Code generation? Don’t generate to Java

I tried to write a program using Java; it all seemed to be going well but then I hit a ridiculous limit. Java cannot be used for this type of problem. I have now completely re-written it in a different programming language, and that works fine.

Be aware of this limit. I was unaware of it when I started this project. But it makes Java completely unsuitable for a whole class of problem.

My customer supplies me with a config file from time to time, this specifies a certain algorithm. When the user enters data, this algorithm must be applied. The algorithm is complex, so performance is an issue.

The solution I chose was to generate code to execute the algorithm, based on the information in the config file. This is a valid computer-science approach, and is used for similar problems. For example, language parsers are often expressed as a grammar, and code to parse documents in the grammar are generated. JSPs are turned into Java classes which are then compiled and executed. WebTek pre-compiles HTML templates containing macros into code which produces the resulting HTML when executed.

However, don’t try this in Java, unless you are only working with small problems. A single method in Java can only be 64KB in size, once compiled.

This means, JSPs can only be of a certain length, parsers can only parse languages of a certain complexity, if WebTek were written in Java then templates could only be of a certain length and complexity, and so on. Do you want to place such restrictions on the software you produce?

My specific problem involves simulating one million variations to a particular solution. How can I fit that into 64K?

  • That is 0.06 bytes per solution variation; yet the simulation of a single variation involves many lines of code (i.e. in total compiling to more than 0.06 bytes!).
  • I could put each variation into its own method, and have a big method which calls them all—but a method call takes more than 0.06 bytes!
  • I could have a hierarchy of methods: one main method which calls, say, 100 sub-methods, each of those call 100 sub-sub-methods, and finally those methods call the methods for the individual variations.

It’s not even possible to know how many bytes a method will generate to! So, as the complexity of the simulation of a variation is expressed in the config file, I would have to essentially have to do a “trial and error” approach: generate a method, compile it, if I get the error concerning the 64KB limit, split the problem up into slightly smaller methods, try the compilation again, repeat, etc. (And the Java compiler is not even very fast.)

This is all so wrong! This is complexity, which isn’t solving the customer’s problem. This complexity costs me time (and thus my customer money), complexity leads to bugs and difficulty of maintenance, etc.

So I have changed the language. Rather than generate Java, I generate C and compile it using the GNU gcc compiler. From the GNU coding standards:

Avoid arbitrary limits on the length or number of any data structure, including file names, lines, files, and symbols, by allocating all data structures dynamically.

This is a good standard! I like it. All programs should be written with this in mind. Your program may well be online in 10 or 20 years, and the hardware may well have changed: a 64KB limit may seem reasonable one year but is a real limitation 10 or 20 years later in software which would otherwise still be useful.

So, if you are solving this type of problem, don’t use Java.

P.S. On a separate project I used a similar approach using Perl, and that worked out fine too.

  • m35

    Ran into that limitation when trying to include a large constant array in my code. Such arrays declarations are apparently compiled into a method that fills the array one value at a time. Too many elements will create a too large method. Of course this case has the work-around of including the static data as a separate resource.

    I assume your situation didn’t lend itself well to automatically breaking the config algorithm up into separate methods?