In this paper, we present a novel compiler technique that provides the capability to guarantee a different memory consistency model for different variables within the same shared memory parallel program. Programmers need only focus on their parallel algorithm and which variables can tolerate the use of old values, while the compiler automatically guarantees sequential consistency for all remaining variables and identifies where unnecessary synchronization overhead can be avoided. Program slicing is exploited to address the interactions between variables that can tolerate old values and those that require the most recent value to be read.