Trait objects like `Box` can only be constructed when certain requirements are satisfied by the trait in question. Trait objects are a form of dynamic dispatch and use a dynamically sized type for the inner type. So, for a given trait `Trait`, when `Trait` is treated as a type, as in `Box`, the inner type is 'unsized'. In such cases the boxed pointer is a 'fat pointer' that contains an extra pointer to a table of methods (among other things) for dynamic dispatch. This design mandates some restrictions on the types of traits that are allowed to be used in trait objects, which are collectively termed as 'object safety' rules. Attempting to create a trait object for a non object-safe trait will trigger this error. There are various rules: ### The trait cannot require `Self: Sized` When `Trait` is treated as a type, the type does not implement the special `Sized` trait, because the type does not have a known size at compile time and can only be accessed behind a pointer. Thus, if we have a trait like the following: ``` trait Foo where Self: Sized { } ``` We cannot create an object of type `Box` or `&Foo` since in this case `Self` would not be `Sized`. Generally, `Self: Sized` is used to indicate that the trait should not be used as a trait object. If the trait comes from your own crate, consider removing this restriction. ### Method references the `Self` type in its parameters or return type This happens when a trait has a method like the following: ``` trait Trait { fn foo(&self) -> Self; } impl Trait for String { fn foo(&self) -> Self { "hi".to_owned() } } impl Trait for u8 { fn foo(&self) -> Self { 1 } } ``` (Note that `&self` and `&mut self` are okay, it's additional `Self` types which cause this problem.) In such a case, the compiler cannot predict the return type of `foo()` in a situation like the following: ```compile_fail,E0038 trait Trait { fn foo(&self) -> Self; } fn call_foo(x: Box) { let y = x.foo(); // What type is y? // ... } ``` If only some methods aren't object-safe, you can add a `where Self: Sized` bound on them to mark them as explicitly unavailable to trait objects. The functionality will still be available to all other implementers, including `Box` which is itself sized (assuming you `impl Trait for Box`). ``` trait Trait { fn foo(&self) -> Self where Self: Sized; // more functions } ``` Now, `foo()` can no longer be called on a trait object, but you will now be allowed to make a trait object, and that will be able to call any object-safe methods. With such a bound, one can still call `foo()` on types implementing that trait that aren't behind trait objects. ### Method has generic type parameters As mentioned before, trait objects contain pointers to method tables. So, if we have: ``` trait Trait { fn foo(&self); } impl Trait for String { fn foo(&self) { // implementation 1 } } impl Trait for u8 { fn foo(&self) { // implementation 2 } } // ... ``` At compile time each implementation of `Trait` will produce a table containing the various methods (and other items) related to the implementation. This works fine, but when the method gains generic parameters, we can have a problem. Usually, generic parameters get _monomorphized_. For example, if I have ``` fn foo(x: T) { // ... } ``` The machine code for `foo::()`, `foo::()`, `foo::()`, or any other type substitution is different. Hence the compiler generates the implementation on-demand. If you call `foo()` with a `bool` parameter, the compiler will only generate code for `foo::()`. When we have additional type parameters, the number of monomorphized implementations the compiler generates does not grow drastically, since the compiler will only generate an implementation if the function is called with unparametrized substitutions (i.e., substitutions where none of the substituted types are themselves parameterized). However, with trait objects we have to make a table containing _every_ object that implements the trait. Now, if it has type parameters, we need to add implementations for every type that implements the trait, and there could theoretically be an infinite number of types. For example, with: ``` trait Trait { fn foo(&self, on: T); // more methods } impl Trait for String { fn foo(&self, on: T) { // implementation 1 } } impl Trait for u8 { fn foo(&self, on: T) { // implementation 2 } } // 8 more implementations ``` Now, if we have the following code: ```compile_fail,E0038 # trait Trait { fn foo(&self, on: T); } # impl Trait for String { fn foo(&self, on: T) {} } # impl Trait for u8 { fn foo(&self, on: T) {} } # impl Trait for bool { fn foo(&self, on: T) {} } # // etc. fn call_foo(thing: Box) { thing.foo(true); // this could be any one of the 8 types above thing.foo(1); thing.foo("hello"); } ``` We don't just need to create a table of all implementations of all methods of `Trait`, we need to create such a table, for each different type fed to `foo()`. In this case this turns out to be (10 types implementing `Trait`)\*(3 types being fed to `foo()`) = 30 implementations! With real world traits these numbers can grow drastically. To fix this, it is suggested to use a `where Self: Sized` bound similar to the fix for the sub-error above if you do not intend to call the method with type parameters: ``` trait Trait { fn foo(&self, on: T) where Self: Sized; // more methods } ``` If this is not an option, consider replacing the type parameter with another trait object (e.g., if `T: OtherTrait`, use `on: Box`). If the number of types you intend to feed to this method is limited, consider manually listing out the methods of different types. ### Method has no receiver Methods that do not take a `self` parameter can't be called since there won't be a way to get a pointer to the method table for them. ``` trait Foo { fn foo() -> u8; } ``` This could be called as `::foo()`, which would not be able to pick an implementation. Adding a `Self: Sized` bound to these methods will generally make this compile. ``` trait Foo { fn foo() -> u8 where Self: Sized; } ``` ### The trait cannot contain associated constants Just like static functions, associated constants aren't stored on the method table. If the trait or any subtrait contain an associated constant, they cannot be made into an object. ```compile_fail,E0038 trait Foo { const X: i32; } impl Foo {} ``` A simple workaround is to use a helper method instead: ``` trait Foo { fn x(&self) -> i32; } ``` ### The trait cannot use `Self` as a type parameter in the supertrait listing This is similar to the second sub-error, but subtler. It happens in situations like the following: ```compile_fail,E0038 trait Super {} trait Trait: Super { } struct Foo; impl Super for Foo{} impl Trait for Foo {} fn main() { let x: Box; } ``` Here, the supertrait might have methods as follows: ``` trait Super { fn get_a(&self) -> &A; // note that this is object safe! } ``` If the trait `Trait` was deriving from something like `Super` or `Super` (where `Foo` itself is `Foo`), this is okay, because given a type `get_a()` will definitely return an object of that type. However, if it derives from `Super`, even though `Super` is object safe, the method `get_a()` would return an object of unknown type when called on the function. `Self` type parameters let us make object safe traits no longer safe, so they are forbidden when specifying supertraits. There's no easy fix for this, generally code will need to be refactored so that you no longer need to derive from `Super`.