← board

Why this is worse than a refusal

The sizeof door's own rule is that a name it cannot size is a DIAGNOSTIC, not a default — its comment says so, and for the non-quiet path it is true: it raises sizeof cannot size this type here. The pre-pass path is different. Under PasCondQuiet an unresolved sizeof sets PasCondUnresolved and answers 0, and 0 is neither 2 nor 4, so every arm of a width ladder tests false and control reaches the {$else}.

In ordinary source that arm is a fallback. In a compiler's source it is routinely {$error} — the author's way of saying "this configuration is not supported" — so the failure presents as a confident statement about an unsupported width, and the reader goes looking at widths.

What is measured, and what is not

Measured 2026-09-16 with tools/fpc_compiler_corpus_probe.sh:

Measured and does NOT reproduce — three separate reductions, all green:

probe result
$elseif chain, local alias size 2, matches fpc
sizeof of a local = word alias correct
cross-unit = word alias in a usesd unit size 2, matches fpc

So the three obvious causes are all excluded and the remaining variable is the CONTEXT the corpus probe creates, not the construct.

The residual question, and who owns it

"Not the construct" is half a finding. The owner of "then what" is whoever takes this: confirm or refute the probe-nesting hypothesis by instrumenting whether PasCondSizeOfNameOrAlias is reached at ProbeDepth > 0 for tcompilerwidechar, and if so decide whether the pre-pass should answer a width ladder at all — a directive it cannot evaluate arguably must not silently pick the fallback arm. Refusing to choose an arm is available and is probably right, since the non-quiet path already refuses.

Behind it sits [[feature-b-rtl-has-no-tdoublerec]] for ncnv, already p85.

FIXED 2026-09-16 (frankb-56) — and this ticket's own hypothesis was WRONG

Filed hours earlier by me with "the reduction does not reproduce" and a probe-nesting hypothesis. Both halves needed correcting, and the correction came from one more probe rather than from more thought.

IT REPRODUCES. The missing variable was a THIRD LEVEL, twenty lines:
prepasserr_base type tprepasswidechar = word;
prepasserr_mid uses prepasserr_base + the {$if sizeof} / {$elseif} / {$else} {$error} ladder
program uses prepasserr_mid

fpc 3.2.2 answers 2; pxx raised Unsupported tprepasswidechar size. Move the type into prepasserr_mid and the same ladder answers correctly — that is why all three of the original reductions came back green: every one of them had the type within one hop.

THE STATED MECHANISM WAS WRONG IN A CHECKABLE WAY. This ticket said an unresolvable sizeof "answers 0, and 0 is neither 2 nor 4". Probed directly — {$if sizeof(tprepasswidechar) = 0} — it takes the ELSE arm too. So it does not answer 0 and no comparison happens at all: PasCondUnresolved turns the WHOLE expression False, and False for every arm is what walks control to the {$else}. The distinction matters for the fix: nothing is wrong with the arithmetic, and a fixture asserting = 0 would have pinned the wrong mechanism.

THE FIX IS THE FILE'S OWN CONTRACT, NOT A NEW RULE. PasCondQuiet's declaration already says the pre-pass "must not HALT there either -- the real pass is the authority and will raise the error itself", and its RESIDUAL paragraph predicted this exact class while naming the tell as an INCLUDE inside such an arm. An {$error} is the same residual through a louder door. {$error} is now gated on PasCondQuiet; the real pass evaluates the same text and raises it with a position.

THE CONTROL GUARDS THE WORSE DIRECTION. Swallowing a real {$error} would be a worse defect than the false halt, since the directive exists to stop a build. Asserted as a must-not-compile row that also checks the message survives. And the fixture is a guard rather than a decoration: the PINNED compiler refuses it with Unsupported tprepasswidechar size.

CORPUS EFFECT, and it is forward progress rather than a unit. nld now advances past this wall to conditional directive: the right operand of 'in' is not a set constant this pass can read: 'supported_optimizerswitches' — which is the documented decline of the in-over-a-set work, its nested cross-unit term limit. A LOUD decline in place of a wrong branch that read like a true statement about the program. ncnv is unchanged at TDoubleRec.

Gate: gate.sh quick GREEN. Byte-identical to fpc 3.2.2.

Log