F# Partially-Applied Unions

The other day I was trying to think how I could group together discriminated union values in a way to let me later determine if values were mutually-exclusive. For a simple discriminated union, I came up with a pretty simple solution: store a Map<SomeUnion, SomeUnion list>. Each key in the map would refer to a SomeUnion value, and the corresponding value would be a list of all the other SomeUnion values that it was mutually-exclusive with.

As I began to work up the data model for this, I quickly realized that the union cases in the SomeUnion type would need to contain additional metadata that was not relevant to whether or not two values were mutually-exclusive. Certain cases of the SomeUnion type might be sub-categorized in some way that was relevant, but they might also contain file paths, or ID numbers corresponding to database records. So to really solve my problem using unions, I needed to find a way to leave out some of the values when constructing a SomeUnion value, and treat the omitted values as a sort of wildcard.

For those who do not work with F#, or have not paid close enough attention, union cases in F# can only contain a single value. To include multiple values, you do so by providing a tuple. The compiler takes the time to provide these out as named members which are useful when working with a union from C# or VB, but regardless, there is no way for one to "omit" values when creating a union case.

On one hand, if I were working with a dynamically-typed language like Python, I would probably have just passed None for all of the values I wanted to treat like wildcards, and made sure that the types (which I control anyway) do not take the time to be picky about the values they are given. On the other hand, I am not working with a dynamically-typed language, and while I do not have a firm opinion as to whether dynamic or static typing is better, I do like working with static types. So this is a brief explanation of what I came up.

Quick Summary

In my finished code, the basic concept is that for each union case which contains two or more "fields" (i.e. its constructor requires passing a tuple), the union case gets a wrapper function inside of a quotation. This quotation is then used within another quotation later, where it can be treated like a normal F# function, allowing parameters to be omitted. Static functions on a helper type, UnionMetadata<'T>, take the finished quotation, and return a UnionMetadata<'T> value, where 'T is the union type being worked with. An example:

[<StructuralComparison; StructuralEquality>]
type SomeUnion =
    | CaseA of bool * int * string
    | CaseB of int
    | CaseC
    ...

let CaseAWrapper = <@ (fun x y z -> CaseA(x, y, z)) @>

With the above code in place, the CaseAWrapper value can be used in another quotation, which can then be passed to the FromQuotation function on the UnionMetadata type. In F# Interactive, it would look something like this:

let meta = UnionMetadata<SomeUnion>.FromQuotation <@ (%CaseAWrapper) true @>;;

val meta : UnionMetadata<SomeUnion> =
    {
        UnionCase = "CaseA";
        SuppliedArguments =
            [
                (Some true, "Item1");
                (null, "Item2");
                (null, "Item3");
            ]
    }

From there, the UnionMetadata<'T> type provides a few helper functions that can be used to determine whether or not a given value matches the metadata. In this case, one could call meta.ValueMatches (CaseA (true, 5, "test")) which would return true; passing CaseA (false, 5, "test") would result in false instead, matching the given contents of the meta value.

To further test this, the CaseAWrapper value could be rewritten, and the meta value updated:

let CaseAWrapper = <@ (fun y x z -> CaseA(x, y, z)) @>
let meta = UnionMetadata<SomeUnion>.FromQuotation <@ (%CaseAWrapper) 10 @>;;

val CaseAWrapper : FSharp.Quotations.Expr<(bool -> string -> SomeUnion)>
...
val meta : UnionMetadata<SomeUnion> =
    {
        UnionCase = "CaseA";
        SuppliedArguments =
            [
                (null, "Item1");
                (Some 10, "Item2");
                (null, "Item3");
            ]
    }

Using this updated definition for meta, any CaseA value where the int field was set to 10 would cause meta.ValueMatches to return true, while any other int value cause meta.ValueMatches to return false. And that's about all you need to know to use the UnionMetadata<'T> type.

Quotations?

Okay, so if you want more information, lets continue using the SomeUnion type as an example to work with. The basic concept that the UnionMetadata<'T> type uses to make all of this work is F# quotations. Quotations are a sort of "meta-programming" capability built into F#, which allows you to embed F# code as a value, in a sort of custom AST (abstract syntax tree) that they created just for this language feature. The full output from F# Interactive for CaseAWrapper looks something like this:

Lambda (y, Lambda (x, Lambda (z, NewUnionCase (CaseA, x, y, z)))

At some point during its work, a compiler might generate something like the above for code like (fun y x z -> CaseA(x, y, z)). Each Lambda value corresponds to one of the function parameters because of how F# handles currying (a function that takes two arguments is actually a function that takes one argument and returns a new function that takes the other argument), and the NewUnionCase value obviously corresponds to the creation of the CaseA value. When stuffed into <@ (%CaseAWrapper) 10 @>, the value looks more like this:

Application (Lambda (y, Lambda (x, Lambda (z, NewUnionCase (CaseA, x, y, z))), Value (10))

The added Application on the left side, and the Value (10) on the right side, correspond to the fact that you have performed "function application" by supplying an argument for the function's first parameter. The UnionMetadata<'T> type has a bunch of overloaded functions all named FromQuotation which each take a slightly different "shape" of quotation. they take quotation shapes like Expr<'T> and Expr<_ -> 'T>, which would be used to process <@ CaseC @> and <@ CaseB @> respectively; Expr<'T> is a quotation representing a value of type 'T, while Expr<_ -> 'T> is a quotation representing a function that takes one argument and returns a 'T value. The current code for UnionMetadata<'T> extends this up to functions that take ten arguments.

It takes a little bit of juggling and rearranging to do, but the private functions on the UnionMetadata<'T> type takes these three types of quotations, Application, Lambda, and NewUnionCase, and processes them into the SuppliedArguments collection. The NewUnionCase quotation also includes a FSharp.Reflection.UnionCaseInfo value, which includes the union case name; this is where the UnionCase value is populated from.

So what can I do with it?

Well, I mean, whatever you want to? Here's an example of how I'm using it:

type Microsoft.FSharp.Collections.Map<'Key, 'TValue when 'Key : comparison> with
    static member addOrUpdate (key : 'Key) (value : 'TValue) (update : 'TValue -> 'TValue) (oldMap : Map<'Key, 'TValue>) : Map<'Key, 'TValue> =
        if Map.containsKey key oldMap then
            let oldVal = oldMap.[key]
            oldMap
            |> Map.remove key
            |> Map.add key (update oldVal)
        else
            Map.add key value oldMap


type ExclusivityRules =
    {
        MutualExclusions : Map<UnionMetadata<SomeUnion>, UnionMetadata<SomeUnion> list>;
    }
    static member Empty =
        {
            MutualExclusions = Map.empty;
        }

    static member private AddMutualExclusion (x1, x2) (rules : ExclusivityRules) =
        let exclusions =
            rules.MutualExclusions
            |> Map.addOrUpdate x1 [x2] (fun xs -> x2::xs)
            |> Map.addOrUpdate x2 [x1] (fun xs -> x1::xs)
        { rules with MutualExclusions = exclusions }

    static member inline MutuallyExclusive ((e1 : Expr<_>), (e2 : Expr<_>)) : ExclusivityRules -> ExclusivityRules =
        (UnionMetadata<SomeUnion>.FromQuotation<_> e1, UnionMetadata<SomeUnion>.FromQuotation<_> e2)
        |> ExclusivityRules.AddMutualExclusion

    static member inline MutuallyExclusive ((e1 : Expr<SomeUnion>), (e2 : Expr<_>)) : ExclusivityRules -> ExclusivityRules =
        (UnionMetadata<SomeUnion>.FromQuotation e1, UnionMetadata<SomeUnion>.FromQuotation<_> e2)
        |> ExclusivityRules.AddMutualExclusion

    static member inline MutuallyExclusive ((e1 : Expr<_>), (e2 : Expr<SomeUnion>)) : ExclusivityRules -> ExclusivityRules =
        (UnionMetadata<SomeUnion>.FromQuotation<_> e1, UnionMetadata<SomeUnion>.FromQuotation e2)
        |> ExclusivityRules.AddMutualExclusion

    static member inline MutuallyExclusive ((e1 : Expr<SomeUnion>), (e2 : Expr<SomeUnion>)) : ExclusivityRules -> ExclusivityRules =
        (UnionMetadata<SomeUnion>.FromQuotation e1, UnionMetadata<SomeUnion>.FromQuotation e2)
        |> ExclusivityRules.AddMutualExclusion


    member this.GetMutualExclusions (newValue : SomeUnion) (currentValues : SomeUnion list) : SomeUnion list =
        let unionInfo = FSharp.Reflection.FSharpValue.GetUnionFields(newValue, typeof<SomeUnion>)
        let ourExclusionList =
            this.MutualExclusions
            |> Map.toSeq
            |> Seq.filter (fst >> (UnionMetadata<_>.UnionInfoMatchesSpec unionInfo))
            |> Seq.collect snd
            |> Seq.groupBy (fun x -> x.UnionCase)
            |> Seq.map (fun (x, ys) -> x, (ys |> Seq.toList))
            |> Map.ofSeq

        if ourExclusionList.IsEmpty then [] else

        currentValues
        |> Seq.filter
            (fun thisValue ->
                let unionInfo = FSharp.Reflection.FSharpValue.GetUnionFields(thisValue, typeof<SomeUnion>)
                let unionCase = fst unionInfo
                if ourExclusionList.ContainsKey unionCase.Name |> not then false else
                ourExclusionList.[unionCase.Name]
                |> Seq.exists (UnionMetadata<_>.UnionInfoMatchesSpec unionInfo)
            )
        |> Seq.toList

The first bit of code adds an extension function addOrUpdate to the type Map<_,_>, (roughly matching the method of the same name on the ConcurrentDictionary type) followed by a new type to hold my collection of rules about what values are mutually-exclusive. The ExclusivityRules type includes an Empty member to help me start, because I have other thoughts in the back of my mind that would require adding additional fields. The GetMutualExclusions function takes a new value and a "current" list of values, and returns the current values which are considered mutually-exclusive with the new value. Finally, there are four versions of a MutuallyExclusive function for defining new mutual exclusions.

The four MutuallyExclusive functions are almost identical, and are enough to make the F# compiler happy regardless of what values I use. Each of the Expr<_> parameters is passed to UnionMetadata<SomeUnion>.FromQuotation<_>, and each of the Expr<SomeUnion> values is passed to UnionMetadata<SomeUnion>.FromQuotation. This minor difference is enough to convince the compiler to infer whether it needs to use the FromQuotation overload that takes an Expr<'T>, or one of the many FromQuotation overloads that deal with quotations for functions.

The private function AddMutualExclusion takes two UnionMetadata<SomeUnion> values, and does most of the legwork, so the MutuallyExclusive overloads just need to call the right FromQuotation overload and pass the results off to the AddMutualExclusion function.

From there, I can do this:

let rules =
    ExclusivityRules.Empty
    |> ExclusivityRules.MutuallyExclusive (<@ CaseB @>, <@ (%CaseAWrapper) 10 @>)
    |> ...


let current = [ CaseB 5 ]

let test = rules.GetMutualExclusions (CaseA true 10 "Test") current

...

val test : SomeUnion list = [CaseB 5]

In my actual code I'm not using the SomeUnion type, but this is the same basic concept. Depending on what else you want to do, you could extend this to track certain types of behavior that vary slightly between cases of a union. Or perhaps certain SomeUnion values can be combined together, and UnionMetadata<'T> values could be mapped to functions that handle merging two values together.

Where do I get this amazing code?

Source code is available on GitHub. As of this writing, the project is a bit incomplete, but the code needed to start working with the samples from this post is there, should work, and has a fair bit of XML documentation written.