m2compiler-V3 — language report

Status at tag v3-fwdvars-resultsfx (2026-10-01). Suite 175/175, self-hosting fixpoint OK. This is a reader-facing report: what the compiler accepts today, what it deliberately does not, and what is missing. It complements docs/features.md (feature-by-feature) and docs/analysis_gaps.md (gap taxonomy).

Legend: ✅ done · 🔄 partial · ⏸ not implemented · 🚫 deliberately out

1. What it is

A from-scratch Modula-2 compiler: a Coco/R-generated frontend (compiler/src/M2.atg, 1,7xx lines, ~75 productions) lowered to QBE SSA (QbeGen.mod), assembled and linked to a small C runtime shim (runtime/syslib/shim.c). Classic identifiers (no _), # for "not equal", 0x hex, // comments; the whole image is buffered and written in one pass, and the compiler self-hosts byte-identically.

2. Lexis and syntax ✅

Identifiers letter { letter | digit }, no underscores.
Integers (decimal, 0x/0X hex, char codes NN C), reals with E exponents, '…'/"…" strings, U'…'/U"…" codepoint literals.
#, <, <=, >, >=, =; NOT/AND/OR only.
// line comments and nested (* … *) block comments.

3. Types ✅

INTEGER, CARDINAL, SHORTINT, LONGINT, SHORTCARD, REAL, LONGREAL, CHAR, BOOLEAN, NIL, BITSET.
gm2 sized names: LONGCARD, SHORTREAL, INTEGER8/16/32/64, CARDINAL8/16/32/64, and SYSTEM storage names BYTE, SBYTE, WORD, SWORD, LONGWORD, SHORTADDR, LONGADDR, PROC, PROCESS, LOC, ADDRESS.
Named types and aliases; integer/char/BOOLEAN/enum subranges (bounds folded, anchored T[lo..hi]); enumerations.
ARRAY [lo..hi, …] OF T (multi-dimensional), open ARRAY OF T formals, index-type arrays ARRAY <ordinalType> OF T.
RECORD, including variant records (CASE tag : T OF … END); whole-record assignment is a flat byte blit.
SET OF bool/char/subrange/enum (multi-word, span ≤ 65536).
POINTER TO, opaque TYPE T; + completion, ADDRESS ↔ pointer.
PROCEDURE … : T types and procedure-typed values.
UCHAR (32-bit codepoint) and UString (U"…").

4. Expressions and statements ✅

Full precedence; + - * / DIV MOD, comparisons, eager AND/OR/ NOT, IN, set union/diff/inter/symdiff, pointer =/#, indexing, field access, ^ deref, qualified L.x / L.P.
String =/# (content), + (concat), assignment (content copy, NUL-terminated), LEN/LENGTH; UString +/LEN likewise.
IF/ELSIF/ELSE, CASE (labels, ranges, ELSE), WHILE, REPEAT/UNTIL, LOOP/EXIT, FOR/TO/BY (static sign), WITH, RETURN, HALT, NEW/DISPOSE.
Builtins VAL, ADR, TSIZE, SIZE, HIGH, LEN, INC, DEC, INCL/EXCL, SHIFT/ROTATE, MIN/MAX (any ordinal type, array or set).
Result suffixes (expression context): F()^, F()[i], F().field.
FROM libc IMPORT … — permissive external C procedures.
Array/record constructors T{…} (static and runtime), expr BY n.
Forward module-level variables: a procedure body may reference a global declared further down the module.

5. Modules and procedures ✅

Program modules, DEFINITION/IMPLEMENTATION modules, IMPORT / FROM … IMPORT, EXPORT (incl. QUALIFIED), nested local modules.
Separate compilation of a whole session into one image with module init bodies; opaque completion; strict 235 signature checks.
Value and VAR parameters, function results, recursion, mutual recursion (FORWARD), nested procedures (static-link display), open-array VAR formals.
Clarion-form CLASS: fields, methods (value/VAR, results), single inheritance, VIRTUAL dynamic dispatch via vtables, WITH, hidden THIS, class BEGIN … END init bodies.

6. Coroutines ✅

PIM SYSTEM.PROCESS, NEWPROCESS, TRANSFER, IOTRANSFER (ucontext-backed). IOTRANSFER degrades to a plain transfer.

7. Backend ✅

Length-prefixed arrays (count + elements), deep record copy, count-checked whole-array assign, runtime bounds/span traps ($abort, rc 134), multi-word sets, alloc8 heap (malloc/free), deterministic %tN/@LN, whole-image buffering, REAL const flush, string literals as descriptors, vtables.

8. Library

runtime/syslib (always linked): SysShim, FileIO, Utf8.

stdlib (opt-in): SysIO, TextIO (+ UTF-8), WholeIO, Math, Strings, CharClass, Files, Conversions, RealIO, ProgramArgs, SysClock, IOChan, Storage, plus the gm2/ISO set added this session:

Area	Modules
conversion	`ConvTypes`, `WholeStr`, `RealStr`, `LongStr`, `LongMath`, `LongIO`, `LongWholeIO`
text IO	`StrIO`, `InOut`, `StdIO`, `STextIO`, `SWholeIO`, `SRealIO`, `SIOResult`, `IOConsts`
runtime	`DynamicStrings` (core), `SysStorage`, `TERMINATION`

9. Missing

9.1 Language

GOTO and labels — parse error. ⏸
Generics — not started. ⏸ (docs/analysis_generics.md).
Statement-context result suffixes — F()[i] := x and F()^.f := x are not accepted (expression suffixes are). ⏸
Chained post-call calls F().M() — not accepted. ⏸
Multiple inheritance — 🚫 out of scope (docs/OOP.txt).
Full forward variables — supported at module scope; forward references from a procedure to a procedure declared later relies on the DEFINITION being present (as before). 🔄
GOTO-free result suffixes only; no EXIT from a nested named loop etc.

9.2 Library

Channel/device framework: ChanConsts, GenDevIF, DeviceTablePtr, StdChans, StreamFile, FIO. ⏸
C wrappers wrapclock/wraptime/wrapc, TimeString. ⏸
DynamicStrings is the core subset (no *DB debug variants, no InitStringCharStar). 🔄
ISO read-result bookkeeping is minimal (a single default-channel result via SIOResult). 🔄

9.3 Architecture / robustness

No AST, single pass — declaration-before-use is enforced, so forward references beyond the two supported kinds fail.
Single image only — no per-module object emission, no import cycles.
Fixed-size tables — 64 formals/locals, 64 call args, 4096 string literals, 256 pooled constructors, 128 session source files, 4 MiB image buffer.
Weak error recovery — one bad token can cascade.
Traps are $abort (rc 134), not a catchable Trap; DISPOSE is shallow.

10. Reproduce

cd .../m2compiler-V3
./bootstrap/fixpoint.sh                       # FIXPOINT OK
cd compiler && ./build.sh && ./run_tests.sh   # 175/175
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" \
  tools/v3-corpus/corpus.sh

language-report.md 6.7 KB 永久連結 文件歷史 原始文件