we should change the syntax to `for<'a, T: Clone> fn(&'a T) -> (i32, bool) cool` (I'm not sure if we should put generics in the for)
+
+# Drawbacks
+[drawbacks]: #drawbacks
+
+# Rationale and alternatives
+[rationale-and-alternatives]: #rationale-and-alternatives
+
+The type behind functions already exists, we just need to expose it to the user.
+
+# Prior art
+[prior-art]: #prior-art
+
+i dont know any
+
+# Unresolved questions
+[unresolved-questions]: #unresolved-questions
+
+- Is the syntax good? It could create confusion between a function pointer.
+- What about closures? They don't even have names so targetting them would be quite difficult. I wouldn't want to use the compiler generated mess of a name like `[closure@src/main.rs:13:18: 13:20]`. It would also contain line numbers which would be changing quite often so thats not ideal.
+
+# Future possibilities
+[future-possibilities]: #future-possibilities
+
+- Also expose the type of closures