Version 3 (modified by benl, 5 years ago)

--

< Overview

Regions and the Class System

Regions

In short, two pieces of data are in different regions if they are never substituted for each other. This property, or lack thereof, is sometimes called aliasing.

In the "extended" version of a type, data type constructors have a region annotation as their first argument, which indicates what region they're in. Due to type elaboration, we usually don't see the region annotations, but we can write them in signatures if we want to:

succ :: forall %r1 %r2. Int %r1 -> Int %r2
succ x = x + 1

sameInt :: forall %r1. Int %r1 -> Int %r1
sameInt x = x

pi :: Float %r1
pi = 3.1415926535

Region variables can be quantified with forall much like type variables. If a region variable in the return type of a function is quantified it means the region is fresh, ie the data was allocated by the function itself.

Notice that in the type of succ, both %r1 and %r2 are quantified, this means that succ accepts data from any region and returns a freshly allocated Int.

sameInt just passes its data though, so the same region is on both argument and return types.

pi is just static Float and not a function that does allocation, so it doesn't have a forall.

Region classes

In Haskell we use type classes on type variables to restrict how these variables can be instantiated.

(==) :: forall a. Eq a => a -> a -> Bool

The Eq a context restricts forall a to just the types that support equality.

In Disciple, we can do a similar thing with regions:

csucc :: forall %r1 %r2
      .  Int %r1 -> Int %r2
      :- Const %r1

The region class constraint Const %r1 restricts csucc so that it only accepts arguments that are constant. Data in Const regions is guaranteed by the type system never to be destructively updated. In Disciple we write the class constraints at the end of the type for clarity, though there is a plan to allow the standard Haskell syntax as well.

The opposite of Const is Mutable and we can explicitly define data values to have this property.

counter :: Int %r1 :- Mutable %r1
counter = 0

Any Int that is passed to csucc is required to be Const. If you try and pass a Mutable Int instead it will be caught by the type system.

main () = putStr $ show $ succ counter
   ./Main.ds:...
       Conflicting region constraints.
                 constraint: Base.Mutable %1
            from the use of: counter
                         at: ./Main.ds:...

        conflicts with,
                 constraint: Base.Const %2
            from the use of: succ
                         at: ./Main.ds:...

Effect purification

Besides manually added annotations, the main source of Const constraints in a Disciple program is the use of laziness. Remember from EvaluationOrder that the suspension operator (@) maps onto a set of primitive suspend functions.

If we wanted a lazy version of succ that only incremented its argument when the result was demanded, then we could write it like this:

lazySucc x = (+) @ x 1

this is desugared to:

lazySucc x = suspend2 (+) x 1

where suspend2 has the following extended type:

suspend2
	:: forall a b c !e1 !e2
	.  (a -(!e1)> b -(!e2)> c) -> a -> b -> c
	,  Pure  !e1
	,  Pure  !e2

Now (+) for Int has this extended type:

(+)  :: forall %r1 %r2 %r3 !e1
     .  Int %r1 -> Int %r2 -(!e1)> Int %r3
     :- !e1 = !{ !Read %r1; !Read %r2 }

Which says it reads the two arguments and returns a freshly allocated Int

When we pass (+) as the first argument of suspend2 the type system ends up with an effect class constraint Pure !{!Read %r1; !Read %r2;} which it satisfies by forcing %r1 and %r2 to be Const. If a value is in a Const region then it is guaranteed never to be destructively updated. This means that it doesn't matter when we read it, so the operation is pure.

To say this another way, when a function application is suspended the type system purifies its visible Read effects by requiring the data being read to be Const.