{$H-} is accepted and ignored, and it is a record-layout change

Split from [[feature-p-packenum-and-h-minus-for-the-fpc-compiler-corpus]] when the {$PACKENUM} half landed. That ticket said "whether that is a problem for pxx specifically needs measuring — pxx's string model is not FPC's — and that measurement is half this ticket", and said to split if only one half turned out real. Both are real. Here is the measurement.

Measured, fpc 3.2.2 vs pxx, 2026-09-05, compiler 89f51a99f0b3

type TR = record s: string; n: Integer; end;

So {$H+} already agrees and {$H-} does not.

The first probe could not fail, and that is worth recording

Measured WITHOUT {$mode objfpc} first, and fpc answered 256 in both arms — so {$H-} looked like a no-op and the gap was invisible. In plain mode a bare string is ALREADY a ShortString, i.e. {$H-} is the default there and the directive changes nothing it could be caught changing. The discriminating probe needs {$mode objfpc}, which is what fpcdefs.inc actually sets one line earlier. A probe run in the wrong mode returns a real number, from a real compiler, about a configuration where the question does not exist.

Why it matters at corpus scale

/usr/share/fpcsrc/3.2.2/compiler/fpcdefs.inc lines 1-3 are {$mode objfpc}, {$asmmode default}, {$H-} — and that include is pulled into essentially every unit of the FPC compiler. So under --mimic-fpc-compiler every bare string in those sources is a managed handle where the source declared a 256-byte inline buffer. Any record crossing between a pxx-built unit and a differently-built one disagrees about where every field after the string lives. Same failure shape as the enum half, and a much bigger field.

Where to start

BareStringKind (util.inc) is the one function that answers what a bare scalar string IS, and its header already says it was consolidated there precisely so there is one answer. tyShortString (ordinal 25) already exists. The shape of the fix is the one {$PACKENUM} just used: a per-token TokHMinus snapshot, because directives run in the LEX pass — a global read at parse time gives the last {$H} in the file to every declaration in it, and a file with a single directive cannot tell the two apart.

Gate

SizeOf of a bare string and of a record containing one, under {$mode objfpc} with {$H+} and {$H-}, against fpc 3.2.2 — both arms, since the {$H+} arm is the control that already passes and the {$H-} arm is the claim. A single-arm test cannot distinguish "implemented" from "the default already agreed".

Related

• [[feature-p-packenum-and-h-minus-for-the-fpc-compiler-corpus]] — the other half, landed.
• [[goal-compile-fpc-compiler]] / [[feature-mimic-fpc-compiler-define-profile]]

Landed (frankB, 2026-09-05, compiler ba573b6cf02a)

{$H-} / {$H+} / {$LONGSTRINGS OFF} / {$LONGSTRINGS ON} all implemented, positionally, on the same per-token discipline {$PACKENUM} uses (TokHMinus[]). BareStringKind — already the single answer to what a bare scalar string IS — now consults the token being parsed.

The kind was right and the type was still wrong

The first cut set the KIND correctly at every declaration site — instrumenting BareStringKind showed it returning tyShortString on every call — and SizeOf still answered 8 for a variable. A ShortString's CAPACITY has no carrier in the kind: string[N] sets LastTypeStrCap from N and the shortstring NAME sets 255, so a bare string under {$H-} is the third spelling of the same type and had to set it too. The variable was allocated with capacity 0.

SizeOf(string) then needed a second fix for the same missing fact: TypeSlotSize takes a kind and no capacity, and the shortstring name never reaches it (BuiltinTypeNameNeedsDecl routes that spelling through ParseTypeKind). Two sites, one omission — the kind and the capacity are one fact and neither half is sufficient.

Measured

Five declaration paths, all against fpc 3.2.2, all 256 under {$H-}: a global, a type TS = string alias, a record field, a local, and SizeOf(string) itself. Values survive both models — concatenation, Length, a const string parameter and result, Copy, Pos, and a record field of each model in one program. test_h_minus_shortstring, .expected is fpc's own output.

A separate divergence, measured and deliberately NOT changed

{$mode objfpc} does not imply {$H+} in fpc — only {$mode delphi} does:

So pxx's default bare string matches fpc only in delphi mode. That is the managed-string model this compiler chose, pre-dates this ticket, and is not touched here — flipping a default is not what a directive ticket may do. It is recorded because the row is easy to meet later and misread as this fix leaking. The corpus is unaffected either way: fpcdefs.inc says {$H-} explicitly.

On the corpus it was filed for

program fd2; {$i fpcdefs.inc}  var s: string;
  pxx --mimic-fpc-compiler : 256      fpc 3.2.2 : 256

Correctness, not reach — same as the enum half. The march is where it was: cutils, globtype, constexp, version, cstreams COMPILE; four stop at TFPCHeapStatus, cmsgs at its object constructor. Those five compiled before too, with a managed handle where their own source said ShortString.

Gate

test_h_minus_shortstring in test-core, BOTH arms — the {$H+} rows would pass against a discarded directive, so only the {$H-} rows are the claim — plus a switch back to {$H+} half way down the file, since a single-directive file cannot catch a global read. gate.sh quick GREEN, FPC seed canary PASS.

Log

• 2026-09-05 — resolved; this names the commit that carried the resolve, which is not always the one that carried the change — commit 8a3a62258.