Sources: background card; Julia developer documentation
The representation Julia code takes after lowering — desugaring — and before type inference. Reachable from the REPL with Meta.lower(mod, expr) or @code_lowered, and represented as a CodeInfo object.
Lowering desugars everything syntactic into a small core: for becomes iterate calls, do blocks become anonymous functions, destructuring becomes indexed_iterate, comprehensions become generator closures, operators become function calls, and control flow becomes explicit goto/gotoifnot on a flat statement vector with SSAValue and SlotNumber operands. It is not quite SSA — slots are mutable — but the later typed IR (@code_typed, after inference) is.
For Sophia this is the natural ingestion point for the Julia frontend, and the reason is that it is the last representation before type inference and dispatch resolution. Ingesting surface syntax would mean reimplementing Julia’s desugaring; ingesting typed IR would mean the inference results are baked in and the ingestion is specific to one argument-type tuple. Lowered IR is generic, canonical, and produced by Julia itself rather than by a reimplementation.
The pipeline positions, for reference:
| Stage | Access | What it has |
|---|---|---|
| surface AST | Meta.parse | macros unexpanded, all sugar present |
| lowered | @code_lowered | desugared, untyped, slot-based, explicit control flow |
| typed / inferred | @code_typed | one specialisation, inferred types, inlining decisions |
| LLVM IR | @code_llvm | LLVM IR |
| native | @code_native | machine code |
A frontend should record ELABORATES_TO edges from lowered IR to SC, and — separately — the typed IR per specialisation together with its dispatch facts, since that is what Content-Addressed Precompilation needs to key the code cache on.