← board

SetLength(a, x, y) one-call multidim allocation not parsed

Symptom

var a: array of array of Integer;
begin SetLength(a, 2, 3); ... end.

→ pascal26:2: error: unexpected token. FPC allows SetLength(arr, d1, d2, ...) to allocate a rectangular jagged array in one call (each extra dimension sizes the sub-arrays). PXX's SetLength parser takes exactly (lvalue, count) and chokes on the second size.

Workaround (works today)

Per-row: SetLength(a, 2); for i := 0 to 1 do SetLength(a[i], 3); — the torture test uses this. So the capability exists; only the one-call sugar is missing.

Fix sketch

ParseFactor/statement SetLength branch (parser.inc ~8122): after the first size expr, accept further , <expr> dimensions and lower to a per-dimension init loop (or a runtime helper) that sizes the nested dynarrays. Rectangular only (FPC semantics). Front-end + a small lowering.

Acceptance

SetLength(a, 2, 3) allocates a 2×3 jagged array; Length(a)=2, Length(a[i])=3; existing 2-arg SetLength unchanged; self-host byte-identical.

Done (2026-06-30, Track A)

SetLength(a, d0, d1, ..., dk) now allocates a rectangular jagged array in one call. Implemented as parse-time AST synthesis (parser.inc): the SetLength statement branch collects all size exprs into SetLenDimNode[], then BuildMultiSetLen / BuildSetLenNest build SetLength(a, d0); for i:=0 to d0-1 do (SetLength(a[i], d1); for j:=... ) — recursive over dimensions, the sub-array targets cloned via CloneAST (the defs.inc "AN_INDEX.Left = sym idx" comment is stale; it is a node). MkForAsc synthesises the ascending AN_FOR. Multidim on a non-dynamic-array target (incl. a managed string) errors clearly at parse time.

Verified: 2-D / 3-D allocate correct extents + contents; managed-element rows (array of array of string); record-field target (SetLength(r.m, 2, 3)); single-dim unchanged; string-target multidim rejected. Self-host byte-identical; make test green (torture test cases 25-27).

Known limitation (documented, low impact): the size expressions are re-cloned per use (the outer call + the loop bound), so a dimension with side effects — SetLength(a, f(), g()) — is evaluated more than once. Real code uses constants/variables (side-effect-free), the FPC norm. A once-eval-into-temps refinement is a small follow-up if ever needed.

Cross-target note

Multidim (and per-row) nested dynarrays work on x86-64 but segfault on arm32 / aarch64 / i386 — a pre-existing nested-dynarray codegen gap on the cross backends, NOT caused by this feature (a per-row 2-D dynarray with no multidim sugar crashes there too). Filed separately: [[bug-nested-dynarray-cross-segfault]]. The torture test exercises multidim on x86-64 only (in make test); single-level dynarrays remain green cross.