I mentioned that the regular expressions can be parsed using a state machine. In most compiler texts, and indeed in most compilers as well, you will find this taken literally. There is typically a real implementation of the state machine, with integers used to define the current state, and a table of actions to take for each combination of current state and input character. If you write a compiler front end using the popular Unix tools LEX and YACC, that's what you'll get. The output of LEX is a state machine implemented in C, plus a table of actions corresponding to the input grammar given to LEX. The YACC output is similar … a canned table-driven parser, plus the table corresponding to the language syntax.
That is not the only choice, though. In our previous installments, you have seen over and over that it is possible to implement parsers without dealing specifically with tables, stacks, or state variables. In fact, in Chapter 5, Control Constructs I warned you that if you find yourself needing these things you might be doing something wrong, and not taking advantage of the power of Pascal. There are basically two ways to define a state machine's state: explicitly, with a state number or code, and implicitly, simply by virtue of the fact that I'm at a certain place in the code (if it's Tuesday, this must be Belgium). We've relied heavily on the implicit approaches before, and I think you'll find that they work well here, too.
In practice, it may not even be necessary to have a well-defined lexical scanner. This isn't our first experience at dealing with multi-character tokens. In Chapter 3, More Expressions, we extended our parser to provide for them, and we didn't even need a lexical scanner. That was because in that narrow context, we could always tell, just by looking at the single lookahead character, whether we were dealing with a number, a variable, or an operator. In effect, we built a distributed lexical scanner, using procedures GetName and GetNum.
With keywords present, we can't know anymore what we're dealing with, until the entire token is read. This leads us to a more localized scanner; although, as you will see, the idea of a distributed scanner still has its merits.