Applications of a graph grammar for program control flow analysis

• Authors: Ken Kennedy;Linda Zucconi
• Venue: POPL '77 Proceedings of the 4th ACM SIGACT-SIGPLAN symposium on Principles of programming languages
• Year: 1977

A standard approach to the analysis of program structure for the purpose of code optimization is to construct the "control flow graph" which models the possible execution paths through the program. Various graph algorithms can be applied to the control flow graph to produce data flow information, possible optimizations, etc. [A1,A2,AC,AU2,AU3,CS,HU1,HU2.HU3,Ke1,Ke2,Ke3,Ke4,Sc,U]. Studies of the form of typical control flow graphs indicate that such graphs tend to fall into a restricted subclass of general graphs. For example, empirical investigations have shown that the vast majority of program graphs have no multiple-entry loops [AC,HU2,HU3,Kn1].The recent work on "structured programming" has suggested that "good" programs fall into an even more restricted subclass. In fact, purists recommend that all programs be synthesized from three basic control structures: sequential statements, if-then-else statements, and single-entry single-exit loops [Di,Wi].Formal language theory [HoU] has given us a practical way to specify the set of strings which comprise a given language: via a grammar. It is then a natural idea to extend grammars from the strings to graphs in hopes of getting the same power of expression. Several researchers have used this approach [FKZ,J2,Ro].In this paper we study the applicability of a grammatical approach to describing the set of control flow graphs which arise from "good" programs in the sense proposed by many programming practitioners. The resulting flow graph language contains all those programs constructed according to the purists' rules and also admits programs with multiple-exit loops if such loops are constructed sensibly. The grammar we use is the "semi-structured flow graph" grammar GSSFG which was studied originally in [FKZ]. There are several appealing properties of this grammar; perhaps the most important, from the point-of-view of a compiler-writer, is the existence of a linear-time parsing algorithm which leads directly to a linear-time data flow analysis method [FKZ].In the present work we summarize the results from [FKZ] and address several new questions. First, how often do programs written by people with no knowledge of the SSFG rules fall into the language defined by GSSFG? In other words, is the language a natural one for programming? Second, once a program has been parsed according to GSSFG do benefits other than fast data flow analysis accrue?The paper is organized into three main sections. Section II introduces GSSFG and the parsing algorithm from [FKZ]. Section III is devoted to an empirical study conducted by the authors in an attempt to answer the question of naturalness, described above. Section IV discusses several applications of the graph parse in a "graph attribute grammar" framework. The summary at the end of the paper includes suggestions for further investigation.