I am trying to learn more about inline and SRTP. Unfortunately I don't have the understanding to make this sample more minimal, but it's not too big:
type ComponentCollection<'props, 'comp when 'props : comparison> =
{
Components : Map<'props, 'comp>
}
with
static member AddComponent (collection, props, comp) =
{
Components =
collection.Components
|> Map.add props comp
}
module ComponentCollection =
let inline add< ^t, ^props, ^comp
when ^t : (static member AddComponent : ((^t * ^props * ^comp) -> ^t)) >
(props : ^props) (comp : ^comp) (xs : ^t) =
(^t : (static member AddComponent : (^t * ^props * ^comp -> ^t ) ) () ) (xs, props, comp)
type Component<'t> =
{
Props : 't
}
type Foo =
{
Foo : int
}
[<EntryPoint>]
let main argv =
let foo = { Foo = 1 }
let a : ComponentCollection<Foo, Component<Foo>> =
{ Components = Map.empty }
|> ComponentCollection.add foo { Props = foo }
0
The gives a compilation error:
error FS0001: The type 'ComponentCollection<'a,'b>' does not support the operator 'get_AddComponent'
This is very misleading to me, since I am using static member AddComponent, not get_AddComponent. I am not sure where the compiler gets get_AddComponent from.
What am I doing wrong here?
Additionally, it would be very helpful if someone could explain this bit of the code:
(^t : (static member AddComponent : (^t * ^props * ^comp -> ^t ) ) () )
I understand that it somehow provides the static member AddComponent of ^t as a free-function, but I cobbled it together from other examples, and the documentation for this syntax is lacking. For example, why is a () required?
As far as I can see, the only thing that is wrong with your snippet is that you have some excessive parentheses which, unfortunately in this case, actually have semantic meaning. Changing the code as follows resolves the problem:
module ComponentCollection =
let inline add< ^t, ^props, ^comp
when ^t : (static member AddComponent : ^t * ^props * ^comp -> ^t) >
(props : ^props) (comp : ^comp) (xs : ^t) =
(^t : (static member AddComponent : ^t * ^props * ^comp -> ^t) (xs, props, comp))
The issue is that F# actually makes a difference between:
static member AddComponent : ^t * ^props * ^comp -> ^t
static member AddComponent : (^t * ^props * ^comp) -> ^t
static member AddComponent : ((^t * ^props * ^comp) -> ^t)
The first is a regular static method with three arguments
The second is a static method taking three-element tuple as a single argument
The last is a static property returning a function value taking a tuple
Regarding your second question - in the working version, the call looks like this:
(^t : (static member AddComponent : ^t * ^props * ^comp -> ^t) (xs, props, comp))
This tells the compiler to access the static member (which is required by the constraint) and invoke it with arguments xs, props and comp. In your original version, this was invoking a getter method of the property (confusingly though, it somehow allowed you to do this without arguments - I suspect this would lead to another type error later).
From experience, SRTP's are complex and hard to use. They require a lot of inlines and Error messages tend to be very obscure.
I would recommend limiting their use.
Also, they would not be C# compatible if that is something you are interested in.
I also changed your code to fix the issues:
type ComponentCollection<'props, 'comp when 'props : comparison> =
{
Components : Map<'props, 'comp>
}
with
member collection.AddComponent(props, comp) =
{
Components =
collection.Components
|> Map.add props comp
}
module ComponentCollection =
let inline add< ^t, ^props, ^comp
when 'props : comparison and ^t : (member AddComponent : ^props * ^comp -> ^t) >
(props : ^props) (comp : ^comp) (xs : ^t) =
(^t : (member AddComponent : ^props * ^comp -> ^t ) (xs, props, comp ))
type Component<'t> =
{
Props : 't
}
type Foo =
{
Foo : int
}
[<EntryPoint>]
let main argv =
let foo = { Foo = 1 }
let a =
{ Components = Map.empty }
|> ComponentCollection.add foo { Props = foo }
0
given this code
type Baz = Baz of int with
static member bar f (Baz(b)) = f b
let inline foo< ^T, ^U when ^T : (static member bar : (^U -> ^T) -> ^T -> ^T)>
(f:(^U -> ^T)) (t:^T) : ^T =
(^T : (static member bar : (^U -> ^T) -> ^T -> ^T) f, t )
let x = foo (fun x -> (Baz 0)) (Baz 1)
I get this error
error FS0043: Method or object constructor 'bar' not found
I assume that signature of my static member can not really be unified to (^U -> ^T) -> ^T -> ^T
How can I solve this?
Looking at the previous question (i.e. switching back to the member function) and your comments, perhaps this would work:
type Baz = Baz of int with
member this.bar (f: 'a -> 'b): 'b = match this with
| Baz i -> f i
let inline foo (f: ^U -> ^T) (t:^T) =
let foo' = (^T : (member bar : (^U -> ^T) -> ^T) (t, f))
foo'
let x = foo (fun x -> (Baz 0)) (Baz 1)
// This returns Baz 0
printfn "%A" x
Is it possible to do an F# type test pattern with a member constraint?
Such as:
let f x =
match x with
| :? (^T when ^T : (static member IsInfinity : ^T -> bool)) as z -> Some z
| _ -> None
or
let g x =
match x with
| (z : ^T when ^T : (static member IsInfinity : ^T -> bool)) -> Some z
| _ -> None
Neither which work.
You cannot do this, as Petr said, statically resolved type parameters are resolved at compile time. They're actually a feature of the F# compiler rather than being a .NET feature, hence why this kind of information isn't available at runtime.
If you wish to check this at runtime, you could use reflection.
let hasIsInfinity (x : 'a) =
typeof<'a>.GetMethod("IsInfinity", [|typeof<'a>|])
|> Option.ofObj
|> Option.exists (fun mi -> mi.ReturnType = typeof<bool> && mi.IsStatic)
This will check for a static method called IsInfinity with type sig: 'a -> bool
Compiles:
let inline f< ^T when ^T : (static member (<<<) : ^T * int -> ^T) > (x : ^T) = x <<< 1
Does not compile:
let inline f< ^T when ^T : (static member (>>>) : ^T * int -> ^T) > (x : ^T) = x >>> 1
Errors:
Attempted to parse this as an operator name, but failed
Unexpected symbol '>' in member signature. Expected ')' or other token.
A type parameter is missing a constraint 'when ^T : (static member ( >>> ) : ^T * int32 -> ^T)'
Adding spaces doesn't help; this line yields the same compiler errors:
let inline f< ^T when ^T : (static member ( >>> ) : ^T * int -> ^T) > (x : ^T) = x >>> 1
I've searched both the documentation and the specification, to no avail. Is this a bug? Is there some way to include the > characters in the member signature?
Sure looks like a bug. It's ugly, but one workaround is to use the long form of the operator name:
let inline f< ^T when ^T : (static member op_RightShift : ^T * int -> ^T)> (x : ^T) =
x >>> 1
Do you even need an explicit constraint? This works just as well:
let inline f (x: ^T) : ^T = x >>> 1
I'm quite new to F# and find type inference really is a cool thing. But currently it seems that it also may lead to code duplication, which is not a cool thing. I want to sum the digits of a number like this:
let rec crossfoot n =
if n = 0 then 0
else n % 10 + crossfoot (n / 10)
crossfoot 123
This correctly prints 6. But now my input number does not fit int 32 bits, so I have to transform it to.
let rec crossfoot n =
if n = 0L then 0L
else n % 10L + crossfoot (n / 10L)
crossfoot 123L
Then, a BigInteger comes my way and guess what…
Of course, I could only have the bigint version and cast input parameters up and output parameters down as needed. But first I assume using BigInteger over int has some performance penalities. Second let cf = int (crossfoot (bigint 123)) does just not read nice.
Isn't there a generic way to write this?
Building on Brian's and Stephen's answers, here's some complete code:
module NumericLiteralG =
let inline FromZero() = LanguagePrimitives.GenericZero
let inline FromOne() = LanguagePrimitives.GenericOne
let inline FromInt32 (n:int) =
let one : ^a = FromOne()
let zero : ^a = FromZero()
let n_incr = if n > 0 then 1 else -1
let g_incr = if n > 0 then one else (zero - one)
let rec loop i g =
if i = n then g
else loop (i + n_incr) (g + g_incr)
loop 0 zero
let inline crossfoot (n:^a) : ^a =
let (zero:^a) = 0G
let (ten:^a) = 10G
let rec compute (n:^a) =
if n = zero then zero
else ((n % ten):^a) + compute (n / ten)
compute n
crossfoot 123
crossfoot 123I
crossfoot 123L
UPDATE: Simple Answer
Here's a standalone implementation, without the NumericLiteralG module, and a slightly less restrictive inferred type:
let inline crossfoot (n:^a) : ^a =
let zero:^a = LanguagePrimitives.GenericZero
let ten:^a = (Seq.init 10 (fun _ -> LanguagePrimitives.GenericOne)) |> Seq.sum
let rec compute (n:^a) =
if n = zero then zero
else ((n % ten):^a) + compute (n / ten)
compute n
Explanation
There are effectively two types of generics in F#: 1) run-type polymorphism, via .NET interfaces/inheritance, and 2) compile time generics. Compile-time generics are needed to accommodate things like generic numerical operations and something like duck-typing (explicit member constraints). These features are integral to F# but unsupported in .NET, so therefore have to be handled by F# at compile time.
The caret (^) is used to differentiate statically resolved (compile-time) type parameters from ordinary ones (which use an apostrophe). In short, 'a is handled at run-time, ^a at compile-time–which is why the function must be marked inline.
I had never tried to write something like this before. It turned out clumsier than I expected. The biggest hurdle I see to writing generic numeric code in F# is: creating an instance of a generic number other than zero or one. See the implementation of FromInt32 in this answer to see what I mean. GenericZero and GenericOne are built-in, and they're implemented using techniques that aren't available in user code. In this function, since we only needed a small number (10), I created a sequence of 10 GenericOnes and summed them.
I can't explain as well why all the type annotations are needed, except to say that it appears each time the compiler encounters an operation on a generic type it seems to think it's dealing with a new type. So it ends up inferring some bizarre type with duplicated resitrictions (e.g. it may require (+) multiple times). Adding the type annotations lets it know we're dealing with the same type throughout. The code works fine without them, but adding them simplifies the inferred signature.
In addition to kvb's technique using Numeric Literals (Brian's link), I've had a lot of success using a different technique which can yield better inferred structural type signatures and may also be used to create precomputed type-specific functions for better performance as well as control over supported numeric types (since you will often want to support all integral types, but not rational types, for example): F# Static Member Type Constraints.
Following up on the discussion Daniel and I have been having about the inferred type signatures yielded by the different techniques, here is an overview:
NumericLiteralG Technique
module NumericLiteralG =
let inline FromZero() = LanguagePrimitives.GenericZero
let inline FromOne() = LanguagePrimitives.GenericOne
let inline FromInt32 (n:int) =
let one = FromOne()
let zero = FromZero()
let n_incr = if n > 0 then 1 else -1
let g_incr = if n > 0 then one else (zero - one)
let rec loop i g =
if i = n then g
else loop (i + n_incr) (g + g_incr)
loop 0 zero
Crossfoot without adding any type annotations:
let inline crossfoot1 n =
let rec compute n =
if n = 0G then 0G
else n % 10G + compute (n / 10G)
compute n
val inline crossfoot1 :
^a -> ^e
when ( ^a or ^b) : (static member ( % ) : ^a * ^b -> ^d) and
^a : (static member get_Zero : -> ^a) and
( ^a or ^f) : (static member ( / ) : ^a * ^f -> ^a) and
^a : equality and ^b : (static member get_Zero : -> ^b) and
( ^b or ^c) : (static member ( - ) : ^b * ^c -> ^c) and
( ^b or ^c) : (static member ( + ) : ^b * ^c -> ^b) and
^c : (static member get_One : -> ^c) and
( ^d or ^e) : (static member ( + ) : ^d * ^e -> ^e) and
^e : (static member get_Zero : -> ^e) and
^f : (static member get_Zero : -> ^f) and
( ^f or ^g) : (static member ( - ) : ^f * ^g -> ^g) and
( ^f or ^g) : (static member ( + ) : ^f * ^g -> ^f) and
^g : (static member get_One : -> ^g)
Crossfoot adding some type annotations:
let inline crossfoot2 (n:^a) : ^a =
let (zero:^a) = 0G
let (ten:^a) = 10G
let rec compute (n:^a) =
if n = zero then zero
else ((n % ten):^a) + compute (n / ten)
compute n
val inline crossfoot2 :
^a -> ^a
when ^a : (static member get_Zero : -> ^a) and
( ^a or ^a0) : (static member ( - ) : ^a * ^a0 -> ^a0) and
( ^a or ^a0) : (static member ( + ) : ^a * ^a0 -> ^a) and
^a : equality and ^a : (static member ( + ) : ^a * ^a -> ^a) and
^a : (static member ( % ) : ^a * ^a -> ^a) and
^a : (static member ( / ) : ^a * ^a -> ^a) and
^a0 : (static member get_One : -> ^a0)
Record Type Technique
module LP =
let inline zero_of (target:'a) : 'a = LanguagePrimitives.GenericZero<'a>
let inline one_of (target:'a) : 'a = LanguagePrimitives.GenericOne<'a>
let inline two_of (target:'a) : 'a = one_of(target) + one_of(target)
let inline three_of (target:'a) : 'a = two_of(target) + one_of(target)
let inline negone_of (target:'a) : 'a = zero_of(target) - one_of(target)
let inline any_of (target:'a) (x:int) : 'a =
let one:'a = one_of target
let zero:'a = zero_of target
let xu = if x > 0 then 1 else -1
let gu:'a = if x > 0 then one else zero-one
let rec get i g =
if i = x then g
else get (i+xu) (g+gu)
get 0 zero
type G<'a> = {
negone:'a
zero:'a
one:'a
two:'a
three:'a
any: int -> 'a
}
let inline G_of (target:'a) : (G<'a>) = {
zero = zero_of target
one = one_of target
two = two_of target
three = three_of target
negone = negone_of target
any = any_of target
}
open LP
Crossfoot, no annotations required for nice inferred signature:
let inline crossfoot3 n =
let g = G_of n
let ten = g.any 10
let rec compute n =
if n = g.zero then g.zero
else n % ten + compute (n / ten)
compute n
val inline crossfoot3 :
^a -> ^a
when ^a : (static member ( % ) : ^a * ^a -> ^b) and
( ^b or ^a) : (static member ( + ) : ^b * ^a -> ^a) and
^a : (static member get_Zero : -> ^a) and
^a : (static member get_One : -> ^a) and
^a : (static member ( + ) : ^a * ^a -> ^a) and
^a : (static member ( - ) : ^a * ^a -> ^a) and ^a : equality and
^a : (static member ( / ) : ^a * ^a -> ^a)
Crossfoot, no annotations, accepts precomputed instances of G:
let inline crossfootG g ten n =
let rec compute n =
if n = g.zero then g.zero
else n % ten + compute (n / ten)
compute n
val inline crossfootG :
G< ^a> -> ^b -> ^a -> ^a
when ( ^a or ^b) : (static member ( % ) : ^a * ^b -> ^c) and
( ^c or ^a) : (static member ( + ) : ^c * ^a -> ^a) and
( ^a or ^b) : (static member ( / ) : ^a * ^b -> ^a) and
^a : equality
I use the above in practice since then I can make precomputed type specific versions which don't suffer from the performance cost of Generic LanguagePrimitives:
let gn = G_of 1 //int32
let gL = G_of 1L //int64
let gI = G_of 1I //bigint
let gD = G_of 1.0 //double
let gS = G_of 1.0f //single
let gM = G_of 1.0m //decimal
let crossfootn = crossfootG gn (gn.any 10)
let crossfootL = crossfootG gL (gL.any 10)
let crossfootI = crossfootG gI (gI.any 10)
let crossfootD = crossfootG gD (gD.any 10)
let crossfootS = crossfootG gS (gS.any 10)
let crossfootM = crossfootG gM (gM.any 10)
Since the question of how to make the type signatures less hairy when using the generalized numeric literals has come up, I thought I'd put in my two cents. The main issue is that F#'s operators can be asymmetric so that you can do stuff like System.DateTime.Now + System.TimeSpan.FromHours(1.0), which means that F#'s type inference adds intermediary type variables whenever arithmetic operations are being performed.
In the case of numerical algorithms, this potential asymmetry isn't typically useful and the resulting explosion in the type signatures is quite ugly (although it generally doesn't affect F#'s ability to apply the functions correctly when given concrete arguments). One potential solution to this problem is to restrict the types of the arithmetic operators within the scope that you care about. For instance, if you define this module:
module SymmetricOps =
let inline (+) (x:'a) (y:'a) : 'a = x + y
let inline (-) (x:'a) (y:'a) : 'a = x - y
let inline (*) (x:'a) (y:'a) : 'a = x * y
let inline (/) (x:'a) (y:'a) : 'a = x / y
let inline (%) (x:'a) (y:'a) : 'a = x % y
...
then you can just open the SymmetricOps module whenever you want have the operators apply only to two arguments of the same type. So now we can define:
module NumericLiteralG =
open SymmetricOps
let inline FromZero() = LanguagePrimitives.GenericZero
let inline FromOne() = LanguagePrimitives.GenericOne
let inline FromInt32 (n:int) =
let one = FromOne()
let zero = FromZero()
let n_incr = if n > 0 then 1 else -1
let g_incr = if n > 0 then one else (zero - one)
let rec loop i g =
if i = n then g
else loop (i + n_incr) (g + g_incr)
loop 0 zero
and
open SymmetricOps
let inline crossfoot x =
let rec compute n =
if n = 0G then 0G
else n % 10G + compute (n / 10G)
compute x
and the inferred type is the relatively clean
val inline crossfoot :
^a -> ^a
when ^a : (static member ( - ) : ^a * ^a -> ^a) and
^a : (static member get_One : -> ^a) and
^a : (static member ( % ) : ^a * ^a -> ^a) and
^a : (static member get_Zero : -> ^a) and
^a : (static member ( + ) : ^a * ^a -> ^a) and
^a : (static member ( / ) : ^a * ^a -> ^a) and ^a : equality
while we still get the benefit of a nice, readable definition for crossfoot.
I stumbled upon this topic when I was looking for a solution and I am posting my answer, because I found a way to express generic numeral without the less than optimal implementation of building up the number by hand.
open System.Numerics
// optional
open MathNet.Numerics
module NumericLiteralG =
type GenericNumber = GenericNumber with
static member instance (GenericNumber, x:int32, _:int8) = fun () -> int8 x
static member instance (GenericNumber, x:int32, _:uint8) = fun () -> uint8 x
static member instance (GenericNumber, x:int32, _:int16) = fun () -> int16 x
static member instance (GenericNumber, x:int32, _:uint16) = fun () -> uint16 x
static member instance (GenericNumber, x:int32, _:int32) = fun () -> x
static member instance (GenericNumber, x:int32, _:uint32) = fun () -> uint32 x
static member instance (GenericNumber, x:int32, _:int64) = fun () -> int64 x
static member instance (GenericNumber, x:int32, _:uint64) = fun () -> uint64 x
static member instance (GenericNumber, x:int32, _:float32) = fun () -> float32 x
static member instance (GenericNumber, x:int32, _:float) = fun () -> float x
static member instance (GenericNumber, x:int32, _:bigint) = fun () -> bigint x
static member instance (GenericNumber, x:int32, _:decimal) = fun () -> decimal x
static member instance (GenericNumber, x:int32, _:Complex) = fun () -> Complex.op_Implicit x
static member instance (GenericNumber, x:int64, _:int64) = fun () -> int64 x
static member instance (GenericNumber, x:int64, _:uint64) = fun () -> uint64 x
static member instance (GenericNumber, x:int64, _:float32) = fun () -> float32 x
static member instance (GenericNumber, x:int64, _:float) = fun () -> float x
static member instance (GenericNumber, x:int64, _:bigint) = fun () -> bigint x
static member instance (GenericNumber, x:int64, _:decimal) = fun () -> decimal x
static member instance (GenericNumber, x:int64, _:Complex) = fun () -> Complex.op_Implicit x
static member instance (GenericNumber, x:string, _:float32) = fun () -> float32 x
static member instance (GenericNumber, x:string, _:float) = fun () -> float x
static member instance (GenericNumber, x:string, _:bigint) = fun () -> bigint.Parse x
static member instance (GenericNumber, x:string, _:decimal) = fun () -> decimal x
static member instance (GenericNumber, x:string, _:Complex) = fun () -> Complex(float x, 0.0)
// MathNet.Numerics
static member instance (GenericNumber, x:int32, _:Complex32) = fun () -> Complex32.op_Implicit x
static member instance (GenericNumber, x:int32, _:bignum) = fun () -> bignum.FromInt x
static member instance (GenericNumber, x:int64, _:Complex32) = fun () -> Complex32.op_Implicit x
static member instance (GenericNumber, x:int64, _:bignum) = fun () -> bignum.FromBigInt (bigint x)
static member instance (GenericNumber, x:string, _:Complex32) = fun () -> Complex32(float32 x, 0.0f)
static member instance (GenericNumber, x:string, _:bignum) = fun () -> bignum.FromBigInt (bigint.Parse x)
let inline genericNumber num = Inline.instance (GenericNumber, num) ()
let inline FromZero () = LanguagePrimitives.GenericZero
let inline FromOne () = LanguagePrimitives.GenericOne
let inline FromInt32 n = genericNumber n
let inline FromInt64 n = genericNumber n
let inline FromString n = genericNumber n
this implementation comes by without complicated iteration during the cast. It uses FsControl for the Instance module.
http://www.fssnip.net/mv
Is crossfoot exactly what you want to do, or is it just summing the digits of a long number?
because if you just want to sum the digits, then:
let crossfoot (x:'a) = x.ToString().ToCharArray()
|> (Array.fold(fun acc x' -> if x' <> '.'
then acc + (int x')
else acc) 0)
... And you are done.
Anyways,
Can you convert stuff to a string, drop the decimal point, remember where the decimal point is, interpret it as an int, run crossfoot?
Here is my solution. I am not sure exactly how you want "crossfoot" to work when you have a decimal point added.
For instance, do you want: crossfoot(123.1) = 7 or crossfoot(123.1) = 6.1? (I'm assuming you want the latter)
Anyways, the code does allow you to work with numbers as generics.
let crossfoot (n:'a) = // Completely generic input
let rec crossfoot' (a:int) = // Standard integer crossfoot
if a = 0 then 0
else a%10 + crossfoot' (a / 10)
let nstr = n.ToString()
let nn = nstr.Split([|'.'|]) // Assuming your main constraint is float/int
let n',n_ = if nn.Length > 1 then nn.[0],nn.[1]
else nn.[0],"0"
let n'',n_' = crossfoot'(int n'),crossfoot'(int n_)
match n_' with
| 0 -> string n''
| _ -> (string n'')+"."+(string n_')
If you need to input big integers or int64 stuff, the way crossfoot works, you can just split the big number into bitesize chunks (strings) and feed them into this function, and add them together.