Typed Compilation Against Non-manifest Base Classes

Abstract

Abstract. Much recent work on proof-carrying code aims to build certifying compilers for single-inheritance object-oriented languages, such as Java or C#. Some advanced object-oriented languages support compiling a derived class without complete information about its base class. This strategy-though necessary for supporting features such as mixins, traits, and first-class classes-is not wellsupported by existing typed intermediate languages. We present a low-level IL with a type system based on the Calculus of Inductive Constructions. It is an appropriate target for efficient, type-preserving compilation of various forms of inheritance, even when the base class is unknown at compile time. Languages (such as Java) that do not require such flexibility are not penalized for it at run time. MotivationIn most object-oriented languages, programmers factor their solutions over a hierarchy of classes. Since the classes in a hierarchy may appear in different compilation units, one question that the language designer (or implementer) must address is: how much information about a base class is needed to compile its derived class?With its emphasis on efficient object layout and method dispatch, C++ [32] requires complete information about the base class: the number, locations, and types of all its fields and methods. Indeed, it is because C++ depends on this information that a seemingly minor change to a base class triggers recompilation of all its descendents. Java [23] is somewhat more flexible. To support binary compatibility, its class files are not committed to a particular object layout. A derived class depends only on the names and types of the base class fields and methods that it uses. Nevertheless, most Java implementations ultimately compile classes to lower-level code using the same layouts and techniques as C++.A few modern object-oriented languages allow classes as module parameters (Moby [15], OCaml [28]) or as first-class values (Loom [5]). Other languages support more flexible forms of inheritance, such as mixins [24,3] and traits [29]. If a base class is not available for inspection when a derived class is compiled, we say the base class is not manifest. Implementations of these languages use a dictionary data structure to map method and field names to their locations in the object layout. The dictionary may be applied at link time or at run time, as required by the language.Here is a simple example in OCaml (although it could be expressed just as easily in Moby). We declare a signature for modules containing a circle class that implements Below, CircleBBox declares a class bbox that inherits from a (non-manifest) base class circle, overrides the area method (using a super call), and defines a new method bounds.module CircleBBox = functor (C : CIRCLE) -> struct class bbox arg = object (self) inherit C.circle arg as super method area = super#area * 4.0 / pi (* area of bbox *) method bounds = let (x,y) = self#center in let r = self#radius in ((x-r,y-r), (x+r,y+r)) end endTo compile this functor, we must mak...

The swiss coercion

“…Of course, all the work surrounding Shao et al's type system for certified binaries , League and Monnier, 2006, Monnier, 2004 was the basis of this work as well.…”

“…This kind of construction can occur for example when building dictionaries in the implementation of object-oriented features [League et al, 2002, League andMonnier, 2006]. But this is no match to our coercion since we can now first construct the dictionary without instantiating each function and then use the Swiss coercion to introduce the outer type abstraction and to instantiate the inner polymorphic elements.…”

“…This amounts to lifting values to the level of types and hence provides the limited form of dependency necessary to enable powerful reasoning: e.g. [League and Monnier, 2006] uses a variant of λH to express and verify the invariants used in an encoding of sophisticated object oriented features such as non-manifest base classes.…”

Schema-Based Compression of XML Data with Relax NG

“…It is clear from the previous section that rngzip is most effective on highly tagged, nested data. The coq benchmark in particular is encouraging, because we aim to use this technology for producing compact representations of sophisticated intermediate languages [17] including ones based, like Coq, on the Calculus of Constructions [35].…”

Singleton types here, singleton types there, singleton types everywhere

“…λH was designed as an internal language for compilers that want to generate certified binaries, and has been shown to be able to handle these needs very well: the original article [Shao et al, 2002] shows how to perform the usual CPS and closure conversion while preserving types; then League and Monnier [2006] shows essentially that λH can be seen as a type system for λinkς [Fisher et al, 2000], and is hence a good target language for compilation of most object-oriented features; and Monnier [2004] showed how a variant of λH extended with regions can be used to write a type-preserving generational garbage collector. Still, λH only provides singleton types as a poor man's substitute to dependent types which, while apparently sufficient for those particular cases, might lead one to believe that it is not up to the task of handling truly dependent types.…”