Types, type checking, and contracts for Ruby
Ruby REXX

README.md

Gem Version Build Status

Table of Contents

Introduction

RDL is a lightweight system for adding types, type checking, and contracts to Ruby. In RDL, types can be used to decorate methods:

require 'rdl'

type '(Fixnum, Fixnum) -> String'
def m(x,y) ... end

This indicates that m returns a String if given two Fixnum arguments. When written as above, RDL enforces this type as a contract checked at run time: When m is called, RDL checks that m is given exactly two arguments and both are Fixnums, and that m returns an instance of String.

RDL can also statically type check method bodies against their signatures. For example:

file.rb:
  require 'rdl'

  type '(Fixnum) -> Fixnum', typecheck: :now
  def id(x)
    "forty-two"
  end
$ ruby file.rb
.../lib/rdl/typecheck.rb:32:in `error':  (RDL::Typecheck::StaticTypeError)
.../file.rb:5:5: error: got type `String' where return type `Fixnum' expected
.../file.rb:5:     "forty-two"
.../file.rb:5:     ^~~~~~~~~~~

Passing typecheck: :now to type checks the method body immediately or as soon as it is defined. Passing typecheck: :call to type statically type checks the method body whenever it is called. Passing typecheck: sym for some other symbol statically type checks the method body when rdl_do_typecheck sym is called.

Note that RDL tries to follow the philosophy that you get what you pay for. Methods with type annotations can be checked dynamically or statically; methods without type annotations are unaffected by RDL. See the performance discussion for more detail.

RDL supports many more complex type annotations; see below for a complete discussion and examples.

RDL types are stored in memory at run time, so it's also possible for programs to query them. RDL includes lots of contracts and types for the core and standard libraries. Since those methods are generally trustworthy, RDL doesn't actually enforce the contracts (since that would add overhead), but they are available to search, query, and use during type checking. RDL includes a small script rdl_query to look up type information from the command line. Note you might need to put the argument in quotes depending on your shell.

$ rdl_query String#include?            # print type for instance method of another class
$ rdl_query Pathname.glob              # print type for singleton method of a class
$ rdl_query Array                      # print types for all methods of a class
$ rdl_query "(Fixnum) -> Fixnum"       # print all methods that take a Fixnum and return a Fixnum
$ rdl_query "(.) -> Fixnum"            # print all methods that take a single arg of any type
$ rdl_query "(..., Fixnum, ...) -> ."  # print all methods that take a Fixnum as some argument

See below for more details of the query format. The rdl_query method performs the same function as long as the gem is loaded, so you can use this in irb.

$ irb
> require 'rdl'
 => true
> require 'types/core'
 => true

> rdl_query '...' # as above

RDL also supports more general contracts, though these can only be enforced at run time and are not statically checked. These more general contracts take the form of preconditions, describing what a method assumes about its inputs, and postconditions, describing what a method guarantees about its outputs. For example:

require 'rdl'

pre { |x| x > 0 }
post { |r,x| r > 0 }
def sqrt(x)
  # return the square root of x
end

Given this program, RDL intercepts the call to sqrt and passes its argument to the pre block, which checks that the argument is positive. Then when sqrt returns, RDL passes the return value (as r) and the initial argument (as x) to the post block, which checks that the return is positive. (Let's ignore complex numbers to keep things simple...) RDL contracts are enforced at method entry and exit. For example, if we call sqrt(49), RDL first checks that 49 > 0; then it passes 49 to sqrt, which (presumably) returns 7; then RDL checks that 7 > 0; and finally it returns 7. Note that pre- and postconditions can't be searched for using rdl_query.

Using RDL

Supported versions of Ruby

RDL currently supports Ruby 2.x. It may or may not work with other versions.

Installing RDL

gem install rdl should do it.

Loading RDL

Use require 'rdl' to load the RDL library. If you want to use the core and standard library type signatures that come with RDL, follow it with require 'types/core'. This will load the types based on the current RUBY_VERSION. Currently RDL has types for the following versions of Ruby:

  • 2.x

(Currently all 2.x versions are assumed to have the same library type signatures, which may not be correct.)

Disabling RDL

For performance reasons you probably don't want to use RDL in production code. To disable RDL, replace require 'rdl' with require 'rdl_disable'. This will cause all invocations of RDL methods to either be no-ops or to do the minimum necessary to preserve the program's semantics (e.g., if the RDL method returns self, then so does the rdl_disable method.)

Rails

To add types to a Ruby on Rails application, add

gem 'rdl'

to your Gemfile to get the latest stable version, or

gem 'rdl', git: 'https://github.com/plum-umd/rdl'

to get the head version from github.

In development and test mode, you will now have access to rdl, types/core from RDL, and extra type annotations for Rails and some related gems. In production mode, RDL will be disabled (by loading rdl_disable).

Warning: Rails support is currently extremely limited, not well tested, and generally needs more work...please send bug reports/pull requests/etc and we will fix things.

Currently, RDL has types for the following versions of Rails:

  • Rails 5.x support - limited to the following:
    • Models
      • Generates type annotations for model column getters and setters
      • find_by and find_by!
      • Generates types annotations for methods added by belongs_to, has_one, has_many, has_and_belongs_to_many.
        • Note currently RDL doesn't do very precise type checking of relations
    • Controllers
      • errors

Preconditions and Postconditions

The pre method takes a block and adds that block as a precondition to a method. When it's time to check the precondition, the block will be called with the method's arguments. If the block returns false or nil the precondition is considered to have failed, and RDL will raise a ContractError. Otherwise, if the block returns a true value, then the method executes as usual. The block can also raise its own error if the contract fails.

The pre method can be called in several ways:

  • pre { block } - Apply precondition to the next method to be defined
  • pre(mth) { block } - Apply precondition to method mth of the current class, where mth is a Symbol or String
  • pre(cls, mth) { block } - Apply precondition to method mth of class cls, where cls is a Class, Symbol, or String, and mth is a Symbol or String

The post method is similar, except its block is called with the return value of the method (in the first position) followed by all the method's arguments. For example, you probably noticed that for sqrt above the post block took the return value r and the method argument x.

(Note: RDL does not clone or dup the arguments at method entry. So, for example, if the method body has mutated fields stored inside those argument objects, the post block or any other check evaluated afterwards will see the mutated field values rather than the original values.)

The post method can be called in the same ways as pre.

Methods can have no contracts, pre by itself, post by itself, both, or multiple instances of either. If there are multiple contracts, RDL checks that all contracts are satisfied, in the order that the contracts were bound to the method.

Type Annotations

The type method adds a type contract to a method. It supports the same calling patterns as pre and post, except rather than a block, it takes a string argument describing the type. More specifically, type can be called as:

  • type 'typ'
  • type m, 'typ'
  • type cls, mth, 'typ'

A type string generally has the form (typ1, ..., typn) -> typ indicating a method that takes n arguments of types typ1 through typn and returns type typ. Below, to illustrate the various types RDL supports, we'll use examples from the core library type annotations.

The type method can be called with wrap: false so the type information is stored but the type is not enforced. For example, due to the way RDL is implemented, the method String#=~ can't have a type or contract on it because then it won't set the correct $1 etc variables:

type :=~, '(Object) -> Fixnum or nil', wrap: false # Wrapping this messes up $1 etc

For consistency, pre and post can also be called with wrap: false, but this is generally not as useful.

RDL Types

Nominal Types

A nominal type is simply a class name, and it matches any object of that class or any subclass.

type String, :insert, '(Fixnum, String) -> String'

Nil Type

The nominal type NilClass can also be written as nil. The only object of this type is nil:

type IO, :close, '() -> nil' # IO#close always returns nil

Currently, nil is treated as if it were an instance of any class.

x = "foo"
x.insert(0, nil) # RDL does not report a type error

We chose this design based on prior experience with static type systems for Ruby, where not allowing this leads to a lot of false positive errors from the type system. However, we may change this in the future.

Top Type (%any)

RDL includes a special "top" type %any that matches any object:

type Object, :=~, '(%any) -> nil'

We call this the "top" type because it is the top of the subclassing hierarchy RDL uses. Note that %any is more general than Object, because not all classes inherit from Object, e.g., BasicObject does not.

Union Types

Many Ruby methods can take several different types of arguments or return different types of results. The union operator or can be used to indicate a position where multiple types are possible.

type IO, :putc, '(Numeric or String) -> %any'
type String, :getbyte, '(Fixnum) -> Fixnum or nil'

Note that for getbyte, we could leave off the nil, but we include it to match the current documentation of this method.

Intersection Types

Sometimes Ruby methods have several different type signatures. (In Java these would be called overloaded methods.) In RDL, such methods are assigned a set of type signatures:

type String, :[], '(Fixnum) -> String or nil'
type String, :[], '(Fixnum, Fixnum) -> String or nil'
type String, :[], '(Range or Regexp) -> String or nil'
type String, :[], '(Regexp, Fixnum) -> String or nil'
type String, :[], '(Regexp, String) -> String or nil'
type String, :[], '(String) -> String or nil'

We say the method's type is the intersection of the types above.

When this method is called at run time, RDL checks that at least one type signature matches the call:

"foo"[0]  # matches first type
"foo"[0,2] # matches second type
"foo"[0..2] # matches third type
"foo"[0, "bar"] # error, doesn't match any type
# etc

Notice that union types in arguments could also be written as intersection types of methods, e.g., instead of the third type of [] above we could have equivalently written

type String, :[], '(Range) -> String or nil'
type String, :[], '(Regexp) -> String or nil'

Optional Argument Types

Optional arguments are denoted in RDL by putting ? in front of the argument's type. For example:

type String, :chomp, '(?String) -> String'

This is actually just a shorthand for an equivalent intersection type:

type String, :chomp, '() -> String'
type String, :chomp, '(String) -> String'

but it helps make types more readable.

Like Ruby, RDL allows optional arguments to appear anywhere in a method's type signature.

Variable Length Argument Types

In RDL, * is used to decorate an argument that may appear zero or more times. Currently in RDL this annotation may only appear on the rightmost argument. For example, String#delete takes one or more String arguments:

type String, :delete, '(String, *String) -> String'

Named Argument Types

RDL allows arguments to be named, for documentation purposes. Names are given after the argument's type, and they do not affect type contract checking in any way. For example:

type Fixnum, :to_s, '(?Fixnum base) -> String'

Here we've named the first argument of to_s as base to give some extra hint as to its meaning.

Dependent Types

RDL allows for refinement predicates to be attached to named arguments. The predicates on the arguments are checked when the method is called, and the predicates on the return is checked when then method returns. For instance:

type '(Float x {{ x>=0 }}) -> Float y {{ y>=0 }}'
def sqrt(x)
    # return the square root of x
end

Here, RDL will check that the sqrt method is called on an argument of type Float which is greater than or equal to 0, and it will check the same of the return value of the method. Note that, in effect, dependent type contracts can be used in place of pre and post contracts.

Dependencies can also exist across a method's arguments and return value:

type '(Fixnum x {{ x>y }}, Fixnum y) -> Float z {{ z==(x+y) }}'
def m(x,y) ... end

Any arbitrary code can be placed between the double braces of a type refinement, and RDL will dynamically check that this predicate evaluates to true, or raise a type error if it evaluates to false.

Most pre- and postconditions can be translated into a dependent type by attaching the precondition to one of the arguments and the postcondition to the return. Note, however, that dependently typed positions must always have a name, even if the associated refinment doesn't refer to that name:

type '(Fixnum x {{ $y > 0 }}) -> nil'    # argument name must be present even though refinment doesn't use it.

Higher-order Types

RDL supports types for arguments or return values which are themselves Proc objects. Simply enclose the corresponding argument's type with braces to denote that it is a Proc. For example:

type '(Fixnum, {(Fixnum) -> Fixnum}) -> Fixnum'
def m(x, y) ... end

The type annotation above states that the method m takes two arguments: one of type Fixnum, and another which is a Proc which itself takes a Fixnum and returns a Fixnum. A Proc may be the return value of a method as well:

type '(Fixnum) -> {(Float) -> Float}'
def m(x) ... end

These higher-order types are checked by wrapping the corresponding Proc argument/return in a new Proc which checks that the type contract holds.

A type contract can be provided for a method block as well. The block's type should be included after the method argument types:

type '(Fixnum, Float) {(Fixnum, String) -> String } -> Float'
def m(x,y,&blk) ... end

Note that this notation will work whether or not a method block is explicitly referenced in the parameters, i.e., whether or not &blk is included above. Finally, dependent types work across higher order contracts:

type '(Fixnum x, Float y) -> {(Fixnum z {{ z>y }}) -> Fixnum}'
def m(x,y,&blk) ... end

The type contract above states that method m returns a Proc which takes a Fixnum z which must be greater than the argument Float y. Whenever this Proc is called, it will be checked that this contract holds.

Class/Singleton Method Types

RDL method signatures can be used both for instance methods and for class methods (often called singleton methods in Ruby). To indicate a type signature applies to a singleton method, prefix the method name with self.:

type File, 'self.dirname', '(String file) -> String dir'

(Notice also the use of a named return type, which we haven't seen before.)

Type signatures can be added to initialize by giving a type signature for self.new:

type File, 'self.new', '(String file, ?String mode, ?String perm, ?Fixnum opt) -> File'

Structural Types

Some Ruby methods are intended to take any object that has certain methods. RDL uses structural types to denote such cases:

type IO, :puts, '(*[to_s: () -> String]) -> nil'

Here IO#puts can take zero or more arguments, all of which must have a to_s method that takes no arguments and returns a String.

The actual checking that RDL does here varies depending on what type information is available. Suppose we call puts(o). If o is an instance of a class that has a type signature t for to_s, then RDL will check that t is compatible with () -> String. On the other hand, if o is an instance of a class with no type signature for to_s, RDL only checks that o has a to_s method, but it doesn't check its argument or return types.

Singleton Types

Not to be confused with types for singleton methods, RDL includes singleton types that denote positions that always have one particular value; this typically happens only in return positions. For example, Dir#mkdir always returns the value 0:

type Dir, 'self.mkdir', '(String, ?Fixnum) -> 0'

In RDL, any integer or floating point number denotes a singleton type. Arbitrary values can be turned into singleton types by wrapping them in ${.}. For example, Float#angle always returns 0 or pi.

type Float, :angle, '() -> 0 or ${Math::PI}'

RDL checks if a value matches a singleton type using equal?. As a consequence, singleton string types aren't currently possible.

Note that the type nil is actually implemented as a singleton type with the special behavior that nil is a treated as a member of any class. However, while nil can in general be used anywhere any type is expected, it cannot be used where a different singleton type is expected. For example, nil could not be a return value of Dir#mkdir or Float#angle.

Self Type

Consider a method that returns self:

class A
  def id
    self
  end
end

If that method might be inherited, we can't just give it a nominal type, because it will return a different object type in a subclass:

class B < A
end

type A, :id, '() -> A'
A.new.id # okay, returns an A
B.new.id # type error, returns a B

To solve this problem, RDL includes a special type self for this situation:

type A, :id, '() -> self'
A.new.id # okay, returns self
B.new.id # also okay, returns self

Thus, the type self means "any object of self's class."

Type Aliases

RDL allows types to be aliases to make them faster to write down and more readable. All type aliases begin with %. RDL has one built-in alias, %bool, which is shorthand for TrueClass or FalseClass:

type String, :==, '(%any) -> %bool'

Note it is not a bug that == is typed to allow any object. Though you would think that developers would generally only compare objects of the same class (since otherwise == almost always returns false), in practice a lot of code does compare objects of different classes.

Method type_alias(name, typ) can be used to create a user-defined type alias, where name must begin with %:

type_alias '%real', 'Integer or Float or Rational'
type_alias '%string', '[to_str: () -> String]'
type_alias '%path', '%string or Pathname'

Type aliases have to be created before they are used (so above, %path must be defined after %string).

Generic Class Types

RDL supports parametric polymorphism for classes, a.k.a. generics. The type_params method names the type parameters of the class, and those parameters can then be used inside type signatures:

class Array
  type_params [:t], :all?

  type :shift, '() -> t'
end

Here the first argument to type_params is a list of symbols or strings that name the type parameters. In this case there is one parameter, t, and it is the return type of shift.

Generic types are applied to type arguments using <...> notation, e.g., Array<Fixnum> is an Array class where t is replaced by Fixnum. Thus, for example, if o is an Array<Fixnum>, then o.shift returns Fixnum. As another example, here is the type for the [] method of Array:

type Array, :[], '(Range) -> Array<t>'
type Array, :[], '(Fixnum or Float) -> t'
type Array, :[], '(Fixnum, Fixnum) -> Array<t>'

Thus if o is again an Array<Fixnum>, then o[0] returns a Fixnum and o[0..5] returns an Array<Fixnum>.

In general it's impossible to assign generic types to objects without knowing the programmer's intention. For example, consider code as simple as x = [1,2]. Is it the programmer's intention that x is an Array<Fixnum>? Array<Numeric>? Array<Object>?

Thus, by default, even though Array is declared to take type parameters, by default RDL treats array objects at the raw type Array, which means the type parameters are ignored whenever they appear in types. For our example, this means a call such as x.push("three") would not be reported as an error (the type signature of Array#push is '(?t) -> Array<t>').

To fully enforce generic types, RDL requires that the developer instantiate! an object with the desired type parameters:

x = [1,2]
x.instantiate!('Fixnum')
x.push("three") # type error

Note that the instantiated type is associated with the object, not the variable:

y = x
y.push("three") # also a type error

When RDL instantiates an object with type parameters, it needs to ensure the object's contents are consistent with the type. Currently this is enforced using the second parameter to type_params, which must name a method that behaves like Array#all?, i.e., it iterates through the contents, checking that a block argument is satisfied. As seen above, for Array we call type_params(:t, :all?). Then at the call x.instantiate('Fixnum'), RDL will call Array#all? to iterate through the contents of x to check they have type Fixnum.

RDL also includes a deinstantiate! method to remove the type instantiation from an object:

x.deinstantiate!
x.push("three") # no longer a type error

Finally, type_params can optionally take a third argument that is an array of variances, which are either :+ for covariance, :- for contravariance, or :~ for invariance. If variances aren't listed, type parameters are assumed to be invariant, which is a safe default.

Variances are only used when RDL checks that one type is a subtype of another. This only happens in limited circumstances, e.g., arrays of arrays where all levels have instantiated types. So generally you don't need to worry much about the variance.

The rules for variances are standard. Let's assume A is a subclass of B. Also assume there is a class C that has one type parameter. Then:

  • C<A> is a subtype of C<A> always
  • C<A> is a subtype of C<B> if C's type parameter is covariant
  • C<B> is a subtype of C<A> if C's type parameter is contravariant

Tuple Types

A type such as Array<Fixnum> is useful for homogeneous arrays, where all elements have the same type. But Ruby programs often use heterogeneous arrays, e.g., [1, "two"]. The best generic type we can give this is Array<Fixnum or String>, but that's imprecise.

RDL includes special tuple types to handle this situation. Tuple types are written [t1, ..., tn], denoting an Array of n elements of types t1 through tn, in that order. For example, [1, "two"] has type [Fixnum, String]. As another example, here is the type of Process#getrlimit, which returns a two-element array of Fixnums:

type Process, 'self.getrlimit', '(Symbol or String or Fixnum resource) -> [Fixnum, Fixnum] cur_max_limit'

Finite Hash Types

Similarly to tuple types, RDL also supports finite hash types for heterogeneous hashes. Finite hash types are written {k1 => v1, ..., kn => vn} to indicate a Hash with n mappings of type ki maps to vi. The ki may be strings, integers, floats, or constants denoted with ${.}. If a key is a symbol, then the mapping should be written ki: vi. In the latter case, the {}'s can be left off:

type MyClass, :foo, '(a: Fixnum, b: String) { () -> %any } -> %any'

Here foo, takes a hash where key :a is mapped to a Fixnum and key :b is mapped to a String. Similarly, {'a'=>Fixnum, 2=>String} types a hash where keys 'a' and 2 are mapped to a Fixnum and String, respectively. Both syntaxes can be used to define hash types.

RDL also allows a "rest" type in finite hashes (of course, they're not so finite if they use it!):

type MyClass, :foo, '(a: Fixnum, b: String, **Float) -> %any'

In this method, a is a Fixnum, b is a String, and any number (zero or more) remaining keyword arguments can be passed where the values have type Float, e.g., a call foo(a: 3, b: 'b', pi: 3.14) is allowed.

Type Casts

Sometimes RDL does not have precise information about an object's type (this is most useful during static type checking). For these cases, RDL supports type casts of the form o.type_cast(t). This call returns a new object that delegates all methods to o but that will be treated by RDL as if it had type t. If force: true is passed to type_cast, RDL will perform the cast without checking whether o is actually a member of the given type. For example, x = "a".type_cast('nil', force: true) will make RDL treat x as if it had type nil, even though it's a String.

Bottom Type (%bot)

RDL also includes a special bottom type %bot that is a subtype of any type, including any class and any singleton types. In static type checking, the type %bot is given to so-called void value expressions, which are return, break, next, redo, and retry (notice that these expressions perform jumps rather than producing a value, hence they can be treated as having an arbitrary type). No Ruby objects have type %bot.

Non-null Type

Types can be prefixed with ! to indicate the associated value is not nil. For example:

type :x=, '(!Fixnum) -> !Fixnum' # x's argument must not be nil

Warning: This is simply documentation of non-nullness, and is not checked by the static type checker. The contract checker might or might not enforce non-nullness. (For those who are curious: RDL has this annotation because it seems useful for descriptive purposes. However, it's quite challenging to build a practical analysis that enforces non-nilness without reporting too many false positives.)

Static Type Checking

As mentioned in the introduction, calling type with typecheck: :now statically type checks the body of the annotated method body against the given signature. If the method has already been defined, RDL will try to check the method immediately. Otherwise, RDL will statically type check the method as soon as it is loaded.

Often method bodies cannot be type checked as soon as they are loaded because they refer to classes, methods, and variables that have not been created yet. To support these cases, some other symbol can be supplied as typecheck: sym. Then when rdl_do_typecheck sym is called, all methods typechecked at sym will be statically checked.

Additionally, type can be called with typecheck: :call, which will delay checking the method's type until the method is called. Currently these checks are not cached, so expect a big performance hit for using this feature.

To perform type checking, RDL needs source code, which it gets by parsing the file containing the to-be-typechecked method. Hence, static type checking does not work in irb since RDL has no way of getting the source. Typechecking does work in pry (this feature has only limited testing) as long as typechecking is delayed until after the method is defined:

[2] pry(main)> require 'rdl'
[3] pry(main)> require 'types/core'
[4] pry(main)> type '() -> Fixnum', typecheck: :later    # note: typecheck: :now doesn't work in pry
[5] pry(main)> def f; 'haha'; end
[6] pry(main)> rdl_do_typecheck :later
RDL::Typecheck::StaticTypeError:
(string):2:3: error: got type `String' where return type `Fixnum' expected
(string):2:   'haha'
(string):2:   ^~~~~~
from .../typecheck.rb:158:in `error'

RDL currently uses the parser Gem to parse Ruby source code. (And RDL uses the parser gem's amazing diagnostic output facility to print type error messages.)

Next we discuss some special features of RDL's type system and some of its limitations.

Types for Variables

In a standard type system, local variables have one type throughout a method or function body. For example, in C and Java, declaring int x means x can only be used as an integer. However, in Ruby, variables need not be declared before they are used. Thus, by default, RDL treats local variables flow-sensitively, meaning at each assignment to a local variable, the variable's type is replaced by the type of the right hand side. For example:

x = 3       # Here `x` is a `Fixnum`
x = "three" # Now `x` is a `String`

(Note this is a slight fib, since after the first line, x will actually have the singleton type 3. But we'll ignore this just to keep the discussion a bit simpler, especially since 3 is a subtype of Fixnum.)

After conditionals, variables have the union of the types they have along both branches:

if (some condition) then x = 3 else x = "three" end
# x has type `Fixnum or String`

RDL also provides a method var_type that can be used to force a local variable to have a single type through a method body, i.e., to treat it flow-insensitively like a standard type system:

var_type :x, 'Fixnum'
x = 3       # okay
x = "three" # type error

The first argument to var_type is a symbol with the local variable name, and the second argument is a string containing the variable's type. Note that var_type is most useful at the beginning of method or code block. Using it elsewhere may result in surprising error messages, since RDL requires variables with fixed types to have the same type along all paths. Method parameters are treated as if var_type was called on them at the beginning of the method, fixing them to their declared type. This design choice may be revisited in the future.

There is one subtlety for local variables and code blocks. Consider the following code:

x = 1
m() { x = 'bar' }
# what is x's type here?

If m invokes the code block, x will be a String after the call. Otherwise x will be 1. Since RDL can't tell whether the code block is ever called, it assigns x type 1 or String. It's actually quite tricky to do very precise reasoning about code blocks. For example, m could (pathologically) store its block in a global variable and then only call it the second time m is invoked. To keep its reasoning simple, RDL treats any local variables captured (i.e., imported from an outer scope) by a code block flow-insensitively for the lifetime of the method. The type of any such local variable is the union of all types that are ever assigned to it.

RDL always treats instance, class, and global variables flow-insensitively, hence their types must be defined with var_type:

class A
  var_type :@f, 'Fixnum'
  def m
    @f = 3       # type safe
    @f = "three" # type error, incompatible type in assignment
    @g = 42      # type error, no var_type for @g
  end
end

The var_type method may also be called as var_type klass, :name, typ to assign a type to an instance or class variable of class klass.

As a short-hand, RDL defines methods attr_accessor_type, attr_reader_type, and attr_writer_type to behave like their corresponding non-_type analogs but assign types to the attributes. For example, attr_accessor_type :f, 'Fixnum', :g, 'String' is equivalent to:

var_type :@f, 'Fixnum'
var_type :@g, 'String'
type :f, '() -> Fixnum'
type :f=, '(Fixnum) -> Fixnum'
type :g, '() -> String'
type :g=, '(String) -> String'

Tuples, Finite Hashes, and Subtyping

When RDL encounters a literal array in the program, it assigns it a tuple type, which allows, among other things, precise handling of multiple assignment. For example:

x = [1, 'foo']  # x has type [1, String]
a, b = x        # a has type 1, b has type String  

RDL also allows a tuple [t1, ..., tn] to be used where Array<t1 or ... or tn> is expected. This means both when a tuple is passed to an Array position, and when any method is invoked on the tuple (even if RDL could safely apply that method to the tuple; this may change in the future):

var_type @f, 'Array<Fixnum or String>'
@f = [1, 'foo'] # okay
@f.length       # also okay

To maintain soundness, a tuple that is used as an Array is treated as if it were always an array. For example:

x = [1, 'foo']  # at this point, x has type [1, String]
var_type @f, '[1, String]'
@f = x          # okay so far
var_type @g, 'Array<Fixnum or String>'
@g = x          # uh oh

When RDL encounters the assignment to @g, it retroactively changes x to have type Array<Fixnum or String>, which is incompatible with type [1, String] of @f, so the assignment to @g signals an error.

RDL uses the same approach for hashes: hash literals are treated as finite hashes. A finite hash {k1=>v1, ..., kn=>vn} can be used where Hash<k1 or ... or kn, v1 or ... or vn> is expected. And if a finite hash is used as a Hash (including invoking methods on the finite hash; this may change in the future), then it is retroactively converted to a Hash.

Other Features and Limitations

Displaying types. As an aid to debugging, the method rdl_note_type e will display the type of e during type checking. At run time, this method returns its argument. Note that in certain cases RDL may type check the same code repeatedly, in which case an expression's type could be printed multiple times.

  • Conditional guards and singletons. If an if or unless guard has a singleton type, RDL will typecheck both branches but not include types from the unrealizable branch in the expression type. For example, if true then 1 else 'two' end has type 1. RDL behaves similarly for && and ||. However, RDL does not implement this logic for case.

  • Case analysis by class. If the guard of a case statement is a variable, and then when branches compare against classes, RDL refines the type of the guard to be those classes within the corresponding when branch. For example, in case x when Fixnum ...(1)... when String ...(2)... end, RDL will assume x is a Fixnum within (1) and a String within (2).

  • Multiple Assignment and nil. In Ruby, extra left-hand sides of multiple assignments are set to nil, e.g., x, y = [1] sets x to 1 and y to nil. However, RDL reports an error in this case; this may change in the future.

  • Block formal arguments. Similarly, RDL reports an error if a block is called with the wrong number of arguments even though Ruby does not signal an error in this case.

  • Caching. If typecheck: :call is specified on a method, Ruby will type check the method every time it is called. In the future, RDL will cache these checks.

  • Dependent Types. RDL ignores refinements in checking code with dependent types. E.g., given a Fixnum x {{ x > 0 }}, RDL will simply treat x as a Fixnum and ignore the requirement that it be positive.

  • Unsupported Features. There are several features of Ruby that are currently not handled by RDL. Here is a non-exhaustive list:

    • super is not supported.
    • lambda has special semantics for return; this is not supported. Stabby lambda is also not supported.
    • Only simple for iteration variables are supported.
    • Control flow for exceptions is not analyzed fully soundly; some things are not reported as possibly nil that could be.
    • Only simple usage of constants is handled.

Assumptions

RDL's static type checker makes some assumptions that should hold unless your Ruby code is doing something highly unusual. RDL assumes:

  • Class#=== is not redefined
  • Proc#call is not redefined

(More assumptions will be added here as they are added to RDL...)

Other RDL Methods

RDL also includes a few other useful methods:

  • rdl_alias new_name, old_name tells RDL that method new_name is an alias for method old_name, and therefore they should have the same contracts and types. This method is only needed when adding contracts and types to method that have already been aliased; it's not needed if the method is aliased after the contract or type has been added.

  • rdl_nowrap, if called at the top-level of a class, causes RDL to behave as if wrap: false were passed to all type, pre, and post calls in the class. This is mostly used for the core and standard libraries, which have trustworthy behavior hence enforcing their types and contracts is not worth the overhead.

  • rdl_remove_type klass, meth removes the type annotation for meth in klass. Fails if meth does not have a type annotation.

  • rdl_query prints information about types; see below for details.

  • rdl_do_at(sym, &blk) invokes blk.call(sym) when rdl_do_typecheck(sym) is called. Useful when type annotations need to be generated at some later time, e.g., because not all classes are loaded.

Performance

RDL supports some tradeoffs between safety and performance. There are three main sources of overhead in using RDL:

  • When a method is wrapped by RDL, invoking it is more expensive because it requires calling the wrapper which then calls the underlying method. This is the most significant run-time cost of using RDL, and it can be eliminated by adding wrap: false to an annotation. (But, this only makes sense for types, since those are the only annotations that can be statically checked.)

  • When type checking is performed, RDL parses the program's source code and type checks method bodies. This source of overhead only happens once per method (unless typecheck: :call is used), so the overhead is probably not too bad (though we have not measured it).

  • RDL adds an implementation of method_added and singleton_method_added to carry out some of its work, and those are called on every method addition. This source of overhead is again likely not too significant.

For uses of pre and post, there's not a lot of choice: those contracts are enforced at run-time and will incur the costs of wrapped methods. However, note that any methods that are not annotated with pre or post will not incur the cost of wrapping. (Similarly, methods not annotated with type never incur any wrapping cost.)

For uses of type, there are more choices, which can be split into two main use cases. First, suppose there's a method m that we want a type for but don't want to type check (for example, it may come from some external library). So suppose we read the documentation and give m type (Fixnum) -> Fixnum. We now have to decide whether to wrap m. If we don't wrap m, then we incur no overhead on calls to m, but we are trusting the type. If we do wrap m, then on every call to it RDL will check that we call it with a Fixnum and it actually returns a Fixnum. So if we're not completely sure of m's type, it might be useful to wrap it and then run test cases against it to see if the type annotation is every violated. (For example, the RDL developers did this to test out many of the core library annotations in RDL.)

Second, suppose there's a method m that we do want to type check, and again m has type (Fixnum) -> Fixnum. Now RDL will use type checking to ensure that if m is given a Fixnum then it returns Fixnum. But now we again have to decide whether to wrap m. If we don't wrap m, then we get the most efficiency, since typechecking (assuming we do not do it at calls) will only happen once and calls will incur no overhead. On the other hand, it could be that some non-typechecked code calls m with something that's not a Fixnum, in which case m might do anything, including report a type error. (Notice the type checking of m assumed its input was a Fixnum, and it doesn't say anything about the case when its argument is not.) To protect against this case, we can wrap m. Then if a caller violates m's type, we'll get an error in the caller code when it tries to call m.

(Side note: If typed methods are wrapped, then their type contracts are checked at run time for all callers, including ones that are were statically type checked and hence couldn't call methods at incorrect types. A future version of RDL will fix this, but it will require some significant changes to RDL's implementation strategy.)

Queries

As discussed above, RDL includes a small script, rdl_query, to look up type information. (Currently it does not support other pre- and postconditions.) The script takes a single argument, which should be a string. Note that when using the shell script, you may need to use quotes depending on your shell. Currently several queries are supported:

  • Instance methods can be looked up as Class#method.
$ rdl_query String#include?
String#include?: (String) -> TrueClass or FalseClass
  • Singleton (class) methods can be looked up as Class.method.
$ rdl_query Pathname.glob
Pathname.glob: (String p1, ?String p2) -> Array<Pathname>
  • All methods of a class can be listed by passing the class name Class.
$ rdl_query Array
&: (Array<u>) -> Array<t>
*: (String) -> String
... and a lot more
  • Methods can also be search for by their type signature:
$ rdl_query "(Fixnum) -> Fixnum"      # print all methods of type (Fixnum) -> Fixnum
BigDecimal.limit: (Fixnum) -> Fixnum
Dir#pos=: (Fixnum) -> Fixnum
... and a lot more

The type signature uses the standard RDL syntax, with two extensions: . can be used as a wildcard to match any type, and ... can be used to match any sequence of arguments.

$ rdl_query "(.) -> ."                 # methods that take one argument and return anything
$ rdl_query "(Fixnum, .) -> ."         # methods that take two arguments, the first of which is a Fixnum
$ rdl_query "(Fixnum, ...) -> ."       # methods whose first argument is a Fixnum
$ rdl_query "(..., Fixnum) -> ."       # methods whose last argument is a Fixnum
$ rdl_query "(..., Fixnum, ...) -> ."  # methods that take a Fixnum somewhere
$ rdl_query "(Fixnum or .) -> ."       # methods that take a single argument that is a union containing a Fixnum
$ rdl_query "(.?) -> ."                # methods that take one, optional argument

Note that aside from . and ..., the matching is exact. For example (Fixnum) -> Fixnum will not match a method of type (Fixnum or String) -> Fixnum.

Configuration

To configure RDL, execute the following shortly after RDL is loaded:

RDL.config { |config|
  # use config to configure RDL here
}

RDL supports the following configuration options:

  • config.nowrap - Array<Class> containing all classes whose methods should not be wrapped.
  • config.gather_stats - currently disabled.
  • config.report - if true, then when the program exits, RDL will print out a list of methods that were statically type checked, and methods that were annotated to be statically type checked but weren't.
  • config.guess_types - List of classes (of type Array<Symbol>). For every method added to a listed class after this configuration option is set, RDL will record the types of its arguments and returns at run-time. Then when the program exits, RDL will print out a skeleton for each class with types for the monitored methods based on what RDL recorded at run-time and based on what Ruby knows about the methods' signatures. This is probably not going to produce the correct method types, but it might be a good starting place.
  • config.type_defaults - Hash containing default options for type. Initially { wrap: true, typecheck: false }.
  • config.pre_defaults - Hash containing default options for pre. Initially { wrap: true }.
  • config.post_defaults - same as pre_defaults, but for post.

Bibliography

Here are some research papers we have written exploring types, contracts, and Ruby.

  • Brianna M. Ren and Jeffrey S. Foster. Just-in-Time Static Type Checking for Dynamic Languages. In ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI), Santa Barbara, CA, June 2016.

  • T. Stephen Strickland, Brianna Ren, and Jeffrey S. Foster. Contracts for Domain-Specific Languages in Ruby. In Dynamic Languages Symposium (DLS), Portland, OR, October 2014.

  • Brianna M. Ren, John Toman, T. Stephen Strickland, and Jeffrey S. Foster. The Ruby Type Checker. In Object-Oriented Program Languages and Systems (OOPS) Track at ACM Symposium on Applied Computing, pages 1565–1572, Coimbra, Portugal, March 2013.

  • Jong-hoon (David) An, Avik Chaudhuri, Jeffrey S. Foster, and Michael Hicks. Dynamic Inference of Static Types for Ruby. In ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages (POPL), pages 459–472, Austin, TX, USA, January 2011.

  • Jong-hoon (David) An. Dynamic Inference of Static Types for Ruby. MS thesis, University of Maryland, College Park, 2010.

  • Michael Furr. Combining Static and Dynamic Typing in Ruby. PhD thesis, University of Maryland, College Park, 2009.

  • Michael Furr, Jong-hoon (David) An, Jeffrey S. Foster, and Michael Hicks. The Ruby Intermediate Langauge. In Dynamic Languages Symposium (DLS), pages 89–98, Orlando, Florida, October 2009.

  • Michael Furr, Jong-hoon (David) An, and Jeffrey S. Foster. Profile-Guided Static Typing for Dynamic Scripting Languages. In ACM SIGPLAN International Conference on Object-Oriented Programming, Systems, Languages and Applications (OOPSLA), pages 283–300, Orlando, Floria, October 2009. Best student paper award.

  • Michael Furr, Jong-hoon (David) An, Jeffrey S. Foster, and Michael Hicks. Static Type Inference for Ruby. In Object-Oriented Program Languages and Systems (OOPS) Track at ACM Symposium on Applied Computing (SAC), pages 1859–1866, Honolulu, Hawaii, March 2009.

Copyright

Copyright (c) 2014-2016, University of Maryland, College Park. All rights reserved.

Authors

Contributors

Discussion group

RDL Users