← board

{$IFOPT X+} is hardwired False, so the wrong arm compiles

{$IFOPT} is a CONDITIONAL. Getting it wrong does not produce a diagnostic — it compiles the other branch. That is the same class as [[bug-p-macpas-conditional-directives-are-ignored-so-both-arms-compile]] and it is worse than a mis-set switch, because nothing downstream can tell that a branch was chosen rather than written.

Found by censusing fpc 3.2.2's own sources for directive VALUES rather than NAMES (the unknown-directive warning keys on the name, so a known directive with an unhandled value is invisible to it — see [[bug-p-a-spurious-unknown-directive-warning-cannot-fail-any-test-we-have]]). {$IFOPT} is 139 uses there.

What was fixed, and why the rest was not

ProcessPasDirective said

if CaseEqual(command, 'ifopt') then
  cond := False { compiler option switches are not modelled; branch off }

and the comment was stale — RChecksVal, QChecksVal, IChecksVal and AssertionsVal are set a few lines below it, in the same procedure. Those four now answer from their own state. They are also where the corpus is: {$IFOPT R+} alone is 113 of the 139 uses.

{$IFOPT R+} was RIGHT BEFORE THE FIX, by coincidence — fpc's R defaults off, and "always False" agrees with "off". The uses that dominate the corpus are exactly the ones a broken implementation gets right, which is why this survived: the discriminating case is {$R+} followed by {$IFOPT R+}, where fpc says TRUE and pxx said FALSE.

The open part: letters with no variable but a fixed behaviour

Measured against fpc 3.2.2 on x86-64, 2026-09-05 — one probe per letter for the default, one for whether an explicit {$L+} moves it:

default ON default OFF not tracked by fpc at all
letters I G J X Z R Q C B D H M P S T V A L O

Z was closed 2026-09-05 — {$PACKENUM} landed and gave the letter a variable, so {$IFOPT Z+} now answers PackEnumVal = 4. Measured across every spelling against fpc 3.2.2, both signs, with the dead A beside it as the control: (none) ON, $Z1 OFF, $Z2 OFF, $Z4 ON, $Z+ ON, $Z- OFF, $PACKENUM 1 OFF, $PACKENUM 4 ON. A and Z are both NUMERIC switches and only a probe separates them — A is dead and Z is live — so the analogy that would have settled either one settles neither. Rows are in test_ifopt_tracks_the_switch_it_names, whose .expected is fpc's own output.

G, J and X remain, and they are the ones the table above was really about.

A not being tracked is the useful negative: it is a NUMERIC switch, so {$IFOPT A+} is false however {$A8} was set. Modelling it from PackRecordsVal would have been wrong, and only the measurement says so.

pxx answers False for every letter, so it is wrong by default for the five in the first column. {$IFOPT X+} guarding extended-syntax code takes the else arm on a compiler that has extended syntax.

Each of those needs its own claim — does pxx behave as X+? — and that is a per-letter determination about this compiler, not a lookup in fpc's table. Two that are not obvious: H (ansistrings) is mode-dependent in fpc, and G (286 code generation) is a switch whose subject no longer exists.

Where pxx will DIFFER from fpc on purpose

{$IFOPT C+} answers TRUE by default here and FALSE in fpc, because AssertionsVal defaults ON — a decision already recorded at its declaration in defs.inc (flipping it to fpc's default would turn every existing pxx assertion into dead code silently). IFOPT reports this compiler's state truthfully; matching fpc's answer would mean lying about ours.

Not part of this, deliberately

fpc diagnoses a malformed {$IFOPT} — {$IFOPT R} and {$IFOPT R*} are Wrong switch toggle, {$IFOPT RR+} and {$IFOPT 9+} are Illegal compiler switch — and pxx silently answers False for all of them. Us accepting what fpc rejects is not a defect, and a differing diagnostic is deferred; recorded so the next reader does not have to re-measure it. The letter IS case-insensitive in fpc ({$IFOPT r+} works), which the fix does honour.

Still reproduces at HEAD — G, J and X all answer OFF (frankS, 2026-09-05)

Measured at 0bbd82cd7, compiler/pascal26 sha 7fca108e4b85 (converged after 1 round(s)). One program per letter, {$ifopt L+} WriteLn('L ON') {$else} WriteLn('L OFF') {$endif}:

letter HEAD
G G OFF the residual — fpc defaults it ON
J J OFF residual
X X OFF residual
A A OFF the documented negative: numeric and untracked, so OFF is CORRECT
R R ON with {$R+}, R OFF with {$R-} already fixed, tracks both ways

Read the R row as a warning, not as a result. A staleness probe written from this ticket's SLUG picks R or Q — a letter the summary already records as fixed — and comes back green on a ticket whose residual is untouched. It is a control drawn from the wrong population, and it very nearly closed this ticket today. The summary is where the residual lives; the slug is a year out of date by construction, because a slug cannot be edited without breaking citations.

The residual stands as the summary states it: G, J, X, each needing its own measured behavioural claim about pxx.

Log