Fork me on GitHub

Futhark 0.14.1 released

Posted on January 24, 2020 by Troels Henriksen

Another release of Futhark is out (full changelog). This release has a large number of user-visible changes, but with one major exception, most are fiddly things intended to ease the eventual transition to size types.

The exception is that tuples in Futhark are now 0-indexed rather than 1-indexed. This was one of the design flaws I discussed in a previous blog post and the resulting discussion gave me the confidence necessary to carry out the change. My reasoning was helped along by the fact that tuples are also 0-indexed in Rust. Generally, when in doubt about some relatively superficial syntactical or lexical detail, do what Rust does. They’ve generally put significant thought into these details, and so will not have any hidden pitfalls. (Similarly, think hard when deviating from the Standard ML type system!) It will probably take some time to chase down all the old Futhark code and examples that still uses 1-indexed tuples, but it will happen eventually.

Inching towards size types

Size types are all about statically ensuring that arrays that should have the same size (such as the inputs to a dot product) will have the same size. Currently, this is done with dynamic checks.

A design for size types could quickly decay into full dependent typing with arbitrary type-level arithmetic, which makes for a rather complicated language. To keep Futhark simple, we restrict array sizes to being either variable names or constants. For example, [n]i32 is an array type, as is [2]bool, but [n+1]i32 is not. Whenever the result size of an operation cannot be expressed as a constant or variable name, e.g. when concatenating or filtering arrays, the type checker will invent an “unknowable size variable” to describe the size. We are still working out the details of size typing, in particular with respect to making the errors understandable. It will take some time to make the switch, so our plan is to make it possible to write code that is type-correct under both the current lax rules, and the new stricter rules. This gives us the chance to experiment on real code, without having to set up an alternative universe of libraries and test programs.

To a large extent, this has worked well. Size types have not required any new language constructs or syntax (until now, as I’ll get to later), as they were merely extra rules imposed by the type checker. However, some of the existing semantics of the language had to be tightened up.

As an example, consider the range expression x..<y. This constructs an array of the integers x to y-1. In general, the size of this array is unknowable, but for the common case 0..<y, it will have size y. But what if y is a negative number at run-time? In our previous semantics, invalid ranges just produced empty arrays, but an empty array has size 0, not y (which is negative). Clearly this won’t fly in a language where sizes actually matter. Instead, invalid ranges now produce a run-time error, which makes the types not a lie.

Changes to value specs

Another thing we changed was value specs. These are used in module types to describe interfaces. For example, we can declare that a module provides a function reverse that given an array with elements t, produces an array of the same size:

val reverse [n] : [n]t -> [n]t

We use a size parameter [n] to give a name to the size of the input array. Previously, such sizes were optional, and we could also have written:

val reverse : []t -> []t

This has strictly less information, because it doesn’t encode the fact that reverse is size-preserving. However it’s useful for functions where the return type does not mention the size:

val sum : []t -> t

Without size types, we can have “anonymous” dimension sizes, since they don’t really matter except for documentation. With size types, it’s important that every dimension has a size. It’s no technical problem for the compiler to find any empty dimensions and invent new names for them, and for value definitions, that is what we do. However, module types are supposed to be for describing interfaces, so we think it is cleaner to require that all sizes must be given a name (except for return types, where an anonymous dimension means that the size cannot be expressed statically).

Size-lifted types

An irregular array is a multidimensional array where the rows have different sizes:

[[1,2,3],
 [4,5],
 [6]]

For various efficiency reasons, these are not supported in Futhark. Currently, this is enforced with dynamic checks. One of the promises of size typing is to enforce this with static checks instead.

Fortunately, we get this property almost for free. By making sizes part of the type, an array expression [x, y] is regular if x and y have the same type. The only thing we have to make certain is that sizes cannot be “hidden” through abstract types. There are two places in Futhark where the definition of a type can be hidden: through parametric polymorphism (type parameters), and via the module system. The latter is the more interesting case.

With a module type, we can declare that some module must define a type, and a few values of that type:

module type mt = {
  type t
  val x : t
  val y : t
}

With a module type ascription we can declare that a module implements mt, but hide the specific definition of the type t:

module m : mt = {
  type t = i32
  let x = 1
  let y = 2
}

Now we can construct an array [m.x, m.y], which will work fine, although it’s not very useful. But now consider this definition:

module m : mt = {
  type t = []i32
  let x = [1]
  let y = [2,2]
}

What happens when we do [m.x, m.y]? Since the definition of m.t is abstract, the type checker has no reason to object, but we are really constructing an irregular array here! That’s a problem. The solution is to forbid type abbreviations that have anonymous sizes in their definition. It is still possible to write size-parametric types, as this just propagates the inner sizes outwards:

type vector [n] = [n]i32

Unfortunately, this solution is also a bit too restrictive. Sometimes it is useful to have abstract type that can internally contain an arrays of hidden sizes. That’s what abstraction is all about! To handle this case, we introduce the notion of size-lifted types:

module type mt = {
  type~ t
  val x : t
  val y : t
}

module m : mt = {
  type~ t = []i32
  let x = [1]
  let y = [2,2]
}

When a type is declared with type~ (rather than type), it is permitted to have anonymous sizes in the definition. On the other hand, size-lifted types cannot be put in arrays, so now [m.x, m.y] will be a type error.

Size-lifted types are similar to fully lifted types (defined with type^), which we previously used to handle higher-order functions. In fact, size-lifted types are a subset of fully lifted types, as fully lifted types also cannot be put in arrays.

While I’m not terribly excited about adding another kind of types to the type system, it’s the simplest solution we could think of, and the mechanism has precedent in Futhark, where it has worked well.

Size coercion

Since our size type system is so simple, there are many situations where the programmer will have to insert explicit coercions. For example, if you know somehow that some filter expression will always produce an array of size k, you’d write:

filter f xs : [k]i32

This repurposes the type ascription operator to perform a kind of dynamic coercion on sizes, where a run-time check verifies that the result of filter is of length k.

While this has the advantage of not requiring any new syntax, it has the downside of changing type ascription from an operator that you insert to improve clarity or disambiguate a type, to one that might perform a dynamic check. So instead, this version of the compiler introduces size coercion:

filter f xs :> [k]i32

It currently means exactly the same thing as type ascription. With size types, it will perform a dynamic coercion (and type ascription will go back to enforcing the type statically).