← board

A unit's {$define} leaks into the units it uses

The reduction, both compilers

unit ua;            unit ub;
{$define LEAKED}    interface function Which: string;
interface           implementation
uses ub;            function Which: string;
                    begin
                    {$ifdef LEAKED} Which := 'LEAKED-IS-DEFINED';
                    {$else}         Which := 'not-defined';
                    {$endif}
                    end;



fpc: via ua: not-defined
pxx: via ua: LEAKED-IS-DEFINED

FPC resets conditional symbols per unit; only -d symbols are global. pxx carries one table across the whole compilation, so whether a used unit sees a symbol depends on who compiled it first and in what order.

Why this is worse than a wrong answer

It does not error. It picks the other arm — so the used unit is built with a different interface, a different field set, or different types than FPC would build, and every later diagnostic is about the consequence. That is the same shape as the enumerated-predicate family: the observable is somewhere other than the cause. Compilation ORDER is also now semantically significant, which it is not in FPC, so the same unit can compile differently depending on which program pulls it in.

What has NOT been established

Nothing beyond the reduction. It was found while walking fcl-passrc rung 7, where pparser.pp defines four symbols that pscanner.pp also tests — but leaking those four into a driver that uses pscanner was measured NOT to reproduce any current corpus wall, so do not close this against a corpus green. The reduction is the specification; a fix needs a per-unit save and restore around a used unit's compilation, and a test that asserts the arm taken in BOTH directions (a symbol defined in the parent and absent in the child, and one defined in the child and absent in the parent afterwards).

Gate

Assert a VALUE from each arm, in both directions, and include the reverse leak: a symbol defined inside ub must not still be defined in ua after the uses. A one-directional test passes on a fix that clears the table instead of saving and restoring it.

Log

Resolution

Fixed by a COMMAND-LINE BASELINE: PasSnapshotDefineBaseline freezes the define table once, after -d and the target/frontend defines and before a byte of source is read; ParseUsesUnitBody calls PasResetDefinesToBaseline on the way IN, in addition to the save/restore it already did on the way out. {$CLAIM} survives by construction rather than by a site anyone remembered to patch -- it was deliberately built outside this table for exactly this reason.

The ticket's title names half of it. {$PACKRECORDS} leaked identically and that half is worse: a used unit's record a: Byte; b: LongInt was FIVE bytes where fpc builds eight. An ABI, in a unit that never asked, with no {$ifdef} anywhere in it. Found by pointing this ticket's own six-line reduction at the OTHER unsaved per-file globals -- the existing save list around a used unit covers CaseSensitiveMode, NestedComments, DelphiMode and PyExprMode and stops there. {$SCOPEDENUMS} was probed and did not leak; the rest of that list ({$H+}, {$ASSERTIONS}, {$R+}, {$Q+}, {$I-}) is unprobed and is a named blank, not a clear.

The root source is a DELIBERATE exception, and it is not FPC's rule

A MAIN program's or main unit's defines still reach the units it uses. pxx's own RTL is configured that way: {$undef PXX_MANAGED_STRING} on line 1 of a program selects the frozen-string model and compiler/builtin/*.pas -- units the compiler appends, that the user never wrote -- read it. Cutting it turned test_frozen_string_reentrant.pas into "call to a runtime stub that was never emitted" inside builtinheap.pas, which is what a configuration define silently losing its effect looks like from the far end. That file is the positive control for the exception and it is already wired in the QUICK tier.

The defect is UNIT-to-unit, where uses ORDER decides the answer. The root is one source, lexically ordered, and order-independent by construction. Anyone wanting a genuinely cross-unit symbol has {$CLAIM}.

Gate, as the ticket asked for it

test/test_a_units_define_and_packing_do_not_reach_the_units_it_uses.pas, seven rows, all byte-identical to fpc 3.2.2 -Mobjfpc -dCLIDEF -Futest/units. FOUR directions, because a fix that CLEARS the table instead of saving and restoring it passes the obvious two: down (the bug), up (the reverse leak the ticket asked for), in (a -d must still reach the child), and self (a unit's own define and its own packing must survive its own uses). The two pre-existing test_pascal_define_unit_scope_order* rows cover SIBLINGS and stay green; this is the parent-to-child direction they could not see.

gate.sh quick GREEN including both lib/rtl rows; the native tier at this tree is 2382/2387 with four reds, none of them this change (test_libwriteln_parity and test_promoint_bitwise are older open regressions, and the two generic rows are b2a41d5f4fb9, which does not contain this commit).