Browse Theses

Object-oriented programming languages promise improved programmer productivity through abstract data types, inheritance, and message passing. Unfortunately, traditional implementations of object-oriented language features are much slower than traditional implementations of their non-object-oriented counterparts. This dissertation describes a collection of implementation techniques that can improve the run-time performance of object-oriented languages. The new techniques identify those messages whose receiver can only be of a single representation and eliminate the overhead of message passing by replacing the message with a normal direct procedure call; these direct procedure calls are then amenable to traditional inline-expansion. Type analysis extracts representation-level type information about the receivers of messages. Customization transforms a single source method into several compiled versions, each version specific to a particular inheriting receiver type; customization allows all messages to self to be inlined away (or at least replaced with direct procedure calls). To avoid generating too much compiled code, the compiler is invoked at run-time, generating customized versions only for those method/receiver type pairs used by a particular program. Splitting transforms a single path through a source method into multiple separate fragments of compiled code, each fragment specific to a particular combination of run-time types. Messages to expressions of these discriminated types can then be optimized away in the split versions. These techniques have been implemented as part of the compiler for the SELF language, a purely object-oriented language designed as a refinement of Smalltalk-80. If only pre-existing implementation technology were used for SELF, programs in SELF would run one to two orders of magnitude slower than their counterparts written in a traditional non-object-oriented language. However, by applying the techniques described in this dissertation, the performance of the SELF system is five times better than the fastest Smalltalk-80 system, better than that of an optimizing Scheme implementation, and close to half that of an optimizing C implementation. These techniques could be applied to other object-oriented languages to boost their performance or enable a more object-oriented programming style.