bug: a metaclass ARRAY ELEMENT is not accepted as a receiver — silent garbage

• Track: P (Pascal frontend)
• Found: 2026-07-13, while writing b317's regression test (fcl-json corpus). The test itself tripped over it, which is how it surfaced.

Repro (16 lines, verified against FPC)

program vc;
{$mode objfpc}{$H+}
type
  TBase = class class function Tag: string; virtual; end;
  TBaseClass = class of TBase;
  TA = class(TBase) class function Tag: string; override; end;
class function TBase.Tag: string; begin Tag := 'base'; end;
class function TA.Tag: string;    begin Tag := 'A'; end;
const
  Map : array[0..0] of TBaseClass = (TA);
var
  c: TBaseClass;
begin
  c := Map[0];
  writeln('via var:  ', c.Tag);      { FPC: A    pxx: A    }
  writeln('via elem: ', Map[0].Tag); { FPC: A    pxx: 4227577  <-- GARBAGE }
end.

Map[0].ClassName happens to work; Map[0].Tag (a virtual class method) returns an integer-looking value. Assigning the element to a metaclass VARIABLE first and calling through that is correct — so the value in the array is right; it is the RECEIVER handling that is wrong.

Silent: no error, no warning, a plausible-looking number.

Cause

The metaclass-receiver checks in compiler/parser.inc (the class-reference-op and metaclass-call paths, e.g. around the Syms[ASTIVal[node]].PtrElemTk = tyClass tests) only accept:

• ASTKind[node] = AN_IDENT — a class of T VARIABLE, and
• ASTKind[node] = AN_PTR_CAST — an inline metaclass cast.

An AN_INDEX (an element of an array of TSomeClass) matches neither, falls through, and the call is built as if the receiver were an ordinary pointer value.

Fix direction

Generalise the receiver test to any node whose STATIC type is a metaclass, rather than enumerating AST kinds. The array case needs the element's class id, which the symbol already carries for an array-of-metaclass (element tyPointer with a class element rec). ResolveNodeRec-style resolution is the natural home — the same "which class is this expression" question the record path already answers, and adding kinds one at a time here is exactly what left this hole (see also b289, the selector-after-indexed-property fix).

Note b317's regression test deliberately reads the const through ClassName and says why, so it does not depend on this being fixed.

RESOLVED 2026-07-13 (b328)

Exactly the fix direction above: the metaclass-receiver detection now also accepts an AN_INDEX over an array symbol whose element is a metaclass (the sym carries PtrElemTk=tyClass / PtrElemRec — a clean discriminator against arrays of ordinary record pointers). Virtual class methods, virtual constructors and the class-ref ops all dispatch through the element now. Pinned: test/test_metaclass_array_element_b328.pas. fpjson suite stays 203/203.

Log

• 2026-07-13 — resolved, commit HEAD.