crosscompiler

SWC crosscompiler

Current Windows selfhost status (2026-11-02): active SWC is the V45 multi-interface Semantic build. Stage2 and Stage3 built from the active manifest source are byte-identical; the active compiler's smoke output exits 42. Full UAST lowering, production E-Graph optimization, GPU kernels and the complete Go frontend remain incomplete.

• bin/swc.exe: 1,981,440 bytes; SHA-256 06C8DEF92D9BDBCAA16C5E46B4EB1157E71F20C72F36840FDE15A94A3B2D269B.
• Active source graph: source/windows/se/swc-selfhost-windows-v45-multi-interface-active.se; source SHA-256 02785102E98EB0C34C026ACE11C0CC4364326E93BAD8BE64250B9C2F957A021B. manifests/swc.smod names this exact source revision.
• Multi-interface UDP discovery finds the local worker on multiple interfaces. SWCR1 has no stable worker identity, so duplicate endpoints for one worker are not safely deduplicated yet.
• Cluster support is experimental. Pairing/authentication and transfer of the exact compiler image and source snapshot reach the worker. The current active build has not delivered an SFJ1 compiler job in the authenticated end-to-end test; remote compilation and result merging are therefore not yet validated for this build. Do not treat --cluster=auto as a working distributed compile path yet.
• The build still supports local compilation when no worker is found. Full UAST lowering, automatic native E-Graph optimization, OpenCL matrix/vector kernels and the complete Go frontend remain incomplete.

The experimental CLI surface is cluster-pair, cluster-worker --cluster-listen, cluster-discover, and compile ... --cluster=auto. For example:

swc cluster-pair --cluster-pair-init --cluster-key cluster.key
swc cluster-worker --cluster-listen --cluster-port 24731 --cluster-key cluster.key
swc compile input.se -o input.exe --cluster=auto --cluster-port 24731 --cluster-key cluster.key

This command sequence currently demonstrates the interface only; the active-build test reached authenticated transfer but observed no compiler-job frame. Historical V35–V43 integration notes below describe earlier candidates and are not evidence that their behavior passes on the active V45 binary.

V20/V21 native integration

Integer switches and their lexical control targets, seven V20 contracts, and the 20-function V21 packed-byte runtime are integrated directly in Semantic. Marked Aggregate/Index/Slice/ForEach UAST nodes compile natively with automatic runtime inclusion. All 321 build/run/rejection checks and 221 runtime vectors pass. See V20/V21 implementation and boundaries.

Typed native Boolean lowering

Typed Rich &&, ||, !, == and != are compiled directly by SWC. Short circuiting preserves left-to-right effects and evaluates operands once. All 39 Boolean, 24 V19/Core and seven parallel checks pass. See the Boolean implementation and evidence

V19 native integration

V19 loop targets, closure boundaries, sparse statement ordering and six callable contracts are integrated directly in Semantic. All 24 native V19/Core checks pass, including nine rejection cases and deterministic parallel loop output. The full archived Go frontend remains unsupported. See V19 status

Parallel compilation

Function compilation automatically uses the available logical processors, limited by the number of functions and available worker memory. Workers claim functions from an atomic queue. A parallel sizing pass and deterministic prefix sum assign disjoint final code regions; a second parallel pass emits code at its real offsets. Labels and relocations are then merged in function order. Reading, validation and the final merge remain sequential.

UAST-Diagnosen

python crosscompiler/audit-uast.py <graph.se> --output <report.json> --strict-graph prüft doppelte Knoten-IDs, defekte Records und hängende Referenzen als Fehler. Identische Kanten sowie mehrere syntax.child-Eltern werden mit Knoten-/Kanten-ID als Warnung ausgegeben; allgemeine Mehrfachreferenzen werden als Information ausgewiesen, weil graphweite Wiederverwendung zulässig sein kann. JSON-/Klammerfehler enthalten Datei, Zeile, Spalte, Ausschnitt und Reparaturhinweis. Vollständige Diagnosedetails stehen im JSON-Bericht; die Konsolenausgabe ist begrenzt.

The implementation is native Semantic in the complete selfhost graph and swc-parallel-functions.se. It requires no external worker program, Go runtime, Python runtime or build wrapper. Windows thread callbacks use an explicit ABI adapter and checked native function addresses. The existing Linux/macOS compiler roots are not changed by this update.

On the validation machine, the selfhost build selected 12 workers and reached 13 native threads including the main thread. Stage 1/2/3 are byte-identical. See implementation and evidence.

Selfhost build

Generated executables and validation.json are saved in the ignored crosscompiler/proof/selfhost directory. Both rebuilds can take several minutes. The optional scripts/build-current-state.ps1 runner performs the strict UAST graph-integrity audit before starting those builds; that preflight requires Python. It also sets SWC_THREADS=0 and enables the parallel trace so every cache-miss build uses the maximum available worker count and records the selection.

The equivalent manual build commands are:

./crosscompiler/bin/swc.exe compile ./crosscompiler/manifests/swc.smod -o ./swc-stage2.exe --embed-all
./swc-stage2.exe compile ./crosscompiler/manifests/swc.smod -o ./swc-stage3.exe --embed-all
Get-FileHash ./crosscompiler/bin/swc.exe, ./swc-stage2.exe, ./swc-stage3.exe

All three hashes must match the hash above. Generated proof executables are temporary.

See selfhost validation, parallel validation, UAST status, Go helper integration and CLI limitations, and native Core Boolean results.

Earlier Windows binaries, manifests, sources and historical crosscompiler documents are preserved in the existing ../old_compiler_artifacts/SWC archive. Linux and macOS artifacts retain their previous validation status; this update validates the Windows selfhost only.

Ordinary direct calls now evaluate arguments left to right; see call-order evidence.

Six V22 selector closures are integrated; raw V22 lowering remains incomplete. See V22 selector status.

Explicit zero-represented 64-bit V22 pointer nil is handled natively. See pointer-null status.

Eleven V23 lowered Core functions and the V23 conditional reader are integrated; source-UAST lowering remains incomplete. See V23 status.

SWC now directly compiles original V23 source-UAST snapshots (arrays, tuples, checked indices and explicit-copy slices); 93 source assertions pass. The full 109-kind catalogue is inventoried, with general coverage still incomplete. See native raw UAST status.

Explicit binary64 literals, arithmetic, comparisons and internal scalar calls now lower natively to SSE2. The float matrix has 59 assertions plus 16 parallel function-result assertions. See native float64 status.

V24 heap-backed arrays and alias-preserving slice reads/writes now compile natively on Windows. See V24 implementation.

UAST93 parameterized lowering reference

The UAST93 archive and checksum-verified extraction are preserved under reference/source-archives/ and reference/uast-v93/. Its 15 new parameterized compile-time metadata witnesses are Linux GNU as/ld reference-backend results, not SWC-native lowerers. UAST93 itself reports 83 restricted witnesses but 0 kinds with complete semantics. See UAST93-V39-IMPORT-STATUS-2026-10-28.md for scope and integration findings.

Fail-closed native lowering

The fail-closed lowering edits remain in a separate source candidate; its --embed-all Stage 2 currently emits a PE but returns exit 3. The active V42-policy/V43-GPU-SE candidate passes byte-identical Stage2/Stage3 and the return42 probe. E-Graph optimization is represented and selfhosted, but not yet connected to native compile_expr. See SWC-FAIL-CLOSED-LOWERING-2026-10-28.md.

UAST94 composition reference

The UAST94 Linux reference archive and checksum-verified extraction are preserved under reference/source-archives/ and reference/uast-v94/. It demonstrates restricted dynamic memory/composition witnesses, not SWC-native or complete UAST lowering. Unsupported-node rejection remains a known limitation in the active V42-policy/V43-GPU-SE compiler. See UAST94-V35-IMPORT-STATUS-2026-10-28.md.