Skip to content

Latest commit

 

History

History
278 lines (190 loc) · 7.69 KB

README.md

File metadata and controls

278 lines (190 loc) · 7.69 KB
Raw

lamdu-calculus

An extended typed Lambda Calculus AST.

Used by Lamdu as the underlying unsugared language.

The Lamdu Calculus language is a statically typed polymorphic non-dependent lambda calculus extension.

The language does not include a parser, as it is meant to be used structurally via its AST directly. It does include a pretty-printer, for ease of debugging.

Type inference for Lamdu Calculus is implemented in a separate package.

Values

The term language supports lambda abstractions and variables, literal values of custom types, holes, nominal type wrapping/unwrapping, extensible variant types and extensible records.

Types

The type language includes nominal types with polymoprhic components, extensible record types (row types) and extensible variant types (polymorphic variants / column types).

Composite types

In Lamdu Calculus, both records and variants are "composite" types and share an underlying AST, using a phantom type to distinguish them.

Records are sets of typed and named fields.

Variants are a value of a set of possible typed data constructors.

Both records and variants are order-agnostic, and disallow repeated appearance of the same names for fields or data constructors.

Both records and variants are extensible. This is usually called "row types" for records and "polymorphic variants" for variants. We use the terminology of "row" and "column" types instead.

Records

In the term language, records are constructed using these AST constructions:

BLeaf LRecEmpty denotes the empty record (denote in pseudo-syntax as ()). Its type is TRecord CEmpty.

BRecExtend (RowExtend tag (value : T) (rest : R))1 denotes a record extension of an existing record rest.

Its type is TRecord (CExtend tag T R).

A sequence of BRecExtend followed by a BLeaf LRecEmpty is denoted in pseudo-syntax as { x : T, y : U, ... }.

To deconstruct a record, we use the GetField operation (denoted in pseudo-syntax as .).

1. The (value : Type) notation is used to denote values' types

Row types

Row types are denoted using a record type variable.

For example, lets examine a lambda abstraction. Assume (+) : Int → Int → Int.

vector → vector.x + vector.y

We can infer the type:

{ x : Int, y : Int } → Int

But the most general type is:

forall r1. {x, y} ∉ r1 => { x : Int, y : Int | r1 } → Int

2

The lambda may be applied with a record that contains more fields besides x and y, and that is what the | r1 denotes. Note that to enforce the no-duplication requirement Lamdu Calculus uses a constraint on composite type variables, for each field that they may not duplicate.

2. The | symbol in the pseudo-syntax is used to indicate that some set of fields denoted by the variable that follows (in this case, r1) is concatenated to the set)

Variants

In the term language, variants are constructed using value injection.

For example, let's look at the type:

+{ Nothing : (), Just : Int }

This pseudo syntax means the structural variant type isomorphic to Haskell's Maybe Int.

The value 5 : Int is not of the type: +{ Nothing : (), Just : Int }.

In Haskell, we use Just to "lift" 5 from Int to Maybe Int. In Lamdu Calculus, we inject via BInject (Inject "Just" 5), which we denote in pseudo-syntax as Just: 5.

BInject (Inject tag (value : T)) injects a given value into a variant type. Its type is forall alts. tag ∉ alts => TVariant (CExtend tag T alts).

This type means that the injected value allows any set of typed alternatives to exist in the larger variant.

Case expressions

Deconstructing a variant requires a case statement. Lamdu Calculus case statements peel one alternate case at a time, so need to be composed to create an ordinary full case.

BCase (Case tag handler rest) denoted in pseudo syntax as

\case
  tag: handler
  rest

The above case expression creates a function with a variant parameter. It analyzes its argument, and if it is a tag, the given handler is invoked with the typed content of the tag.

If the argument is not a tag, then the rest handler is invoked. The rest handler is given a smaller variant type as an argument. A variant type which no longer has the tag case inside it, as that was ruled out.

\case
  Just: x → x + 1
  \case
    Nothing: () → 0
    ?

The above composed case expression will match against Just:

  • Match: it will evaluate to the content of the Just added to 1.

  • Mismatch: it will match against Nothing:

    • Match: it will ignore the empty record contained in the Nothing case, and evaluate to 0

    • Mismatch: Evaluate to a hole (denoted by ?) which is like Haskell's undefined, and takes on any type.

The type of the above case statement would be inferred to:

forall v1. (Just, Nothing) ∉ v1 =>
+{ Just : Int , Nothing : () | v1 } → Int

Note that the case statement allows any structural variant type that has the proper Nothing and Just cases (and it would reach a hole if the value happens to be neither Nothing nor Just). This is not typically what we want. We'd like to close the variant type so it is not extensible.

To do that, we must also support the empty case statement that matches no possible alternatives, allowing us to close the composition. The empty case statement AST construction is:

BLeaf LAbsurd, and is denoted as absurd (as it is the analogue of the absurd function in Haskell and Agda). The type of absurd is: forall r. +{} → r (+{} is the empty variant type, aka Void).

We can now close the above case expression:

\case
  Just: x → x + 1
  \case
    Nothing: () → 0
    absurd

And now our inferred type will be simpler:

+{ Just : Int, Nothing : () } → Int

As a short-hand for nested/composed cases, we'll use Haskell-like syntax for cases as well, meaning the expanded, composed case.

\case
Just: x → x + 1
Nothing: () → 0

Example use-case: Composing interpreters

We can define single-command interpreters:

interpretAdd default =
  \case
  Add: {x, y} → x + y
  default
interpretMul default =
  \case
  Mul {x, y} → x * y
  default

And then we can compose them:

interpreterArithmetic = interpretAdd . interpretMul

interpreter = (interpreterArithmetic . interpreterConditionals) absurd

Yielding a single interperter that can handle the full set of cases in its composition.

Example use-case: Exception Monads

See the Exception Monad use-case for a more compelling example use-case of extensible variants.

Nominal types

Purposes

Recursive Types

Infinite structural types are not allowed, so you can use a nominal type to “tie the knot”. For example:

Stream a = +{ Empty : (), NonEmpty : { head: a, tail: () -> Stream a } }

A Stream a may contain a Stream a within, therefore the type is recursive.

Safety

Often one wants to distinguish types whose representations/structure are the same, but their meanings are very different

Rank-N types

Nominal types may bind type variables (i.e: the structural type within a nominal type is wrapped in “forall”s)

This is being used by the Mut type (equivalent to Haskell's ST)

Future plans

  • Unifying with the underlying AST in AlgoWMutable which implements type inference more efficiently than Algorithm-W-Step-By-Step which is compatible with this package.

  • Support for type-classes