RFC 1214: projections-lifetimes-and-wf

lang (typesystem | associated-types | lifetimes)

Summary

Type system changes to address the outlives relation with respect to projections, and to better enforce that all types are well-formed (meaning that they respect their declared bounds). The current implementation can be both unsound (#24622), inconvenient (#23442), and surprising (#21748, #25692). The changes are as follows:

The proposed changes here have been tested and found to cause only a modest number of regressions (about two dozen root regressions were previously found on crates.io; however, that run did not yet include all the provisions from this RFC; updated numbers coming soon). In order to minimize the impact on users, the plan is to first introduce the changes in two stages:

  1. Initially, warnings will be issued for cases that violate the rules specified in this RFC. These warnings are not lints and cannot be silenced except by correcting the code such that it type-checks under the new rules.
  2. After one release cycle, those warnings will become errors.

Note that although the changes do cause regressions, they also cause some code (like that in #23442) which currently gets errors to compile successfully.

Motivation

TL;DR

This is a long detailed RFC that is attempting to specify in some detail aspects of the type system that were underspecified or buggily implemented before. This section just summarizes the effect on existing Rust code in terms of changes that may be required.

Warnings first, errors later. Although the changes described in this RFC are necessary for soundness (and many of them are straight-up bugfixes), there is some impact on existing code. Therefore the plan is to first issue warnings for a release cycle and then transition to hard errors, so as to ease the migration.

Associated type projections and lifetimes work more smoothly. The current rules for relating associated type projections (like T::Foo) and lifetimes are somewhat cumbersome. The newer rules are more flexible, so that e.g. we can deduce that T::Foo: 'a if T: 'a, and similarly that T::Foo is well-formed if T is well-formed. As a bonus, the new rules are also sound. ;)

Simpler outlives relation. The older definition for the outlives relation T: 'a was rather subtle. The new rule basically says that if all type/lifetime parameters appearing in the type T must outlive 'a, then T: 'a (though there can also be other ways for us to decide that T: 'a is valid, such as in-scope where clauses). So for example fn(&'x X): 'a if 'x: 'a and X: 'a (presuming that X is a type parameter). The older rules were based on what kind of data was actually reachable, and hence accepted this type (since no data of &'x X is reachable from a function pointer). This change primarily affects struct declarations, since they may now require additional outlives bounds:

// OK now, but after this RFC requires `X: 'a`:
struct Foo<'a, X> {
    f: fn(&'a X) // (because of this field)
}

More types are sanity checked. Generally Rust requires that if you have a type like SomeStruct<T>, then whatever where clauses are declared on SomeStruct must hold for T (this is called being "well-formed"). For example, if SomeStruct is declared like so:

struct SomeStruct<T:Eq> { .. }

then this implies that SomeStruct<f32> is ill-formed, since f32 does not implement Eq (just PartialEq). However, the current compiler doesn't check this in associated type definitions:

impl Iterator for SomethingElse {
    type Item = SomeStruct<f32>; // accepted now, not after this RFC
}

Similarly, WF checking was skipped for trait object types and fn arguments. This means that fn(SomeStruct<f32>) would be considered well-formed today, though attempting to call the function would be an error. Under this RFC, that fn type is not well-formed (though sometimes when there are higher-ranked regions, WF checking may still be deferred until the point where the fn is called).

There are a few other places where similar requirements were being overlooked before but will now be enforced. For example, a number of traits like the following were found in the wild:

trait Foo {
    // currently accepted, but should require that Self: Sized
    fn method(&self, value: Option<Self>);
}

To be well-formed, an Option<T> type requires that T: Sized. In this case, though T=Self, and Self is not Sized by default. Therefore, this trait should be declared trait Foo: Sized to be legal. The compiler is currently attempting to enforce these rules, but many cases were overlooked in practice.

Impact on crates.io

This RFC has been largely implemented and tested against crates.io. A total of 43 (root) crates are affected by the changes. Interestingly, the vast majority of warnings/errors that occur are not due to new rules introduced by this RFC, but rather due to older rules being more correctly enforced.

Of the affected crates, 40 are receiving future compatibility warnings and hence continue to build for the time being. In the remaining three cases, it was not possible to isolate the effects of the new rules, and hence the compiler reports an error rather than a future compatibility warning.

What follows is a breakdown of the reason that crates on crates.io are receiving errors or warnings. Each row in the table corresponds to one of the explanations above.

ProblemFuture-compat. warningsErrors
More types are sanity checked353
Simpler outlives relation5

As you can see, by far the largest source of problems is simply that we are now sanity checking more types. This was always the intent, but there were bugs in the compiler that led to it either skipping checking altogether or only partially applying the rules. It is interesting to drill down a bit further into the 38 warnings/errors that resulted from more types being sanity checked in order to see what kinds of mistakes are being caught:

CaseProblemNumber
1Self: Sized required26
2Foo: Bar required11
3Not object safe1

An example of each case follows:

Cases 1 and 2. In the compiler today, types appearing in trait methods are incompletely checked. This leads to a lot of traits with insufficient bounds. By far the most common example was that the Self parameter would appear in a context where it must be sized, usually when it is embedded within another type (e.g., Option<Self>). Here is an example:

trait Test {
    fn test(&self) -> Option<Self>;
    //                ~~~~~~~~~~~~
    //            Incorrectly permitted before.
}

Because Option<T> requires that T: Sized, this trait should be declared as follows:

trait Test: Sized {
    fn test(&self) -> Option<Self>;
}

Case 2. Case 2 is the same as case 1, except that the missing bound is some trait other than Sized, or in some cases an outlives bound like T: 'a.

Case 3. The compiler currently permits non-object-safe traits to be used as types, even if objects could never actually be created (#21953).

Projections and the outlives relation

RFC 192 introduced the outlives relation T: 'a and described the rules that are used to decide when one type outlives a lifetime. In particular, the RFC describes rules that govern how the compiler determines what kind of borrowed data may be "hidden" by a generic type. For example, given this function signature:

fn foo<'a,I>(x: &'a I)
    where I: Iterator
{ ... }

the compiler is able to use implied region bounds (described more below) to automatically determine that:

When associated types were introduced in RFC 195, some new rules were required to decide when an "outlives relation" involving a projection (e.g., I::Item: 'a) should hold. The initial rules were very conservative. This led to the rules from RFC 192 being adapted to cover associated type projections like I::Item. Unfortunately, these adapted rules are not ideal, and can still lead to annoying errors in some situations. Finding a better solution has been on the agenda for some time.

Simultaneously, we realized in #24622 that the compiler had a bug that caused it to erroneously assume that every projection like I::Item outlived the current function body, just as it assumes that type parameters like I outlive the current function body. This bug can lead to unsound behavior. Unfortunately, simply implementing the naive fix for #24622 exacerbates the shortcomings of the current rules for projections, causing widespread compilation failures in all sorts of reasonable and obviously correct code.

This RFC describes modifications to the type system that both restore soundness and make working with associated types more convenient in some situations. The changes are largely but not completely backwards compatible.

Well-formed types

A type is considered well-formed (WF) if it meets some simple correctness criteria. For builtin types like &'a T or [T], these criteria are built into the language. For user-defined types like a struct or an enum, the criteria are declared in the form of where clauses. In general, all types that appear in the source and elsewhere should be well-formed.

For example, consider this type, which combines a reference to a hashmap and a vector of additional key/value pairs:

struct DeltaMap<'a, K, V> where K: Hash + 'a, V: 'a {
    base_map: &'a mut HashMap<K,V>,
    additional_values: Vec<(K,V)>
}

Here, the WF criteria for DeltaMap<K,V> are as follows:

Let's look at those K:'a bounds a bit more closely. If you leave them out, you will find that the the structure definition above does not type-check. This is due to the requirement that the types of all fields in a structure definition must be well-formed. In this case, the field base_map has the type &'a mut HashMap<K,V>, and this type is only valid if K: 'a and V: 'a hold. Since we don't know what K and V are, we have to surface this requirement in the form of a where-clause, so that users of the struct know that they must maintain this relationship in order for the struct to be interally coherent.

An aside: explicit WF requirements on types

You might wonder why you have to write K:Hash and K:'a explicitly. After all, they are obvious from the types of the fields. The reason is that we want to make it possible to check whether a type like DeltaMap<'foo,T,U> is well-formed without having to inspect the types of the fields -- that is, in the current design, the only information that we need to use to decide if DeltaMap<'foo,T,U> is well-formed is the set of bounds and where-clauses.

This has real consequences on usability. It would be possible for the compiler to infer bounds like K:Hash or K:'a, but the origin of the bound might be quite remote. For example, we might have a series of types like:

struct Wrap1<'a,K>(Wrap2<'a,K>);
struct Wrap2<'a,K>(Wrap3<'a,K>);
struct Wrap3<'a,K>(DeltaMap<'a,K,K>);

Now, for Wrap1<'foo,T> to be well-formed, T:'foo and T:Hash must hold, but this is not obvious from the declaration of Wrap1. Instead, you must trace deeply through its fields to find out that this obligation exists.

Implied lifetime bounds

To help avoid undue annotation, Rust relies on implied lifetime bounds in certain contexts. Currently, this is limited to fn bodies. The idea is that for functions, we can make callers do some portion of the WF validation, and let the callees just assume it has been done already. (This is in contrast to the type definition, where we required that the struct itself declares all of its requirements up front in the form of where-clauses.)

To see this in action, consider a function that uses a DeltaMap:

fn foo<'a,K:Hash,V>(d: DeltaMap<'a,K,V>) { ... }

You'll notice that there are no K:'a or V:'a annotations required here. This is due to implied lifetime bounds. Unlike structs, a function's caller must examine not only the explicit bounds and where-clauses, but also the argument and return types. When there are generic type/lifetime parameters involved, the caller is in charge of ensuring that those types are well-formed. (This is in contrast with type definitions, where the type is in charge of figuring out its own requirements and listing them in one place.)

As the name "implied lifetime bounds" suggests, we currently limit implied bounds to region relationships. That is, we will implicitly derive a bound like K:'a or V:'a, but not K:Hash -- this must still be written manually. It might be a good idea to change this, but that would be the topic of a separate RFC.

Currently, implied bound are limited to fn bodies. This RFC expands the use of implied bounds to cover impl definitions as well, since otherwise the annotation burden is quite painful. More on this in the next section.

NB. There is an additional problem concerning the interaction of implied bounds and contravariance (#25860). To better separate the issues, this will be addressed in a follow-up RFC that should appear shortly.

Missing WF checks

Unfortunately, the compiler currently fails to enforce WF in several important cases. For example, the following program is accepted:

struct MyType<T:Copy> { t: T }

trait ExampleTrait {
    type Output;
}

struct ExampleType;

impl ExampleTrait for ExampleType {
    type Output = MyType<Box<i32>>;
    //            ~~~~~~~~~~~~~~~~
    //                   |
    //   Note that `Box<i32>` is not `Copy`!
}

However, if we simply naively add the requirement that associated types must be well-formed, this results in a large annotation burden (see e.g. PR 25701). For example, in practice, many iterator implementation break due to region relationships:

impl<'a, T> IntoIterator for &'a LinkedList<T> {
   type Item = &'a T;
   ...
}

The problem here is that for &'a T to be well-formed, T: 'a must hold, but that is not specified in the where clauses. This RFC proposes using implied bounds to address this concern -- specifically, every impl is permitted to assume that all types which appear in the impl header (trait reference) are well-formed, and in turn each "user" of an impl will validate this requirement whenever they project out of a trait reference (e.g., to do a method call, or normalize an associated type).

Detailed design

This section dives into detail on the proposed type rules.

A little type grammar

We extend the type grammar from RFC 192 with projections and slice types:

T = scalar (i32, u32, ...)        // Boring stuff
  | X                             // Type variable
  | Id<P0..Pn>                    // Nominal type (struct, enum)
  | &r T                          // Reference (mut doesn't matter here)
  | O0..On+r                      // Object type
  | [T]                           // Slice type
  | for<r..> fn(T1..Tn) -> T0     // Function pointer
  | <P0 as Trait<P1..Pn>>::Id     // Projection
P = r                             // Region name
  | T                             // Type
O = for<r..> TraitId<P1..Pn>      // Object type fragment
r = 'x                            // Region name

We'll use this to describe the rules in detail.

A quick note on terminology: an "object type fragment" is part of an object type: so if you have Box<FnMut()+Send>, FnMut() and Send are object type fragments. Object type fragments are identical to full trait references, except that they do not have a self type (no P0).

Syntactic definition of the outlives relation

The outlives relation is defined in purely syntactic terms as follows. These are inference rules written in a primitive ASCII notation. :) As part of defining the outlives relation, we need to track the set of lifetimes that are bound within the type we are looking at. Let's call that set R=<r0..rn>. Initially, this set R is empty, but it will grow as we traverse through types like fns or object fragments, which can bind region names via for<..>.

Simple outlives rules

Here are the rules covering the simple cases, where no type parameters or projections are involved:

OutlivesScalar:
  --------------------------------------------------
  R ⊢ scalar: 'a

OutlivesNominalType:
  ∀i. R ⊢ Pi: 'a
  --------------------------------------------------
  R ⊢ Id<P0..Pn>: 'a

OutlivesReference:
  R ⊢ 'x: 'a
  R ⊢ T: 'a
  --------------------------------------------------
  R ⊢ &'x T: 'a

OutlivesObject:
  ∀i. R ⊢ Oi: 'a
  R ⊢ 'x: 'a
  --------------------------------------------------
  R ⊢ O0..On+'x: 'a

OutlivesFunction:
  ∀i. R,r.. ⊢ Ti: 'a
  --------------------------------------------------
  R ⊢ for<r..> fn(T1..Tn) -> T0: 'a

OutlivesFragment:
  ∀i. R,r.. ⊢ Pi: 'a
  --------------------------------------------------
  R ⊢ for<r..> TraitId<P0..Pn>: 'a

Outlives for lifetimes

The outlives relation for lifetimes depends on whether the lifetime in question was bound within a type or not. In the usual case, we decide the relationship between two lifetimes by consulting the environment, or using the reflexive property. Lifetimes representing scopes within the current fn have a relationship derived from the code itself, while lifetime parameters have relationships defined by where-clauses and implied bounds.

OutlivesRegionEnv:
  'x ∉ R               // not a bound region
  ('x: 'a) in Env      // derivable from where-clauses etc
  --------------------------------------------------
  R ⊢ 'x: 'a

OutlivesRegionReflexive:
  --------------------------------------------------
  R ⊢ 'a: 'a

OutlivesRegionTransitive:
  R ⊢ 'a: 'c
  R ⊢ 'c: 'b
  --------------------------------------------------
  R ⊢ 'a: 'b

For higher-ranked lifetimes, we simply ignore the relation, since the lifetime is not yet known. This means for example that for<'a> fn(&'a i32): 'x holds, even though we do not yet know what region 'a is (and in fact it may be instantiated many times with different values on each call to the fn).

OutlivesRegionBound:
  'x ∈ R               // bound region
  --------------------------------------------------
  R ⊢ 'x: 'a

Outlives for type parameters

For type parameters, the only way to draw "outlives" conclusions is to find information in the environment (which is being threaded implicitly here, since it is never modified). In terms of a Rust program, this means both explicit where-clauses and implied bounds derived from the signature (discussed below).

OutlivesTypeParameterEnv:
  X: 'a in Env
  --------------------------------------------------
  R ⊢ X: 'a

Outlives for projections

Projections have the most possibilities. First, we may find information in the in-scope where clauses, as with type parameters, but we can also consult the trait definition to find bounds (consider an associated type declared like type Foo: 'static). These rule only apply if there are no higher-ranked lifetimes in the projection; for simplicity's sake, we encode that by requiring an empty list of higher-ranked lifetimes. (This is somewhat stricter than necessary, but reflects the behavior of my prototype implementation.)

OutlivesProjectionEnv:
  <P0 as Trait<P1..Pn>>::Id: 'b in Env
  <> ⊢ 'b: 'a
  --------------------------------------------------
  <> ⊢ <P0 as Trait<P1..Pn>>::Id: 'a

OutlivesProjectionTraitDef:
  WC = [Xi => Pi] WhereClauses(Trait)
  <P0 as Trait<P1..Pn>>::Id: 'b in WC
  <> ⊢ 'b: 'a
  --------------------------------------------------
  <> ⊢ <P0 as Trait<P1..Pn>>::Id: 'a

All the rules covered so far already exist today. This last rule, however, is not only new, it is the crucial insight of this RFC. It states that if all the components in a projection's trait reference outlive 'a, then the projection must outlive 'a:

OutlivesProjectionComponents:
  ∀i. R ⊢ Pi: 'a
  --------------------------------------------------
  R ⊢ <P0 as Trait<P1..Pn>>::Id: 'a

Given the importance of this rule, it's worth spending a bit of time discussing it in more detail. The following explanation is fairly informal. A more detailed look can be found in the appendix.

Let's begin with a concrete example of an iterator type, like std::vec::Iter<'a,T>. We are interested in the projection of Iterator::Item:

<Iter<'a,T> as Iterator>::Item

or, in the more succint (but potentially ambiguous) form:

Iter<'a,T>::Item

Since I'm going to be talking a lot about this type, let's just call it <PROJ> for now. We would like to determine whether <PROJ>: 'x holds.

Now, the easy way to solve <PROJ>: 'x would be to normalize <PROJ> by looking at the relevant impl:

impl<'b,U> Iterator for Iter<'b,U> {
    type Item = &'b U;
    ...
}

From this impl, we can conclude that <PROJ> == &'a T, and thus reduce <PROJ>: 'x to &'a T: 'x, which in turn holds if 'a: 'x and T: 'x (from the rule OutlivesReference).

But often we are in a situation where we can't normalize the projection (for example, a projection like I::Item where we only know that I: Iterator). What can we do then? The rule OutlivesProjectionComponents says that if we can conclude that every lifetime/type parameter Pi to the trait reference outlives 'x, then we know that a projection from those parameters outlives 'x. In our example, the trait reference is <Iter<'a,T> as Iterator>, so that means that if the type Iter<'a,T> outlives 'x, then the projection <PROJ> outlives 'x. Now, you can see that this trivially reduces to the same result as the normalization, since Iter<'a,T>: 'x holds if 'a: 'x and T: 'x (from the rule OutlivesNominalType).

OK, so we've seen that applying the rule OutlivesProjectionComponents comes up with the same result as normalizing (at least in this case), and that's a good sign. But what is the basis of the rule?

The basis of the rule comes from reasoning about the impl that we used to do normalization. Let's consider that impl again, but this time hide the actual type that was specified:

impl<'b,U> Iterator for Iter<'b,U> {
    type Item = /* <TYPE> */;
    ...
}

So when we normalize <PROJ>, we obtain the result by applying some substitution Θ to <TYPE>. This substitution is a mapping from the lifetime/type parameters on the impl to some specific values, such that <PROJ> == Θ <Iter<'b,U> as Iterator>::Item. In this case, that means Θ would be ['b => 'a, U => T] (and of course <TYPE> would be &'b U, but we're not supposed to rely on that).

The key idea for the OutlivesProjectionComponents is that the only way that <TYPE> can fail to outlive 'x is if either:

Now, the only way that <TYPE> can refer to a parameter P is if it is brought in by the substitution Θ. So, if we can just show that all the types/lifetimes that in the range of Θ outlive 'x, then we know that Θ <TYPE> outlives 'x.

Put yet another way: imagine that you have an impl with no parameters, like:

impl Iterator for Foo {
    type Item = /* <TYPE> */;
}

Clearly, whatever <TYPE> is, it can only refer to the lifetime 'static. So <Foo as Iterator>::Item: 'static holds. We know this is true without ever knowing what <TYPE> is -- we just need to see that the trait reference <Foo as Iterator> doesn't have any lifetimes or type parameters in it, and hence the impl cannot refer to any lifetime or type parameters.

Implementation complications

The current region inference code only permits constraints of the form:

C = r0: r1
  | C AND C

This is convenient because a simple fixed-point iteration suffices to find the minimal regions which satisfy the constraints.

Unfortunately, this constraint model does not scale to the outlives rules for projections. Consider a trait reference like <T as Trait<'X>>::Item: 'Y, where 'X and 'Y are both region variables whose value is being inferred. At this point, there are several inference rules which could potentially apply. Let us assume that there is a where-clause in the environment like <T as Trait<'a>>::Item: 'b. In that case, if 'X == 'a and 'b: 'Y, then we could employ the OutlivesProjectionEnv rule. This would correspond to a constraint set like:

C = 'X:'a AND 'a:'X AND 'b:'Y

Otherwise, if T: 'a and 'X: 'Y, then we could use the OutlivesProjectionComponents rule, which would require a constraint set like:

C = C1 AND 'X:'Y

where C1 is the constraint set for T:'a.

As you can see, these two rules yielded distinct constraint sets. Ideally, we would combine them with an OR constraint, but no such constraint is available. Adding such a constraint complicates how inference works, since a fixed-point iteration is no longer sufficient.

This complication is unfortunate, but to a large extent already exists with where-clauses and trait matching (see e.g. #21974). (Moreover, it seems to be inherent to the concept of assocated types, since they take several inputs (the parameters to the trait) which may or may not be related to the actual type definition in question.)

For the time being, the current implementation takes a pragmatic approach based on heuristics. It first examines whether any region bounds are declared in the trait and, if so, prefers to use those. Otherwise, if there are region variables in the projection, then it falls back to the OutlivesProjectionComponents rule. This is always sufficient but may be stricter than necessary. If there are no region variables in the projection, then it can simply run inference to completion and check each of the other two rules in turn. (It is still necessary to run inference because the bound may be a region variable.) So far this approach has sufficed for all situations encountered in practice. Eventually, we should extend the region inferencer to a richer model that includes "OR" constraints.

The WF relation

This section describes the "well-formed" relation. In previous RFCs, this was combined with the outlives relation. We separate it here for reasons that shall become clear when we discuss WF conditions on impls.

The WF relation is really pretty simple: it just says that a type is "self-consistent". Typically, this would include validating scoping (i.e., that you don't refer to a type parameter X if you didn't declare one), but we'll take those basic conditions for granted.

WfScalar:
  --------------------------------------------------
  R ⊢ scalar WF

WfParameter:
  --------------------------------------------------
  R ⊢ X WF                  // where X is a type parameter

WfTuple:
  ∀i. R ⊢ Ti WF
  ∀i<n. R ⊢ Ti: Sized       // the *last* field may be unsized
  --------------------------------------------------
  R ⊢ (T0..Tn) WF

WfNominalType:
  ∀i. R ⊢ Pi Wf             // parameters must be WF,
  C = WhereClauses(Id)      // and the conditions declared on Id must hold...
  R ⊢ [P0..Pn] C            // ...after substituting parameters, of course
  --------------------------------------------------
  R ⊢ Id<P0..Pn> WF

WfReference:
  R ⊢ T WF                  // T must be WF
  R ⊢ T: 'x                 // T must outlive 'x
  --------------------------------------------------
  R ⊢ &'x T WF

WfSlice:
  R ⊢ T WF
  R ⊢ T: Sized
  --------------------------------------------------
  [T] WF

WfProjection:
  ∀i. R ⊢ Pi WF             // all components well-formed
  R ⊢ <P0: Trait<P1..Pn>>   // the projection itself is valid
  --------------------------------------------------
  R ⊢ <P0 as Trait<P1..Pn>>::Id WF

WF checking and higher-ranked types

There are two places in Rust where types can introduce lifetime names into scope: fns and trait objects. These have somewhat different rules than the rest, simply because they modify the set R of bound lifetime names. Let's start with the rule for fn types:

WfFn:
  ∀i. R, r.. ⊢ Ti WF
  --------------------------------------------------
  R ⊢ for<r..> fn(T1..Tn) -> T0 WF

Basically, this rule adds the bound lifetimes to the set R and then checks whether the argument and return type are well-formed. We'll see in the next section that means that any requirements on those types which reference bound identifiers are just assumed to hold, but the remainder are checked. For example, if we have a type HashSet<K> which requires that K: Hash, then fn(HashSet<NoHash>) would be illegal since NoHash: Hash does not hold, but for<'a> fn(HashSet<&'a NoHash>) would be legal, since &'a NoHash: Hash involves a bound region 'a. See the "Checking Conditions" section for details.

Note that fn types do not require that T0..Tn be Sized. This is intentional. The limitation that only sized values can be passed as argument (or returned) is enforced at the time when a fn is actually called, as well as in actual fn definitions, but is not considered fundamental to fn types themselves. There are several reasons for this. For one thing, it's forwards compatible with passing DST by value. For another, it means that non-defaulted trait methods to do not have to show that their argument types are Sized (this will be checked in the implementations, where more types are known). Since the implicit Self type parameter is not Sized by default (RFC 546), requiring that argument types be Sized in trait definitions proves to be an annoying annotation burden.

The object type rule is similar, though it includes an extra clause:

WfObject:
  rᵢ = union of implied region bounds from Oi
  ∀i. rᵢ: r
  ∀i. R ⊢ Oi WF
  --------------------------------------------------
  R ⊢ O0..On+r WF

The first two clauses here state that the explicit lifetime bound r must be an approximation for the the implicit bounds rᵢ derived from the trait definitions. That is, if you have a trait definition like

trait Foo: 'static { ... }

and a trait object like Foo+'x, when we require that 'static: 'x (which is true, clearly, but in some cases the implicit bounds from traits are not 'static but rather some named lifetime).

The next clause states that all object type fragments must be WF. An object type fragment is WF if its components are WF:

WfObjectFragment:
  ∀i. R, r.. ⊢ Pi
  TraitId is object safe
  --------------------------------------------------
  R ⊢ for<r..> TraitId<P1..Pn>

Note that we don't check the where clauses declared on the trait itself. These are checked when the object is created. The reason not to check them here is because the Self type is not known (this is an object, after all), and hence we can't check them in general. (But see unresolved questions.)

WF checking a trait reference

In some contexts, we want to check a trait reference, such as the ones that appear in where clauses or type parameter bounds. The rules for this are given here:

WfTraitReference:
  ∀i. R, r.. ⊢ Pi
  C = WhereClauses(Id)      // and the conditions declared on Id must hold...
  R, r0...rn ⊢ [P0..Pn] C   // ...after substituting parameters, of course
  --------------------------------------------------
  R ⊢ for<r..> P0: TraitId<P1..Pn>

The rules are fairly straightforward. The components must be well formed, and any where-clauses declared on the trait itself much hold.

Checking conditions

In various rules above, we have rules that declare that a where-clause must hold, which have the form R ̣⊢ WhereClause. Here, R represents the set of bound regions. It may well be that WhereClause does not use any of the regions in R. In that case, we can ignore the bound-regions and simple check that WhereClause holds. But if WhereClause does refer to regions in R, then we simply consider R ⊢ WhereClause to hold. Those conditions will be checked later when the bound lifetimes are instantiated (either through a call or a projection).

In practical terms, this means that if I have a type like:

struct Iterator<'a, T:'a> { ... }

and a function type like for<'a> fn(i: Iterator<'a, T>) then this type is considered well-formed without having to show that T: 'a holds. In terms of the rules, this is because we would wind up with a constraint like 'a ⊢ T: 'a.

However, if I have a type like

struct Foo<'a, T:Eq> { .. }

and a function type like for<'a> fn(f: Foo<'a, T>), I still must show that T: Eq holds for that function to be well-formed. This is because the condition which is geneated will be 'a ⊢ T: Eq, but 'a is not referenced there.

Implied bounds

Implied bounds can be derived from the WF and outlives relations. The implied bounds from a type T are given by expanding the requirements that T: WF. Since we currently limit ourselves to implied region bounds, we we are interesting in extracting requirements of the form:

Some caution is required around projections when deriving implied bounds. If we encounter a requirement that e.g. X::Id: 'r, we cannot for example deduce that X: 'r must hold. This is because while X: 'r is sufficient for X::Id: 'r to hold, it is not necessary for X::Id: 'r to hold. So we can only conclude that X::Id: 'r holds, and not X: 'r.

When should we check the WF relation and under what conditions?

Currently the compiler performs WF checking in a somewhat haphazard way: in some cases (such as impls), it omits checking WF, but in others (such as fn bodies), it checks WF when it should not have to. Partly that is due to the fact that the compiler currently connects the WF and outlives relationship into one thing, rather than separating them as described here.

Constants/statics. The type of a constant or static can be checked for WF in an empty environment.

Struct/enum declarations. In a struct/enum declaration, we should check that all field types are WF, given the bounds and where-clauses from the struct declaration. Also check that where-clauses are well-formed.

Function items. For function items, the environment consists of all the where-clauses from the fn, as well as implied bounds derived from the fn's argument types. These are then used to check that the following are well-formed:

These WF requirements are imposed at each fn or associated fn definition (as well as within trait items).

Trait impls. In a trait impl, we assume that all types appearing in the impl header are well-formed. This means that the initial environment for an impl consists of the impl where-clauses and implied bounds derived from its header. Example: Given an impl like impl<'a,T> SomeTrait for &'a T, the environment would be T: Sized (explicit where-clause) and T: 'a (implied bound derived from &'a T). This environment is used as the starting point for checking the items:

Inherent impls. In an inherent impl, we can assume that the self type is well-formed, but otherwise check the methods as if they were normal functions. We must check that all items are well-formed, along with the where clauses declared on the impl.

Trait declarations. Trait declarations (and defaults) are checked in the same fashion as impls, except that there are no implied bounds from the impl header. We must check that all items are well-formed, along with the where clauses declared on the trait.

Type aliases. Type aliases are currently not checked for WF, since they are considered transparent to type-checking. It's not clear that this is the best policy, but it seems harmless, since the WF rules will still be applied to the expanded version. See the Unresolved Questions for some discussion on the alternatives here.

Several points in the list above made use of implied bounds based on assuming that various types were WF. We have to ensure that those bounds are checked on the reciprocal side, as follows:

Fns being called. Before calling a fn, we check that its argument and return types are WF. This check takes place after all higher-ranked lifetimes have been instantiated. Checking the argument types ensures that the implied bounds due to argument types are correct. Checking the return type ensures that the resulting type of the call is WF.

Method calls, "UFCS" notation for fns and constants. These are the two ways to project a value out of a trait reference. A method call or UFCS resolution will require that the trait reference is WF according to the rules given above.

Normalizing associated type references. Whenever a projection type like T::Foo is normalized, we will require that the trait reference is WF.

Drawbacks

N/A

Alternatives

I'm not aware of any appealing alternatives.

Unresolved questions

Best policy for type aliases. The current policy is not to check type aliases, since they are transparent to type-checking, and hence their expansion can be checked instead. This is coherent, though somewhat confusing in terms of the interaction with projections, since we frequently cannot resolve projections without at least minimal bounds (i.e., type IteratorAndItem<T:Iterator> = (T::Item, T)). Still, full-checking of WF on type aliases seems to just mean more annotation with little benefit. It might be nice to keep the current policy and later, if/when we adopt a more full notion of implied bounds, rationalize it by saying that the suitable bounds for a type alias are implied by its expansion.

For trait object type fragments, should we check WF conditions when we can? For example, if you have:

trait HashSet<K:Hash>

should an object like Box<HashSet<NotHash>> be illegal? It seems like that would be inline with our "best effort" approach to bound regions, so probably yes.

Appendix

The informal explanation glossed over some details. This appendix tries to be a bit more thorough with how it is that we can conclude that a projection outlives 'a if its inputs outlive 'a. To start, let's specify the projection <PROJ> as:

<P0 as Trait<P1...Pn>>::Id

where P can be a lifetime or type parameter as appropriate.

Then we know that there exists some impl of the form:

impl<X0..Xn> Trait<Q1..Qn> for Q0 {
    type Id = T;
}

Here again, X can be a lifetime or type parameter name, and Q can be any lifetime or type parameter.

Let Θ be a suitable substitution [Xi => Ri] such that ∀i. Θ Qi == Pi (in other words, so that the impl applies to the projection). Then the normalized form of <PROJ> is Θ T. Note that because trait matching is invariant, the types must be exactly equal.

RFC 447 and #24461 require that a parameter Xi can only appear in T if it is constrained by the trait reference <Q0 as Trait<Q1..Qn>>. The full definition of constrained appears below, but informally it means roughly that Xi appears in Q0..Qn somewhere outside of a projection. Let's call the constrained set of parameters Constrained(Q0..Qn).

Recall the rule OutlivesProjectionComponents:

OutlivesProjectionComponents:
  ∀i. R ⊢ Pi: 'a
  --------------------------------------------------
  R ⊢ <P0 as Trait<P1..Pn>>::Id: 'a

We aim to show that ∀i. R ⊢ Pi: 'a implies R ⊢ (Θ T): 'a, which implies that this rule is a sound approximation for normalization. The argument follows from two lemmas ("proofs" for these lemmas are sketched below):

  1. First, we show that if R ⊢ Pi: 'a, then every "subcomponent" P' of Pi outlives 'a. The idea here is that each variable Xi from the impl will match against and extract some subcomponent P' of Pi, and we wish to show that the subcomponent P' extracted by Xi outlives 'a.
  2. Then we will show that the type θ T outlives 'a if, for each of the in-scope parameters Xi, Θ Xi: 'a.

Definition 1. Constrained(T) defines the set of type/lifetime parameters that are constrained by a type. This set is found just by recursing over and extracting all subcomponents except for those found in a projection. This is because a type like X::Foo does not constrain what type X can take on, rather it uses X as an input to compute a result:

Constrained(scalar) = {}
Constrained(X) = {X}
Constrained(&'x T) = {'x} | Constrained(T)
Constrained(O0..On+'x) = Union(Constrained(Oi)) | {'x}
Constrained([T]) = Constrained(T),
Constrained(for<..> fn(T1..Tn) -> T0) = Union(Constrained(Ti))
Constrained(<P0 as Trait<P1..Pn>>::Id) = {} // empty set

Definition 2. Constrained('a) = {'a}. In other words, a lifetime reference just constraints itself.

Lemma 1: Given R ⊢ P: 'a, P = [X => P'] Q, and X ∈ Constrained(Q), then R ⊢ P': 'a. Proceed by induction and by cases over the form of P:

  1. If P is a scalar or parameter, there are no subcomponents, so P'=P.
  2. For nominal types, references, objects, and function types, either P'=P or P' is some subcomponent of P. The appropriate "outlives" rules all require that all subcomponents outlive 'a, and hence the conclusion follows by induction.
  3. If P' is a projection, that implies that P'=P.
    • Otherwise, Q must be a projection, and in that case, Constrained(Q) would be the empty set.

Lemma 2: Given that FV(T) ∈ X, ∀i. Ri: 'a, then [X => R] T: 'a. In other words, if all the type/lifetime parameters that appear in a type outlive 'a, then the type outlives 'a. Follows by inspection of the outlives rules.

Edit History

RFC1592 - amend to require that tuple fields be sized