Handoff — L3 via option 2: 183/183 but self-host blocked by V3 codegen

Branch ast-stage-c, HEAD edf24d9. Suite 182/182; fixpoint OK (3,147,849 bytes). Clean except toto.mod.

Headline

Option 2 (zero SymTab surface) is functionally complete and byte-verified: t_lower3 matches the legacy image and the suite reaches 183/183 with lower_ok t_lower3. It still does not self-host, and the cause is now a different V3 codegen bug, not the table design.

What was built and verified this session

compiler/src/Lower.def / Lower.mod:

Owned, flat tables (no arrays of records — see the bug below): symName : ARRAY OF CHAR (stride 64), symKind, symTy, and procs : ARRAY OF ProcEnt with a flat pname (stride 64). Helpers Reset, NoteVar, NoteParam, NoteProc, SetProcRes, no-op ScopeEnter/ScopeLeave, CopyN/CopySpan, FindSym/ FindProc/Resolve/ProcInfo/ParamInfo. No SymTab edits.
All SymTab.ProcUid/ProcNPar/ParamType/... call-sites switched to ProcInfo/ParamInfo; LowerDesign/IsSimpleScalar/IsCall to Resolve/FindProc.
Grammar hooks (inert): M2 root Lower.Reset; VarDecl/ConstDecl Lower.NoteVar; ProcHeading Lower.ScopeEnter+NoteProc+ SetProcRes; ParamSection Lower.NoteParam/NoteVar; ProcDecl Lower.ScopeLeave + AST.SetOp(astNode, 0).
ScopeLeave is a no-op: parse-time notes must survive until the end-of-unit walk.

Result: t_lower3 (function + VAR-param procedure + calls + RETURN) byte-matches; suite 183/183.

New finding — nested-array record field miscompiles

Repro docs/wip/nested-array-field-bug.mod:

TYPE PE = RECORD
  name  : ARRAY [0..63] OF CHAR;
  pname : ARRAY [0..3] OF ARRAY [0..63] OF CHAR;   (* nested *)
END;
VAR procs : ARRAY [0..3] OF PE;
...
procs[nProcs-1].pname[j][i] := name[i];

./M2 compiles it, but the generated program segfaults (rc 139; want 42). A flat pname : ARRAY [0..2047] OF CHAR (manual stride) works. This is a real user-facing bug — likely the element stride / descriptor-pointer setup for a nested-array field (same family as the Stage-A array-of-record stride fix in QbeGen.ElemSize, and docs/wip/local-array-bug.mod).

It was hit directly: the V3-built compiler (m2s2) crashed on every input when Lower's formal-name table used a nested-array field. Flattening fixed that crash.

Why self-host still fails

After flattening, m2s2 works on small programs (t_exit, a CONST program) but still crashes (SIGABRT) on the full compiler input (empty output). Shrinking the tables (MaxSym 2048→512, MaxProc 512→64) did not help, so it is not table size — it is a size/complexity-triggered latent V3 codegen bug in the compiled Lower/M2P code. HEAD self-hosts; the ~250 added lines tip it over.

This is the third instance of the same compiler fragility (local arrays, nested-array fields, and now this). Finding the exact construct needs a bisect of the added code with a V3-built compiler — a dedicated task.

Next steps

Fix the nested-array-field bug in QbeGen (like the Stage-A stride fix): make arr[i].field[j][k] correct. Add the repro to the suite. This is a real, self-contained bug and may also matter for the size-triggered failure.
Bisect the size-triggered self-host crash: build m2s2, then compile subsets of the added Lower/M2P code with it to find the construct that crashes. Candidates: the flat-stride indexing, the CopySpan offsets, or a specific grammar hook.
Once self-host is green: keep lower_ok t_lower3, expect 183/183 + FIXPOINT OK.

Do NOT

No FORWARD (breaks gm2 whole-program pass).
No SymTab edits.
No arrays of records with aggregate (nested-array) fields in Lower.
No committed ASTDUMP.

Resume

cd ~/Projets/Projets-Modula2/MyWork/m2compiler-V3
git checkout ast-stage-c
cd compiler && ./build.sh && ./run_tests.sh   # 182/182
cd .. && ./bootstrap/fixpoint.sh              # FIXPOINT OK

handoff-l3-option2.md 4.0 KB Историја Датотека