Type systems are a very effective way to improve programming language reliability. By grouping the data manipulated by the program into classes called types, and ensuring that operations are never applied to types over which they are not defined (e.g. accessing an integer as if it were an array, or calling a string as if it were a function), a tremendous number of programming errors can be detected and avoided, ranging from the trivial (mis-spelled identifier) to the fairly subtle (violation of data structure invariants). These restrictions are also very effective at thwarting basic attacks on security vulnerabilities such as buffer overflows.
The enforcement of such typing restrictions is called type checking, and can be performed either dynamically (through run-time type tests) or statically (at compile-time, through static program analysis). We favour static type checking, as it catches bugs earlier and even in rarely-executed parts of the program, but note that not all type constraints can be checked statically if static type checking is to remain decidable (i.e. not degenerate into full program proof). Therefore, all typed languages combine static and dynamic type-checking in various proportions.
Static type checking amounts to an automatic proof of partial correctness of the programs that pass the compiler. The two key words here are partial, since only type safety guarantees are established, not full correctness; and automatic, since the proof is performed entirely by machine, without manual assistance from the programmer (beyond a few, easy type declarations in the source). Static type checking can therefore be viewed as the poor man's formal methods: the guarantees it gives are much weaker than full formal verification, but it is much more acceptable to the general population of programmers.
Type systems and language design.
Unlike most other uses of static program analysis, static type-checking rejects programs that it cannot analyze safe. Consequently, the type system is an integral part of the language design, as it determines which programs are acceptable and which are not. Modern typed languages go one step further: most of the language design is determined by the type structure (type algebra and typing rules) of the language and intended application area. This is apparent, for instance, in the XDuce and CDuce domain-specific languages for XML transformations whose design is driven by the idea of regular expression types that enforce DTDs at compile-time. For this reason, research on type systems -- their design, their proof of semantic correctness (type safety), the development and proof of associated type checking and inference algorithms -- plays a large and central role in the field of programming language research, as evidenced by the huge number of type systems papers in conferences such as Principles of Programming Languages.
Polymorphism in type systems.
There exists a fundamental tension in the field of type systems that
drives much of the research in this area. On the one hand, the desire
to catch as many programming errors as possible leads to type systems
that reject more programs, by enforcing fine distinctions between
related data structures (say, sorted arrays and general arrays). The
downside is that code reuse becomes harder: conceptually identical
operations must be implemented several times (say, copying a general array
and a sorted array). On the other hand, the desire to support code
reuse and to increase expressiveness leads to type
systems that accept more programs, by assigning a common type to
broadly similar objects (for instance, the
Object type of
all class instances in Java). The downside is a loss of precision in
static typing, requiring more dynamic type checks (downcasts in Java)
and catching fewer bugs at compile-time.
Polymorphic type systems offer a way out of this dilemma by combining precise, descriptive types (to catch more errors statically) with the ability to abstract over their differences in pieces of reusable, generic code that is concerned only with their commonalities. The paradigmatic example is parametric polymorphism, which is at the heart of all typed functional programming languages. Many forms of polymorphic typing have been studied since then. Taking examples from our group, the work of Rémy, Vouillon and Garrigue on row polymorphism, integrated in Objective Caml, extended the benefits of this approach (reusable code with no loss of typing precision) to object-oriented programming, extensible records and extensible variants. Another example is the work by Pottier on subtype polymorphism, using a constraint-based formulation of the type system.
Another crucial issue in type systems research is the issue of type inference: how many type annotations must be provided by the programmer, and how many can be inferred (reconstructed) automatically by the typechecker? Too many annotations make the language more verbose and bother the programmer with unnecessary details. Too little annotations make type checking undecidable, possibly requiring heuristics, which is unsatisfactory. Objective Caml requires explicit type information at data type declarations and at component interfaces, but infers all other types.
In order to be predictable, a type inference algorithm must be complete. That is, it must not find one, but all ways of filling in the missing type annotations to form an explicitly typed program. This task is made easier when all possible solutions to a type inference problem are instances of a single, principal solution.
Maybe surprisingly, the strong requirements -- such as the existence of principal types -- that are imposed on type systems by the desire to perform type inference sometimes lead to better designs. An illustration of this is row variables. The development of row variables was prompted by type inference for operations on records. Indeed, previous approaches were based on subtyping and did not easily support type inference. Row variables have proved simpler than structural subtyping and more adequate for typechecking record update, record extension, and objects.
Type inference encourages abstraction and code reuse. A programmer's understanding of his own program is often initially limited to a particular context, where types are more specific than strictly required. Type inference can reveal the additional generality, which allows making the code more abstract and thus more reuseable.
B.C.Pierce. Types and
Programming Languages, MIT Press.
F.Pottier and D.Rémy. The Essence of ML Type Inference, in Advanced Topics in Types and Programming Languages, B.C.Pierce (ed), MIT Press.
F.Pottier. Simplifying subtyping constraints: a theory, 2001.
D.Rémy and J.Vouillon. Objective ML: An effective object-oriented extension to ML, 1998.
D.Le Botlan and D.Rémy. MLF: Raising ML to the power of System F, 2003.