That's a great point! Currently, there would indeed be an instant where both the ndarray and the input SV needed to have the full amount allocated, in part because get_dataref physicalises its ndarray (i.e. allocates its memory). COW semantics do open a bit of a can of worms, because PDL currently assumes it has full ownership of the block of memory pointed at by the PV. I am going to just hope that all works fine.

PDL's File::Map-using code (via PDL::IO::FastRow) is an alternative approach, which would avoid the use of RAM entirely.

This being a long-standing issue suggests to me there isn't a huge demand for it. However, I am very open to adding a PDL->from_sv method that does what your code does (and would also set the datasv member to the passed-in SV rather than use a char * and SvREFCNT_inc it) - it would also need to deal with the COW situation correctly, which I don't know how to do. Would that help? I think it would actually provide a more generalised implementation of the File::Map stuff in any case.