source: vendors/XMLF90/current/src/xpath/Developer.Guide @ 1963

Developer Notes for the XPATH-lite API

This API is built on top of the more fundamental SAX one. The basic
idea is to have a set of programable handlers that communicate among
themselves and with the parser via module variables.

A new pseudo-handler, "signal_handler", has been added to the optional
argument list of xml_parse(). In the current implementation it just
checks whether a stop signal has been raised by the user side of the
program (for example, after a "begin tag" event, or after going out
of the scope of a parent element).

Since the SAX parser is stream-oriented, and XPATH searches can be
done in any order, there are new routines to rewind the XML file and
to synchronize the physical file reader with a previously saved point
in the XML tree. 

These tools, and the rather sophisticated path-matching routine
provided (which allows for wildcards and the "//" construction) would
be enough for standard use, as implemented in the routine
"get_node". However, there is also the possibility of performing
searches constrained to a given ancestor element ("context" searches)
so that blocks of logically related information can be processed
together. A context is implicitly created by a call to
"mark_node". Contexts can be saved and "synched" to, allowing for
repeated constrained searches (calls with relative paths). Contexts
can even be passed to subroutines to package the parsing of common
elements once and for all. (See for example Examples/xpath/pseudo.f90).
This feature is nevertheless in need of a more rigorous specification:

* What should be the behavior if a "mark_node" is followed by a
  call to "get_node" with an absolute path?
* Should there always be an "automatic rewind" to the beginning of the
  context before any successive calls to "get_node"?


LIMITATIONS

The pcdata buffer provided by the user as a character variable could
overflow.  Note that the parser itself uses a string of length
MAX_PCDATA_SIZE (currently 65536) as a buffer to hold PCDATA. A
warning is issued if there is not enough space (in the user or in the
system buffer) to hold the data.

Support for converting PCDATA characters to numerical arrays "on the
fly" is planned for a forthcoming version.


The coding style is that of the F subset of Fortran90. I strongly
believe that it makes for better coding and fewer errors.
Go to http://www.fortran.com/imagine1/ and get a feel for it. You can
download free implementations for Linux and Windows, or get an
inexpensive CD+Book combination to help support the project. Of course,
F *is* Fortran, so you can always compile it with a Fortran compiler.