Abstract

Observations of values computed during program execution suggests that compilers miss numerous optimization opportunities: even in highly optimized programs, up to 40% of executed instructions compute the same value as their previous dynamic instance. Very likely, this redundant recomputation reflects expression equivalence that holds only along some---but frequent---execution paths. Because not all executions of an expression are optimizable, this some-paths redundancy is beyond conventional optimizers: conservative analyzers fail to expose it, and inflexible transformations fail to remove it. To harness the some-paths potential, this research proposes an optimization framework for exploiting redundancies that may not exist along all execution paths, or may differ along some paths. The underlying framework algorithms will be applicable to a broad class of optimization problems that are based on value flow, i.e. on the paradigm of reusing the flow of previously computed val...

Citations