[][src]Crate voile

Voile is a dependently-typed programming language evolved from minitt.



The focus of Voile is extensible algebraic data types on top of dependent types. It can solve the expression problem without using any design patterns (like visitor or object-algebra in Java, or finally-tagless or DTALC in Haskell).

We're pretty much inspired by the coexistence of guarded recursion, coinductive data types, sized types and inductive types in Agda, which is nice to have all of them but they do not work very well as we can see in a discussion here about guarded recursion checker or in a GitHub issue about the incompatibility between size and guarded recursion. We can observe that sums, records, and (dependent) pattern matching in Agda only work well when being top-level bindings.


  • First-class language components
    • Sum/Record
  • Dependent type goodies
    • Pi/Sigma

Extensible ADTs

Voile supports sum-types (coproduct), record-types (product) and their instances as first-class language components.

First-class record support is related to Record Calculus, or Extensible Records, or "first-class labels". The idea of "extensible record" is that the creation, manipulation and destruction of "records" can be done locally as an expression, without the need of declaring a record type globally, like:

$$ \newcommand{\Bool}[0]{(\texttt{True}\mid\texttt{False})} \begin{alignedat}{2} &\texttt{not}&&:\Bool \rarr \Bool \\ \space &\texttt{ifThenElse}&&:\forall A. \Bool \rarr A \rarr A \rarr A \end{alignedat} $$

Universal quantification should also support generalizing over a part of the records (in other words, row-polymorphism), like:

$$ \newcommand{\xx}[0]{\texttt{X}} \newcommand{\T}[0]{\lBrace\xx:A, ...=r\rBrace} \begin{alignedat}{2} &\texttt{getX}&&:\forall A. \T \rarr A \\ \space &\texttt{getX}&&=\lambda s. (s.\xx) \\ \space &\texttt{setX}&&:\forall A. \T \rarr A \rarr \T \\ \space &\texttt{setX}&&= [\textnormal{syntax undecided yet}] \end{alignedat} $$

Existing row-polymorphism implementation divides into two groups according to how they support such generalization, either by making "record type"/"record value" first-class expressions, or by introducing a standalone "row" (the type of $r$ above) type.

The study on extensible records has a long history, but there isn't much research and implementations on extensible sums and the combination of bidirectional type-checking and row-polymorphism yet.
Currently, I can only find one programming language, MLPolyR (there's a language spec, a paper first-class cases, a PhD thesis type-safe extensible programming and an IntelliJ plugin), whose pattern matching is first-class (in the papers it's called "first-class cases").

First-class pattern matching is useful because it solves the expression for free, which means that library authors using such languages can split their library features into several sub-libraries -- one core library with many extensions. Library users can combine the core with extensions they want like a Jigsaw and exclude everything else (to avoid unwanted dependencies) or create their own extensions without touching the original codebase.

Exceptions (undecided yet)

MLPolyR exploits first-class sums in error-handling -- exceptions are treated as first-class sums while try-catch clauses are pattern matching on them ("consumes" a variant in a sum). Putting exceptions into function signatures looks like Java's checked exceptions, but with full type-inference. In this way, we can have cross-control-flow exceptions (instead of monadic) safely, because it's easy to ensure that an expression is exception-free simply by looking at its type.

We can denote function from $A$ to $B$ that may throw exception $E$ like this:

$$ A \xrightarrow{E} B $$

A higher-order function taking an exception-throwing function and handles one exception will have signature like this (convert langauge-level exceptions to monadic exceptions):

$$ (A \xrightarrow{E \mid r} B) \rarr (E \rarr B) \rarr (A \xrightarrow{r} B) $$

However, without the help of dependent types, there can be false positives such as conditionally-thrown exceptions -- consider a function f like this (assume $e : E$):

fun f a = if a then throw e else 0

Invocation f false is not going to raise any exception, but the compiler disagrees because of the type of f inferred (something like $\mathbb{B} \xrightarrow{E} \mathbb{Z}$) is irrelevant to the argument applied. In the industry of dependent types, there isn't much development on exceptions. We might have a try here.

Induction and Coinduction

According to elementary-school discrete math, induction has something to do with recursion. However, recursion on sum types is a huge problem against the design of first-class sum types.

In the cliché programming languages with non-first-class sums (where the sum types need to be declared globally before usage), type-checking against recursive sums can be easily supported because the types are known to the type-checker -- there's no need of reduction on an already-resolved sum type. Once a term is known to be some sum type, it becomes a canonical value.

For Voile, arbitrary expressions can appear inside of a sum type term, which means reduction is still needed before checking some other terms against this sum-type-term. If there's recursion on a sum-type-term, like the natural number definition:

$$ \mathbb{N}=\texttt{Zero} \mid \texttt{Suc}\ \mathbb{N} $$

The type-checker will infinitely loop on its reduction.

The above definition will actually be rejected by the termination checker, but recursion on sum-types have to be supported because we have been using it for a long time -- we shouldn't sacrifice this fundamental language feature.

We choose to annotate types with the so-called sized types (present in MiniAgda) to inform the compiler about how data size are changed in a function. This will also help recursive type definitions, because it acts as an argument to the recursive call in the type definition!

$$ \mathbb{N}^{(\texttt{s}\ i)}=\texttt{Zero} \mid \texttt{Suc}\ \mathbb{N}^i $$

This function now perfectly terminates. We may also write it in a more inductive way:

$$ \begin{alignedat}{2} &\mathbb{N}^0&&=\texttt{Zero}\\ \space &\mathbb{N}^{(\texttt{s}\ i)}&&=\texttt{Suc}\ \mathbb{N}^i \end{alignedat} $$

Depends on how sized types are designed in Voile.


Voile's implementation is inspired from Agda, mlang, MiniAgda and its prototype, minitt.

MiniAgda supports induction, coinduction with sized types.

TODO Something needs to be written here.

About the name, Voile This name is inspired from a friend whose username is Voile (or Voileexperiments). However, this is also the name of a library in the Scarlet Devil Mansion.
The librarian of Voile, the Magic Library



Type-Checking module.


Abstract syntax, surface syntax, parser and well-typed terms (core language).