Specifying and Executing Optimizations for Parallel Programs

Abstract

Compiler optimizations, usually expressed as rewrites on program graphs, are a core part of all modern compilers. However, even production compilers have bugs, and these bugs are difficult to detect and resolve. The problem only becomes more complex when compiling parallel programs; from the choice of graph representation to the possibility of race conditions, optimization designers have a range of factors to consider that do not appear when dealing with single-threaded programs. In this paper we present PTRANS, a domain-specific language for formal specification of compiler transformations, and describe its executable semantics. The fundamental approach of PTRANS is to describe program transformations as rewrites on control flow graphs with temporal logic side conditions. The syntax of PTRANS allows cleaner, more comprehensible specification of program optimizations; its executable semantics allows these specifications to act as prototypes for the optimizations themselves, so that candidate optimizations can be tested and refined before going on to include them in a compiler. We demonstrate the use of PTRANS to state, test, and refine the specification of a redundant store elimination optimization on parallel programs.Comment: In Proceedings GRAPHITE 2014, arXiv:1407.767

Specifying and executing optimizations for generalized control flow graphs

“…Our work builds on the TRANS approach of expressing optimizations as rewrites on control flow graphs with temporal logic side conditions due to Lacey et al [2] and Kalvala et al [3], and the extension of that work to PTRANS due to Mansky and Gunter [12,13,14]. We differ from these in the use of a basic action language that respects the structure of control flow graphs, and in the use of a strategy language with explicit scoping of variables and with the structure of a Kleene algebra.…”