Trealla with Fil-C

The Fil-C package tracks upstream commit 98a821db66715618e358c36c7d7164213c0f6820, HEAD when checked on November 10, 2026. The previous pins were 12f4cbd7 (September 5, 2026, one commit after v3.9.40; 630 commits earlier) and v2.84.14 (November 5, 2025). All were built with the existing Fil-C toolchain; LLVM and the runtime were unchanged.

nix build .#pkgsFilc.trealla
./result/bin/tpl
nix build .#checks.x86_64-linux.trealla

Packaging changes

• Adapt nixpkgs' makefile substitutions to GNUmakefile.
• Select READLINE=1 explicitly; the new default is editline.
• Set the installed library search path and install the Prolog source library tree, including subdirectories. The hosted build also embeds many libraries.
• Refresh the existing FFI handle patch against the new source layout.
• Preserve tabling iterator pointers across backtracking.

The new tabling patch fixes two pointer-to-integer-to-pointer round trips in bif_tbl_get_answer_2 and bif_tbl_wkl_work_3. Each now stores its cursor and generation in a pointer-typed member of the existing saved-state union. Choice-point copies consequently preserve the cursor's capability. The generation checks that reject invalidated enumerations remain intact. This is ordinary C and does not require a Fil-C-specific escape hatch.

November 2026 update

• Both Fil-C patches are redone against the new source. The tabling patch now covers a third cursor, and the FFI handle patch covers the new '$dlsym_deref'/3.
• trealla-runtime-atom-arity-unify.patch is dropped: upstream's unify.c now refuses to unify a run-time atom with a compound of the same name.
• trealla-goal-expansion-share-cells.patch stays. Unpatched HEAD still reads a freed bigint in the runtime check's bigint_expansion.pl.
• trealla-heap-page-mark.patch is new and fixes wrong answers in unpatched HEAD. A clause that builds a term and tail-calls with it could have its frame reused and the term cut short when the term began a new heap page; msort(Xs, Ys), merged(Ys, Zs) then failed for some list lengths, which stopped experiments/regex-nanofil/rx.pl on the pattern word. Upstream ccb10b7d (October 12) exposed it by allowing more tail calls; the cause is an older off-by-one in how a heap mark names its page. The patch describes it, the runtime check has the reproduction, and upstream's 688 tests pass with it. It is not reported upstream yet.
• Upstream added foreign callbacks (foreign_callback/4, libffi closures) and passes more C pointers through Prolog integers. Those paths are not converted to handles and are untested under Fil-C; a pointer that travels through an integer loses its capability, so they would trap rather than work.
• (1<<64) /\ \1 now evaluates correctly and atom_number/2 fails on a non-number instead of throwing, two defects nanofil had worked around (nanofil/docs/archive/hosts.md).

Validation

Before the cursor patch, the latest revision built successfully but failed five of 402 upstream tests: dcg_tabling, tabling, tabling_incremental, tabling_reconstruct, and tabling_subsumption. Fil-C reported a pointer with a null capability in the tabling implementation. The other 397 passed. After the cursor patch, all 402 upstream tests passed in 2 minutes 10 seconds on the development host.

The Nix runtime check exercises all six tabling regression files and five DCG files, plus CLP(Z), installed-library association maps, exact integer arithmetic, a tail call whose argument starts a new heap page, and a call from Prolog into a Fil-C-built shared library through use_foreign_module. It checks both exit status and expected output; Trealla can report an initialization error while exiting successfully. Package build-time checks remain disabled in favor of this explicit check.

This is not comprehensive FFI validation. An exploratory call through the private $register_function evaluable-function interface reached a Fil-C trap in ffi_prep_cif; the public use_foreign_module predicate interface passed. That private-interface problem remains outside this refresh.

The full upstream suite can also be run from an extracted source checkout:

ln -s /absolute/path/to/result/bin/tpl tpl
./tests/run.sh

What is developing upstream

The recent history is predominantly Andrew Davison's work and combines language/runtime fixes with substantial new facilities. The following are source-history observations, not a claim that every feature is production ready:

• Tabling is new. Native tabling landed in July 2026. Late August and early September added resource restraints, answer subsumption, incremental invalidation, shared completed tables, and answer reconstruction changes. The phase-two design records tests, measurements, and revisions to the design.
• DCG implementation is changing. August introduced native translation and subsequent refinements. The regression suite compares the native translator with a Prolog reference and exercises tabled DCGs.
• Embedding is expanding. August added libtrealla.a and Janus work; the README explicitly labels the Python interface experimental. Freestanding ports now cover embedded boards, and the pinned HEAD adds Raspberry Pi Ethernet support. These suggest active interest in uses beyond an interactive Prolog executable.
• Core correctness work continues. Recent commits fix term expansion, Logtalk behavior, indexing races, and process/socket portability. CLP(Z)'s most recent source update was in July. Passing its focused tests is useful evidence, but does not establish the stability of the whole constraint solver.

For a compiler experiment, DCGs, structured terms, and explicit worklists are a reasonable starting point. Tabling deserves dedicated completeness and error-propagation tests: the pinned driver still catches worker exceptions as branch failure and resets incomplete tables. A caller must not infer that a returned answer set is complete merely because no exception escaped.