[exabrial]

His other project "Byte Me", along with judicious javap usage, has been super useful for me learning JVM bytecode so I could make a machine learning model compiler for the JVM (basically compile your ML models as native code; ONNX, tree ensembles, regressors, classifiers, etc as native JVM classes with no massive runtime needed)

still in the works, but its here for those interested: Petrify: https://github.com/exabrial/petrify

[deepsun]

Why not compiling it to Java source code (not bytecode)? Users would use their own Java compiler then.

Same as, say, ANTLR generates code to parse various texts to AST.

[exabrial]

Great question, actually I tried that! m2cgen is a project that does that in fact.

It works fine for simple models, but breaks down for production-sized tree ensembles. The JVM has a hard 64KB method size limit, and javac controls how your deeply nested if/else trees get laid out. m2cgen's own FAQ says to reduce estimators when you hit recursion limits during generation. With direct bytecode emission I control the method structure precisely, I can split across methods exactly where needed and manage the constant pool directly. I also wrote much more efficient bytecode than m2cgen creates as equivalent source.

The source code is also a pretty useless step, sets off all kinds of static analysis alarms in your stack, and also I worry about source code injection (not that can't happen with petrify, it's just a lot harder).

Finally, I'm grateful for the sweat the authors of m2cgen have put in, but the project has gone without updates for 4 years. That doesn't mean it's useless (some mature software never sees updates), but it's not a positive sign either.