← board

Multidimensional fixed arrays

Done (commits 48401d9 2-D, 121c913 N-D)

2-D fixed arrays in var sections landed via the flatten-to-1-D design below: both m[i,j] and m[i][j] fold into one linear AN_INDEX at parse time (dims in parallel arrays SymArr2*), so the 1-D path handles codegen unchanged. test_cross_multidim byte-identical on all 4 targets, matches FPC. The single chokepoint was ParseLValueAST (both reads and writes of identifier indexing route through it), so the "~6 sites" worry was overstated.

Follow-ups landed same day:

ALL landed (2026-06-16): param arrays (const+var, commit), N-D to arbitrary rank (121c913), dynamic-array type aliases (separate commit). The "## Gap / ## Scope / ## Recommended design" sections below are the original plan, kept for history — fully delivered.

Gap

Neither multidim fixed-array syntax parses:

var m: array[0..2, 0..2] of Integer;   { Expected: ] }
var m: array[0..2] of array[0..2] of Integer;   { Expected: begin }

and the corresponding m[i, j] / m[i][j] indexing. FPC accepts both and they are common in user code. Single-dimension fixed arrays and nested dynamic arrays both work today; only nested fixed arrays are missing.

Scope

Recommended design (flatten to 1D at parse time — zero codegen change)

Storage and indexing both reduce to the existing 1D array path, so no backend work is needed (works on all four targets for free):

  1. Type parse (ParseVarSection / typed-const / type-decl array branches): parse array[lo1..hi1, lo2..hi2] of T and array[lo1..hi1] of array[lo2..hi2] of T. Allocate a 1D array via AllocArray(name, T, 0, span1*span2 - 1) where spanK = hiK-loK+1. Record the dims in parallel arrays keyed by sym (do NOT add TSymbol fields — MAX_UFIELD landmine; mirror SymDynDepth): SymArr2[sym]:Boolean, SymArr2Lo1/Span1/Lo2/Span2[sym]:Integer.
  2. Index m[i,j] and m[i][j]: build the flattened 1D index expression (i - lo1) * span2 + (j - lo2) and a single AN_INDEX(m, flatExpr). Both forms collapse to the same thing (FPC treats them identically).

Index-parse sites to touch (the spread-out risk — must be consistent)

All build AN_INDEX and currently parse exactly one [expr]; each needs to accept a comma ([i, j]) and, when the base sym has SymArr2, emit the flattened index. Centralise the flatten in one helper BuildArr2Index(baseNode, sym) called from each:

For m[i][j] (nested AN_INDEX), detect a 2-level index on a SymArr2 base and fold the two into one flattened AN_INDEX at parse time (the inner m[i] row value is not separately representable under flattening — error if used alone).

Build a test_cross_multidim (read+write, both syntaxes, non-zero los, Int64 elems) and wire into all four suites. Because flattening keeps it on the 1D path, expect byte-identical across targets immediately.

Notes