Repro:
trait O {
type M;
}
trait U<A: O> {
const N: A::M;
}
impl<D> O for D {
type M = u8;
}
impl<C: O> U<C> for u16 {
const N: C::M = 4;
}
This caused E0277 "the trait bound is not satisfied" error:
error[E0277]: the trait bound `C: std::marker::Sized` is not satisfied
--> src/main.rs:11:5
|
11 | const N: C::M = 4;
| ^^^^^^^^^^^^^^^^^^ `C` does not have a constant size known at compile-time
|
= help: the trait `std::marker::Sized` is not implemented for `C`
= help: consider adding a `where C: std::marker::Sized` bound
= note: required because of the requirements on the impl of `O` for `C`
The error doesn't make sense because C is obviously Sized. It seems the associated const type is evaluated before the generic bounds are added, causing this error.
Changing both A::M and C::M to u8 makes the error go away.
Reproducible on all 3 versions on playground (1.23.0, 1.24.0-beta, 1.25.0-nightly).
This exact issue just occured to @CryZe and me when trying to use associated constants to create a constant struct in a library with user-provided data (without const fns) (actual use case: playground). A minimized examples is this (playground):
use std::marker::PhantomData;
trait Const {
type T;
const VAL: Self::T;
}
struct MakeTuple<U>(PhantomData<U>);
impl<U: Const + Sized> Const for MakeTuple<U> {
type T = (U::T, U::T);
const VAL: Self::T = (U::VAL, U::VAL);
}
Error:
error[E0277]: the trait bound `U: std::marker::Sized` is not satisfied
--> src/main.rs:11:5
|
11 | const VAL: Self::T = (U::VAL, U::VAL);
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ `U` does not have a constant size known at compile-time
|
= help: the trait `std::marker::Sized` is not implemented for `U`
= help: consider adding a `where U: std::marker::Sized` bound
= note: required because of the requirements on the impl of `Const` for `MakeTuple<U>`
error[E0277]: the trait bound `U: Const` is not satisfied
--> src/main.rs:11:5
|
11 | const VAL: Self::T = (U::VAL, U::VAL);
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ the trait `Const` is not implemented for `U`
|
= help: consider adding a `where U: Const` bound
= note: required because of the requirements on the impl of `Const` for `MakeTuple<U>`
The interesting thing is that even though both bounds are specified, the compiler doesn't seem to take them into account.
Edit: The issue is not using a trait with an associated type and const as generic bound. Instead the problem is implementing a trait with an associated type and const in general: playground
I ran into the same problem. It's also in 1.26 stable.
Here's a further minimized example (play):
trait EmptyTrait {}
trait Const {
type T;
const VAL: Self::T;
}
impl<U: EmptyTrait + std::marker::Sized> Const for U
{
type T = u32;
const VAL: u32 = 5;
}
We can also use the type parameter in the trait rather than the type (play):
trait EmptyTrait {}
trait Const<U> {
type T;
const VAL: Self::T;
}
impl<U: EmptyTrait + Sized> Const<U> for () {
type T = u32;
const VAL: Self::T = 5;
}
Both code fragments give an error saying that in the blanket impl, Rust cannot figure out that U implements EmptyTrait and Sized:
error[E0277]: the trait bound `U: std::marker::Sized` is not satisfied
--> src/lib.rs:12:5
|
12 | const VAL: Self::T = 5;
| ^^^^^^^^^^^^^^^^^^^^^^^ `U` does not have a constant size known at compile-time
|
= help: the trait `std::marker::Sized` is not implemented for `U`
= help: consider adding a `where U: std::marker::Sized` bound
= note: required because of the requirements on the impl of `Const<U>` for `()`
error[E0277]: the trait bound `U: EmptyTrait` is not satisfied
--> src/lib.rs:12:5
|
12 | const VAL: Self::T = 5;
| ^^^^^^^^^^^^^^^^^^^^^^^ the trait `EmptyTrait` is not implemented for `U`
|
= help: consider adding a `where U: EmptyTrait` bound
= note: required because of the requirements on the impl of `Const<U>` for `()`
Note that in both cases, if you remove the EmptyTrait constraint on the quantified type U, the error about EmptyTrait goes away! This is very strange. However, like in #50824, the Sized error message stays.
The problem goes away if you declare the type of VAL in Const to be u32, rather than some _type_ that is computed potentially based on parameters.
I'm not familiar with the compiler internals, but my hypothesis is that associated types and constants depending on a type parameter are not evaluated properly in the constant expression evaluator. In this case, it's associated types that do not reduce well: const VAL: Self::T = 5; forces Rust to do some fancy type of computation at compile time in order to type check, but there's a bug in the code for such computations.
I feel there are a couple of different issues in the same spirit:
[u8; ::std::mem::size_of::<T>()]. This fails. The error message shows that Rust cannot decide that T : Sized, even though T : Sized was in the where clause. Same in #43408, where they get a warning (!) about a "constant evaluation error".Copying over my minimal example from #53908 as another data point (playground):
use std::marker::PhantomData;
pub trait FromPest: Sized {
type Rule;
const RULE: Self::Rule;
}
impl<T> FromPest for PhantomData<T>
where
T: FromPest,
{
type Rule = T::Rule;
const RULE: T::Rule = T::RULE;
}
Here I've added a Sized supertrait to my trait to get rid of the Sized issue but it still can't determine the required trait implementation (of my trait in this case) to make this work.
All of these cases now compile, except for the original report, which has a type error:
error[E0308]: mismatched types
--> src/lib.rs:11:21
|
11 | const N: C::M = 4u8;
| ^^^ expected associated type, found u8
|
= note: expected type `<C as O>::M`
found type `u8`
Current output:
error[E0308]: mismatched types
--> src/lib.rs:11:21
|
11 | const N: C::M = 4u8;
| ^^^ expected associated type, found u8
|
= note: expected type `<C as O>::M`
found type `u8`
= note: consider constraining the associated type `<C as O>::M` to `u8` or calling a method that returns `<C as O>::M`
= note: for more information, visit https://doc.rust-lang.org/book/ch19-03-advanced-traits.html
The correct code would be
trait O {
type M;
}
trait U<A: O> {
const N: A::M;
}
impl<D> O for D {
type M = u8;
}
impl<C> U<C> for u16 where C: O<M=u32>{
const N: C::M = 4;
}
The outstanding work would be providing a suggestion explaining what "consider constraining the associated type <C as O>::M to u8" means. This will be hard to do given the different cases that need to be accounted for, for the case above you need to move the C: O constraint to where, for example (although that's not _needed_, just nicer stylistically).
Most helpful comment
I ran into the same problem. It's also in 1.26 stable.
Here's a further minimized example (play):
We can also use the type parameter in the trait rather than the type (play):
Both code fragments give an error saying that in the blanket impl, Rust cannot figure out that
UimplementsEmptyTraitandSized:Note that in both cases, if you remove the
EmptyTraitconstraint on the quantified typeU, the error aboutEmptyTraitgoes away! This is very strange. However, like in #50824, theSizederror message stays.The problem goes away if you declare the type of
VALinConstto beu32, rather than some _type_ that is computed potentially based on parameters.I'm not familiar with the compiler internals, but my hypothesis is that associated types and constants depending on a type parameter are not evaluated properly in the constant expression evaluator. In this case, it's associated types that do not reduce well:
const VAL: Self::T = 5;forces Rust to do some fancy type of computation at compile time in order to type check, but there's a bug in the code for such computations.I feel there are a couple of different issues in the same spirit:
[u8; ::std::mem::size_of::<T>()]. This fails. The error message shows that Rust cannot decide thatT : Sized, even thoughT : Sizedwas in the where clause. Same in #43408, where they get a warning (!) about a "constant evaluation error".