# m2comp step 3 — AST layer (tag: `m2comp-step3`)

Coco/R frontend + SymTab checks (steps 1–2) + an abstract syntax
tree built alongside parsing, per H. Mössenböck, "How to Build
Abstract Syntax Trees with Coco/R" (`docs/AST.pdf`): every
nonterminal carries an output attribute returning its subtree, built
with plain semantic actions. No codegen yet (step 4: MC64 `MGen`
walking this tree).

## What was built

- `src/AST.def/.mod` — 30 node kinds in three families:
  decls (`Module/Block/Import/ImpName/Export/Const/Type/Var/
  Member/Param/Proc`), statements (`Assign/Call/If/While/Repeat/
  Loop/For/Return/Exit`), expressions (`Bin/Un/Name/Int/Real/Char/
  Str/Field/Index/Deref/CallExpr`).
  Nodes are heap pointer records (`NEW`) with child links
  (`left/right/extra/more`) and a sibling link (`next`) — the
  `dsc`/`next` shape borrowed from the Oberon `SymTable`
  (`docs/oberon-master`), while `SymTab` itself stays flat.
  Types are stored as descriptor indices (no type subtree);
  unary ops use `opNot/opNeg/opPos`, binary ops reuse `SymTab`
  codes. Error-tolerant: nodes build even with `InvalidType`, so
  rejected sources still dump. Includes `Append`, `SetChainType`
  (identifier lists typed after the fact), `ParseInt`
  (decimal + `H` hex), and a NIL-safe depth-capped `DumpTree`.
- `src/M2comp.atg` — full attribution with step-2 checks untouched:
  `MkModule/MkBlock` roots dumped by the `Module` action;
  `ELSIF` chains attach at the deepest else (standard
  `If(a,A,If(b,B,C))` nesting); `MkFor` carries optional BY;
  calls share the step-2 frame machinery (`MkCall`/`MkCallExpr`
  with pnum); designators become `Name/Field/Index/Deref` chains
  (indices appended per bracket); literals folded
  (`0FFH`→255, `'q'`→113, reals/strings kept as text).
- `run_tests.sh` — new `expect_dump` assertions on tree content
  (`Module Showcase`, `Proc Work`, `CallExpr`, `Field q`,
  `Int 255`, `Proc P`).

## Tests — 25/25

19/19 accept/reject from step 2 unchanged + 6 AST content checks.

## Bugs found (both fixed, both covered)

- `ParseInt` detected hex only when the digit scan stopped at `H`,
  so `0FFH` dumped as `Int 0`. Fixed by checking the trailing `H`
  first, then hex-parsing the prefix.
- `NEW` on `POINTER TO RECORD` needs `FROM Storage IMPORT
  ALLOCATE` under `gm2 -fiso` (ISO rule).

## Notes

- Dump verified by hand on `showcase`: nested procs, recursion,
  `p^`, `m[1,2]`, `Local.q`, `ELSIF` nesting, `LOOP/EXIT`,
  `FOR` with/without `BY`, precedence (`Min(3,4)+(2*3-4/2)`).
- Known simplifications for step 4: string/real literals kept as
  text (truncated at 63 chars); record members build nodes that
  the `Type` alternative discards (types resolve via `SymTab`).

## Next

Step 4: MC64 `MGen` backend — tree-walking emitter (`mc64-spec.md`
§11 opcodes, frames §5.2, `.MC4` images runnable under `mcint`),
reusing the `m2c` lowering as reference. `SymTab` (`ProcNum`,
`ByNum` queries, export tables) and `AST` (pnum/typ on nodes) are
already backend-ready.