DDC/ClassSystem

From HaskellWiki

Revision as of 05:11, 20 March 2008 (edit) Benl23 (Talk | contribs) m ← Previous diff		Revision as of 21:11, 8 April 2008 (edit) (undo) DavidLazar (Talk | contribs) m (typo) Next diff →
Line 1:		Line 1:
	== Regions ==		== 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 sometimes called ''aliasing''.	+	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''.
	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:		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:

---

Revision as of 21:11, 8 April 2008

1 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.

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:

Region variables can be quantified with 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 , both and are quantified, this means that accepts data from any region and returns a freshly allocated . just passes its data though, so the same region is on both argument and return types. is just static and not a function that does allocation, so it doesn't have a .

2 Region classes

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

The here restricts 'forall a' to just the types that support equality.

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

The region class constraint restricts so that it only accepts arguments which are constant. Data in 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 (ticket) to allow the standard Haskell syntax as well. The opposite of is and we can explicitly define data values to have this property. Any that is passed to is required to be . If you try and pass a instead it will be caught by the type system.

   ./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:...

3 Effect purification

Besides manually added annotations, the main source of constraints in a Disciple program is the use of laziness. Remember from EvaluationOrder that the suspension operator maps onto a set of primitive functions. If we wanted a lazy version of that only incremented its argument when the result was demanded, then we could write it like this:

this is desugared to:

where has the following type (ignoring closure information once again) Now (+) for has this type: Which says it reads the two arguments and returns a freshly allocated . When we pass as the first argument of the type system ends up with an effect class constraint which it satisfies by forcing and to be . If a value is in a region then it is guaranteed never to be destructively updated, which means it doesn't matter when we read it, which means that operation is pure. To say this another way, when a function application is suspended the type system purifies its visible effects by requiring the data being read to be .