Date of Award
Doctor of Philosophy (PhD)
J. Eliot B. Moss
Charles C. Weems
Architecture, Code Generation, Compilers, Instruction Selection, Intermediate Representation, Programming Languages
One of the most difficult tasks a compiler writer faces is the construction of the instruction selector. The instruction selector is that part of the compiler that translates compiler intermediate representation (IR) into instructions for a target machine. Unfortunately, implementing an instruction selector “by hand” is a difficult, time consuming, and error prone task. The details of both the IR and target instruction set must be carefully considered in order to generate correct and efficient code. This, in turn, requires an expert in both the compiler internals as well as the target machine. Even an expert, however, can implement an instruction selector that is difficult to verify and debug. In this dissertation we describe the instruction selector problem, cover previous attempts at solving it, and identify what we believe to be the most prominent factor inhibiting their widespread adoption. This dissertation proposes a generalized approach toward generating instruction selectors automatically. In particular, we propose CISL, a common machine description language for specifying the semantics of compiler IR and target instructions, and GIST, a machine independent heuristic search procedure that can find equivalent instruction sequences between compiler IR and target instructions. CISL is an object-oriented-based language leveraging modern programming language constructs (e.g., classes, inheritance, mixins) and is capable of describing instructions for a variety of IR and target ISAs (Instruction Set Architecture). GIST leverages CISLs well-defined semantics and a canonicalization process to discover automatically instruction selector patterns: target instruction sequences that implement IR semantics. These instruction selector patterns are then generated in a compiler implementation independent format (XML). Small adapter programs use the generated instruction selector patterns to generate compiler specific implementation code. Our experiments show that instruction selector patterns can be discovered automatically and independent of a particular compiler framework or target machine. In addition, experience proved that adapter programs are easy to implement and instruction selector code is easy to generate from generated patterns. Furthermore, the generated instruction selectors are comparable in performance to the original compilers.
Richards, Timothy David, "Generalized Instruction Selector Generation: The Automatic Construction of Instruction Selectors from Descriptions of Compiler Internal Forms and Target Machines" (2010). Open Access Dissertations. Paper 178.