# Within-module forward procedure calls — M2S and M2P compile clean (110/110)

Tag `v3-forward-calls`.

## The finding

CR's parser output **relies on gm2 hoisting**: by default it wraps its
forward declarations in a comment (`(* ----- FORWARD not needed in
multipass compilers … *)`). The `-m` flag toggles that
(`M  Suppresses FORWARD declarations in parser (for multipass
compilers)`), and our `build.sh` passed `-m`.

V3 is single-pass, so with `-m` the generated `M2P.mod` had ~102
`undeclared identifier` errors for its own procedures.

## The fix (two halves)

1. **Emit real `FORWARD` declarations for V3-facing generation** — run
   CR **without `-m`**. V3 already supports `FORWARD` (including
   mutual recursion); verified with a direct test.
2. **Keep `-m` for the gm2 stage-1 build.** gm2 *compiles* `FORWARD`
   fine (`gm2 -c` OK) but its link-time whole-program/“pass 3”
   recompilation rejects `FORWARD` in an implementation module
   (`too many errors in pass 3`); reproduced minimally, and
   `-fscaffold-*`, `-fno-whole-program` etc. do not help.

So the bootstrap is stage-specific: stage 1 (gm2) generates with `-m`;
stage 2+ (V3) generates without `-m`. The `.ssa` fixpoint is between
V3 stages, both using the `FORWARD` form.

## Other fixes in this step

- **Empty `ELSIF … THEN` body** — the earlier `[ StatSeq ]` change had
  missed the `ELSIF` site (different indentation); CR-generated code
  hits it (comment-only `THEN`).
- **`BITSET`** — added as a predefined type (`SET OF [0..15]`).
- **Typed set constructor** `TypeName{ … }` (`BITSET{0}`,
  `BITSET{}`) — new `TypedSetLit` rule; `Fact` uses it when a type
  name designator is followed by `{`.
- **File-handle leak** — `compiler.frm` never closed the source file,
  so each session source consumed a handle and the host `FileIO` cap
  (`MaxFiles = BitSetSize` = 16) tripped once a session had >16
  sources (exactly the compiler's own build). The driver now closes
  `src` after each file.

## Result

Under V3 (sources generated without `-m`):

| Source | Errors |
|---|---|
| `SymTab` | 0 |
| `QbeGen` | 0 |
| `M2S` | 0 |
| `M2P` | **0** (was 107 before FORWARDs, then 5, then 0) |

Main suite **110/110**.

## Remaining: the driver `M2.mod` (12 errors)

- **Nested modules** — the driver's `ListHandler` is a Wirth-style
  local `MODULE … EXPORT StoreError, PrintListing, …`, which V3 parses
  but marks 230; its exports (`StoreError`, `PrintListing`,
  `PrintConsoleErrors`, `ResetErrors`) are therefore undeclared at the
  enclosing scope.
- **String building** — the driver uses `Strings.Delete`/`Append` to
  construct file names. V3's descriptor header is *both* the
  open-array count and the logical string length (`m2setlen` sets it),
  so once `Assign` shrinks it, `Append` writing past the logical
  length traps. `Delete` alone is fine; incremental `Append` needs a
  model decision (or a driver rewrite).

## Build note

Generating for V3 (no `-m`) and for gm2 (`-m`) are different commands;
this wants a small stage-aware build script (or a `build_v3.sh`),
which also feeds `bootstrap/fixpoint.sh`.

## Files

`compiler/src/M2.atg`, `compiler/src/SymTab.mod`,
`compiler/src/compiler.frm`, `compiler/src/M2.lst`.