← board

A record-shaped var initialiser is refused

type TRec = record n: Integer; ev: TEv; end;
var R: TRec = (n: 7; ev: nil);



pascal26:6: error: parenthesised initializer requires an array variable

FPC 3.2.2 accepts it and initialises the fields. Found 2026-08-21 while sweeping the sibling positions for [[bug-a-nil-is-not-accepted-as-a-method-pointer-argument]] — and deliberately NOT folded into that fix, because it is not about nil at all: the same declaration with (n: 7; ev: SomeMethod) or with no reference field whatever is refused just the same. The nil ticket only found it.

Why this is a small job

const R: TRec = (n: 7; ev: nil) already works, including the local-const and array-of-record forms, and lands its fields through the PendingInit* / LocalInit* channel (pasparser_decl.inc, the "Record typed constant" arm). An initialised var is the same parse and the same channel — FPC treats a var initialiser as a writable typed const — so this is a routing gap, not new machinery.

The diagnostic is the giveaway: "requires an ARRAY variable" says the var path learned parenthesised initialisers for arrays only and the record arm was never wired up beside it. That is the double case devdocs/dev/normalise-dont-special-case.md describes, in its usual form — the second path is the one that stayed broken.

Check the siblings before closing: a var of an ARRAY OF RECORD with a parenthesised initialiser, and a LOCAL var with one.

Gate

A test asserting field values for a record var, an array-of-record var and a local one, each against FPC 3.2.2. Self-host byte-identical + tools/gate.sh quick.

RESOLVED 2026-08-25 — and the sibling check turned it into a de-duplication

The ticket's read was right: a routing gap, not new machinery. What it did not know is that the record-field loop had been hand-written three times:

  1. the scalar typed-const arm in ParseConstSection (~110 lines),
  2. the ARRAY-of-record element arm in the same routine (~85 more, subtly different — it handled strings explicitly instead of through TryParseInitValForm, and had no array-valued-field support),
  3. …and the var arm, which is the one that did not exist at all.

Three mechanisms for one concept is the line devdocs/dev/root-cause-over-microfix.md draws between a smell and a design flaw, so the fix is an extraction rather than a fourth copy:

procedure ParseRecordInitializerInto(symIdx, recId, elemIdx: Integer; isLocal: Boolean);

elemIdx is PI_ELEM_NONE for a plain record and the flat element index inside an array — the emitter already combined Elem + FOff, so the array case was never anything but a parameter. Four call sites now: const scalar, const array-of-record element, var scalar, var array-of-record element.

The sibling check the ticket demanded found a real second bug. With the scalar var form fixed, var A: TArr = ((n: 1; s: 'a'), (n: 2; s: 'b')) failed with "not a constant": the var section's initialiser loop tracks paren DEPTH for N-D arrays, so it read the element's ( as another dimension and then met n where it wanted a value. The new arm has to be tested before the depth-tracking one, and is.

Local var L: TP = (x: 9; y: 8) works through the LocalInit* channel, same as a local typed const.

Known limitation, inherited and unchanged: a LOCAL record initialiser with a STRING field is refused — "record constant with string fields must be global" — because a local init slot has no ValAux and its FLen is already spent on the field-name span. That is the local typed-const rule, it is loud, and lifting it means widening the local init row rather than touching this path. The test uses an all-ordinal record for its local row and says why.

Test: test/test_a_record_typed_var_initialiser.pas, wired into test-core, .expected = fpc 3.2.2's own output. It asserts the two var forms, the two CONST forms (the extraction rewrote the path they take, so they are as much at risk as the rows it adds), an uninitialised record var still zeroed, the local form, and a written row — writing through R and A[1] afterwards, which is the whole difference between an initialised var and a const.

Gate: make compiler/pascal26 converged in 1 round, tools/gate.sh quick GREEN, fpc-testsuite unmoved.

Log