Reduce consists of two kernels. The first one reads input data and accumulates partial sums in a temporary storage. The second one reads the temporary storage and writes the final result. Therefore, it's safe to alias input and output arrays.
Nonetheless, allowing in-place execution would limit our abilities to optimize algorithm later. Since aliasing doesn't provide significant memory saving, I'd rather not allow it.

Regarding ByKey variant, it relies on decoupled look back, so it should be safe to alias in/out data as long as value types for input/output iterators match exactly. The only limiting factor is LOAD_LDG which makes aliasing in this case an UB. In-place execution in this case would provide significant memory savings. If there's a request, I suggest we add an overload that would allow in-place execution.

Regarding Segmented version, one block is assigned per segment. Results are written without synchronization between blocks, therefore, any aliasing with output array would introduce a data race.