Implied bounds

6 July 2014

I am on vacation for a few weeks. I wanted to take some time to jot down an idea that’s been bouncing around in my head. I plan to submit an RFC at some point on this topic, but not yet, so I thought I’d start out by writing a blog post. Also, my poor blog has been neglected for some time. Consider this a draft RFC. Some important details about references are omitted and will come in a follow-up blog post.

The high-level summary of the idea is that we will take advantage of bounds declared in type declarations to avoid repetition in fn and impl declarations.

Summary and motivation

Recent RFCs have introduced the ability to declare bounds within type declarations. For example, a HashMap type might be defined as follows:

struct HashMap<K:Hash,V> { ... }
trait Hash : Eq { ... }

These type declarations indicate that every hashmap is parameterized by a key type K and a value type V. Furthermore, K must be a hashable type. (The trait definition for Hash, meanwhile, indicates that every hashable type must also be equatable.)

Currently, the intention with these bounds is that every time the user writes HashMap<SomeKey,SomeValue>, the compiler will run off and verify that, indeed, SomeKey implements the trait Hash. (Which in turn implies that SomeKey implements Eq.)

This RFC introduces a slight twist to this idea. For the types of function parameters as well as the self types of impls, we will not verify their bounds immediately, but rather attach those bounds as [where clauses][where] on the fn. This shifts the responsibility for proving the bounds are satisfied onto the fn’s caller; in turn, it allows the fn to assume that the bounds are satisfied. The net result is that you don’t have to write as many duplicate bounds.

As applied to type parameter bounds

Let me give an example. Here is a generic function that inserts a key into a hashmap if there is no existing entry for the key:

fn insert_if_not_already_present<K,V>(
    hashmap: &mut HashMap<K,V>
    key: K,
    value: V)
    if hashmap.contains_key(&key) { return; }
    hashmap.insert(key, value);

Today this function would not type-check because the type K has no bounds. Instead one must declare K:Hash. But this bound feels rather pointless – after all, the fact that the function takes a hashmap as argument implies that K:Hash. With the proposed change, however, the fn above is perfectly legal.

Because impl self types are treated the same way, it will also be less repititious to define methods on a type. Whereas before one would have to write:

impl<K:Hash,V> HashMap<K,V> {

it is now sufficient to leave off the Hash bound, since it will be inferred from the self-type:

impl<K,V> HashMap<K,V> {

As applied to lifetimes

In fact, we already have a similar rule for lifetimes. Specifically, in some cases, we will infer a relationship between the lifetime parameters of a function. This is the reason that the following function is legal:

struct Foo { field: uint }
fn get_pointer<'a,'b>(x: &'a &'b Foo) -> &'a int {

Here, the lifetime of (**x).field (when all dereferences are written in full) is most properly 'b, but we are returning a reference with lifetime 'a. The compiler permits this because there exists a parameter of type &'a &'b Foo – from this, the compiler infers that 'a <= 'b. The basis for this inference is a rule that you cannot have a reference that outlives its referent. This is very helpful for making some programs typecheck: this is particularly true with generic traits, as described in this blog post.

Detailed design

Well-formed types and the BOUNDS function

We say that a type is well-formed if all of its bounds are met. We define a function BOUNDS(T) that maps from a type T to the set of bounds that must be satisfied for T to be called well-formed.

For the scalar types like int or float, BOUNDS just returns the empty set:

BOUNDS(int) = {}
BOUNDS(uint) = {}
BOUNDS(...) = {}

For struct types like HashMap<SomeKey,SomeValue>, the function combines the bounds declared on the HashMap type with those declared on SomeKey and SomeValue. (The SUBST() function is used to substitute the actual type parameters T1 ... Tn for their formal counterparts.)

                              BOUNDS(T1), ..., BOUNDS(Tn))

Enum and object types are handled in precisely the same way as struct types.

For vector types, the element type must be sized:

BOUNDS([T, ..N]) = UNION({T : Sized}, BOUNDS(T))

Well-formed references

For references, the type must have a suitable lower bound:

BOUNDS(&'a mut T) = UNION({'a <= LOWER-BOUND(T)}, BOUNDS(T))

Note that I have not defined the LOWER-BOUND function. The proper definition of this function is important and I have been working on it, but I prefer to defer that subject to a post/RFC of its own. (Clarifying the lower-bound function, however, is the heart of #5723 along with a number of other recent bugs being filed on lifetimes.) Note that this definition subsumes the existing rule for references described in my prior blog post.