We have implemented the algorithmic system $\vdashA$. Our
implementation is written in OCaml and uses CDuce as a library to
provide the semantic subtyping machinery. Besides the type-checking
algorithm defined on the base language, our implementation supports
record types of Section \ref{ssec:struct}) and the refinement of
function types
\iflongversion%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
described in Section \ref{sec:refining}. Furthermore, our implementation uses for the inference of arrow types
the following improved rule:
\input{optimize}
\else
of Section \ref{sec:refining} with the rule of
Appendix~\ref{app:optimize}.
\fi%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
The implementation is rather crude and
consists of 2000 lines of OCaml code, including parsing, type-checking
of programs, and pretty printing of types. We demonstrate the output
of our type-checking implementation in Table~\ref{tab:implem} by
listing some examples none of which can be typed by current
systems (even though some systems such as Flow and TypeScript
can type some of these examples by adding explicit type annotations, the code 6,
7, 9, and 10 in Table~\ref{tab:implem} and, even more, the \code{and\_} and \code{xor\_} functions at the end of this
section are out of reach of current systems, even when using the right
explicit annotations). It should be noted that for all the examples we discuss,
the the time for the type inference process is less than 5ms, hence we do not
report precise timings in the table.
These and other examples can be tested in the
\ifsubmission
anonymized
\else
\fi
online toplevel available at
\url{https://occtyping.github.io/}%
\input{code_table}
\input{code_table2}
In Table~\ref{tab:implem}, the second column gives a code fragment and the third
column the type deduced by our implementation. Code~1 is a
straightforward function similar to our introductory example \code{foo} in (\ref{foo},\ref{foo2}). Here the
programmer annotates the parameter of the function with a coarse type
$\Int\vee\Bool$. Our implementation first type-checks the body of the
function under this assumption, but doing so it collects that the type of
$\texttt{x}$ is specialized to \Int{} in the ``then'' case and to \Bool{}
in the ``else'' case. The function is thus type-checked twice more
under each hypothesis for \texttt{x}, yielding the precise type
$(\Int\to\Int)\land(\Bool\to\Bool)$. Note that w.r.t.\ rule \Rule{AbsInf+} of Section~\ref{sec:refining}, our implementation improved the output of the computed
type. Indeed, using rule~[{\sc AbsInf}+] we would obtain the
type
$(\Int\to\Int)\land(\Bool\to\Bool)\land(\Bool\vee\Int\to\Bool\vee\Int)$
with a redundant arrow. Here we can see that, since we deduced
the first two arrows $(\Int\to\Int)\land(\Bool\to\Bool)$, and since
the union of their domain exactly covers the domain of the third arrow, then
the latter is not needed. Code~2 shows what happens when the argument
of the function is left unannotated (i.e., it is annotated by the top
type \Any, written ``\texttt{Any}'' in our implementation). Here
type-checking and refinement also work as expected, but the function
only type checks if all cases for \texttt{x} are covered (which means
that the function must handle the case of inputs that are neither in \Int{}
nor in \Bool).
The following examples paint a more interesting picture. First
(Code~3) it is
easy in our formalism to program type predicates such as those
hard-coded in the $\lambda_{\textit{TR}}$ language of \citet{THF10}. Such type
predicates, which return \texttt{true} if and only if their input has
a particular type, are just plain functions with an intersection
type inferred by the system of Section~\ref{sec:refining}. We next define Boolean connectives as overloaded
functions. The \texttt{not\_} connective (Code~4) just tests whether its
argument is the Boolean \texttt{true} by testing that it belongs to
the singleton type \True{} (the type whose only value is
\texttt{true}) returning \texttt{false} for it and \texttt{true} for
any other value (recall that $\neg\True$ is equivalent to
$\texttt{Any\textbackslash}\True$). It works on values of any type,
but we could restrict it to Boolean values by simply annotating the
parameter by \Bool{} (which in CDuce is syntactic sugar for
\True$\vee$\False) yielding the type
$(\True{\to}\False)\wedge(\False{\to}\True)$.
The \texttt{or\_} connective (Code~5) is straightforward as far as the
code goes, but we see that the overloaded type precisely captures all
possible cases. Again we use a generalized version of the
\texttt{or\_} connective that accepts and treats any value that is not
\texttt{true} as \texttt{false} and again, we could easily restrict the
domain to \Bool{} if desired.\\
\indent
To showcase the power of our type system, and in particular of
the ``$\worra{}{}$''
type operator, we define \texttt{and\_} (Code~6) using De Morgan's
Laws instead of
using a direct definition. Here the application of the outermost \texttt{not\_} operator is checked against type \True. This
allows the system to deduce that the whole \texttt{or\_} application
has type \False, which in turn leads to \texttt{not\_\;x} and
\texttt{not\_\;y} to have type $\lnot \True$ and therefore both \texttt{x}
and \texttt{y} to have type \True. The whole function is typed with
the most precise type (we present the type as printed by our
implementation, but the first arrow of the resulting type is
equivalent to
$(\True\to\lnot\True\to\False)\land(\True\to\True\to\True)$).
All these type predicates and Boolean connectives can be used together
to write complex type tests, as in Code~7. Here we define a function
\texttt{f} that takes two arguments \texttt{x} and \texttt{y}. If
\texttt{x} is an integer and \texttt{y} a Boolean, then it returns the
integer \texttt{1}; if \texttt{x} is a character or
\texttt{y} is an integer, then it returns \texttt{2}; otherwise the
function returns \texttt{3}. Our system correctly deduces a (complex)
intersection type that covers all cases (plus several redundant arrow
types). That this type is as precise as possible can be shown by the fact that
when applying
\texttt{f} to arguments of the expected type, the \emph{type} deduced for the
whole expression is the singleton type \texttt{1}, or \texttt{2},
or \texttt{3}, depending on the type of the arguments.
Code~8 allows us to demonstrate the use and typing of record paths. We
model, using open records, the type of DOM objects that represent XML
or HTML documents. Such objects possess a common field
\texttt{nodeType} containing an integer constant denoting the kind of
the node (e.g., \p{1} for an element node, \p{3} for a text node, \ldots). Depending on the kind, the object will have
different fields and methods. It is common practice to perform a test
on the value of the \texttt{nodeType} field. In dynamic languages such
as JavaScript, the relevant field can directly be accessed
after having checked for the appropriate \texttt{nodeType}, whereas
in statically typed languages such as Java, a downward cast
from the generic \texttt{Node} type to the expected precise type of
the object is needed. We can see that using the extension presented in
Section~\ref{ssec:struct} we can deduce the correct type for
\texttt{x} in all cases. Of particular interest is the last case,
since we use a type case to check the emptiness of the list of child
nodes. This splits, at the type level, the case for the \Keyw{Element}
type depending on whether the content of the \texttt{childNodes} field
is the empty list or not.
Code~9 shows the usefulness of the rule \Rule{OverApp}.
Consider the definition of the \texttt{xor\_} operator.
Here the rule~[{\sc AbsInf}+] is not sufficient to precisely type the
function, and using only this rule would yield a type
$\Any\to\Any\to\Bool$.
\iflongversion
Let us follow the behavior of the
``$\worra{}{}$'' operator. Here the whole \texttt{and\_} is requested
to have type \True, which implies that \texttt{or\_ x y} must have
type \True. This can always happen, whether \texttt{x} is \True{} or
not (but then depends on the type of \texttt{y}). The ``$\worra{}{}$''
operator correctly computes that the type for \texttt{x} in the
``\texttt{then}'' branch is $\True\vee\lnot\True\lor\True\simeq\Any$,
and a similar reasoning holds for \texttt{y}.
\fi%%%%%%%%%%%%%%
However, since \texttt{or\_} has type
%\\[.7mm]\centerline{%
$(\True\to\Any\to\True)\land(\Any\to\True\to\True)\land
(\lnot\True\to\lnot\True\to\False)$
%}\\[.7mm]
then the rule \Rule{OverApp} applies and \True, \Any, and $\lnot\True$ become candidate types for
\texttt{x}, which allows us to deduce the precise type given in the table. Finally, thanks to rule \Rule{OverApp} it is not necessary to use a type case to force refinement. As a consequence, we can define the functions \texttt{and\_} and \texttt{xor\_} more naturally as:
\begin{alltt}\color{darkblue}\morecompact
let and_ = fun (x : Any) -> fun (y : Any) -> not_ (or_ (not_ x) (not_ y))
let xor_ = fun (x : Any) -> fun (y : Any) -> and_ (or_ x y) (not_ (and_ x y))
\end{alltt}
for which the very same types as in Table~\ref{tab:implem} are deduced.
Last but not least Code~10 (corresponding to our introductory
example~\eqref{nest1}) illustrates the need for iterative refinement of
type environments, as defined in Section~\ref{sec:typenv}. As
explained, a single pass analysis would deduce
for {\tt x}
a type \Int{} from the {\tt f\;x} application and \Any{} from the {\tt g\;x}
application. Here by iterating a second time, the algorithm deduces
that {\tt x} has type $\Empty$ (i.e., $\textsf{Empty}$), that is that the first branch can never
be selected (and our implementation warns the user accordingly). In hindsight, the only way for a well-typed overloaded function to have
type $(\Int{\to}\Int)\land(\Any{\to}\Bool)$ is to diverge when the
argument is of type \Int: since this intersection type states that
whenever the input is \Int, {\em both\/} branches can be selected,
yielding a result that is at the same time an integer and a Boolean.
This is precisely reflected by the case $\Int\to\Empty$ in the result.
Indeed our {\tt example10} function can be applied to an integer, but
at runtime the application of {\tt f ~x} will diverge.
Although these experiments are still preliminary, they show how the
combination occurrence typing and set-theoretic types, together
with the type inference for overloaded function types presented in
Section~\ref{sec:refining} goes beyond what languages like
TypeScript and Flow do, since they can only infer single arrow types.
Our refining of overloaded
functions is also future-proof and resilient to extensions: since it ``retypes'' functions
using information gathered by the typing of occurrences in the body,
its precision will improve with any improvement of
our occurrence typing framework.