source: trunk/third/libiconv/DESIGN @ 16937

While some other iconv(3) implementations - like FreeBSD iconv(3) - choose
the "many small shared libraries" and dlopen(3) approach, this implementation
packs everything into a single shared library. Here is a comparison of the
two designs.

* Run-time efficiency
  1. A dlopen() based approach needs a cache of loaded shared libraries.
  Otherwise, every iconv_open() call will result in a call to dlopen()
  and thus to file system related system calls - which is prohibitive
  because some applications use the iconv_open/iconv/iconv_close sequence
  for every single filename, string, or piece of text.
  2. In terms of virtual memory use, both approaches are on par. Being shared
  libraries, the tables are shared between any processes that use them.
  And because of the demand loading used by Unix systems (and because libiconv
  does not have initialization functions), only those parts of the tables
  which are needed (typically very few kilobytes) will be read from disk and
  paged into main memory.
  3. Even with a cache of loaded shared libraries, the dlopen() based approach
  makes more system calls, because it has to load one or two shared libraries
  for every encoding in use.

* Total size
  In the dlopen(3) approach, every shared library has a symbol table and
  relocation offset. All together, FreeBSD iconv installs more than 200 shared
  libraries with a total size of 2.3 MB. Whereas libiconv installs 0.45 MB.

* Extensibility
  The dlopen(3) approach is good for guaranteeing extensibility if the iconv
  implementation is distributed without source. (Or when, as in glibc, you
  cannot rebuild iconv without rebuilding your libc, thus possibly
  destabilizing your system.)
  The libiconv package achieves extensibility through the LGPL license:
  Every user has access to the source of the package and can extend and
  replace just libiconv.so.
  The places which have to be modified when a new encoding is added are as
  follows: add an #include statement in iconv.c, add an entry in the table in
  iconv.c, and of course, update the README and iconv_open.3 manual page.

* Use within other packages
  If you want to incorporate an iconv implementation into another package
  (such as a mail user agent or web browser), the single library approach
  is easier, because:
  1. In the shared library approach you have to provide the right directory
  prefix which will be used at run time.
  2. Incorporating iconv as a static library into the executable is easy -
  it won't need dynamic loading. (This assumes that your package is under
  the LGPL or GPL license.)


All conversions go through Unicode. This is possible because most of the
world's characters have already been allocated in the Unicode standard.
Therefore we have for each encoding two functions:
- For conversion from the encoding to Unicode, a function called xxx_mbtowc.
- For conversion from Unicode to the encoding, a function called xxx_wctomb,
  and for stateful encodings, a function called xxx_reset which returns to
  the initial shift state.


All our functions operate on a single Unicode character at a time. This is
obviously less efficient than operating on an entire buffer of characters at
a time, but it makes the coding considerably easier and less bug-prone. Those
who wish best performance should install the Real Thing (TM): GNU libc 2.1
or newer.