Cross targets: a string[N] field poisons every pointer store to its record

• Type: bug (backend gap — loud, compile-time). Track A (codegen: ir_codegen386.inc, ir_codegen_aarch64.inc, ir_codegen_arm32.inc).
• Status: done
• Opened: 2026-07-14
• Found by: Track T — tools/pasmith_run.py --cross on the widened grammar ([[feature-pasmith-widen-grammar]]); every generated program with the record rung was rejected by all three cross targets. T owns the tool, never the bug.
• Related: [[bug-frozen-string-unsupported-riscv32-xtensa]] (b345) — same family: string[N] support outside the x86-64 backend.

Repro

program p4;
{$mode objfpc}
type
  PR = ^TR;
  TR = record a: longint; s: string[8]; end;   { the string[8] field is the trigger }
var r: TR; p: PR;
begin
  p := @r;
  r.s := '';
  p^.a := 1;              { storing the LONGINT field -- not the string }
  writeln(r.a);
end.

What is and isn't broken

Removing the string[8] field makes it compile everywhere. So it is not the store's own type that is unsupported — p^.a is a plain longint:

The presence of the field is enough. The guard is at ir_codegen386.inc:3278 (and the two ARM equivalents): it rejects any store whose type kind is outside {ordinal, boolean, pointer, char, class, unknown}, and a shortstring-bearing record apparently reaches it with the record's own kind rather than the stored field's.

Impact

Fuzzing the record rung against the cross-target oracle is impossible today — every program is rejected before it runs, so the backend-divergence oracle sees nothing at all for the entire class of linked/record-heavy programs. Beyond the fuzzer, it means no 32-bit or ARM program can hold a string[N] inside a record it reaches through a pointer, which is an ordinary shape (a node in a list, a record on the heap).

Acceptance

• The repro compiles and prints 1 on i386, aarch64 and arm32.
• p^.s := 'abc' (storing the shortstring itself through the pointer) works on all three.
• A cross-target regression test covers a heap record with a string[N] field.
• tools/pasmith_run.py --cross --wide then runs with --shorts on (it currently has to drop the shortstring rung for cross runs).

Log

• 2026-07-14 — resolved, commit 7716bd2a.