We need to say how you define new types in Carbon. This proposal is specifically about record types. The proposal is not intended to be a complete story for record types, but enough to get agreement on direction. It primarily focuses on:
- use cases including: data classes, encapsulated types with virtual and non-virtual methods and optional single inheritance, interfaces as base classes that support multiple inheritance, and mixins for code reuse;
- anonymous structural data types for record literals used to initialize class values and ad-hoc parameter and return types with named components; and
- future work, including the provisional syntax in use for features that have not been decided.
This is a replacement for earlier proposal #98.
This proposal adds an initial design for record types called "classes", including structural data classes called struct types as well as struct literals. The design is replacing the skeletal design for what were called "struct" types with a new document on classes.
This particular proposal is focusing on the Carbon goal that code is "easy to read, understand, and write." Future proposals will address other aspects of the class type design such as performance.
There was an earlier proposal #98, that made a number of different choices, including:
- Tuples were given named components instead of having separate struct literals.
- No form of multiple inheritance was proposed.
- Operators were define using methods like C++ instead of by implementing interfaces like Rust.
- Constructors had a special form like C++ instead of being regular functions like Rust.
- Tuples and classes were both considered example of record types.
- Members of classes could be individually left uninitialized.
- Coverage of nominal types, inheritance, etc. were considered in much more detail.
Whether we would support implementing interfaces for specific anonymous data classes was discussed on Discord. The conclusion was "yes", reasoning that we would support that for the same reason as a number of other cases such as tuple and pointer types. A specific use case would be implementing interface
interface ConstructWidgetFrom { fn Construct(Self) -> Widget; }
for type {.kind: WidgetKind, .size: Int}.
Issue #665 decided that by default members of a class would be publicly accessible. There were a few reasons:
- The readability of public members is the most important, since we expect most readers to be concerned with the public API of a type.
- The members that are most commonly private are the data fields, which have relatively less complicated definitions that suffer less from the extra annotation.
Additionally, there is precedent for this approach in modern object-oriented languages such as Kotlin and Python, both of which are well regarded for their usability.
It further decided that members would be given more restricted access using a local annotation on the declaration itself rather than a block or region approach such as used in C++. This is primarily motivated by a desire to reduce context sensitivity, following the principle introduced in #646. It helps readers to more easily determine the accessibility of a member in large classes, say when they have jumped to a specific definition in their IDE.
Issue #653 discussed whether structural data class types should have an introducer to distinguish them from structural data class literals. Ultimately we decided no introducer was needed:
- Outside of
{}, types could be distinguished from literal values by the presence of a:after the first field name. - This creates a sort of consistency: introducers are frequently used when
introducing new names, as in
fn,var,interface, and so on. Struct type declarations don't introduce new names so they don't require an introducer. - It avoids having a different introducer for things that are still treated as classes for many purposes. This means we won't have to frequently say "struct or class" in documentation.
- We do want to use these type expressions in contexts that will benefit from being more concise, such as inside function and variable declarations.
This does cause an issue that {} is both an empty struct literal and empty
struct type. However, we've already accepted that complexity with tuples, so
this choice is more consistent. If we find that we need an introducer for tuple
types to distinguish the empty tuple from its type, we expect to find that we
have the same problem with empty struct literals, and the other way around. We
are explicitly to choosing to accept the risk that this won't work out in order
to have a more concise syntax in case it does.
Do literals have "class" type or are they some other kind of type? Issue #651 decided that all of these types were different kinds of classes:
- Literals like
{.a = 2}are "structural data class literals" or "struct literals" for short. Here "structural" means that two types are considered equal if their fields match. They are not "nominal" since they don't have a name to use for type equality. - The types of those literals like
{.a: i64}would be "structural data classes" or "struct types" for short. - There would also be "nominal data classes" that are declared with a syntax more similar to other nominal classes.
We preferred to refer to all of these as class types, rather than have to frequently refer to "struct or class types", adding additional words to name more specific subsets, like "data classes". In contrast, tuple types are not considered classes, but classes and tuples together form product types.
The term "class" was chosen over "struct" since they generally support object-oriented features like encapsulation, inheritance, and dynamic dispatch, and how C++ programmers generally refer to their record types. These were considered more significant than the C++ distinction that classes default to private access control. Since we plan to use a different syntax in Carbon to specify access restrictions, the different default seemed straightforward to teach.