EN FR
EN FR


Section: New Results

Liveness Analysis in Explicitly-Parallel Programs

Participants : Alain Darte, Alexandre Isoard, Tomofumi Yuki.

In the light of the parallel specifications encountered in our other work – kernel offloading with pipelined communications [10], automatic parallelization, analysis of X10 [22], [23] and of OpenStream (see Section 7.2), intra-array reuse (see Section 7.4) – we revisited scalar and array element-wise liveness analysis for programs with parallel specifications. In earlier work on memory allocation/contraction (register allocation or intra- and inter-array reuse in the polyhedral model), a notion of “time” or a total order among the iteration points was used to compute the liveness of values. In general, the execution of parallel programs is not a total order, and hence the notion of time is not applicable.

We first revised how conflicts are computed by using ideas from liveness analysis for register allocation, studying the structure of the corresponding conflict/interference graphs. Instead of considering the conflict between two pairs of live ranges, we only consider the conflict between a live range and a write. This simplifies the formulation from having four instances involved in the test down to three, and also improves the precision of the analysis in the general case. Then we extended the liveness analysis to work with partial orders so that it can be applied to many different parallel languages/specifications with different forms of parallelism. An important result is that the complement of the conflict graph with partial orders is directly connected to memory reuse, even in presence of races. However, programs with conditionals do not even have a partial order, and our next step will be to handle such cases with more accuracy. Details of this work have been published at the international workshop IMPACT'16 [3].

Some new developments are in progress to explore even further the properties of such liveness analysis and the construction of conflict sets, in the general case (with connections with the concept of trace monoid) or for some common situations such as series-parallel graphs, appearing in languages such as X10 or OpenMP.