LN (aLaN) | AMM (Alan Minus Minus, simplified syntax) / \ JS AGA (Alan Graphcode Assembler, dependency graph code in a text format) | AGC (Alan Graphcode, binary format for the AVM) | AGZ (Alan Graphcode GZipped, compressed binary format for the AVM)
LN is the high level language meant to be easily accessible, like a cleaned-up Typescript/JS. Automatic memory management, automatic IO concurrency, automatic event-level and array-level CPU parallelism (two tracks in the runtime right now), and a strict type system but with interfaces for defining generic functions that we believe we can write 100% automatic type inference without the compiler requiring a fallback to an explicit type.
AMM is a heavily stripped-down version of the language but technically syntactically compatible. It only has constants, events, event handlers, functions, closures, function calls, and assignments. Everything else, including conditionals, is stripped away.
AGA looks very similar to assembly, but with annotation on events, closures, and the dependency graph that the Rust AVM uses to reorganize for concurrency and parallelism benefits.
AGC is the binary version of AGA and is a near direct translation of the text to a 64-bit little-endian integer packed array, with keywords being given numeric values that are clearly interpretable as 8 UTF-8/ASCII characters.
AGZ is a gzipped AGC file that can also be loaded directly by the AVM.
While overkill at the moment, the compiler solves the shortest path from the specified input file to output file and constructs the pipeline to do so on the fly. (This is much faster than it sounds and has allowed us to add new inputs/outputs and translation formats quickly.)
The first stage of the compiler dominates with the most complex grammar rules in parsing and the most flexibility in interpreting generics and interface-based functions into "real" types and functions. Once the AST is generated, the first layer transforms this into a scope-oriented collection of elements and then generates AMM code on a per-event-handler basis, finally it looks for constants in these handlers and hoists them to the global scope and deduplicates them, then emits this in the AMM format.
If targeting to AGC, this stage of the compiler converts the AGA to AGC with a very straightforward parse and re-emit into 64-bit integers appended to an array and then written to disk. This format doesn't work in the browser due to depending on Node.js's Buffer API, which no browser shim supports, yet, but thankfully that is not needed to build runnable code in the browser!
If the final output target is AGZ, the final stage of the compiler takes the AGC binary, compresses it, and writes out the compressed data. This stage depends on Node.js' built-in zlib bindings so it also does not work in the browser, but is also not necessary there.