Choosing an appropriate data layout can have a significant impact on the performance of any application. In this presentation we outline the importance of having a flexible way of modifying the data layout to enable more aggressive optimizations. First, we discuss how changing the data layout for complex numbers may enable more efficient SIMD vectorization for Fourier transforms and complex matrix-matrix multiplication. Then, we focus on how changing the data distribution helps in scaling three dimensional Fourier transforms on thousands of compute nodes. Finally, we look at how modifying the data layout enables memory bound operations like Fourier transforms and block sparse computations to be merged to improve data locality and reduce data movement. Using these three examples, we emphasize the need for methods and languages that allow data layouts and data layout changes to be easily expressed in order to achieve higher performance gains.