roc/roc-for-elm-programmers.md

Roc for Elm programmers

Roc is a direct descendant of the Elm programming language. The two languages are similar, but not the same!

This is a guide to help Elm programmers learn what's different between Elm and Roc.

NOTE: Almost all what's in this document has been implemented - but not quite all of it!

Comments

Like Python, Ruby, Perl, and Elixir, inline comments in Roc begin with # instead of --.

Doc comments begin with ## instead.

Like Python, Roc does not have multiline comment syntax.

String Interpolation

Roc strings work like Elm strings except that they support string interpolation. Here's a Roc string which uses interpolation:

"Hi, my name is \(name)!"

The Elm equivalent would be:

"Hi, my name is " ++ name ++ "!"

This interpolation syntax comes from Swift. Only a single identifier can go inside the parentheses (like (name) here), and the identifier needs to be a string already. Arbitrary expressions are not allowed, which means weird situations like string literals inside string literals don't come up.

Roc strings also have the type Str rather than Elm's String. This is to make common qualified operations like Str.join more concise; the idea is that you'll use the abbreviation often enough that you'll quickly get used to it. (I got used to str in Rust very quickly.)

Type Aliases

Rather than a type alias keyword, in Roc you define type aliases with : like so:

Username : Str

You can also define type aliases anywhere, not just at the top level. Their scoping rules work as normal.

Separately, Roc also allows standalone type annotations with no corresponding implementation. So I can write this as an annotation with no implementation:

getUsername : User -> Username

Roc will automatically fill in the implementation of this as the equivalent of a Debug.todo. If it ever gets run, it will crash, but for debugging purposes or sketching out APIs, you don't need to bother writing getUsername = Debug.todo "implement".

let syntax

Imagine if Elm's let...in worked exactly the same way, except you removed the let and in keywords. That's how it works in Roc.

For example, this Elm code computes someNumber to be 1234:

someNumber =
    let
        foo =
            1000

        blah =
            234
    in
    foo + blah

Here's the equivalent Roc code:

someNumber =
    foo =
        1000

    blah =
        234

    foo + blah

Like let...in in Elm, this is indentation-sensitive. Each of the definitions ("defs" for short) must have the same indentation as the ending expression.

Function definitions

Roc only has one syntax for defining functions, and it looks almost exactly like Elm's anonymous functions. The one difference is that multiple arguments are separated by commas.

So where in Elm you might write foo a b = in Roc you'd write foo = \a, b -> instead.

One minor benefit of the comma is that you don't need to use parentheses to destructure arguments inline. For example, in Elm, you always need to use parens to destructure variants inline in function declarations, like in these two examples:

\(UserId id1) (UserId id2) ->

\(UserId id) ->

Without the parentheses, it wouldn't be clear where one argument ended and the next one began.

In Roc, the commas make argument boundaries unambiguous, so no parens are needed. You can write the above like so in Roc:

\UserId id1, UserId id2 ->

\UserId id ->

Unbound type variables

In Elm, every type variable is named. For example:

List.reverse : List a -> List a

[] : List a

The a in List.reverse is a bound type variable, because it appears more than once in the type. Whatever the first list's a is, that's what the second list's a must be as well.

The a in [] : List a is an unbound type variable. It has no restrictions, which is why [] can be passed to any function that expects a List.

In Roc, this distinction between bound and unbound type variables is reflected at the syntax level. Here are those types in Roc:

List.reverse : List a -> List a

[] : List *

The * is the "wildcard" type variable. It is only for unbound type variables like this. Like the wildcard * in path globs like *.txt, it matches anything.

You can still choose names for unbound type variables if you like, but the compiler will infer them as * by default.

In Elm, the type of always is a -> (b -> a). The equivalent Roc type would be:

always : a -> (* -> a)

This makes unbound type variables easier to talk about out loud. Rather than saying (for example) "List a" or "Html msg with a lowercase m" you can say "List star" or "Html star".

Record Syntax

Roc uses Rust/JavaScript syntax for record literals, e.g. { x: 1, y: 2 }.

It also allows omitting the value; { x, y } is sugar for { x: x, y: y }.

You can pattern match on exact record values, e.g. { x: 5 } ->.

Roc does not have the "a type alias for a record creates a convenience constructor function" feature that Elm has. However, it does allow trailing commas, in both values and type annotations. Roc's formatter (which is built into the compiler, and which is zero-configuration like elm-format) formats multi-line record literals (and record types) with a comma at the end of each line, like so:

user =
    {
        firstName: "Sam",
        lastName: "Sample",
        email: "sam@example.com",
    }

This is easy to read and leads to tidy version control diffs; no matter how you reorder these fields, or add or remove fields, the diffs will only be of the relevant fields in question, not any adjacent fields or tokens.

Lists are formatted similarly to records, except of course they don't have labeled fields.

Closed record annotations look the same as they do in Elm, e.g. { name : Str, email : Str }. Open record annotations look a bit different.

In Elm:

{ a | name : Str, email : Str } -> Str

In Roc:

{ name : Str, email : Str }* -> Str

Here, the open record's type variable appears immediately after the }.

In the Elm example, the a is unbound, which in Roc means it appears as *.

This syntax makes it easier to write a function that accepts an open record with an unbound type variable (e.g. "this record, plus other fields if you like"). This is a totally reasonable thing to do - as often as you like! It has multiple upsides: it makes "named arguments" work with data model records more often, and makes it easier to change functions in backwards-compatible ways. It has no major downsides.

The syntax encourages doing this. "Just add a star" like so:

{ name : Str, email : Str }* -> Str

You can also use bound type variables too. In Elm:

{ a | name : Str, email : Str } -> { a | name : Str, email : Str }

In Roc:

{ name : Str, email : Str }a -> { name : Str, email : Str }a

Like in Elm, using records with bound variables should be extremely rare. They need to exist for the type system to work, and they aren't useless, but any time you find yourself reaching for them, there is a very high chance that there's a better way to write that code!

Record Update

Elm has "record update" syntax like this:

{ user | firstName = "Sam", lastName = "Sample" }

Roc has the same feature, but its syntax looks like this:

{ user & firstName: "Sam", lastName: "Sample" }

The record before the & can be qualified, like so:

{ Foo.defaultConfig & timeZone: utc }

However, it cannot involve record field access. So this would not compile:

{ Foo.defaults.config & timeZone: utc }

Optional Record Fields

There's a pattern in Elm where you pass a function a record of configuration values, some of which you don't really care about and want to leave as defaults. To incorporate the default config options, you call the function like so:

table { defaultConfig | height = 800, width = 600 }

This way, as the caller I'm specifying only the height and width fields, and leaving the others to whatever is inside defaultConfig. Perhaps it also has the fields x and y.

In Roc, you can do this like so:

table { height: 800, width: 600 }

...and the table function will fill in its default values for x and y. There is no need to use a defaultConfig record.

Here's how table would be defined in Roc:

table = \{ height, width, title ? "", description ? "" } ->

This is using optional field destructuring to destructure a record while also providing default values for any fields that might be missing. Here's the type of table:

table :
    {
        height : Pixels,
        width : Pixels,
        title ? Str,
        description ? Str,
    }
    -> Table
table = \{ height, width, title ? "", description ? "" } ->

This says that table takes a record with two required fields (height and width and two optional fields (title and description). It also says that the height and width fields have the type Pixels (a type alias for some numeric type), whereas the title and description fields have the type Str. This means you can choose to omit title, description, or both, when calling the function...but if you provide them, they must have the type Str.

This is also the type that would have been inferred for table if no annotation had been written. Roc's compiler can tell from the destructuring syntax title ? "" that title is an optional field, and that it has the type Str. These default values can reference other expressions in the record destructure; if you wanted, you could write { height, width, title ? "", description ? Str.concat "A table called " title }.

Destructuring is the only way to implement a record with optional fields. (For example, if you write the expression config.title and title is an optional field, you'll get a compile error.)

This means it's never possible to end up with an "optional value" that exists outside a record field. Optionality is a concept that exists only in record fields, and it's intended for the use case of config records like this. The ergonomics of destructuring mean this wouldn't be a good fit for data modeling.

Pattern matching

Roc's pattern matching conditionals work about the same as how they do in Elm. Here are two differences:

• Roc uses the syntax when...is instead of case...of
• In Roc, you can use | to handle multiple patterns in the same way

For example:

when color is
    Blue -> 1
    Green | Red | Yellow -> 2
    Purple -> 3

Like Rust, you can add an if guard after a pattern:

when color is
    Blue -> 1
    Green | Red | Yellow if totalColors >= 3 -> 2
    Green | Red | Yellow -> 4 # only gets run if totalColors < 3

This gives you a way to use constants in patterns:

pi = 3.14
e = 2.72

when number is
    0 -> "zero"
    1 -> "one"
    v if v == pi -> "pi"
    v if v == e -> "e"
    _ -> ""

Custom Types

This is the biggest semantic difference between Roc and Elm.

Let's start with the motivation. Suppose I'm using a platform for making a web server, and I want to:

• Read some data from a file
• Send a HTTP request containing some of the data from the file
• Write some data to a file containing some of the data from the HTTP response

Assuming I'm writing this on a Roc platform which has a Task-based API, and that Task.await is like Elm's Task.andThen but with the arguments flipped, here's one way I might write this:

doStuff = \filename ->
    Task.await (File.read filename) \fileData ->
        Task.await (Http.get (urlFromData fileData)) \response ->
            File.write filename (responseToData response)

Note that in Elm you'd need to add a <| before the anonymous functions (e.g. <| \response ->) but in Roc you don't.

What would the type of the above expression be? Let's say these function calls have the following types:

File.read : Filename -> Task File.Data File.ReadErr
File.write : Filename, File.Data -> Task File.Data File.WriteErr
Http.get : Url -> Task Http.Response Http.Err

Task.await : Task a err, (a -> Task b err) -> Task b err

If these are the types, the result would be a type mismatch. Those Task values have incompatible error types, so await won't be able to chain them together.

This situation is one of the motivations behind Roc's tags feature. Using tags, not only will this type-check, but at the end we get a combined error type which has the union of all the possible errors that could have occurred in this sequence. We can then handle those errors using a single when, like so:

when error is
    # Http.Err possibilities
    PageNotFound -> ...
    Timeout -> ...
    BadPayload -> ...

    # File.ReadErr possibilities
    FileNotFound -> ...
    ReadAccessDenied -> ...
    FileCorrupted -> ...

    # File.WriteErr possibilities
    DirectoryNotFound -> ...
    WriteAccessDenied -> ...
    DiskFull -> ...

Here is the set of slightly different types that will make the original chained expression compile. (await is unchanged.)

File.read : Filename -> Task File.Data (File.ReadErr *)
File.write : Filename, File.Data -> Task File.Data (File.WriteErr *)
Http.get : Url -> Task Http.Response (Http.Err *)

await : Task a err, (a -> Task b err) -> Task b err

The key is that each of the error types is a type alias for a Roc tag union. Here's how those look:

Http.Err a :
    [
        PageNotFound,
        Timeout,
        BadPayload Str,
    ]a

File.ReadErr a :
    [
        FileNotFound,
        Corrupted,
        BadFormat,
    ]a

File.WriteErr a :
    [
        FileNotFound,
        DiskFull,
    ]a

For a side-by-side comparison, here's how we would implement something similar in Elm:

type Http.Err
    = PageNotFound
    | Timeout
    | BadPayload String

type File.ReadErr
    = FileNotFound
    | Corrupted
    | BadFormat

type File.WriteErr
    = FileNotFound
    | DiskFull

There are a few differences between them, but the most significant one here is that the Roc version has a type variable.

That type variable has a similar purpose to the type variable in Elm's open records (e.g. the a in { a | name : String, email : String } which in Roc would be { name : Str, email : Str }a) - except applied to sum types (such as Elm's custom types) instead of product types (such as records).

If you were to remove the type variables from the Roc declaraionts for Http.Err, File.ReadErr, and File.WriteErr, they would work practically the same way as the Elm one. Roc tag unions can be used as traditional algebraic data types, and they have the usual support for pattern matching, exhaustiveness checking, and so on.

You don't need to declare tag unions before using them. Instead, you can just write a tag (essentially a variant) anywhere you like, and Roc will infer the type of the union it goes in.

Here are some examples of using tags in a REPL:

> True
True : [ True ]*

> False
False : [ False ]*

> Ok "hi"
Ok "hi" : [ Ok Str ]*

> SomethingIJustMadeUp "hi" "there"
SomethingIJustMadeUp "hi" "there" : [ SomethingIJustMadeUp Str Str ]*

> x = Foo
Foo : [ Foo ]*

> y = Foo "hi" Bar
Foo "hi" 5 : [ Foo Str [ Bar ]* ]*

> z = Foo [ "str1", "str2" ]
Foo [ "str1", "str2" ] : [ Foo (List Str) ]*

The [ ]s in the types are tag unions, and they list all the possible different tags that the value could be at runtime. In all of these tag unions, there is only one tag. Notice the * at the end; that's the type variable we saw earlier.

Similarly to how if you put { name = "" } into elm repl, it will infer a type of { a | name : String } - that is, an open record with an unbound type variable and name : Str field - if you put a tag Foo "" into roc repl, it will infer a type of [ Foo Str ]* - that is, an open tag union with one alternative: a Foo tag with a Str payload.

The same tag can be used with different arities and types. In the REPL above, x, y, and z, can all coexist even though they use Foo with different arities - and also with different types within the same arity.

Similarly, you can pattern match on tags without declaring any types, and Roc will infer the type of the tag union being matched.

For example, suppose we don't write any type annotations anywhere, and have this pattern match:

when blah is
    MyStr str -> Str.concat str "!"
    MyBool bool -> Bool.not bool

The inferred type of this expression would be [ MyStr Str, MyBool Bool ].

Exhaustiveness checking is still in full effect here. It's based on usage; if any code pathways led to blah being set to the tag Foo, I'd get an exhaustiveness error because this when does not have a Foo branch.

There's an important interaction here between the inferred type of a when-expression and the inferred type of a tag value. Note which types have a * and which do not.

x : [ Foo ]*
x = Foo

y : [ Bar Str ]*
y = Bar "stuff"

tagToStr : [ Foo, Bar Str ] -> Str
tagToStr = \tag ->
    when tag is
        Foo -> "hi"
        Bar str -> Str.concat str "!"

Each of these type annotations involves a tag union - a collection of tags bracketed by [ and ].

• The type [ Foo, Bar Str ] is a closed tag union.
• The type [ Foo ]* is an open tag union.

You can pass x to tagToStr because an open tag union is type-compatible with any closed tag union which contains its tags (in this case, the Foo tag). You can also pass y to tagToStr for the same reason.

In general, when you make a tag value, you'll get an open tag union (with a *). Using when can get you a closed union (a union without a *) but that's not always what happens. Here's a when in which the inferred type is an open tag union:

alwaysFoo : [ Foo Str ]* -> [ Foo Str ]*
alwaysFoo = \tag ->
    when tag is
        Foo str -> Foo (Str.concat str "!")
        _ -> Foo ""

The return value is an open tag union because all branches return something tagged with Foo.

The argument is also an open tag union, because this when-expression has a default branch; that argument is compatible with any tag union. This means you can pass the function some totally nonsensical tag, and it will still compile.

Note that the argument does not have the type *. That's because you cannot pass it values of any type; you can only pass it tags!

You could, if you wanted, change the argument's annotation to be []* and it would compile. After all, its default branch means it will accept any tag!

Still, the compiler will infer [ Foo Str ]* based on usage.

Just because [ Foo Str ]* is the inferred type of this argument, doesn't mean you have to accept that much flexibility. You can restrict it by removing the *. For example, if you changed the annotation to this...

alwaysFoo : [ Foo Str, Bar Bool ] -> [ Foo Str ]*

...then the function would only accept tags like Foo "hi" and Bar False. By writing out your own annotations, you can get the same level of restriction you get with traditional algebraic data types (which, after all, come with the requirement that you write out their annotations). Using annotations, you can restrict even when-expressions with default branches to accept only the values you define to be valid.

In fact, if you want a traditional algebraic data type in Roc, you can get about the same functionality by making (and then using) a type alias for a closed tag union. Here's exactly how Result is defined using tags in Roc's standard library:

Result ok err : [ Ok ok, Err err ]

You can also use tags to define recursive data structures, because recursive type aliases are allowed as long as the recursion happens within a tag. For example:

LinkedList a : [ Nil, Cons a (LinkedList a) ]

Inferred recursive tags use the as keyword. For example, the inferred version of the above type alias would be:

[ Nil, Cons a b ] as b

The * in open tag unions is actually an unbound ("wildcard") type variable. It can be bound too, with a lowercase letter like any other bound type variable. Here's an example:

exclaimFoo : [ Foo Str ]a -> [ Foo Str ]a
exclaimFoo = \tag ->
    when tag is
        Foo str -> Foo (Str.concat str "!")
        other -> other

The * says "this union can also include any other tags", and here the a says "the return value union includes Foo Str, plus whichever other tags the argument includes in its union."

The Roc type [] is equivalent to Elm's Never. You can never satisfy it!

Opaque Types

The tags discussed in the previous section are globally available, which means they cannot be used to create opaque types.

Private tags let you create opaque types. They work just like the global tags from the previous section, except:

• Private tags begin with an @ (e.g. @Foo instead of Foo)
• Private tags are scoped to the current module, rather than globally scoped
• Private tags can only be instantiated in the current module

For example, suppose I define these inside the Username module:

Username : [ @Username Str ]

fromStr : Str -> Username
fromStr = \str ->
    @Username str

toStr : Username -> Str
toStr = \@Username str ->
    str

I can now expose the Username type alias, which other modules can use as an opaque type.

It's not even syntactically possible for me to expose the @Username tag, because @ tags are not allowed in the exposing list. Only code written in this Username module can instantiate a @Username value.

If I were to write @Username inside another module (e.g. Main), it would compile, but that @Username would be type-incompatible with the one created inside the Username module. Even trying to use == on them would be a type mismatch, because I would be comparing a [ Username.@Username Str ]* with a [ Main.@Username Str ]*, which are incompatible.

Modules and Shadowing

In Elm, my main module (where main lives) might begin like this:

module MyApp exposing (main)

import Parser
import Http exposing (Request)
import Task exposing (Task, await)

Roc application modules (where the equivalent of main lives) begin with the app keyword rather than the module keyword, and the import syntax is a bit different. Here's how the above module header imports section would look in Roc:

app imports [ Parser, Http.{ Request }, Task.{ Task, await } ]

app modules are application entrypoints, and they don't formally expose anything. They also don't have names, so other modules can't even import them!

Modules that can be imported are interface modules. Their headers look like this:

interface Parser
    exposes [ Parser, map, oneOf, parse ]
    imports [ Utf8 ]

The name interface is intended to draw attention to the fact that the interface these expose is very important.

All imports and exports in Roc are enumerated explicitly; there is no .. syntax.

Since neither global tags nor private tags have a notion of "importing variants" (global tags are always available in all modules, and private tags are never available in other modules), there's also no exposing (Foo(..)) equivalent.

Like Elm, Roc does not allow shadowing.

Elm does permit overriding open imports - e.g. if you have import Foo exposing (bar), or import Foo exposing (..), you can still define bar = ... in the module. Roc treats this as shadowing and does not allow it.

Function equality

In Elm, if you write (\val -> val) == (\val -> val), you currently get a runtime exception which links to the == docs, which explain why this is the current behavior and what the better version will look like.

OCaml also has the "runtime exception if you compare functions for structural equality" behavior, but unlike Elm, in OCaml this appears to be the long-term design.

In Roc, function equality is a compile error, tracked explicitly in the type system. Here's the type of Roc's equality function:

'val, 'val -> Bool

Whenever a named type variable in Roc has a ' at the beginning, that means it is a functionless type - a type which cannot involve functions. If there are any functions in that type, you get a type mismatch. This is true whether val itself is a function, or if it's a type that wraps a function, like { predicate: (Str -> Bool) } or List (Bool -> Bool).

So if you write (\a -> a) == (\a -> a) in Roc, you'll get a type mismatch. If you wrap both sides of that == in a record or list, you'll still get a type mismatch.

If a named type variable has a ' anywhere in a given type, then it must have a ' everywhere in that type. So it would be an error to have a type like x, 'x -> Bool because x has a ' in one place but not everywhere.

Standard Data Structures

Elm has List, Array, Set, and Dict in the standard library.

Roc has all of these except Array, and there are some differences in how they work:

• List in Roc uses the term "list" the way Python does: to mean an ordered sequence of elements. Roc's List is more like an array, in that all the elements are sequential in memory and can be accessed in constant time. It still uses the [ ] syntax for list literals. Also there is no :: operator because "cons" is not an efficient operation on an array like it is in a linked list.
• Set in Roc is like Set in Elm: it's shorthand for a Dict with keys but no value, and it has a slightly different API.
• Dict in Roc is like Dict in Elm, except it's backed by hashing rather than ordering. Roc silently computes hash values for any value that can be used with ==, so instead of a comparable constraint on Set elements and Dict keys, in Roc they instead have the functionless constraint indicated with a '.

Roc also has a literal syntax for dictionaries and sets. Here's how to write a Dict literal:

{: "Sam" => True, "Ali" => False, firstName => False :}

This expression has the type Dict Str Bool, and the firstName variable would necessarily be a Str as well.

The Dict literal syntax is for two reasons. First, Roc doesn't have tuples; without tuples, initializing the above Dict would involve an API that looked something like one of these:

Dict.fromList [ { k: "Sam", v: True }, { k: "Ali", v: False }, { k: firstName, v: False } ]

Dict.fromList [ KV "Sam" True, KV "Ali" False KV firstName False

This works, but is not nearly as nice to read.

Additionally, Dict literals can compile directly to efficient initialization code without needing to (hopefully be able to) optimize away the intermediate List involved in fromList.

{::} is an empty Dict.

You can write a Set literal like this:

[: "Sam", "Ali", firstName :]

The Set literal syntax is partly for the initialization benefit, and also for symmetry with the Dict literal syntax.

[::] is an empty Set.

Roc does not have syntax for pattern matching on data structures - not even [ ] like Elm does.

Operators

In Elm, operators are functions. In Roc, all operators are syntax sugar.

This means, for example, that you cannot write (/) in Roc; that would be a syntax error. However, the / operator in Roc is infix syntax sugar for Num.div, which is a normal function you can pass to anything you like.

Elm has one unary operator, namely -. (In Elm, -x means "apply unary negate to x.") Roc has that one too, and also unary !. The expression !foo desugars to Bool.not foo, and !foo bar desugars to Bool.not (foo bar).

This was introduced because Roc does not expose any functions globally by default (the way Elm does with Basics functions like not, round, etc.). In Roc, only operators and standard types (like Str and Bool) are exposed globally. Having to fully qualify not was annoying, and making an exception just for not seemed less appealing than making an operator for it, especially when unary ! is so widely used in other languages.

Because Roc has unary !, its "not equal to" operator is != instead of Elm's /=, for symmetry with unary !.

There's an Operator Desugaring Table at the end of this guide, so you can see exactly what each Roc operator desugars to.

Currying and |>

Roc functions aren't curried. Calling (List.append foo) is a type mismatch because List.append takes 2 arguments, not 1.

For this reason, function type annotations separate arguments with , instead of ->. In Roc, the type of Set.add is:

Set.add : Set 'elem, 'elem -> Set 'elem

You might notice that Roc's Set.add takes its arguments in the reverse order from how they are in Elm; the Set is the first argument in Roc, whereas it would be the last argument in Elm. This is because Roc's |> operator works like Elixir's rather than like Elm's; here is an example of what it does in Roc:

a b c
    |> f x y

# f (a b c) x y

In Roc, the |> operator inserts the previous expression as the first argument to the subsequent expression, rather than as the last argument as it does in Elm.

This makes a number of operations more useful in pipelines. For example, in Roc, |> Num.div 2.0 divides by 2:

2000
  |> Num.div 2.0

# 1000.0

In Elm, where |> inserts 2 as the last argument, 2 ends up being the numerator rather than the denominator:

2000
  |> (/) 2.0

# 0.001

Another example is List.append, which is called List.concat in Roc:

[ 1, 2 ]
  |> List.concat [ 3, 4 ]

# [ 1, 2, 3, 4 ]

In Elm:

[ 1, 2 ]
  |> List.append [ 3, 4 ]

# [ 3, 4, 1, 2 ]

There are various trade-offs here, of course. Elm's |> has a very elegant implementation, and (|>) in Elm can be usefully passed to other functions (e.g. fold) whereas in Roc it's not even possible to express the type of |>.

As a consequence of |> working differently, "pipe-friendly" argument ordering is also different. That's why Set.add has a "flipped" signature in Roc; otherwise, |> Set.add 5 wouldn't work. Here's the type of Roc's Set.add again, and also a pipeline using it:

Set.add : Set 'elem, 'elem -> Set 'elem

[: "a", "b", "c" :]
    |> Set.add "d"

Roc has no << or >> operators, and there are no functions in the standard library for general-purpose pointfree function composition.

The <| operator

Roc has no <| operator. (It does have |> though.)

In Elm, <| is used as a minor convenience for when you want to avoid some parens in a single-line expression (e.g. foo <| bar baz over foo (bar baz)) and as a major convenience when you want to pass an anonymous function, if, or case as an argument.

For example, elm-test relies on it:

test "it works" <|
    \_ -> verify stuff

In Roc, this does not require a <|. This Roc code does the same thing as the preceding Elm code:

test "it works"
    \_ -> verify stuff

This is convenient with higher-order functions which take a function as their final argument. Since many Roc functions have the same type as Elm functions except with their arguments flipped, this means it's possible to end a lot of expressions with anonymous functions - e.g.

modifiedNums =
    List.map nums \num ->
        doubled = num * 2

        modified = modify doubled

        modified / 2

Separately, you don't need parens or an operator to pass when or if as arguments. Here's another example:

foo 1 2 if something then 3 else 4

# Same as `foo 1 2 (if something then 3 else 4)`

CoffeeScript also does this the way Roc does.

Backpassing

Suppose I'm using a platform for making a CLI, and I want to run several Tasks in a row which read some files from the disk. Here's one way I could do that, assuming Task.await is like Elm's Task.andThen with arguments flipped:

readLicense : Filename -> Task License File.ReadErr
readLicense = \filename ->
    Task.await (File.readUtf8 settingsFilename) \settingsYaml ->
        when Yaml.decode settingsDecoder settingsYaml is
            Ok settings ->
                Task.await (File.readUtf8 settings.projectFilename) \csv ->
                    when Csv.decode projectDecoder csv is
                        Ok project ->
                            Task.await (File.readUf8 project.licenseFilename) \licenseStr ->
                                when License.fromStr licenseStr is
                                    Ok license -> Task.succeed license
                                    Err err -> Task.fail (InvalidFormat err)

                        Err err -> Task.fail (InvalidFormat err)

            Err err ->
                Task.fail (InvalidFormat err)

We can write this with |> instead of when like so:

readLicense : Filename -> Task License File.ReadErr
readLicense = \filename ->
    Task.await (File.readUtf8 settingsFilename) \settingsYaml ->
        settingsYaml
            |> Yaml.decode settingsDecoder
            |> Task.fromResult
            |> Task.mapFail InvalidFormat
            |> Task.await \settings ->
                Task.await (File.readUtf8 settings.projectFilename) \projectCsv ->
                    projectCsv
                        |> Csv.decode projectDecoder
                        |> Task.fromResult
                        |> Task.mapFail InvalidFormat
                        |> Task.await \project ->
                            Task.await (File.readUf8 project.licenseFilename) \licenseStr ->
                                License.fromStr licenseStr
                                    |> Task.fromResult
                                    |> Task.mapFail InvalidFormat

We can also write it this way, which is equivalent to the previous two ways:

readLicense : Filename -> Task License File.ReadErr
readLicense = \filename ->
    settingsYaml <- Task.await (File.readUtf8 settingsFilename)

    settings <-
        settingsYaml
            |> Yaml.decode settingsDecoder
            |> Task.fromResult
            |> Task.mapFail InvalidFormat

    projectCsv <- Task.await (File.readUtf8 settings.projectFilename)

    project <-
        projectCsv
            |> Csv.decode projectDecoder
            |> Task.fromResult
            |> Task.mapFail InvalidFormat

    licenseStr <-
        Task.await (File.readUf8 project.licenseFilename)

    License.fromStr licenseStr
        |> Task.fromResult
        |> Task.mapFail InvalidFormat

This uses backpassing syntax to nest anonymous functions without indenting them. Here's a smaller demonstration of backpassing; the second snippet is sugar for the first.

list =
    List.map numbers \num -> num + 1

list =
    num <- List.map numbers

    num + 1

Both snippets are calling List.map passing numbers as the first argument, and a \num -> num + 1 function for the other argument.

The difference is that in the second snippet, the \num -> num + 1 function is written backwards, like this:

    num <-

    num + 1

This is called backpassing because you write the function backwards and then immediately pass it as an argument to another function.

The other function - the one you're passing this one to - goes right after the <- symbol. That function should be called with one argument missing at the end, such as with List.map numbers (which is missing its final argument).

Here's another pair of snippets, this time using two backpassing calls:

incrementedNumbers =
    List.map lists \numbers ->
        List.map numbers \num -> num + 1

incrementedNumbers =
    numbers <- List.map lists
    num <- List.map numbers

    num + 1

In the second snippet, we have two functions defined in the backpassing style. The first function is:

numbers <-
    num <- List.map numbers

    num + 1

This function desugars to \numbers -> … and is being passed as the final argument to List.map lists.

The second function defined in backpassing style is:

    num <-

    num + 1

This function desugars to \numbers -> … and is being passed as the final argument to List.map numbers. That List.map numbers call is the body of the numbers <- function we defined in backpassing style a moment ago - so in a normal function definition, it would be \numbers -> List.map numbers …

Note that backpassing can be combined with the |> operator, which lets you call a function with two arguments missing from the end - one provided by the |> and the other provided by the <-, like so:

incrementedNumbers =
    num <-
        [ 1, 2, 3 ]
            |> List.reverse
            |> List.map

    num + 1

Here, the first argument to List.map is provided by the |> (namely the reversed [ 1, 2, 3 ] list), and the second argument is provided by +the <- (namely the \num -> … function).

Backpassing can also be used with functions that take multiple arguments; for example, you could write key, value <- Dict.map dictionary similarly to how we used List.map here. That would desugar into a Dict.map dictionary \key, value -> … function.

To be clear, backpassing is designed to be used with chaining functions like Task.await which are prone to lots of nesting. It isn't designed to be used with functions like List.map; this is just a simplified example to show that <- can be used with any function...even those where it doesn't improve code clarity!

Finally, here's an example combining backpassing with ordinary = definitions:

task =
    user <- Task.await fetchUser

    url = user.baseUrl

    settings, bio, posts <- Task.map3 (getSettings url) (getBio url) (getPosts url)

    profile = makeProfile settings bio

    Task.succeed { profile, posts }

Here, every new name that's introduced to scope is aligned on the left-hand edge of the expression - regardless of whether it's coming from = or from <-.

Numbers

Like Elm, Roc organizes numbers into integers and floating-point numbers. However, Roc breaks them down even further. For example, Roc has two different sizes of float types to choose from:

• F64 - a 64-bit IEEE 754 binary floating point number
• F32 - a 32-bit IEEE 754 binary floating point number

Both types are desirable in different situations. For example, when doing simulations, the precision of the F64 type is desirable. On the other hand, GPUs tend to heavily prefer 32-bit floats because a serious bottleneck is how long it takes data to transfer from CPU to GPU, so having to send half as many bytes per render (compared to 64-bit floats) can be huge for performance.

Roc also supports D64 and D32, which are IEEE 754 decimal floating point numbers. The upside of these is that they are decimal-based, so 0.1 + 0.2 == 0.3 (whereas in binary floats this is not true), which makes them much better for calculations involving currency, among others. The downside of decimal floats is that they do not have hardware support (except on certain highly uncommon processors), so calculations involving them take longer.

Roc does not let floating point calculations result in Infinity, -Infinity, or NaN. Any operation which would result in one of these (such as sqrt or /) will return a Result.

Similarly to how there are different sizes of floating point numbers, there are also different sizes of integer to choose from:

• I8
• I16
• I32
• I64
• I128

Roc also has unsigned integers which are never negative. They are U8, U16, U32, U64, U128, and Nat.

The size of Nat depends on what target you're building for; on a 64-bit target (the most common), at runtime Nat will be the same as U64, whereas on a 32-bit target (for example, WebAssembly) at runtime it will be the same as U32 instead. Nat comes up most often with collection lengths and indexing into collections. For example:

• List.len : List * -> Nat
• List.get : List elem, Nat -> Result elem [ OutOfBounds ]*
• List.set : List elem, Nat, elem -> List elem

As with floats, which integer type to use depends on the values you want to support as well as your performance needs. For example, raw sequences of bytes are typically represented in Roc as List U8. You could also represent them as List U128, but it's much more efficient to use List U8, since each byte will be at most 255 anyway.

Like Elm, it's possible in Roc to have functions that work on either integers or floating-point numbers. However, the types are different. For example, the type of Num.add (which the + operator desugars to) is:

Num.add : Num a, Num a -> Num a

This accepts any of the numeric types discussed above, from I128 to F32 to D64 and everything in between. This is because those are all type aliases for Num types. For example:

• I64 is a type alias for Num (Integer Signed64)
• U8 is a type alias for Num (Integer Unsigned8)
• F32 is a type alias for Num (FloatingPoint Binary32)
• D64 is a type alias for Num (FloatingPoint Decimal64)

(Those types like Integer, FloatingPoint, and Signed64 are all defined like Never; you can never instantiate one. They are used only as phantom types.)

So Roc does not use number, but rather uses Num - which works more like List. Either way, you get + being able to work on both integers and floats!

Separately, there's also Int a, which is a type alias for Num (Integer a), and Float a, which is a type alias for Num (Float a). These allow functions that can work on any integer or any float. For example, Num.bitwiseAnd : Int a, Int a -> Int a.

In Roc, number literals with decimal points are Float * values. Number literals without a decimal point are Num * values. Almost always these will end up becoming something more specific, but in the unlikely event (most often in a REPL) that you actually do end up with an operation that runs on either an Int * or a Num * value, it will default to being treated as an I64. Similarly, a Float * value will default to being treated as a D64, which means if someone is learning Roc as their first programming language and they type 0.1 + 0.2 into a REPL, they won't be confused by the answer.

If you encounter overflow with either integers or floats in Roc, you get a runtime exception rather than wrapping overflow behavior (or a float becoming Infinity or -Infinity). You can opt into wrapping overflow instead with functions like Num.addWrap : Int a, Int a -> Int a, or use a function that gives Err if it overflows, like Num.addChecked : Num a, Num a -> Result (Num a) [ Overflow ]*.

comparable, appendable, and number

These don't exist in Roc.

• appendable is only used in Elm for the (++) operator, and Roc doesn't have that operator.
• comparable is used for comparison operators (like < and such), plus List.sort, Dict, and Set. Roc's List.sort accepts a "sorting function" argument which specifies how to sort the elements. Roc's comparison operators (like <) only accept numbers; "foo" < "bar" is valid Elm, but will not compile in Roc. Roc's dictionaries and sets are hashmaps behind the scenes (rather than ordered trees), and their keys only need the functionless restriction.
• number is replaced by Num, as described previously.

Also like Python Roc permits underscores in number literals for readability purposes. Roc also supports hexadecimal (0x01), octal (0o01), and binary (0b01) integer literals; these literals all have type Int * instead of Num *.

If you put these into a hypothetical Roc REPL, here's what you'd see:

> 1_024 + 1_024
2048 : Num *

> 1 + 2.14
3.14 : Float *

> 1.0 + 1
2.0 : Float *

> 1.1 + 0x11
<type mismatch between `1.1 : Float *` and `0x11 : Int *`>

> 11 + 0x11
28 : Int *

Phantom Types

Phantom types exist in Elm but not in Roc. This is because phantom types can't be defined using type aliases (in fact, there is a custom error message in Elm if you try to do this), and Roc only has type aliases. However, in Roc, you can achieve the same API and runtime performance characteristics as if you had phantom types, by using phantom values instead.

A phantom value is one which affects types, but which holds no information at runtime. As an example, let's say I wanted to define a units library - a classic example of phantom types. I could do that in Roc like this:

Quantity units data : [ Quantity units data ]

km : Num a -> Quantity [ Km ] (Num a)
km = \num ->
    Quantity Km num

cm : Num a -> Quantity [ Cm ] (Num a)
cm = \num ->
    Quantity Cm num

mm : Num a -> Quantity [ Mm ] (Num a)
mm = \num ->
    Quantity Mm num

add : Quantity u (Num a), Quantity u (Num a) -> Quantity u (Num a)
add = \Quantity units a, Quantity _ b ->
    Quantity units (a + b)

From a performance perspective, it's relevant here that [ Km ], [ Cm ], and [ Mm ] are all unions containing a single tag. That means they hold no information at runtime (they would always destructure to the same tag), which means they can be "unboxed" away - that is, discarded prior to code generation.

During code generation, Roc treats Quantity [ Km ] Int as equivalent to Quantity Int. Then, because Quantity Int is an alias for [ Quantity Int ], it will unbox again and reduce that all the way down to to Int.

This means that, just like phantom types, phantom values affect type checking only, and have no runtime overhead. Rust has a related concept called phantom data.

Standard library

elm/core has these modules:

• Array
• Basics
• Bitwise
• Char
• Debug
• Dict
• List
• Maybe
• Platform
• Platform.Cmd
• Platform.Sub
• Process
• Result
• Set
• String
• Task
• Tuple

In Roc, the standard library is not a standalone package. It is baked into the compiler, and you can't upgrade it independently of a compiler release; whatever version of Roc you're using, that's the version of the standard library you're using too. (This is because Roc doesn't have a concept like Elm's Kernel; it would not be possible to ship Roc's standard library as a separate package!)

Roc's standard library has these modules:

• Str
• Bool
• Num
• List
• Dict
• Set
• Result

Some differences to note:

• All these standard modules are imported by default into every module. They also expose all their types (e.g. Bool, List, Result) but they do not expose any values - not even negate or not. (True, False, Ok, and Err are all global tags, so they do not need to be exposed; they are globally available regardless!)
• In Roc it's called Str instead of String.
• List refers to something more like Elm's Array, as noted earlier.
• No Char. This is by design. What most people think of as a "character" is a rendered glyph. However, rendered glyphs are comprised of grapheme clusters, which are a variable number of Unicode code points - and there's no upper bound on how many code points there can be in a single cluster. In a world of emoji, I think this makes Char error-prone and it's better to have Str be the only first-class unit. For convenience when working with unicode code points (e.g. for performance-critical tasks like parsing), the single-quote syntax is sugar for the corresponding U32 code point - for example, writing '鹏' is exactly the same as writing 40527. Like Rust, you get a compiler error if you put something in single quotes that's not a valid Unicode scalar value.
• No Basics. You use everything from the standard library fully-qualified; e.g. Bool.not or Num.negate or Num.ceiling. There is no Never because [] already serves that purpose. (Roc's standard library doesn't include an equivalent of Basics.never, but it's one line of code and anyone can implmement it: never = \a -> never a.)
• No Tuple. Roc doesn't have tuple syntax. As a convention, Pair can be used to represent tuples (e.g. List.zip : List a, List b -> List [ Pair a b ]*), but this comes up infrequently compared to languages that have dedicated syntax for it.
• No Task. By design, platform authors implement Task (or don't; it's up to them) - it's not something that really could be usefully present in Roc's standard library.
• No Process, Platform, Cmd, or Sub - similarly to Task, these are things platform authors would include, or not.
• No Maybe. This is by design. If a function returns a potential error, use Result with an error type that uses a zero-arg tag to describe what went wrong. (For example, List.first : List a -> Result a [ ListWasEmpty ]* instead of List.first : List a -> Maybe a.) If you want to have a record field be optional, use an Optional Record Field directly (see earlier). If you want to describe something that's neither an operation that can fail nor an optional field, use a more descriptive tag - e.g. for a nullable JSON decoder, instead of nullable : Decoder a -> Decoder (Maybe a), make a self-documenting API like nullable : Decoder a -> Decoder [ Null, NonNull a ]*.

Operator Desugaring Table

Here are various Roc expressions involving operators, and what they desugar to.

Expression	Desugars to
a + b	Num.add a b
a - b	Num.sub a b
a * b	Num.mul a b
a / b	Num.div a b
a // b	Num.divFloor a b
a ^ b	Num.pow a b
a % b	Num.rem a b
a %% b	Num.mod a b
a >> b	Num.shr a b
a << b	Num.shl a b
-a	Num.neg a
-f x y	Num.neg (f x y)
a == b	Bool.isEq a b
a != b	Bool.isNotEq a b
a && b	Bool.and a b
a || b	Bool.or a b
!a	Bool.not a
!f x y	Bool.not (f x y)
a |> b	b a
a b c |> f x y	f (a b c) x y