Logging Messages

Message Storage
===============
Each message is identified by a string identifier, e.g. "INVFILNAM", which is
associated with some text (e.g. "%s is an invalid file name").  These are stored
in a single std::map, in a class called the Dictionary.

The message identifier (along with parameters) is passed through the logging
system to an appender, which uses the identifier to look up the message in
the dictionary.  The message is then formatted and written out.

Message File
============
A message file is a file containing message definitions.  Typically there will
be one message file for each component that declares message symbols.


A example file could
be:

# Example message file
# $ID:$

$PREFIX TEST_
TEST1       message %s is much too large
+ This message is a test for the general message code

UNKNOWN     unknown message
+ Issued when the message is unknown.

Point to note:
* Leading and trailing space are trimmed from the line.
* Blank lines are ignored
* Lines starting with "#" are comments are are ignored.
* Lines starting $ are directives.  At present, the only directive recognised
  is $PREFIX, which has one argument: the string used to prefix symbols.  If
  there is no facility directive, there is no prefix to the symbols.
* Lines starting + indicate an explanation for the preceding message.  These
  are processed by a separate program and used to generate an error messages
  manual.  However they are treated like comments here.
* Message lines.  These comprise a symbol name and a message (which includes
  C-style substitution strings).

Message Compiler
================
The message compiler produces two files:

1) A C++ header file (called <message-file-name>.h) that holds lines of the
form:

namespace {
const char* PREFIX_IDENTIFIER = "identifier";
   :

}

These are just convenience symbols to avoid the need to type in the string in
quotes.  PREFIX_ is the string in the $PREFIX directive and is used to avoid
potential clashes with system-defined symbols.

2) A C++ source file (called <message-file-name>.cpp) that holds the code
to insert the symbols and messages into the map.

This file declares an array of identifiers/messages in the form:

namespace {
const char* messages = {
    identifier1, text1,
    identifier2, text2,
    :
    NULL
};
}

(A more complex structure to group identifiers and their messages could be
imposed, but as the array is generated by code and will be read by code,
it is not needed.)

It then declares an object that will add information to the global dictionary:

DictionaryAppender <message-file-name>_<prefix>_<time>(messages);

(Declaring the object as "static" or in the anonymous namespace runs the risk
of it being optimised away when the module is compiled with optimisation.  But
giving it a standard name would cause a clash when multiple files are used,
hence an attempt at disambiguation.)

The constructor of the DictionaryAppender object retrieves the singleton
global Dictionary object (created using standard methods to avoid the "static
initialization fiasco") and adds each identifier and text to the member
std::map.  A check is made as each is added; if the identifier already exists,
it is added to "overflow" vector; the vector is printed to the main logging
output when logging is finally enabled (to indicate a programming error).

User Message Files
==================
During logging initialization, a search is made for a user message file in a
specific location.  (The specification of the location has yet to be decided -
it will probably be a command-line option.)  These messages are read into a
local Dictionary object (which performs the same checks as the global
Dictionary for duplicate messages).

The local Dictionary is then merged with the global Dictionary.  In this case
though, warnings are issued where a message does not replace one in the global
Dictionary.

An additional check that could be done is to compare the user message string
with the main message string and to check that they have the same number of
"%s" components.  This will avoid potential problems in message formatting. (As
noted in another design document, the intention within logging is to convert
all parameters to strings at the point at which the logging call is made.)

Message Compiler Implementation
===============================
The fact that user files are read in at run-time implies that the code that
reads the files should be C++.  An implication of this is that the message
compiler should be written in C++ (instead of Python, which is probably
better for the task) to avoid two sets of code where message files are parsed.