lang (syntax | traits | typesystem)
Add a conservative form of abstract return types, also known as impl Trait
, that will be compatible with most possible future extensions
by initially being restricted to:
Abstract return types allow a function to hide a concrete return type behind a trait interface similar to trait objects, while still generating the same statically dispatched code as with concrete types.
With the placeholder syntax used in discussions so far, abstract return types would be used roughly like this:
fn foo(n: u32) -> impl Iterator<Item = u32> {
(0..n).map(|x| x * 100)
}
// ^ behaves as if it had return type Map<Range<u32>, Closure>
// where Closure = type of the |x| x * 100 closure.
for x in foo(10) {
// x = 0, 100, 200, ...
}
There has been much discussion around the impl Trait
feature already, with
different proposals extending the core idea into different directions:
This RFC is an attempt to make progress on the feature by proposing a minimal subset that should be forwards-compatible with a whole range of extensions that have been discussed (and will be reviewed in this RFC). However, even this small step requires resolving some of the core questions raised in the blog post.
This RFC is closest in spirit to the original RFC, and we'll repeat its motivation and some other parts of its text below.
Why are we doing this? What use cases does it support? What is the expected outcome?
In today's Rust, you can write a function signature like
fn consume_iter_static<I: Iterator<Item = u8>>(iter: I)
fn consume_iter_dynamic(iter: Box<Iterator<Item = u8>>)
In both cases, the function does not depend on the exact type of the argument. The type is held "abstract", and is assumed only to satisfy a trait bound.
In the _static
version using generics, each use of the function is
specialized to a concrete, statically-known type, giving static dispatch, inline
layout, and other performance wins.
In the _dynamic
version using trait objects, the concrete argument type is
only known at runtime using a vtable.
On the other hand, while you can write
fn produce_iter_dynamic() -> Box<Iterator<Item = u8>>
you cannot write something like
fn produce_iter_static() -> Iterator<Item = u8>
That is, in today's Rust, abstract return types can only be written using trait
objects, which can be a significant performance penalty. This RFC proposes
"unboxed abstract types" as a way of achieving signatures like
produce_iter_static
. Like generics, unboxed abstract types guarantee static
dispatch and inline data layout.
Here are some problems that unboxed abstract types solve or mitigate:
Returning unboxed closures. Closure syntax generates an anonymous type implementing a closure trait. Without unboxed abstract types, there is no way to use this syntax while returning the resulting closure unboxed because there is no way to write the name of the generated type.
Leaky APIs. Functions can easily leak implementation details in their return
type, when the API should really only promise a trait bound. For example, a
function returning Rev<Splits<'a, u8>>
is revealing exactly how the iterator
is constructed, when the function should only promise that it returns some
type implementing Iterator<Item = u8>
. Using newtypes/structs with private fields
helps, but is extra work. Unboxed abstract types make it as easy to promise only
a trait bound as it is to return a concrete type.
Complex types. Use of iterators in particular can lead to huge types:
Chain<Map<'a, (i32, u8), u16, Enumerate<Filter<'a, u8, vec::MoveItems<u8>>>>, SkipWhile<'a, u16, Map<'a, &u16, u16, slice::Items<u16>>>>
Even when using newtypes to hide the details, the type still has to be written out, which can be very painful. Unboxed abstract types only require writing the trait bound.
Documentation. In today's Rust, reading the documentation for the Iterator
trait is needlessly difficult. Many of the methods return new iterators, but
currently each one returns a different type (Chain
, Zip
, Map
, Filter
,
etc), and it requires drilling down into each of these types to determine what
kind of iterator they produce.
In short, unboxed abstract types make it easy for a function signature to promise nothing more than a trait bound, and do not generally require the function's author to write down the concrete type implementing the bound.
As explained at the start of the RFC, the focus here is a relatively narrow introduction of abstract types limited to the return type of inherent methods and free functions. While we still need to resolve some of the core questions about what an "abstract type" means even in these cases, we avoid some of the complexities that come along with allowing the feature in other locations or with other extensions.
Let's start with the bikeshed: The proposed syntax is impl Trait
in return type
position, composing like trait objects to forms like impl Foo + Send + 'a
.
It can be explained as "a type that implements Trait
",
and has been used in that form in most earlier discussions and proposals.
Initial versions of this RFC proposed @Trait
for brevity reasons,
since the feature is supposed to be used commonly once implemented,
but due to strong negative reactions by the community this has been
changed back to the current form.
There are other possibilities, like abstract Trait
or ~Trait
, with
good reasons for or against them, but since the concrete choice of syntax
is not a blocker for the implementation of this RFC, it is intended for
a possible follow-up RFC to address syntax changes if needed.
The core semantics of the feature is described below.
Note that the sections after this one go into more detail on some of the design decisions, and that it is likely for many of the mentioned limitations to be lifted at some point in the future. For clarity, we'll separately categorize the core semantics of the feature (aspects that would stay unchanged with future extensions) and the initial limitations (which are likely to be lifted later).
Core semantics:
If a function returns impl Trait
, its body can return values of any type that
implements Trait
, but all return values need to be of the same type.
As far as the typesystem and the compiler is concerned, the return type outside of the function would not be a entirely "new" type, nor would it be a simple type alias. Rather, its semantics would be very similar to that of generic type parameters inside a function, with small differences caused by being an output rather than an input of the function.
Sized
.Self
type for an impl
block.Because OIBITS like Send
and Sync
will leak through an abstract return
type, there will be some additional complexity in the compiler due to some
non-local type checking becoming necessary.
The return type has an identity based on all generic parameters the function body is parameterized by, and by the location of the function in the module system. This means type equality behaves like this:
fn foo<T: Trait>(t: T) -> impl Trait {
t
}
fn bar() -> impl Trait {
123
}
fn equal_type<T>(a: T, b: T) {}
equal_type(bar(), bar()); // OK
equal_type(foo::<i32>(0), foo::<i32>(0)); // OK
equal_type(bar(), foo::<i32>(0)); // ERROR, `impl Trait {bar}` is not the same type as `impl Trait {foo<i32>}`
equal_type(foo::<bool>(false), foo::<i32>(0)); // ERROR, `impl Trait {foo<bool>}` is not the same type as `impl Trait {foo<i32>}`
The code generation passes of the compiler would not draw a distinction between the abstract return type and the underlying type, just like they don't for generic parameters. This means:
-> impl Any
would return the type id of the underlying type.
Initial limitations:
impl Trait
may only be written within the return type of a freestanding or
inherent-impl function, not in trait definitions or any non-return type position. They may also not appear
in the return type of closure traits or function pointers,
unless these are themselves part of a legal return type.
impl Trait
multiple times in the same return type would be valid,
like for example in -> (impl Foo, impl Bar)
.The type produced when a function returns impl Trait
would be effectively
unnameable, just like closures and function items.
The function body cannot see through its own return type, so code like this would be forbidden just like on the outside:
fn sum_to(n: u32) -> impl Display {
if n == 0 {
0
} else {
n + sum_to(n - 1)
}
}
There has been a lot of discussion about what the semantics of the return type
should be, with the theoretical extremes being "full return type inference" and
"fully abstract type that behaves like a autogenerated newtype wrapper". (This
was in fact the main focus of the
blog post on impl Trait
.)
The design as chosen in this RFC lies somewhat in between those two, since it
allows OIBITs to leak through, and allows specialization to "see" the full type
being returned. That is, impl Trait
does not attempt to be a "tightly sealed"
abstraction boundary. The rationale for this design is a mixture of pragmatics
and principles.
Principles for specialization transparency:
The specialization RFC has given us a basic principle for how to understand bounds in function generics: they represent a minimum contract between the caller and the callee, in that the caller must meet at least those bounds, and the callee must be prepared to work with any type that meets at least those bounds. However, with specialization, the callee may choose different behavior when additional bounds hold.
This RFC abides by a similar interpretation for return types: the signature represents the minimum bound that the callee must satisfy, and the caller must be prepared to work with any type that meets at least that bound. Again, with specialization, the caller may dispatch on additional type information beyond those bounds.
In other words, to the extent that returning impl Trait
is intended to be
symmetric with taking a generic T: Trait
, transparency with respect to
specialization maintains that symmetry.
Pragmatics for specialization transparency:
The practical reason we want impl Trait
to be transparent to specialization is the
same as the reason we want specialization in the first place: to be able to
break through abstractions with more efficient special-case code.
This is particularly important for one of the primary intended usecases:
returning impl Iterator
. We are very likely to employ specialization for various
iterator types, and making the underlying return type invisible to
specialization would lose out on those efficiency wins.
OIBITs leak through an abstract return type. This might be considered controversial, since it effectively opens a channel where the result of function-local type inference affects item-level API, but has been deemed worth it for the following reasons:
Ergonomics: Trait objects already have the issue of explicitly needing to
declare Send
/Sync
-ability, and not extending this problem to abstract
return types is desirable. In practice, most uses of this feature would have
to add explicit bounds for OIBITS if they wanted to be maximally usable.
Low real change, since the situation already somewhat exists on structs with private fields:
In fact, a large part of the point of OIBITs in the first place was to cut across abstraction barriers and provide information about a type without the type's author having to explicitly opt in.
This means, however, that it has to be considered a silent breaking change to
change a function with an abstract return type in a way that removes OIBIT impls,
which might be a problem. (As noted above, this is already the case for struct
definitions.)
But since the number of used OIBITs is relatively small, deducing the return type in a function body and reasoning about whether such a breakage will occur has been deemed as a manageable amount of work.
In the most recent RFC related to this feature, a more "tightly sealed" abstraction mechanism was proposed. However, part of the discussion on specialization centered on precisely the issue of what type abstraction provides and how to achieve it. A particular salient point there is that, in Rust, privacy is already our primary mechanism for hiding ("privacy is the new parametricity"). In practice, that means that if you want opacity against specialization, you should use something like a newtype.
An abstract return type cannot be named in this proposal, which means that it
cannot be placed into structs
and so on. This is not a fundamental limitation
in any sense; the limitation is there both to keep this RFC simple, and because
the precise way we might want to allow naming of such types is still a bit
unclear. Some possibilities include a typeof
operator, or explicit named
abstract types.
There have been various proposed additional places where abstract types
might be usable. For example, fn x(y: impl Trait)
as shorthand for
fn x<T: Trait>(y: T)
.
Since the exact semantics and user experience for these locations are yet
unclear (impl Trait
would effectively behave completely different before and after
the ->
), this has also been excluded from this proposal.
Functions with abstract return types can not see through their own return type, making code like this not compile:
fn sum_to(n: u32) -> impl Display {
if n == 0 {
0
} else {
n + sum_to(n - 1)
}
}
This limitation exists because it is not clear how much a function body can and should know about different instantiations of itself.
It would be safe to allow recursive calls if the set of generic parameters is identical, and it might even be safe if the generic parameters are different, since you would still be inside the private body of the function, just differently instantiated.
But variance caused by lifetime parameters and the interaction with specialization makes it uncertain whether this would be sound.
In any case, it can be initially worked around by defining a local helper function like this:
fn sum_to(n: u32) -> impl Display {
fn sum_to_(n: u32) -> u32 {
if n == 0 {
0
} else {
n + sum_to_(n - 1)
}
}
sum_to_(n)
}
Because impl Trait
defines a type tied to the concrete function body,
it does not make much sense to talk about it separately in a function signature,
so the syntax is forbidden there.
One valid critique for the existing impl Trait
proposal is that it does not
cover more complex scenarios where the return type would implement
one or more traits depending on whether a type parameter does so with another.
For example, an iterator adapter might want to implement Iterator
and
DoubleEndedIterator
, depending on whether the adapted one does:
fn skip_one<I>(i: I) -> SkipOne<I> { ... }
struct SkipOne<I> { ... }
impl<I: Iterator> Iterator for SkipOne<I> { ... }
impl<I: DoubleEndedIterator> DoubleEndedIterator for SkipOne<I> { ... }
Using just -> impl Iterator
, this would not be possible to reproduce.
Since there have been no proposals so far that would address this in a way that would conflict with the fixed-trait-set case, this RFC punts on that issue as well.
One important usecase of abstract return types is to use them in trait methods.
However, there is an issue with this, namely that in combinations with generic trait methods, they are effectively equivalent to higher kinded types. Which is an issue because Rust's HKT story is not yet figured out, so any "accidental implementation" might cause unintended fallout.
HKT allows you to be generic over a type constructor, a.k.a. a
"thing with type parameters", and then instantiate them at some later point to
get the actual type.
For example, given a HK type T
that takes one type as parameter, you could
write code that uses T<u32>
or T<bool>
without caring about
whether T = Vec
, T = Box
, etc.
Now if we look at abstract return types, we have a similar situation:
trait Foo {
fn bar<U>() -> impl Baz
}
Given a T: Foo
, we could instantiate T::bar::<u32>
or T::bar::<bool>
,
and could get arbitrary different return types of bar
instantiated
with a u32
or bool
,
just like T<u32>
and T<bool>
might give us Vec<u32>
or Box<bool>
in the example above.
The problem does not exist with trait method return types today because they are concrete:
trait Foo {
fn bar<U>() -> X<U>
}
Given the above code, there is no way for bar
to choose a return type X
that could fundamentally differ between instantiations of Self
while still being instantiable with an arbitrary U
.
At most you could return a associated type, but then you'd lose the generics
from bar
trait Foo {
type X;
fn bar<U>() -> Self::X // No way to apply U
}
So, in conclusion, since Rust's HKT story is not yet fleshed out, and the compatibility of the current compiler with it is unknown, it is not yet possible to reach a concrete solution here.
In addition to that, there are also different proposals as to whether an abstract return type is its own thing or sugar for a associated type, how it interacts with other associated items and so on, so forbidding them in traits seems like the best initial course of action.
Why should we not do this?
As has been elaborated on above, there are various way this feature could be extended and combined with the language, so implementing it might cause issues down the road if limitations or incompatibilities become apparent. However, variations of this RFC's proposal have been under discussion for quite a long time at this point, and this proposal is carefully designed to be future-compatible with them, while resolving the core issue around transparency.
A drawback of limiting the feature to return type position (and not arguments)
is that it creates a somewhat inconsistent mental model: it forces you to
understand the feature in a highly special-cased way, rather than as a general
way to talk about unknown-but-bounded types in function signatures. This could
be particularly bewildering to newcomers, who must choose between T: Trait
,
Box<Trait>
, and impl Trait
, with the latter only usable in one place.
The fact that specialization and OIBITs can "see through" impl Trait
may be
surprising, to the extent that one wants to see impl Trait
as an abstraction
mechanism. However, as the RFC argued in the rationale section, this design is
probably the most consistent with our existing post-specialization abstraction
mechanisms, and lead to the relatively simple story that privacy is the way to
achieve hiding in Rust.
What other designs have been considered? What is the impact of not doing this?
See the links in the motivation section for detailed analysis that we won't repeat here.
But basically, without this feature certain things remain hard or impossible to do in Rust, like returning a efficiently usable type parameterized by types private to a function body, for example an iterator adapter containing a closure.
What parts of the design are still to be determined?
The precise implementation details for OIBIT transparency are a bit unclear: in general, it means that type checking may need to proceed in a particular order, since you cannot get the full type information from the signature alone (you have to typecheck the function body to determine which OIBITs apply).