Browse Theses

The Smalltalk-80('TM) system makes it possible to write programs quickly by providing object-oriented programming, incremental compilation, run-time type checking, user-extensible data types and control structures, and an interactive graphical interface. However, the potential savings in programming effort have been curtailed by poor performance in widely available computers or high processor cost. Smalltalk-80 systems pose tough challenges for implementors: dynamic data typing, a high-level instruction set, frequent and expensive procedure calls, and object-oriented storage management.

To solve these problems, a group of researchers at U.C. Berkeley has designed and built the SOAR (Smalltalk On A RISC) microprocessor. In order to determine the performance of Smalltalk-80 on SOAR and to evaluate the importance of each of the ideas, simulations of five representative benchmarks have been analyzed. The results suggest that: (1) Six ideas substantially improve performance: compilation to a low-level instruction set, multiple windows of on-chip registers, caching the target of a call instruction in the instruction itself, byte insert and extract instructions, instructions for arithmetic and comparison operations on tagged integers, and our storage management algorithm, Generation Scavenging. (2) Seven features contribute little to performance: shadow registers to simplify trap recovery, hardware assistance for garbage collection, vectored traps, addressable registers, clearing multiple registers in parallel, conditional trap instructions, and load- and store-multiple instructions. (3) The language-specific hardware in SOAR doubles its performance over a RISC II with the same cycle time. (4) Generation Scavenging, a storage reclamation algorithm developed by the author, consumes only 3% of the CPU time, in contrast to the 9% of comparable Smalltalk-80 systems. (5) Despite a five-to-one handicap in basic cycle time, the NMOS SOAR microprocessor should run as fast as an ECL Dorado minicomputer.

The dissertation reports two results that run counter to conventional wisdom: that a reduced instruction set computer can offer excellent performance for a system with dynamic data typing such as Smalltalk-80, and that automatic storage reclamation need not be time-consuming.