Shape declaration syntax is similar to array syntax, but instead of values, types are specified for each key:
Values are also specified in a similar way:
At this point, you might be thinking that this is basically just a struct - why not just use a class with public properties instead?
Shapes differ from objects in that subtypes and supertypes is entirely based on field definitions, not an explicit relationship declaration -
shape() is a supertype of all shapes, and
shape('foo' => string, 'bar' => int) is a subtype of:
shape('foo' => string)
shape('bar' => string)
shape('foo' => ?string)
... among many others - this is referred to as structural subtyping. As a less abstract example:
Managing Untyped Data
Hack code is often mixed with PHP code, either because a project is migrating to Hack, or because PHP libraries are still a valuable resource in greenfield projects; in practice, this means that inside some Hack function, we often have untyped data from PHP that we need to pass to a different Hack function which requires typed parameters.
The initial way to do this was to check the type, then either call the function you need, or throw an exception:
This rapidly gets extremely verbose and annoying; to make things slightly better, you can take advantage of the typechecker's understanding of the
Large projects usually end up creating utility functions for common type operations - in particular, converting nullable types to non-nullable:
fredemmott/type-assert is a library of functions for asserting types in this fashion, including (among others):
TypeAssert::isString(mixed): stringand similar for other scalar types
TypeAssert::matchesTypeStructure<T>(TypeStructure<T>, mixed): T
The last example is where things get interesting for shapes:
TypeStructure<T> is an unsupported API in HHVM 3.12 and above, allowing runtime reflection over shape declarations; while it's unsupported, it's very unlikely to be removed without a replacement (this is my personal opinion, not a guarantee).
Shapes: A Real-World Example
When interacting with third-party APIs, we'll often get back a JSON blob - but there's no way to statically know what the shape of that data will be, and how it can be safely interacted with. We can address this by defining in advance what shape we expect the data to be, then asserting at runtime that it matches - TypeAssert makes this very convenient. For example, you can fetch basic information on the latest GitHub commit to HHVM in a type-safe way:
The output of this is:
This approach gives you a very readable and extensible definition of what fields you expect, what types you expect, and means you don't need to write your own validation logic, as
TypeAssert does it for you based on your definition.
There's one problem I've repeatedly came across when using shapes: the combination of implicit subtypes with nullable types being acceptable for both ommitted and nullable types mean that the typechecker loses its' ability to detect typos in some situations:
While this isn't a big deal, it is annoying when you're used to the typechecker usually doing an excellent job of instantly finding your mistakes :)