darcs-2.11.0: a distributed, interactive, smart revision control system

Safe HaskellSafe-Inferred




Option specifications using continuations with a changing answer type.

Based on


with additional inspiration provided by


which shows how the same format specifiers can be used for both sprintf and sscanf.

The OptSpec type corresponds to the format specifiers for the sprintf and sscanf functions, which I called ounparse and oparse here; they no longer work on Strings but instead on any list (the intention is, of course, that this is a list of flags).

As explained in the original paper by Kenichi Asai, we cannot use Cont, even with the recent additions of the shift and reset combinators, since Cont requires that the answer type remains the same over the whole computation, while the trick used here requires that the answer type can change.

Besides parsing and unparsing, the OptSpec type contains two more members: odesc is the list of OptDescr that getOpt needs as input for parsing the command line and for generating the usage help, while ocheck takes a list of flags and returns a list of error messages, which can be used to check for conflicting options.


Option specifications

data OptSpec d f a b

A type for option specifications.

It consists of four components: a parser, an unparser, a checker, and a list of descriptions.

The parser converts a flag list to some result value. This can never fail: we demand that primitive parsers are written so that there is always a default value (use Maybe with default Nothing as a last resort).

The unparser does the opposite of the parser: a value is converted back to a flag list.

The checker returns a list of error messages (which should be empty if there are no problems found). This can be used to e.g. check whether there are conflicting flags in the list.

Separating the checker and parser is unusual. The reason for this is that we want to support flags coming from multiple sources, such as the command line or a defaults file. Prioritising these sources is done by concatenating the flag lists in the order of precedence, so that earlier flags win over later ones. That means that when parsing the (final) flag list, conflicting flags are resolved by picking the first flag that matches an option. The checker, on the other hand, can be called for each source separately.

The last component is a list of descriptors for each single switch/flag that the option is made of.

The OptSpec type is heavily parameterized. The type arguments are:

The flag type, such as DarcsFlag.
A type that describes an single flag, such as OptDescr or DarcsOptDescr. It should be a Functor.

Abstracting over these types is not technically necessary: for the intended application in Darcs, we could as well fix them as d=DarcsOptDescr, and f=DarcsFlag, saving two type parameters. However, doing that here would only obscure what's going on, making the code harder to understand, not easier. Besides, the resulting more general type signatures give us additional guarantees, known as "free theorems" (free as in beer, not in speak).

In contrast, the type parameters

a, b
are necessary to make chaining of options a la typed printf/scanf possible. In a nutshell, a is the result type of a function that consumes the result of parsing or unparsing an option, while b is the complete type of such a function.

The ounparse and oparse members use continuation passing style, which is the reason for their apparently "inverted" type signature. To understand them, it helps to look at the type of "primitive" (not yet combined) options (see PrimOptSpec below). For a primitive option, b gets instantiated to v -> a, where v is the type of values associated with the option. The whole option spec then has type

  o :: 'OptSpec' d f a (v -> a)

so that the oparse and ounparse members are instantiated to

  ounparse :: forall a. ([f] -> a) -> (x -> a)
  oparse   :: forall a. (x -> a) -> ([f] -> a)

which can be easily seen to be equivalent to

  ounparse :: x -> [f]
  oparse   :: [f] -> x

Chaining such options results in a combined option of type

  o1 ^ o2 ^ ... :: OptSpec d f a (v1 -> v2 -> ... -> a)

that is, b gets instantiated to

  v1 -> v2 -> ... -> a

To use such an option (primitive or combined), you pass in the consumer. A typical consumer of option values is a command implementation. Given

  cmd :: v1 -> v2 -> ... -> [String] -> IO ()

we can parse the flags and pass the results to cmd:

  oparse (o1 ^ o2 ^ ...) cmd flags




ounparse :: ([f] -> a) -> b

Convert option value (back) to flag list, in CPS.

oparse :: b -> [f] -> a

Convert flag list to option value, in CPS. Note: as a pure function, it is not supposed to fail.

ocheck :: [f] -> [String]

Check for erros in a flag list, returns error messages.

odesc :: [d f]

Descriptions, one for each flag that makes up the option.


IsoFunctor (OptSpec d f a) 
Monoid (PrimOptSpec d f a [v])

See oappend and oempty.

Primitive combinators

The type OptSpec d f, together with the operation ^ and the unit oid forms a category. We could express this with an

  instance Category (OptSpec d f) where
    id = oid
    (.) = (^)

I decided against doing that because I like the id and . from the Prelude.

Proving the category laws is easy because the operation and unit are implemented independently for each component. This means OptSpec d f is simply the product of four categories, reducing the problem to proving the laws for each component separately.

This is just list concatenation, which is a monoid, and every monoid is a category (by adding two phantom type arguments).
Same here, noting that ([f] ->) is a monoid homomorphism (as expressed by the instance Monoid b => Monoid (a -> b) in Data.Monoid).
This can be seen by flipping the arguments (which is a functor i.e. preserves category laws), so the type becomes [f] -> b -> a, and noting as before that ([f] ->) is a monoid homomorphism and thus a functor (by adding two phantom type arguments), reducing the operation to simple function composition. If this rather abstract argument doesn't convince you, do the calculations as an exercise.
for this I don't have an easy abstract argument at hand, so I'll do the calculation:
    o1 ^ (o2 ^ o3)
  =   definition outer (^)
    k -> o1 (f1 -> (o2 ^ o3) (f23 -> k (f1 ++ f23)))
  =   definition inner (^)
    k -> o1 (f1 -> (k' -> o2 (f2 -> o3 (f3 -> k' (f2 ++ f3)))) (f23 -> k (f1 ++ f23)))
  =   beta reduce: f1 --> f23 -> k (f1 ++ f23)
    k -> o1 (f1 -> (o2 (f2 -> o3 (f3 -> (f23 -> k (f1 ++ f23)) (f2 ++ f3)))))
  =   beta reduce: f23 --> f2 ++ f3
    k -> o1 (f1 -> (o2 (f2 -> o3 (f3 -> (k (f1 ++ (f2 ++ f3)))))))

and from the other side:

    (o1 ^ o2) ^ o3
  =   definition outer (^)
    k -> (o1 ^ o2) (f12 -> o3 (f3 -> k (f12 ++ f3)))
  =   definition inner (^)
    k -> (k' -> o1 (f1 -> o2 (f2 -> k' (f1 ++ f2)))) (f12 -> o3 (f3 -> k (f12 ++ f3)))
  =   beta reduce: k' --> f12 -> o3 (f3 -> k (f12 ++ f3))
    k -> (o1 (f1 -> o2 (f2 -> (f12 -> o3 (f3 -> k (f12 ++ f3))) (f1 ++ f2))))
  =   beta reduce: f12 --> f1 ++ f2
    k -> (o1 (f1 -> o2 (f2 -> (o3 (f3 -> k ((f1 ++ f2) ++ f3))))))

so again we have reduced the problem to the associativity of (++). Left and right unit laws are left to the reader...

oid :: OptSpec d f a a

Identity OptSpec, unit for ^

(^) :: OptSpec d f b c -> OptSpec d f a b -> OptSpec d f a c

OptSpec composition, associative

Derived combinators

onormalise :: OptSpec d f [f] b -> [f] -> [f]

Normalise a flag list by parsing and then unparsing it. This adds all implicit (default) flags to the list, which is useful as long as there is legacy code that circumvents the OptSpec abstraction and directly tests for flag membership.

onormalise opts = (oparse opts . ounparse opts) id

defaultFlags :: OptSpec d f [f] b -> [f]

The list of default flags for an OptSpec.

defaultFlags opts = onormalise opts []

Lifting isomorphisms

oimap :: Iso b c -> OptSpec d f a b -> OptSpec d f a c

Lift an isomorphism between b and c to one between OptSpec d f a b and OptSpec d f a c.

The forward component of the Iso is needed for ounparse, the backward component for oparse. For the other two components this is the identity.

Primitive options

type PrimOptSpec d f a v = OptSpec d f a (v -> a)

Type of primitive (not yet combined) options. The type parameter b gets instantiated to (v -> a), adding one argument of type v to the answer type of the continuation.

oappend :: PrimOptSpec d f a [v] -> PrimOptSpec d f a [v] -> PrimOptSpec d f a [v]

Combine two list valued options of the same type "in parellel". This is done by concatenating the resulting option values (oparse), flags (ounparse), errors (ocheck), and descriptors (odesc), respectively, of the input options.

oempty :: PrimOptSpec d f a [v]

Unit for oappend.

parseFlags :: (forall a. PrimOptSpec d f a v) -> [f] -> v

Parse a list of flags against a primitive option spec, returning the value associated with the option. As noted above, this cannot fail because options always have a default value.

parseFlags o fs = oparse o id fs