Even "languages built around AOT compilation" are increasingly blurring the phase distinction between "ahead-of-time" and "at runtime", via increased use of metaprogramming and constant evaluation facilities. We may end up with a state of things where even some "AOT-language" programs are written to do much of their work at compile time, perhaps adding more and more of those facilities that were formerly allowed only at "runtime".
"Metaprogramming" is a very broad term. Templates etc. might be nice, but it's hardly anything compared to what you can do with Lisp macros (there's an orthogonal issue that this power leads to unreadable code; see "Lisp Curse", but ignore it for now.).
which I might finish up when I'm done with the report I'm writing. It is possible to create Java expression trees with trees of static method calls that look a lot like S-expressions and stick them together into statements, methods and classes.
You should be able to do the same tricks people do with LISP macros and it could work code generation miracles but it would have that "LISP curse" problem in spades.
The plan is to generate a code generator that is sufficient to generate the full DSL implementation (ferocity0) and use that to generate the full implementation (ferocity.)
I have some tests for ferocity0 writing .java files to get fed to javac and for ferocity0 running expression trees with a primitive interpreter. Already the type system is enriched over the type system because interpreted ferocity0 can handle Java expressions as a type at run time so you get issues like quoting and unquoting in LISP.
How does this compare to Common Lisp or Scheme macros in in terms of amount of code needed to generate things? How does it compare in terms of what kind of functions and instructions are used? Can one use standard Java language to use it? Could one for example throw a for loop in there or directly write something like "class Point { ..."?
