# m2comp step 2 — SymTab semantic analysis (tag: `m2comp-step2`)

Coco/R lexer/parser (step 1) + static type checking, still no codegen.
Follows Blaise Phase 2 (type system) with the semantic pass folded
into grammar actions, Wirth single-pass style.

## What was built

- `src/SymTab.def/.mod` — ported from CocoGm2 `Test1` (proven under
  `gm2 -fiso`) and extended for procedures and local modules:
  - procedures: nested, value/`VAR` params (pending-list entry),
    function results; signatures persist in `procs[]`/`params[]`
    after body scopes pop (no `FORWARD` headings, so no 231);
  - local modules (Wirth form): body scope + `(module, name)` export
    table; `M.x` resolves via exports, or the live body scope from
    inside `M` itself (`SelfKind/SelfType/SelfProc`);
  - call checking with an 8-deep frame stack (nested calls safe);
    invalid callees get a dead frame to suppress cascades;
  - open arrays (formal-only): `NewOpen/WrapArray/IsOpen`;
    `VAR`/`value` rules enforced at declaration (230 otherwise);
  - loop-depth tracking for `EXIT` validation (230).
- `src/M2comp.atg` — full attribution: every `Type`/`Expr`/`Stat`
  production carries `SymTab.TypeIndex` attributes; `Designator`
  returns type + kind + proc number with record/field (`215/216`),
  array (`217/218`, per-index stepping), pointer (`219`) checks;
  `AssignOrCall` checks storability + `Assignable` (210) and call
  arity/formals (233, incl. proper-proc-in-expression);
  `RETURN` checked against function context (232); conditions (214);
  `FOR` control/bounds (220); sets via `222`; cycles (223);
  ordinals (224); `MODULE/END` match (202); dup/undeclared (200/201).
- Checks NOT yet done (no codegen info): `WITH` (no grammar support),
  enum/set-literal construction (kept in `SymTab`, grammar later),
  `HIGH`, range checks, `NIL` dereference.

## Tests — 19/19 (`run_tests.sh`)

Accept: `ok_minimal`, `ok_proc`, `showcase` (extended with real
`y`/`z` declarations — calls to merely-imported names are now
rightly 233 — plus `Chars`/`sa` for a `VAR` open-formal call),
`bad_mismatch` still rejected. Reject (one per code):
`t_dup` 200, `t_undecl` 201, `t_assign` 210, `t_arith` 211,
`t_bool` 212, `t_rel` 213, `t_cond` 214, `t_field` 216,
`t_index` 217, `t_ptr` 219, `t_for` 220, `t_call`/`t_call2` 233,
`t_return` 232, `t_exit` 230.

## Bugs found by the tests (both fixed, both covered)

- EBNF grouping trap: in `A | B | C {loop}` Coco/R binds the loop
  to the LAST alternative only, so `-v + 1` broke after step-2
  rewrite. Fixed with a parenthesized group + shared loop.
- Leaked call frames: `Fact` pushed a dead frame for every bare
  designator but popped only on the call path, silently shifting all
  later arity checks (e.g. `P(i)` vs 2 formals passed). Fixed with
  balanced push/pop (`pushed` flag); verified `P()/P(i)/P(i,TRUE)/
  P(i,TRUE,i)` behave correctly.

## Docs added (reference for next steps)

- `docs/AST.pdf` (Mössenböck: ASTs via plain Coco/R output
  attributes — recipe for the step-3 AST layer),
- `docs/oberon-master/` (freeoberon Oberon compiler: hand-written
  recursive descent + `Item`-technique generator; pointer-linked
  `SymTable` studied but NOT mimicked — flat `TypeIndex` design
  stays, since the future MC64 `MGen` needs it),
- `docs/Document 1 sans titre` renamed to `BenjaminKowarsh.txt`.

## Next

Step 3: AST layer per `AST.pdf` (node modules, output attributes;
pointer-record tree à la Oberon `dsc`/`next`), then MC64 `MGen`
backend behind it. `SymTab` API is already backend-ready
(`ProcNum`, `ByNum` queries, export tables).