ENCODE_FILE and DECODE_FILE turn out to be surprisingly slow, and allocate copious amounts of memory, to the point that they often turn up in both memory and cpu profiles. (This is on macOS; I haven't checked the situation elsewhere.)

For instance, a single call to file-relative-name, with ASCII-only arguments, manages to allocate 140 KiB. There are several conversion steps each involving creating temporary buffers as well as the compilation and execution of very large "quick-check" regexps. Example:

(progn
  (require 'profiler)
  (profiler-reset)
  (garbage-collect)
  (profiler-start 'mem)
  (file-relative-name "abc")
  (profiler-stop)
  (profiler-report))

This applies to just about every function dealing with files or file names.

The attached patch is somewhat conservatively written but at least a starting point. It reduces the memory consumption by file-relative-name in the example above to zero. Perhaps we can assume that file names codings are always ASCII-compatible; if so, the shortcut can be taken in encode_file_name and decode_file_name directly.

There is already a hack in encode_file_name that assumes that no unibyte string ever needs encoding; if so, the shortcut could perhaps be extended to decode_file_name and simplified.