In Carbon, a sum type is a type whose values are grouped into several distinct named cases, called alternatives. A value of a sum type notionally consists of a discriminator tag that identifies which alternative is present, together with that alternative's value if it has one. Sum types are typically handled with pattern matching.
The choice
keyword is used to declare a sum type by specifying its interface,
leaving the implementation to the compiler. A choice
declaration consists of
the choice
keyword followed by the name of the type, and then a list of
alternatives inside curly braces. An alternative declaration consists of the
alternative name, optionally followed by a parameter list in parentheses. If
present, the parameter list has the same syntax as in a
function declaration. For example:
choice OptionalI32 {
Some(value: i32),
None
}
This declares a sum type named OptionalI32
with two alternatives: Some
,
which holds a single i32
value (the parameter name value
has no effect other
than documentation), and None
, which is empty. Choice types can also be
parameterized, like class types:
choice Optional(T:! type) {
Some(value: T),
None
}
A value of a function-like alternative is specified by "calling" it like a
function, and a value of an empty alternative like None
is specified by naming
it:
var my_opt: Optional(i32) = Optional(i32).None;
my_opt = Optional(i32).Some(42);
The value of a choice type can be inspected using a match
statement:
match (my_opt) {
case .Some(the_value: i32) => {
Print(the_value);
}
case .None => {
Print("None");
}
}
choice
declarations are a convenience shorthand for common use cases, but they
have limited flexibility. There is no way to control the representation of a
choice
type, or define methods or other members for it (although you can
extend it to implement interfaces, using an
out-of-line impl
or
adapter). However, a class
type can be
extended to behave like a sum type. This is much more verbose than a choice
declaration, but gives the author full control over the representation and class
members.
The ability to create instances of the sum type can be straightforwardly emulated with factory functions and static constants, and the internal storage layout will presumably involve untagged unions or some other low-level storage primitive which hasn't been designed yet, but the key to defining a sum type's interface is enabling it to support pattern matching. To do that, the sum type has to specify two things:
- The set of all possible alternatives, including their names and parameter
types, so that the compiler can typecheck the
match
body, identify any unreachablecase
s, and determine whether anycase
s are missing. - The algorithm that, given a value of the sum type, determines which alternative is present, and specifies the values of its parameters.
It does so by implementing the Match
interface, which is defined as follows:
interface Match {
interface BaseContinuation {
let ReturnType:! type;
}
let template Continuation:! type;
fn Op[self: Self, C:! Continuation](continuation: C*)
-> C.(MatchContinuation.ReturnType);
}
Continuation
must itself be an interface that extends
Match.BaseContinuation
, and its definition specifies the set of possible
alternatives: each alternative is represented as a method of that interface.
When compiling a proper pattern (or set of patterns that includes a proper
pattern, as with the cases of a match
) whose type is a sum type, the compiler
generates an implementation of Continuation
and passes it to Match.Op
. The
sum type's implementation of Match.Op
is responsible for determining which
alternative is present and what its parameters are, and calling the
corresponding method of continuation
with those parameters. The Match.Op
implementation is required to call exactly one such method exactly once before
returning. The compiler populates the Continuation
method bodies with whatever
code should be executed when the corresponding alternatives match.
TODO: if Carbon has explicit support for tail calls, we should probably
require that Match.Op
invoke the continuation as a tail call.
For example, here's how Optional
can be defined as a class:
class Optional(T:! type) {
// Factory functions
fn Some(value: T) -> Self;
let None:! Self;
private var has_value: bool;
private var value: T;
impl as Match {
interface Continuation {
extend Match.BaseContinuation;
fn Some[addr self: Self*](value: T) -> ReturnType;
fn None[addr self: Self*]() -> ReturnType;
}
fn Op[self: Self, C:! Continuation](continuation: C*) -> C.ReturnType {
if (self.has_value) {
return continuation->Some(self.value);
} else {
return continuation->None();
}
}
}
// Operations like destruction, copying, assignment, and comparison are
// omitted for brevity.
}
And here's how the compiler-generated implementation of
Optional.(Match.Continuation)
for the match
statement shown earlier might
look, if it were written in Carbon:
class __MatchStatementImpl {
extend impl as Match(Optional.MatchContinuation) where .ReturnType = () {
fn Some(the_value: i32) {
Print(the_value);
}
fn None() {
Print("None");
}
}
}
my_opt.(Match.Op)({} as __MatchStatementImpl);
(The name __MatchStatementImpl
is a placeholder for illustration purposes; the
actual generated class will be anonymous.)
The mechanism described above for proper patterns may also be used for
expression patterns if they have the form of an alternative pattern. An
expression pattern of type T
has the form of an alternative pattern if T
implements Match
, and the expression consists of an optional expression that
names T
, followed by a designator that names a method of
T.(Match.Continuation)
, optionally followed by a tuple expression. If an
expression pattern has that form, it may be matched using the mechanism above,
as if it were a proper pattern, rather than by evaluating the expression and
comparing it to the scrutinee using ==
. Both possible implementations must be
well-formed (and this is enforced by the compiler), but it is unspecified which
implementation is used to generate code.
As a result, it is strongly recommended that user-defined sum types ensure
that for every alternative there is a factory function or constant member with
the same name and parameter list, such that pattern-matching on the result will
correctly reproduce the arguments to the factory function. For example, the
definition of Optional
above satisfies this requirement, because for any
regular type T
, the expression Optional(T).None
evaluates to a value that
matches the pattern Optional(T).None
(under both possible matching
mechanisms), and for any x
of type T
, the expression Optional(T).Some(x)
evaluates to a value that matches the pattern Optional(T).Some(y: T)
and binds
y
to a value that's equal to x
. Expression patterns involving a sum type
that doesn't meet this requirement will fail to compile, or have behavior that
observably changes depending on the compiler's implementation choices.
Another corollary of this rule is that if an alternative takes no arguments, its
pattern syntax is the same as its expression syntax. For example,
case Optional(i32).None() => ...
is not well-formed, because
Optional(i32).None()
has the form of an alternative pattern, but the
implementation in terms of ==
is not well-formed because
Optional(i32).None()
is not a well-formed expression. If we had defined
Optional.None
as a factory function instead of a constant,
case Optional(i32).None() => ...
would be well-formed but
case Optional(i32).None => ...
would not be.
Note that the compiler-generated continuation method bodies are not required to
contain the code in the case
body (or whatever code is in the scope of the
pattern). For example, they might only store the parameter values and then
return an index that identifies the case
body to be executed.
- Providing
choice
types only, with no support for user-defined sum types. - Indexing alternatives by type instead of by name.
- Implementing user-defined sum types in terms of
choice
type proxies rather than callbacks. - Implementing user-defined sum types in terms of invertible pattern functions.
- Proposal #157: Design direction for sum types