# Step: V3 compiler corpus runs

Tag `v3-corpus-run`. V3 suite **169/169**, fixpoint OK.

## Why

The TopSpeed corpus validated the *sidecar grammar*, not V3 the
compiler (the sidecar is a separate `TSM2` binary built from
`TopSpeed-V3-M2.atg`; it reuses V3's `SymTab`/`QbeGen`/`FileIO` sources
but never runs `compiler/M2`).  So this work runs **V3 itself** over two
real Modula-2 corpora and fixes what it finds.

Harness: `tools/v3-corpus/corpus.sh` — compiles each file with
`compiler/M2` in a scratch tree, together with its sibling DEFINITION
and the DEFINITION of every module it imports (`$V3_LIBS` search path),
and splits failures into CRASH / LIB (unresolved import) / LANG (real
gap).

## Corpora and results

**1. gm2 testsuite** (`gm2/pim/pass` + `gm2/iso/run/pass`, 612 files):

| | no libs | with `$V3_LIBS` |
| --- | --- | --- |
| **compile-OK** | 242 | **392** |
| CRASH | 0 | 0 |
| LIB | 185 | **9** |
| LANG | 185 | **212** |

**2. gm2 standard libraries** (`m2pim`/`m2iso`/`m2log`/`m2cor`/`m2min`,
336 files):

| | no libs | with `$V3_LIBS` |
| --- | --- | --- |
| **compile-OK** | 78 | **127** |
| LIB | 176 | **0** |
| LANG | 82 | **209** |

(With the import path set, LIB noise is gone and the true LANG counts
are visible.)

## Fixes landed

1. **Long-line buffer overflow** (`v3-fix-longline-crash`) — `GetLine`
   wrote past a 256-byte listing buffer; a long line (an 860-char enum)
   overflowed the stack.  Fixed (`GetLine` truncates, buffer 4096,
   bounded caret loop); CRASH → 0.
2. **`EXPORT QUALIFIED`** in DEFINITION modules (`v3-export-qualified`).
3. **`SYSTEM` facilities** — predefine `BYTE`/`SBYTE`/`WORD`/`SWORD`/
   `LONGWORD`/`SHORTADDR`/`LONGADDR`/`PROC`/`LOC` (mapped onto V3's
   integer family / `ADDRESS`) so `FROM SYSTEM IMPORT BYTE;` and
   `ARRAY OF BYTE` resolve.
4. **Anchored subranges** `T[lo..hi]` (`INTEGER [-1 .. 24]`).
5. **Unnamed procedure-type parameters** (`PROCEDURE (VAR ARRAY OF
   REAL)`) plus **open arrays in procedure types** and structural
   proc-type formal compatibility (`ParamTypesOk`), so assigning a
   procedure with an open-array parameter to a procedure variable works.
6. **Harness import resolution** — `corpus.sh` now finds and passes the
   `.def` of each imported module (`$V3_LIBS`), removing the LIB noise.
7. **Array constructors** `T{…}` (GNU Modula-2).  A `CONST` whose value
   is a constructor (`my = arr {'h','e','l','l','o'}`, `arr {1,2,3,4 BY
   3}`, nested `names { "Thu", "Fri" }`) now lowers to a pooled static
   descriptor (`data $ctoN = { l len, … }`) with the same layout as a
   declared array, and the constant resolves to its address; whole-array
   copy, indexing and `LEN` then work unchanged.  Elements: scalars
   (`b`/`w`/`d`/`l` by class), nested constructors, string literals for
   nested `ARRAY OF CHAR`, and the `expr BY n` repeat form.  New
   `QbeGen.Ctor{Begin,Elem,End}`; the old `TypedSetLit` became
   the unified `TypedBraceLit` (`BITSET{…}` still a set).
8. **Record constructors and runtime construction** — `RecName{…}`
   fills a record's fields in declaration order (static when every
   element is constant, so it works in `CONST`; a nested constructor is
   inlined, and a string becomes a matching `CHAR`-array field).  When
   any element is a runtime value the descriptor is built on the stack
   (`alloc8`) and filled with stores/copies (`ElemStore`, `CopyArray`,
   `CopyRecord`, `CopySet`); this also covers non-`CONST` array
   constructors (`a := arr{i, i+1, i+2}`).  `strcons.mod` now compiles.
9. **Set fields in record constructors** — `R{7, S{1,3,5}}` (both
   runtime and via a typed set constructor); the nested `braceCls`
   dispatch flag is saved/restored around a nested set constructor.
10. **Constant string concatenation in constructor elements** —
    `arraytype {Hello + " " + World}`: `CONST` string values are
    supported, `+` on constant strings folds to one literal
    (`QbeGen.StrFold`), and a string element expands into consecutive
    `CHAR` elements.  `constarray2.mod` compiles.
11. **Index-type arrays** `ARRAY <ordinalType> OF T` (e.g.
    `ARRAY Colour OF T`, `ARRAY Sq OF T`) — the index type's ordinal
    bounds give `[lo..hi]` (`array2`/`array3` now compile).  Enum
    values are accepted as array indices.
12. **`MIN`/`MAX`** builtins over any ordinal type, array, or set
    (`MIN(Colour)`, `MAX(large)`, `ARRAY [MIN(c)..MAX(c)] OF T`);
    `SymTab.TypeBounds` + enum literal counts.
13. **`FROM libc` intrinsic** — `printf`/`exit`/… import as permissive
    external C procedures (varargs, no result) and bind to the C
    symbols; string/CHAR-array actuals pass as `descriptor+8`.
14. **Sets over enums and CHAR ranges** — `SET OF <enum>`,
    `SET OF ['A'..'z']`, enum/`BOOLEAN` subranges; span limit
    256 → 65536.
15. **`MIN`/`MAX` for REAL/LONGINT** and **gm2 sized names**
    (`LONGCARD`, `SHORTREAL`, `INTEGER8/16/32/64`,
    `CARDINAL8/16/32/64`).
16. **Set builtins** `INCL`/`EXCL` (statements) and `SHIFT`/`ROTATE`
    (expressions, bit-by-bit over a runtime amount).
17. **Module-level forward type references** — a global may name a type
    declared further down; emission is deferred until the TYPE block
    completes (TYPE-block forwards already worked).
18. **`String` and opaque/forward types** — an unresolved (opaque or
    forward) alias is pointer-like (assign `NIL`, copy, compare, pass),
    which is what gm2's opaque `String` needs.  `String` is **not**
    built-in: it is imported from a definition module
    (`DynamicStrings`/`opaque`), as in gm2.  An undeclared type name at
    module level is reported (`201`, undeclared identifier) once the
    module's TYPE declarations are known (`FwdPending`/`FwdInfo`).
19. **Variant-record assignment** — whole-record assignment is now a
    flat byte blit of the (max) record size, correct for overlapping
    variant fields; variant records are emitted as a zeroed blob.
    (Previously it compiled but crashed / corrupted data.)
20. **Nested local `MODULE`** — exported names, nested-nested modules
    and bodies are lowered as startup init functions (was already
    largely working; verified and documented).
21. **`ADDRESS` ↔ pointer assignment** (`p := ADR(x)`), checked before
    the pointer-subclass test.

## Remaining LANG gaps

- **`undeclared identifier`** — SYSTEM facilities beyond the set
  above, and small library holes.
- **gm2-specific syntax** — `PROCEDURE __BUILTIN__ name(…)`, C-name
  pragmas, ISO module attributes (not V3's dialect; skip).
- **`not supported yet`** — deliberate `230`s.

## Reproduce

```sh
cd tools/v3-corpus
./corpus.sh                       # testsuite, no libs
G=$HOME/bin/Modula2/Gm2/lib/gcc/x86_64-pc-linux-gnu/16.0.1/m2
V3_LIBS="$G/m2pim:$G/m2iso:$G/m2log:$G/m2cor:$G/m2min" ./corpus.sh
V3_LIBS="$G/m2pim:$G/m2iso:$G/m2log:$G/m2cor:$G/m2min" \
  ./corpus.sh "$G/m2pim" "$G/m2iso" "$G/m2log" "$G/m2cor" "$G/m2min"
```

## Suggested next steps

1. More `SYSTEM` facilities / library holes.
2. Whole-set `LargeInt`-style operations; gm2-only pragmas (skip).

## Files

`tools/v3-corpus/{corpus.sh,README.md}`, `compiler/src/compiler.frm`
(long-line), `compiler/src/M2.atg` (`EXPORT QUALIFIED`, anchored
subranges, unnamed proc-type params, `TypedBraceLit` array/record
constructors, index-type arrays, `MIN`/`MAX`, enum indices),
`compiler/src/SymTab.def`/`.mod` (SYSTEM predefs, `ParamTypesOk`,
`TypeBounds`, enum counts), `compiler/src/QbeGen.def`/`.mod`
(`CtorBegin`/`CtorElem`/`CtorEnd`, `StrFold`, `FlushCtors`, static +
runtime descriptor building), `compiler/tests/t_ctor{1..7}.mod`,
`compiler/tests/t_minmax.mod`, `compiler/tests/t_longline.mod`,
`compiler/run_tests.sh`, this doc.