difference between robust_getcwd and rpl_getcwd

[Eric Blake]

Why are there two implementations of a recursive readdir
implementation of getcwd()?  One (rpl_getcwd) is in
lib/getcwd.c, which fails to use root-dev-ino.h (and is
therefore currently broken on cygwin's distinct //), but uses
fstatat for speed and thread-safety, and which produces
no error message on failure.  The other, robust_getcwd,
is in src/pwd.c, uses fchdir and dies on error (since pwd
is single-threaded), rather than returning a failure, but
since it used root-dev-ino.h it was easy to patch to recognize
cygwin's distinct //.  It just seems odd to me that there
are two implementations, and I wonder if any features
should be factored out between the two.  Furthermore,
since some parts of gnulib, such as canonicalize.c,
depend on getcwd, it would be nice if it were as easy
to patch lib/getcwd.c to recognize // as a distinct root.

-- 
Eric Blake