How do the Java and C versions compare to the ML version? I have (and really like!) that one, but Appel's an ML hacker, and I'd be surprised if it translated 100% to C / Java.
I took a course based on the Java version, and it felt very clunky to me - the sample project code was very verbose and used the Visitor pattern whenever it could possibly be wedged in, making it a bit of a nightmare to understand the flow of the code as a whole. A good IDE would probably clear up a lot of this confusion, but it always felt like we were fighting the language. Take this with a grain of salt, however, since I'm fairly certain my professor edited the projects quite a bit before handing them to the students; perhaps the original book material is less convoluted.
I have the C version, and I personally found it easier going than the Dragon Book. As with the other books in the series, the beginning chapters dealing with the lexer an the parser are very mathematical - but the parts actually dealing with emitting bytecode are quite practical in nature. I ended up using the lexx/yacc tutorials to give me a grounding in parsing, without worrying too much about the mathematical basis for it, and then used the Appel book for everything after having derived an AST.
