Which approach minimizes cache misses when slicing large allocatable arrays?

Fortran provides powerful array slicing capabilities that allow developers to access subsets of large datasets using the start:end syntax. When working with large allocatable arrays, these slices can be used to create processing windows similar to pagination in data-driven applications.

However, slicing often results in non-contiguous memory access patterns. Depending on the array's dimensionality and the stride of the slice, the compiler may struggle to optimize cache utilization, potentially leading to significant performance degradation compared to contiguous memory access.

When implementing a windowing system for a large dataset in Fortran 2008 or later, what are the trade-offs between using array slices versus utilizing pointers with the target attribute to manage memory references? Does the use of PACK to create a contiguous temporary copy of a slice consistently outperform direct slicing for repeated iterations over the same subset?