M2compiler-V3

m2compiler-V3 — modern Modula-2, Coco/R frontend, QBE backend

A from-scratch Modula-2 compiler in the Blaise spirit (phased, test-guarded, self-hosting as the end goal): a Coco/R-generated frontend in GNU Modula-2, a fresh QBE SSA backend, and a length-prefixed runtime data model. Modern syntax follows the Benjamin R10 specs where they don't break classic code (Kowarsch/Redux cleanups: # not <>, NOT/AND/OR only, no B/C octals, 0x hex, // comments).

Prerequisites

• GNU Modula-2 (gm2 -fiso; developed with GCC 16.0.1 experimental)
• Coco/R CR V1.53 (CRBIN env var, defaults to the sibling CocoGm2/CR used here)
• qbe (default target amd64_sysv; pin: vendor/qbe-pin.txt)
• cc, timeout, standard shell tools

Build and test

cd compiler
./build.sh      # M2.atg -> scanner/parser/driver -> ./M2
./run_tests.sh  # M2 -> gen_ssa/*.ssa -> qbe -> cc -> exit codes

Pipeline per test: ./M2 tests/foo.mod writes gen_ssa/FOO.ssa; a global VAR ExitCode : INTEGER selects the process exit code (else 0). Reject tests assert on .LST messages; every error also prints as file:line:col: message on stdout. M2 itself exits 0 on success, 1 on compile errors and 2 on usage/IO problems (no input, unopenable file) — no aborts.

Layout (Blaise-adapted)

compiler/src/      M2.atg, frames, FileIO, SymTab, QbeGen, Lower (+ generated)
compiler/tests/    t_*.mod (check/reject/run) + showcase<N>.mod tours
compiler/gen_ssa/  build outputs (gitignored)
runtime/syslib/    always-linked runtime (SYSTEM, Storage, SysIO, …)
stdlib/            opt-in libraries (Strings, TextIO, …; classic form)
tools/             syntax helpers + v3-corpus (coverage harness)
bootstrap/         stage1/2/3 + fixpoint.sh (self-hosting proof)
vendor/            qbe toolchain pin
docs/              STATUS.md, plan.md, plan-*.md, analysis_*.md, summary_*.md

Status (2026-10-16)

• Suite 219/219; self-hosting fixpoint OK (byte-identical stage2/stage3, 3,714,376 bytes); corpus 372/612 compile-OK, 0 crashes (tools/v3-corpus).
• Active branch resolve-check (holds the AST/Lower/parse refactors and generics).

Lowers today: program modules; CONST/VAR/TYPE (scalars, subranges, enums, arrays, records incl. variants, sets, pointers); ARRAY/RECORD/SET/POINTER, WITH, IN; whole array/record copy; NEW/DISPOSE; all scalar arithmetic, relations, boolean connectives; IF/CASE/WHILE/REPEAT/LOOP/FOR, GOTO/labels; procedures with value/VAR params, functions, recursion, FORWARD, nested procedures (static-link display), procedure values; Clarion-form CLASS with methods, inheritance and virtual dispatch; PIM coroutines (PROCESS/NEWPROCESS/TRANSFER/IOTRANSFER); Unicode (UCHAR codepoints, UString, UTF-8 at the edges); ISO type conversions T(x) (incl. CONST folding); separate compilation (lib.def lib.mod prog.mod in one run → one image, FROM/IMPORT, qualified L.x, opaque-type completion); generics — ISO 10514-2 refinement form (GENERIC … MODULE M(T:TYPE) + MODULE X = M(actuals)) with type/value/procedure parameters, constraints (T : TYPE (ORDINAL|INTEGER|REAL|POINTER|SCALAR)), nested generics, and generic classes.

Architecture: a two-phase parse — each unit's procedure bodies are deferred and re-parsed at unit end, so forward references bind normally and the old forward-reference machinery is gone. There is no separate resolve/check pass; Lower is the only emitter and walks the AST. -notwophase restores the legacy inline parse (debug escape).

Long-term rules already locked: every ARRAY is LONGCARD count

• elements (all arrays, inline, element counts, no NUL); UCHAR = 32-bit codepoint, UString = count + codepoints, UTF-8 at the edges with hard lex errors; self-hosting bar is byte-identical .ssa over three stages.

Working conventions

• Grammar grows section by section in compiler/src/M2.atg (kept LL(1)-clean — Coco/R must report no errors).
• Every step keeps the full suite green; per-session record in docs/summary_*.md; showcase<N>.mod tours each stage. See docs/INDEX.md for the load-bearing docs.
• Tags v3-* mark green states; no pushes without asking (see git.txt for the remote recipe).