Refining data flow information using infeasible paths

Abstract

Experimental evidence indicates that large programs exhibit significant amount of branch correlation amenable to compile-time detection. Branch correlation gives rise to infeasible paths, which in turn make data flow information overly conservative. For example, def-use pairs that always span infeasible paths cannot be tested by any program input, preventing 100% def-use testing coverage. We present an algorithm for identifying infeasible program paths and a data flow analysis technique that improves the precision of traditional def-use pair analysis by incorporating the information about infeasible paths into the analysis. Infeasible paths are computed using branch correlation analysis, which can be performed either intra- or inter-procedurally. The efficiency of our technique is achieved through demand-driven formulation of both the infeasible paths detection and the def-use pair analysis. Our experiments indicate that even when a simple form of intraprocedural branch correlation is considered, more than 2% of def-use pairs in the SPEC95 benchmark programs can be found infeasible.

Supported in part by Hewlett Packard, the National Science Foundation PYI Award CCR-9157371, and Grant CCR-9402226 to the Univ. of Pittsburgh.