lang (traits | typesystem | method-call | resolve | trait-object | machine | traits)
Cleanup the trait, method, and operator semantics so that they are well-defined and cover more use cases. A high-level summary of the changes is as follows:
&self
and &mut self
etc,
so that self-type declarations like self: Rc<Self>
become possible.This RFC excludes discussion of associated types and multidimensional type classes, which will be the subject of a follow-up RFC.
The current trait system is ill-specified and inadequate. Its implementation dates from a rather different language. It should be put onto a surer footing.
Addressed by: New method resolution algorithm.
The deref operator *
is a flexible one. Imagine a pointer p
of
type ~T
. This same *
operator can be used for three distinct
purposes, depending on context.
&*p
.&mut *p
.consume(*p)
.Not all of these operations are supported by all types. In fact,
because most smart pointers represent aliasable data, they will only
support the creation of immutable references (e.g., Rc
, Gc
).
Other smart pointers (e.g., the RefMut
type returned by RefCell
)
support mutable or immutable references, but not moves. Finally, a
type that owns its data (like, indeed, ~T
) might support #3.
To reflect this, we use distinct traits for the various operators. (In fact, we don't currently have a trait for copying/moving the contents, this could be a distinct RFC (ed., I'm still thinking this over myself, there are non-trivial interactions)).
Unfortunately, the method call algorithm can't really reliably choose
mutable vs immutable deref. The challenge is that the proper choice
will sometimes not be apparent until quite late in the process. For
example, imagine the expression p.foo()
: if foo()
is defined with
&self
, we want an immutable deref, otherwise we want a mutable
deref.
Note that in this RFC I do not completely address this issue. In
particular, in an expression like (*p).foo()
, where the dereference
is explicit and not automatically inserted, the sense of the
dereference is not inferred. For the time being, the sense can be
manually specified by making the receiver type fully explicit: (&mut *p).foo()
vs (&*p).foo()
. I expect in a follow-up RFC to possibly
address this problem, as well as the question of how to handle copies
and moves of the referent (use #3 in my list above).
Addressed by: New method resolution algorithm.
Issue #XYZ. When multiple traits define methods with the same name, it is ambiguous which trait is being used:
trait Foo { fn method(&self); }
trait Bar { fn method(&self); }
In general, so long as a given type only implements Foo
or Bar
,
these ambiguities don't present a problem (and ultimately Universal
Function Call Syntax or UFCS will present an explicit resolution).
However, this is not guaranteed. Sometimes we see "blanket" impls
like the following:
impl<A:Base> Foo for A { }
This impl basically says "any type T
that implements Base
automatically implements Foo
". Now, we expect an ambiguity error
if we have a type T
that implements both Base
and Bar
. But in
fact, we'll get an ambiguity error even if a type only implements
Bar
. The reason for this is that the current method resolution
doesn't "recurse" and check additional dependencies when deciding if
an impl
is applicable. So it will decide, in this case, that the
type T
could implement Foo
and then record for later that T
must
implement Base
. This will lead to weird errors.
Addressed by: Expanded coherence rules.
The job of coherence is to ensure that, for any given set of type parameters, a given trait is implemented at most once (it may of course not be implemented at all). Currently, however, coherence is more conservative that it needs to be. This is partly because it doesn't take into account the very property that it itself is enforcing.
The problems arise due to the "blanket impls" I discussed in the previous section. Consider the following two traits and a blanket impl:
trait Base { }
trait Derived { }
impl<A:Base> Derived for A { }
Here we have two traits Base
and Derived
, and a blanket impl which
implements the Derived
trait for any type A
that also implements
Base
.
This implies that if you implement Base
for a type S
, then S
automatically implements Derived
:
struct S;
impl Base for S { } // Implement Base => Implements Derived
On a related note, it'd be an error to implement both Base
and Derived
for the same type T
:
// Illegal
struct T;
impl Base for T { }
impl Derived for T { }
This is illegal because now there are two implements of Derived
for T
. There is the direct one, but also an indirect one. We do not
assign either higher precedence, we just report it as an error.
So far, all is in agreement with the current rules. However, problems
arise if we imagine a type U
that only implements Derived
:
struct U;
impl Derived for U { } // Should be OK, currently not.
In this scenario, there is only one implementation of Derived
. But
the current coherence rules still report it as an error.
Here is a concrete example where a rule like this would be useful. We
currently have the Copy
trait (aka Pod
), which states that a type
can be memcopied. We also have the Clone
trait, which is a more
heavyweight version for types where copying requires allocation. It'd
be nice if all types that could be copied could also be cloned -- it'd
also be nice if we knew for sure that copying a value had the same
semantics as cloning it, in that case. We can guarantee both using a
blanket impl like the following:
impl<T:Copy> Clone for T {
fn clone(&self) -> T {
*self
}
}
Unfortunately, writing such an impl today would imply that no other
types could implement Clone
. Obviously a non-starter.
There is one not especially interesting ramification of
this. Permitting this rule means that adding impls to a type could
cause coherence errors. For example, if I had a type which implements
Copy
, and I add an explicit implementation of Clone
, I'd get an
error due to the blanket impl. This could be seen as undesirable
(perhaps we'd like to preserve that property that one can always add
impls without causing errors).
But of course we already don't have the property that one can always add impls, since method calls could become ambiguous. And if we were to add "negative bounds", which might be nice, we'd lose that property. And the popularity and usefulness of blanket impls cannot be denied. Therefore, I think this property ("always being able to add impls") is not especially useful or important.
Addressed by: Gradual vtable resolution algorithm
In an effort to improve inference, the current implementation has a rather ad-hoc two-pass scheme. When performing a method call, it will immediately attempt "early" trait resolution and -- if that fails -- defer checking until later. This helps with some particular scenarios, such as a trait like:
trait Map<E> {
fn map(&self, op: |&E| -> E) -> Self;
}
Given some higher-order function like:
fn some_mapping<E,V:Map<E>>(v: &V, op: |&E| -> E) { ... }
If we were then to see a call like:
some_mapping(vec, |elem| ...)
the early resolution would be helpful in connecting the type of elem
with the type of vec
. The reason to use two phases is that often we
don't need to resolve each trait bound to a specific impl, and if we
wait till the end then we will have more type information available.
In my proposed solution, we eliminate the phase distinction. Instead, we simply track pending constraints. We are free to attempt to resolve pending constraints whenever desired. In paricular, whenever we find we need more type information to proceed with some type-overloaded operation, rather than reporting an error we can try and resolve pending constraints. If that helps give more information, we can carry on. Once we reach the end of the function, we must then resolve all pending constraints that have not yet been resolved for some other reason.
Note that there is some interaction with the distinction between input
and output type parameters discussed in the previous
example. Specifically, we must never infer the value of the Self
type parameter based on the impls in scope. This is because it would
cause crate concatentation to potentially lead to compilation errors
in the form of inference failure.
There are important properties I would like to guarantee:
Here are some properties I do not intend to guarantee:
undecidable-instances
flag. I don't think decidability is
especially important. The compiler can simply keep a recursion
counter and report an error if that level of recursion is
exceeded. This counter can be adjusted by the user on a
crate-by-crate basis if some bizarre impl pattern happens to require
a deeper depth to be resolved.In general, I won't give a complete algorithmic specification. Instead, I refer readers to the prototype implementation. I would like to write out a declarative and non-algorithmic specification for the rules too, but that is work in progress and beyond the scope of this RFC. Instead, I'll try to explain in "plain English".
Currently methods must be declared using the explicit-self shorthands:
fn foo(self, ...)
fn foo(&self, ...)
fn foo(&mut self, ...)
fn foo(~self, ...)
Under this proposal we would keep these shorthands but also permit any
function in a trait to be used as a method, so long as the type of the
first parameter is either Self
or something derefable Self
:
fn foo(self: Gc<Self>, ...)
fn foo(self: Rc<Self>, ...)
fn foo(self: Self, ...) // equivalent to `fn foo(self, ...)
fn foo(self: &Self, ...) // equivalent to `fn foo(&self, ...)
It would not be required that the first parameter be named self
,
though it seems like it would be useful to permit it. It's also
possible we can simply make self
not be a keyword (that would be my
personal preference, if we can achieve it).
The coherence rules fall into two categories: the orphan restriction and the overlapping implementations restriction.
Orphan check: Every implementation must meet one of the following conditions:
The trait being implemented (if any) must be defined in the current crate.
The Self
type parameter must meet the following grammar, where
C
is a struct or enum defined within the current crate:
T = C
| [T]
| [T, ..n]
| &T
| &mut T
| ~T
| (..., T, ...)
| X<..., T, ...> where X is not bivariant with respect to T
Overlapping instances: No two implementations of the same trait can
be defined for the same type (note that it is only the Self
type
that matters). For this purpose of this check, we will also
recursively check bounds. This check is ultimately defined in terms of
the RESOLVE algorithm discussed in the implementation section below:
it must be able to conclude that the requirements of one impl are
incompatible with the other.
Here is a simple example that is OK:
trait Show { ... }
impl Show for int { ... }
impl Show for uint { ... }
The first impl implements Show for int
and the case implements
Show for uint
. This is ok because the type int
cannot be unified
with uint
.
The following example is NOT OK:
trait Iterator<E> { ... }
impl Iterator<char> for ~str { ... }
impl Iterator<u8> for ~str { ... }
Even though E
is bound to two distinct types, E
is an output type
parameter, and hence we get a coherence violation because the input
type parameters are the same in each case.
Here is a more complex example that is also OK:
trait Clone { ... }
impl<A:Copy> Clone for A { ... }
impl<B:Clone> Clone for ~B { ... }
These two impls are compatible because the resolution algorithm is
able to see that the type ~B
will never implement Copy
, no matter
what B
is. (Note that our ability to do this check relies on the
orphan checks: without those, we'd never know if some other crate
might add an implementation of Copy
for ~B
.)
Since trait resolution is not fully decidable, it is possible to concoct scenarios in which coherence can neither confirm nor deny the possibility that two impls are overlapping. One way for this to happen is when there are two traits which the user knows are mutually exclusive; mutual exclusion is not currently expressible in the type system [7] however, and hence the coherence check will report errors. For example:
trait Even { } // Naturally can't be Even and Odd at once!
trait Odd { }
impl<T:Even> Foo for T { }
impl<T:Odd> Foo for T { }
Another possible scenario is infinite recursion between impls. For example, in the following scenario, the coherence checked would be unable to decide if the following impls overlap:
impl<A:Foo> Bar for A { ... }
impl<A:Bar> Foo for A { ... }
In such cases, the recursion bound is exceeded and an error is conservatively reported. (Note that recursion is not always so easily detected.)
Let us assume the method call is r.m(...)
and the type of the
receiver r
is R
. We will resolve the call in two phases. The first
phase checks for inherent methods [4] and the second phase for
trait methods. Both phases work in a similar way, however. We will
just describe how trait method search works and then express the
inherent method search in terms of traits.
The core method search looks like this:
METHOD-SEARCH(R, m):
let TRAITS = the set consisting of any in-scope trait T where:
1. T has a method m and
2. R implements T<...> for any values of Ty's type parameters
if TRAITS is an empty set:
if RECURSION DEPTH EXCEEDED:
return UNDECIDABLE
if R implements Deref<U> for some U:
return METHOD-SEARCH(U, m)
return NO-MATCH
if TRAITS is the singleton set {T}:
RECONCILE(R, T, m)
return AMBIGUITY(TRAITS)
Basically, we will continuously auto-dereference the receiver type,
searching for some type that implements a trait that offers the method
m
. This gives precedence to implementations that require fewer
autodereferences. (There exists the possibility of a cycle in the
Deref
chain, so we will only autoderef so many times before
reporting an error.)
Once we find a trait that is implemented for the (adjusted) receiver
type R
and which offers the method m
, we must reconcile the
receiver with the self type declared in m
. Let me explain by
example.
Consider a trait Mob
(anyone who ever hacked on the MUD source code
will surely remember Mobs!):
trait Mob {
fn hit_points(&self) -> int;
fn take_damage(&mut self, damage: int) -> int;
fn move_to_room(self: GC<Self>, room: &Room);
}
Let's say we have a type Monster
, and Monster
implements Mob
:
struct Monster { ... }
impl Mob for Monster { ... }
And now we see a call to hit_points()
like so:
fn attack(victim: &mut Monster) {
let hp = victim.hit_points();
...
}
Our method search algorithm above will proceed by searching for an
implementation of Mob
for the type &mut Monster
. It won't find
any. It will auto-deref &mut Monster
to yield the type Monster
and
search again. Now we find a match. Thus far, then, we have a single
autoderef *victims
, yielding the type Monster
-- but the method
hit_points()
actually expects a reference (&Monster
) to be given
to it, not a by-value Monster
.
This is where self-type reconciliation steps in. The reconciliation process works by unwinding the adjustments and adding auto-refs:
RECONCILE(R, T, m):
let E = the expected self type of m in trait T;
// Case 1.
if R <: E:
we're done.
// Case 2.
if &R <: E:
add an autoref adjustment, we're done.
// Case 3.
if &mut R <: E:
adjust R for mutable borrow (if not possible, error).
add a mut autoref adjustment, we're done.
// Case 4.
unwind one adjustment to yield R' (if not possible, error).
return RECONCILE(R', T, m)
In this case, the expected self type E
would be &Monster
. We would
first check for case 1: is Monster <: &Monster
? It is not. We would
then proceed to case 2. Is &Monster <: &Monster
? It is, and hence
add an autoref. The final result then is that victim.hit_points()
becomes transformed to the equivalent of (using UFCS notation)
Mob::hit_points(&*victim)
.
To understand case 3, let's look at a call to take_damage
:
fn attack(victim: &mut Monster) {
let hp = victim.hit_points(); // ...this is what we saw before
let damage = hp / 10; // 1/10 of current HP in damage
victim.take_damage(damage);
...
}
As before, we would auto-deref once to find the type Monster
. This
time, though, the expected self type is &mut Monster
. This means
that both cases 1 and 2 fail and we wind up at case 3, the test for
which succeeds. Now we get to this statement: "adjust R
for mutable
borrow".
At issue here is the
overloading of the deref operator that was discussed earlier.
In this case, the end result we want is Mob::hit_points(&mut *victim)
, which means that *
is being used for a mutable borrow,
which is indicated by the DerefMut
trait. However, while doing the
autoderef loop, we always searched for impls of the Deref
trait,
since we did not yet know which trait we wanted. [2] We need to
patch this up. So this loop will check whether the type &mut Monster
implements DerefMut
, in addition to just Deref
(it does).
This check for case 3 could fail if, e.g., victim
had a type like
Gc<Monster>
or Rc<Monster>
. You'd get a nice error message like
"the type Rc
does not support mutable borrows, and the method
take_damage()
requires a mutable receiver".
We still have not seen an example of cases 1 or 4. Let's use a slightly modified example:
fn flee_if_possible(victim: Gc<Monster>, room: &mut Room) {
match room.find_random_exit() {
None => { }
Some(exit) => {
victim.move_to_room(exit);
}
}
}
As before, we'll start out with a type of Monster
, but this type the
method move_to_room()
has a receiver type of Gc<Monster>
. This
doesn't match cases 1, 2, or 3, so we proceed to case 4 and unwind
by one adjustment. Since the most recent adjustment was to deref from
Gc<Monster>
to Monster
, we are left with a type of
Gc<Monster>
. We now search again. This time, we match case 1. So the
final result is Mob::move_to_room(victim, room)
. This last case is
sort of interesting because we had to use the autoderef to find the
method, but once resolution is complete we do not wind up
dereferencing victim
at all.
Finally, let's see an error involving case 4. Imagine we modified
the type of victim
in our previous example to be &Monster
and
not Gc<Monster>
:
fn flee_if_possible(victim: &Monster, room: &mut Room) {
match room.find_random_exit() {
None => { }
Some(exit) => {
victim.move_to_room(exit);
}
}
}
In this case, we would again unwind an adjustment, going from
Monster
to &Monster
, but at that point we'd be stuck. There are no
more adjustments to unwind and we never found a type
Gc<Monster>
. Therefore, we report an error like "the method
move_to_room()
expects a Gc<Monster>
but was invoked with an
&Monster
".
Inherent methods can be "desugared" into traits by assuming a trait
per struct or enum. Each impl like impl Foo
is effectively an
implementation of that trait, and all those traits are assumed to be
imported and in scope.
Today's algorithm isn't really formally defined, but it works very differently from this one. For one thing, it is based purely on subtyping checks, and does not rely on the generic trait matching. This is a crucial limitation that prevents cases like those described in lack of backtracking from working. It also results in a lot of code duplication and a general mess.
One of the goals of this proposal is to remove the hokey distinction between early and late resolution. The way that this will work now is that, as we execute, we'll accumulate a list of pending trait obligations. Each obligation is the combination of a trait and set of types. It is called an obligation because, for the method to be correctly typed, we must eventually find an implementation of that trait for those types. Due to type inference, though, it may not be possible to do this right away, since some of the types may not yet be fully known.
The semantics of trait resolution mean that, at any point in time, the type checker is free to stop what it's doing and try to resolve these pending obligations, so long as none of the input type parameters are unresolved (see below). If it is able to definitely match an impl, this may in turn affect some type variables which are output type parameters. The basic idea then is to always defer doing resolution until we either (a) encounter a point where we need more type information to proceed or (b) have finished checking the function. At those times, we can go ahead and try to do resolution. If, after type checking the function in its entirety, there are still obligations that cannot be definitely resolved, that's an error.
To ensure crate concentanability, we must only consider the Self
type parameter when deciding when a trait has been implemented (more
generally, we must know the precise set of input type parameters; I
will cover an expanded set of rules for this in a subsequent RFC).
To see why this matters, imagine a scenario like this one:
trait Produce<R> {
fn produce(&self: Self) -> R;
}
Now imagine I have two crates, C and D. Crate C defines two types,
Vector
and Real
, and specifies a way to combine them:
struct Vector;
impl Produce<int> for Vector { ... }
Now imagine crate C has some code like:
fn foo() {
let mut v = None;
loop {
if v.is_some() {
let x = v.get().produce(); // (*)
...
} else {
v = Some(Vector);
}
}
}
At the point (*)
of the call to produce()
we do not yet know the
type of the receiver. But the inferencer might conclude that, since it
can only see one impl
of Produce
for Vector
, v
must have type
Vector
and hence x
must have the type int
.
However, then we might find another crate D that adds a new impl:
struct Other;
struct Real;
impl Combine<Real> for Other { ... }
This definition passes the orphan check because at least one of the
types (Real
, in this case) in the impl is local to the current
crate. But what does this mean for our previous inference result? In
general, it looks risky to decide types based on the impls we can see,
since there could always be more impls we can't actually see.
It seems clear that this aggressive inference breaks the crate concatenation property. If we combined crates C and D into one crate, then inference would fail where it worked before.
If x
were never used in any way that forces it to be an int
, then
it's even plausible that the type Real
would have been valid in some
sense. So the inferencer is influencing program execution to some
extent.
The basis for the coherence check, method lookup, and vtable lookup algorithms is the same function, called RESOLVE. The basic idea is that it takes a set of obligations and tries to resolve them. The result is four sets:
In general, if we ever encounter a NO-IMPL or UNDECIDABLE, it's probably an error. DEFERRED obligations are ok until we reach the end of the function. For details, please refer to the prototype.
The addition of a Deref
trait makes autoderef complicated, because
we may encounter situations where the smart pointer and its
reference both implement a trait, and we cannot know what the user
wanted.
The current rule just decides in favor of the smart pointer; this is somewhat unfortunate because it is likely to not be what the user wanted. It also means that adding methods to smart pointer types is a potentially breaking change. This is particularly problematic because we may want the smart pointer to implement a trait that requires the method in question!
An interesting thought would be to change this rule and say that we
always autoderef first and only resolve the method against the
innermost reference. Note that UFCS provides an explicit "opt-out" if
this is not what was desired. This should also have the (beneficial,
in my mind) effect of quelling the over-eager use of Deref
for types
that are not smart pointers.
This idea appeals to me but I think belongs in a separate RFC. It needs to be evaluated.
Note 1: when combining with DST, the in
keyword goes
first, and then any other qualifiers. For example, in unsized RHS
or
in type RHS
etc. (The precise qualifier in use will depend on the
DST proposal.)
Note 2: Note that the DerefMut<T>
trait extends
Deref<T>
, so if a type supports mutable derefs, it must also support
immutable derefs.
Note 3: The restriction that inputs must precede outputs is not strictly necessary. I added it to keep options open concerning associated types and so forth. See the Alternatives section, specifically the section on associated types.
Note 4: The prioritization of inherent methods could be
reconsidered after DST has been implemented. It is currently needed to
make impls like impl Trait for ~Trait
work.
Note 5: The set of in-scope traits is currently defined as those that are imported by name. PR #37 proposes possible changes to this rule.
Note 6: In the section on autoderef and ambiguity, I
discuss alternate rules that might allow us to lift the requirement
that the receiver be named self
.
Note 7: I am considering introducing mechanisms in a subsequent RFC that could be used to express mutual exclusion of traits.