It is still common to have applications which have some 32-bit parts, and some 16-bit parts. The 32-bit parts try to avoid using 16-bit selector:offset addressing, because of the overhead of loading the selector registers, as well as to avoid the challenge of correctly dealing with storage references in both modes.
A typical example is a 32-bit application calling a 16-bit DLL.
Since storage is (must be) the same for all parts of a process, there has to be a way to convert one form of an address to the other.
Only 16-bit application selectors from the LDT are eligible for this quick form of the conversion, and only linear addresses less than %20000000 can be converted to 16:16 format.
Additionally, addresses in the packed region may NOT be converted by this quick method, but by a search of the LDT descriptor base (linear) addresses, followed by a calculation.
The top of normal application space, at %1BFFFFFF, is mapped to selector DFFF. The top of protected shared addresses at %1FFFFFFF maps to selector FFFF, if used.