rfcs/text/2451-re-rebalancing-coherence.md

• Feature Name: re_rebalancing_coherence
• Start Date: 2018-05-30
• RFC PR: rust-lang/rfcs#2451
• Rust Issue: rust-lang/rust#55437

Summary

This RFC seeks to clarify some ambiguity from RFC #1023, and expands it to allow type parameters to appear in the type for which the trait is being implemented, regardless of whether a local type appears before them. More concretely, it allows impl<T> ForeignTrait<LocalType> for ForeignType<T> to be written.

Motivation

For better or worse, we allow implementing foreign traits for foreign types. For example, impl From<Foo> for Vec<i32> is something any crate can write, even though From is a foreign trait, and Vec is a foreign type. However, under the current coherence rules, we do not allow impl<T> From<Foo> for Vec<T>.

There's no good reason for this restriction. Fundamentally, allowing for Vec<ForeignType> requires all the same restrictions as allowing Vec<T>. Disallowing type parameters to appear in the target type restricts how crates can be extended.

Consider an example from Diesel. Diesel constructs an AST which represents a SQL query, and then provides a trait to construct the final SQL. Because different databases have different syntax, this trait is generic over the backend being used. Diesel wants to support third party crates which add new AST nodes, as well as crates which add support for new backends. The current rules make it impossible to support both.

The Oracle database requires special syntax for inserting multiple records in a single query. However, the impl required for this is invalid today. impl<'a, T, U> QueryFragment<Oracle> for BatchInsert<'a, T, U>. There is no reason for this impl to be rejected. The only impl that Diesel could add which would conflict with it would look like impl<'a, T> QueryFragment<T> for BatchInsert<'a, Type1, Type2>. Adding such an impl is already considered a major breaking change by RFC #1023, which we'll expand on below.

For some traits, this can be worked around by flipping the self type with the type parameter to the trait. Diesel has done that in the past (e.g. T: NativeSqlType<DB> became DB: HasSqlType<T>). However, that wouldn't work for this case. A crate which adds a new AST node would no longer be able to implement the required trait for all backends. For example, a crate which added the LOWER function from SQL (which is supported by all databases) would not be able to write impl<T, DB> QueryFragment<Lower<T>> for DB.

Unless we expand the orphan rules, use cases like this one will never be possible, and a crate like Diesel will never be able to be designed in a completely extensible fashion.

Guide-level explanation

Definitions

Local Trait: A trait which was defined in the current crate. Whether a trait is local or not has nothing to do with type parameters. Given trait Foo<T, U>, Foo is always local, regardless of the types used for T or U.

Local Type: A struct, enum, or union which was defined in the current crate. This is not affected by type parameters. struct Foo is considered local, but Vec<Foo> is not. LocalType<ForeignType> is local. Type aliases and trait aliases do not affect locality.

Covered Type: A type which appears as a parameter to another type. For example, T is uncovered, but the T in Vec<T> is covered. This is only relevant for type parameters.

Blanket Impl: Any implementation where a type appears uncovered. impl<T> Foo for T, impl<T> Bar<T> for T, impl<T> Bar<Vec<T>> for T, and impl<T> Bar<T> for Vec<T> are considered blanket impls. However, impl<T> Bar<Vec<T>> for Vec<T> is not a blanket impl, as all instances of T which appear in this impl are covered by Vec.

Fundamental Type: A type for which you cannot add a blanket impl backwards compatibly. This includes &, &mut, and Box. Any time a type T is considered local, &T, &mut T, and Box<T> are also considered local. Fundamental types cannot cover other types. Any time the term "covered type" is used, &T, &mut T, and Box<T> are not considered covered.

What is coherence and why do we care?

Let's start with a quick refresher on coherence and the orphan rules. Coherence means that for any given trait and type, there is one specific implementation that applies. This is important for Rust to be easy to reason about. When you write <Foo as Bar>::trait_method, the compiler needs to know what actual implementation to use.

In languages without coherence, the compiler has to have some way to choose which implementation to use when multiple implementations could apply. Scala does this by having complex scope resolution rules for "implicit" parameters. Haskell (when a discouraged flag is enabled) does this by picking an impl arbitrarily.

Rust's solution is to enforce that there is only one impl to choose from at all. While the rules required to enforce this are quite complex, the result is easy to reason about, and is generally considered to be quite important for Rust. New features like specialization allow more than one impl to apply, but for any given type and trait, there will always be exactly one which is most specific, and deterministically be chosen.

An important piece of enforcing coherence is restricting "orphan impls". An impl is orphaned if it is implementing a trait you don't own for a type you don't own. Rust's rules around this balance two separate, but related goals:

• Ensuring that two crates can't write impls that would overlap (e.g. no crate other than std can write impl From<usize> for Vec<i32>. If they could, your program might stop compiling just by using two crates with an overlapping impl).
• Restricting the impls that can be written so crates can add implementations for traits/types they do own without worrying about breaking downstream crates.

Teaching users

This change isn't something that would end up in a guide, and is mostly communicated through error messages. The most common one seen is E0210. The text of that error will be changed to approximate the following:

Generally speaking, Rust only permits implementing a trait for a type if either the trait or type were defined in your program. However, Rust allows a limited number of impls that break this rule, if they follow certain rules. This error indicates a violation of one of those rules.

A trait is considered local when {definition given above}. A type is considered local when {definition given above}.

When implementing a foreign trait for a foreign type, the trait must have one or more type parameters. A type local to your crate must appear before any use of any type parameters. This means that impl<T> ForeignTrait<LocalType<T>, T> for ForeignType is valid, but impl<T> ForeignTrait<T, LocalType<T>> for ForeignType is not.

The reason that Rust considers order at all is to ensure that your implementation does not conflict with one from another crate. Without this rule, you could write impl<T> ForeignTrait<T, LocalType> for ForeignType, and another crate could write impl<T> ForeignTrait<TheirType, T> for ForeignType, which would overlap. For that reason, we require that your local type come before the type parameter, since the only alternative would be disallowing these implementations at all.

Additionally, the case of impl<T> ForeignTrait<LocalType> for T should be special cased, and given its own error message, which approximates the following:

This error indicates an attempt to implement a trait from another crate for a type parameter.

Rust requires that for any given trait and any given type, there is at most one implementation of that trait. An important piece of this is that we disallow implementing a trait from another crate for a type parameter.

Rust's orphan rule always permits an impl if either the trait or the type being implemented are local to the current crate. Therefore, we can't allow impl<T> ForeignTrait<LocalTypeCrateA> for T, because it might conflict with another crate writing impl<T> ForeignTrait<T> for LocalTypeCrateB, which we will always permit.

Finally, RFC #1105 states that implementing any non-fundamental trait for an existing type is not a breaking change. This directly condradicts RFC #1023, which is entirely based around "blanket impls" being breaking changes. Regardless of whether the changes proposed to the orphan rules in this proposal are accepted, a blanket impl being a breaking change must be true today. Given that the compiler currently accepts impl From<Foo> for Vec<Foo>, adding impl<T> From<T> for Vec<T> must be considered a major breaking change.

As such, RFC #1105 is amended to remove the statement that implementing a non-fundamental trait is a minor breaking change, and states that adding any blanket impl for an existing trait is a major breaking change, using the definition of blanket impl given above.

Reference-level explanation

Concrete orphan rules

Assumes the same definitions as above.

Given impl<P1..=Pn> Trait<T1..=Tn> for T0, an impl is valid only if at least one of the following is true:

• Trait is a local trait
• All of
  • At least one of the types T0..=Tn must be a local type. Let Ti be the first such type.
  • No uncovered type parameters P1..=Pn may appear in T0..Ti (excluding Ti)

The primary change from the rules defined in RFC #1023 is that we only restrict the appearance of uncovered type parameters. Once again, it is important to note that for the purposes of coherence, #[fundamental] types are special. Box<T> is not considered covered, and Box<LocalType> is considered local.

Under this proposal, the orphan rules continue to work generally as they did before, with one notable exception; We will permit impl<T> ForeignTrait<LocalType> for ForeignType<T>. This is completely valid under the forward compatibility rules set in RFC #1023. We can demonstrate that this is the case with the following:

• Any valid impl of ForeignTrait in a child crate must reference at least one type that is local to the child crate.
• The only way a parent crate can reference the type of a child crate is with a type parameter.
• For the impl in child crate to overlap with an impl in parent crate, the type parameter must be uncovered.
• Adding any impl with an uncovered type parameter is considered a major breaking change.

We can also demonstrate that it is impossible for two sibling crates to write conflicting impls, with or without this proposal.

• Any valid impl of ForeignTrait in a child crate must reference at least one type that is local to the child crate.
• The only way a local type of sibling crate A could overlap with a type used in an impl from sibling crate B is if sibling crate B used a type parameter
• Any type parameter used by sibling crate B must be preceded by a local type
• Sibling crate A could not possibly name a type from sibling crate B, thus that parameter can never overlap.

Effects on parent crates

RFC #1023 is amended to state that adding a new impl to an existing trait is considered a breaking change unless, given impl<P1..=Pn> Trait<T1..=Tn> for T0:

• At least one of the types T0..=Tn must be a local type, added in this revision. Let Ti be the first such type.
• No uncovered type parameters P1..=Pn appear in T0..Ti (excluding Ti)

The more general way to put this rule is: "Adding an impl to an existing trait is a breaking change if it could possibly conflict with a legal impl in a downstream crate".

This clarification is true regardless of whether the changes in this proposal are accepted or not. Given that the compiler currently accepts impl From<Foo> for Vec<Foo>, adding the impl impl<T> From<T> for Vec<T> must be considered a major breaking change.

To be specific, the following adding any of the following impls would be considered a breaking change:

• impl<T> OldTrait<T> for OldType
• impl<T> OldTrait<AnyType> for T
• impl<T> OldTrait<T> for ForeignType

However, the following impls would not be considered a breaking change:

• impl NewTrait<AnyType> for AnyType
• impl<T> OldTrait<T> for NewType
• impl<T> OldTrait<NewType, T> for OldType

Drawbacks

The current rules around coherence are complex and hard to explain. While this proposal feels like a natural extension of the current rules, and something many expect to work, it does make them slightly more complex.

The orphan rules are often taught as "for an impl impl Trait for Type, either Trait or Type must be local to your crate". While this has never been actually true, it's a reasonable hand-wavy explanation, and this gets us even further from it. Even though impl From<Foo> for Vec<()> has always been accepted, impl<T> From<Foo> for Vec<T> feels even less local. While Vec<()> only applies to std, Vec<T> now applies to types from std and any other crate.

Rationale and alternatives

• Rework coherence even more deeply. The rules around the orphan rule are complex and hard to explain. Even --explain E0210 doesn't actually try to give the rationale behind them, and just states the fairly arcane formula from the original RFC. While this proposal is a natural extension of the current rules, and something that many expect to "just work", it ultimately makes them even more complex.

  In particular, this keeps the "ordering" rule. It still serves a purpose with this proposal, but much less of one. By keeping it, we are able to allow impl<T> SomeTrait<LocalType, T> for ForeignType, because no sibling crate can write an overlapping impl. However, this is not something that the majority of library authors are aware of, and requires API designers to order their type parameters based on how likely they are to be overridden by other crates.

  We could instead provide a mechanism for traits to opt into a redesigned coherence system, and potentially default to that in a future edition. However, that would likely cause a lot of confusion in the community. This proposal is a strict addition to the set of impls which are allowed with the current rules, without an increase in risk or impls which are breaking changes. It seems like a reasonably conservative move, even if we eventually want to overhaul coherence.

• Get rid of the orphan rule entirely. A long standing pain point for crates like Diesel has been integration with other crates. Diesel doesn't want to care about chrono, and chrono doesn't want to care about Diesel. A database access library shouldn't dictate your choice of time libraries, vice versa.

  However, due to the way Rust works today, one of them has to. Nobody can create a diesel-chrono crate due to the orphan rule. Maybe if we just allowed crates to have incompatible impls, and set a standard of "don't write orphan impls unless that's the entire point of your crate", it wouldn't actually be that bad.

Unresolved questions

• Are there additional implementations which are clearly acceptable under the current restrictions, which are disallowed with this extension? Should we allow them if so?