Exploiting specifications to improve program performance

Many approaches to programming emphasize the use of interfaces. The basic idea is to decompose programs into modules and specifying how each module's interface behaves. This makes it easier to reason about programs because one can rely on a module's specification rather than examining its implementation, which is more complicated. Although programmers benefit from interface specifications when reasoning about programs, existing compilers do not. In this thesis, I discuss how to incorporate specifications into a programming language to improve performance. I use specifications in two ways: (1) to allow programmers to define new optimizations that make general interfaces more efficient to use, and (2) to enhance conventional optimizations. The specifications can be written incrementally, so programmers can choose to write only the parts of specifications needed to improve performance. I demonstrate my approach using Speckle, a statically typed, imperative programming language that incorporates specifications. Users define optimizations in Speckle by providing multiple implementations for a single procedure. One implementation must be general enough to work in any context. The other implementations are more efficient but require an additional precondition specified by the user. The compiler uses specifications to prove that particular calls to the procedure can use the specialized implementations. The prototype Speckle compiler (PSC) incorporates primitive automated theorem-proving technology to optimize programs. In addition to user-defined optimizations, PSC identifies opportunities to perform three kinds of conventional optimizations: eliminating common subexpressions, moving code out of loops, and eliminating dead code. Because specifications are simpler than code, PSC detects optimizations that most compilers cannot, such as hoisting a procedure call out of a loop. Also, because specifications contain information not found in code, PSC detects optimizations that are impossible without specifications. (Copies available exclusively from MIT Libraries, Rm. 14-0551, Cambridge, MA 02139-4307. Ph. 617-253-5668; Fax 617-253-1690.)