Section: New Results

SSI Revisited: A Program Representation for Sparse Dataflow Analyses

Participants : Andre Tavares [UFMG, Brazil] , Mariza Bigonha [UFMG, Brazil] , Roberto Bigonha [UFMG, Brazil] , Benoit Boissinot, Fernando Pereira [UFMG, Brazil] , Fabrice Rastello.

Dataflow analyses usually associate information about variables to program regions. Informally, if these regions are too small, e.g., a point between two consecutive statements, we call the analysis dense. On the other hand, if these regions include many such points, then we call it sparse. We developed a systematic method to build program representations that support forward and/or backward sparse analyses. To pave the way that leads to this framework, we first clarified the literature on intermediate program representations. We revisited the static single information (SSI) form introduced in the 90s and showed how to simplify the construction of program representations for unidirectional dataflow analyses. We showed how to cope with live-ranges that have multiple uses/definitions without losing the equivalence property with the initial dataflow analysis problem. This allows us to simplify, for unidirectional problems, the SSI construction algorithm.

We also showed that our approach, up to a parameter choice, subsumes other program representations such as the SSA, SSI, and e-SSA forms. We can produce intermediate representations isomorphic to the sparse evaluation graphs (SEGs) of Choi et al. This data structure enables sparse solutions to the class of dataflow problems called partitioned dataflow analysis (PDA). However, contrary to SEGs, we can handle - sparsely - problems that are not PDA. We have implemented this framework in the LLVM compiler and have empirically compared different program representations in terms of size and construction time.

This work is part of a collaboration with the Federal University of Minas Gerais  8.2 and is under reviewing process for the journal Science of Computer Programming.