← board

Pascal corpus rung 2 — fgl (real FPC generic containers)

What this rung is

rtl/objpas/fgl.pp is FPC's generic-container unit — TFPSList, TFPGList<T>, TFPGObjectList<T>, TFPGInterfacedObjectList<T>, TFPGMap<K,V>, TFPGMapObject<K,V>, plus their enumerators. ~2,000 lines of real, widely-used library source that leans on generics × classes × interfaces × pointer-level container plumbing all at once. It is the smallest piece of real Object Pascal that exercises that intersection, which is why it sits directly above the FPC conformance sweep on the ladder.

Why it needed doing (the inversion)

fgl was already named as a compat target and there was already a driver, test/test_fgl_use.pas, wired into test-core. But its guard was:

@if [ -d /usr/share/fpcsrc/3.2.2/rtl/objpas ]; then ... \
else echo "fgl(real FPC source): SKIP (no fpcsrc)"; fi

/usr/share/fpcsrc is a distro FPC-source package that is not installed on the dev box, the watcher box, or in a fresh clone. So the flagship FPC-compatibility check had been printing SKIP (no fpcsrc) and passing — green, and measuring nothing. Nothing fetched the FPC RTL sources either; the installer fetched FPC's test suite from a pinned commit but not its RTL.

What landed

  1. Fetcher — tools/install_lib_candidates.sh fpc-rtl: sparse rtl/objpas + rtl/inc + LICENSE from the same pinned FPC commit the testsuite fetch already used (0d122c49…, release_3_2_2), into the gitignored library_candidates/fpc-rtl/, with PROVENANCE.md. Also added to all.
  2. Drivers — test/fgl/*.pas, one per container: fpslist, list_int, list_str, map_int, map_str, objectlist, ifclist.
  3. Oracle — test/fgl/*.expected, produced by compiling those same drivers against that same fgl.pp with FPC 3.2.2 itself. All seven build and run correctly under FPC, so every red below is a pxx gap and never a bad driver. Note the exact claim: this is behavioural parity with the reference compiler on real library code. It says nothing about the machine code we emit — see the claims-discipline section of CLAUDE.md.
  4. Runner — tools/run_fgl_corpus.sh [compiler] [objpas-dir], modelled on tools/run_c_conformance.sh: self-skips loudly when the tree is absent (falling back to a system fpcsrc if one happens to exist), honours test/fgl/pxx.skip (name.pas<TAB>reason, every entry naming its ticket), prints test-fgl: PASS/FAIL/SKIP.
  5. Make target — test-fgl, spelling out library_candidates/fpc-rtl so testmgr's CORPUS_RE sees it and the job self-skips loudly rather than silently greening. test-core's inline check now prefers the fetched tree, so it runs everywhere instead of skipping.

Baseline 2026-08-25 (pxx VERSION 374 / self-hosted HEAD): 3 pass, 4 known-fail

driver state wall
fpslist PASS —
list_int PASS —
map_int PASS —
list_str red [[bug-p-a-string-typecast-is-a-conversion-and-not-a-cast]] (fgl.pp:892)
map_str red same ticket (fgl.pp:1602)
objectlist red [[bug-p-inherited-ignores-the-parents-default-parameter-values]] (fgl.pp:1061)
ifclist red [[bug-p-a-cast-as-lvalue-does-not-accept-a-builtin-type-name]] (fgl.pp:1189)

Three tickets unlock four of the seven drivers, and each is a narrow double-case defect where the sibling path already works — the shape devdocs/dev/normalise-dont-special-case.md describes. Two of them (String(…) as a cast, inherited + defaults) are between them what stops string-keyed containers and the standard owning-container constructor idiom — i.e. most of what real Object Pascal code does with fgl.

Also found while bringing this up, filed separately because it is not an fgl wall: [[bug-p-stray-tokens-in-a-unit-declaration-section-are-silently-skipped]] and [[bug-p-a-diagnostic-in-a-used-unit-names-the-wrong-source-file]].

Not done here

Gate

make compiler/pascal26 (self-host fixedpoint) + tools/run_fgl_corpus.sh ./compiler/pascal26 + tools/gate.sh quick.

Parent: [[feature-pascal-corpus-expansion]] · mechanism copied from [[feature-c-corpus-expansion]] / tools/run_c_conformance.sh

Log