# m2comp sessions, steps 1–4 — day wrap-up

Date: 2026-09-16. Built in one day, with E. directing grammar
choices against Kowarsch (arXiv:1809.07080) and the Blaise pipeline
as process reference.

## Where we stand (all committed and tagged)

| Step | Tag | Content | Suite |
| ---- | --- | ------- | ----- |
| 1 | m2comp-step1 | Coco/R lexer+parser, `gm2 -fiso`, LL(1)-clean | 3/3 |
| 1.1 | m2comp-step1.1 | Bare subrange type, syntax showcase | 4/4 |
| 2 | m2comp-step2 | SymTab semantic analysis (200–224, 230/232/233) | 19/19 |
| 3 | m2comp-step3 | AST layer (Mössenböck recipe) + dump | 25/25 |
| 4 | m2comp-step4 | MC64 tree-walking backend, `.MC4` under `mcint` | 41/41 |

Layout: all sources in `m2comp/src/` (grammar, frames, FileIO,
SymTab, AST, MGen, generated scanner/parser/driver), binary at
`m2comp/M2comp`, tests in `m2comp/tests/` (check + reject + AST
dump + execution with ExitCode values), docs in `m2comp/docs/`
(per-step summaries, Kowarsch text, `AST.pdf`, `oberon-master/`
reference, this note).

## Hard-won facts (don't rediscover)

- gm2 is multipass: `FORWARD` is rejected ("too many errors in
  pass 3"); CR's `-m` flag comments forwards out; m2c has none.
- Op-code ranges must be globally unique for tree-walk dispatch
  (Add 0–2, Mul 10–14, Rel 20–27); overlap silently misroutes.
- `51H`/`70H` store orders differ; use `70H` (`[addr, value]`).
- Bodies emitted inline need jump-over (`Jmp/DefLabel`).
- `| head` masks VM hangs (rc is head's); harness uses `timeout`
  + rc checks; diverging programs fail fast.
- `showcase.mod` diverges by construction (infinite regress);
  check-only. `ok_proc` fixed to terminate.
- Negative-vs-zero `D5` quirk reproduces byte-identically under
  m2c — VM domain, avoided in tests.
  [Correction (step 5.2): the VM was innocent. Real `< > <= >=`
  never emitted `D5` — bare stack juggling left operand bits as
  the boolean (nonzero reads back TRUE). Fixed in both MGens;
  see `summary_m2comp_step5.2.md`.]

## Tomorrow: step 5 candidates (in priority order)

1. `DEFINITION`/`IMPLEMENTATION` separate compilation
   (multi-`.MC4`, `depCount` loader path) — biggest language gap.
2. Opaque types + procedure types.
3. `WITH` / `CASE` (grammar + SymTab already shaped for them).
4. `LONGINT`/`LONGREAL` quads (`0C5–CC`, `0x12`).
5. Debug listings / OPDF-style debug info.

`SymTab` (`ByNum`, exports, bounds), `AST` (pnum/typ/aux/tag),
`MGen` (scope stack, image assembly) are ready for all of these.
Start with `./build.sh && ./run_tests.sh` (41/41, ~2 min).