Eclair is a statically typed language, meaning that all types are known at compilation time. As part of the compilation process, Eclair checks the types of all values in a program to ensure it is valid. The rest of this page describes how the typesystem works.
Types in a program
At the moment, Eclair supports two primitive types:
string (UTF-8 encoded
u32 (unsigned 32-bit integers). Note that there is no support
for signed 32-bit integers yet, but this will be added in the future.
Literals and operators
Operators are built in to the language and have pre-defined types. Arithmetic
operators and comparisons are only supported on integers. Both left- and
right-hand-side arguments need to be an integer type, and return an
integer-typed value as well. Equality operators (
!=) are supported
on all primitive types.
Variables can be used in rule clauses as arguments to a relation, in a comparison or equality constraint.
The type of a variable is based on how it is used inside a rule. When a variable is used in a rule for the first time, Eclair will assign a type to the variable. If the same variable is later used again in the same rule, then the previously assigned type is checked against the current expected type.
When variables are used in a relation or externally defined function or constraint, they are checked against the corresponding argument in the definition.
When a variable is used in a comparison or equality operator, Eclair tries to infer the type of the variable based on either the operator itself, or on the other argument used in the operator. If Eclair fails to determine the types of one or more variables in a rule, an error is thrown and compilation is aborted.
Wildcards are treated the same as variables, but with the main difference that each wildcard is treated as a unique variable. (Two wildcards can have different types, unlike variables of the same name.)
Typed holes are placeholders for a part of the program that is not complete yet. They can be very useful when you are writing your program and still figuring out part of logical rule, or when debugging if you want to know more information about the types of values in a function.
For example, given the function below:
@def edge(u32, u32) input. @def reachable(u32, u32) output. reachable(x, z) :- edge(x, ?), // <-- typed hole here reachable(y, z).
When you try to compile this Eclair program (try it out!), the compiler will give you an error message listing all possible candidate variables.
The typesystem treats each typed hole as a unique variable. Depending how it is used (in a rule or in an equality clause), the compiler will try to figure out the best matching type for the hole. After this, it collects all information about the rest of the types in a fact or rule and presents the information in an error message.
Relations (both facts and rules) contain one or more arguments. Each of these arguments needs to have the type declared up front. The typesystem uses this information to later check the type of a value against this expected type.
Relations do not have any return type and so cannot be used in e.g. arithmetic expressions or equality constraints.
As an example, in the following snippet “relation” is a relation that consists
@def relation(string, u32).
The main difference is that externally defined functions (those defined with a return type), can be used in arithmetic expressions and equality constraints.
The following is a description in plain English how the typechecker checks an Eclair program:
- The typechecker first collects all type-level information about relations and external definitions.
- Afterwards, it checks the types of each top level declaration (facts and rules) independently.
- Values in a top-level fact are checked from left to right.
- Values in a top-level rule are checked from left to right, and from top to bottom.
- All relations or externally defined functions need to be defined.
- The number of arguments used in a fact or rule is checked against the definition. An error is reported if it does not match the expected number of arguments.
- If type-level information is available (e.g. when using a variable as an argument to a relation), then the value will be checked against that type. Otherwise the type will be inferred. Type inference of values is scoped to a single rule.
- A user-defined function can’t be used in a location where a relation or user-defined constraint is expected, and vice versa.
- Each primitive type is treated as a distinct type. You will get a type-mismatch error if you use a string value where an integer is expected and vice versa.
- The type of literals is inferred.
- Arguments passed to operators are checked against their pre-defined types. Arguments passed to equality constraints are inferred to see if both arguments have the same types. A type-mismatch or unification error will be raised if an inconsistency is detected.
- A typed hole will always raise a type error to show the type information the compiler deduced for that rule.
If no errors are found while checking all these rules, the program is considered to be valid with respect to the types.
For now the typesystem is still relatively simple, but it will be extended to support polymorphic relations (for code reuse), support for other primitive types, user-defined types, … Stay tuned! 😊