Zig frontend — a working SKELETON; 6 of 28 ordinary constructs compile (re-measured 2026-09-12)

• Type: feature — umbrella. Track Z (the Zig frontend's own lane: zlexer/zparser, Zig->IR lowering, lib/zrtl (which does not exist on disk — checked 2026-09-12; the lane owns the name, not a directory), Zig tests). The older "spans Track A + Track B" reading predates the Z lane; what it was pointing at is that the REMAINING items are not frontend work — the comptime VM, record-ABI shapes and real tySlice/tagged-union primitives are Track A shared machinery, and lib/zrtl breadth is Track B on demand. The frontend itself is Z.
• Status: experimental — a SKELETON that compiles its own six tests and a real chess perft, and refuses most ordinary Zig. Re-measured 2026-09-12 (see the section at the foot): 6 of 28 isolated single-construct probes compile. The 2026-07-08 claim that "everything reachable by pure parse-time desugaring is DONE" is FALSE on its own terms — struct methods, top-level const, for (arr) |v|, labeled blocks, multiline strings and [_]T{} are all pure parse-time work and all refused. The compiler's OWN diagnostics call it a skeleton. Still low-prio by the experimental-frontends rule (experimental/README.md); the point is that the gap list is now real instead of asserted.
• Owner: —

Theoretic completion (2026-07-08, Track Z — user-directed)

Sub-tickets 3/4/5/7 landed in one pass, still zero shared-internals edits (zparser.inc + one lexer char + test + Makefile). The trick throughout: every "missing primitive" was faked with vocabulary the IR already has — the same UClass tag+payload shape the Rust skeleton's enums pioneered:

• #7 switch — hidden scrutinee temp + AN_IF chain (arm value lists OR together; blocks / expressions / return arms; else). No ranges.
• #5 defer/errdefer — fn-level parse-time stash, replayed in reverse at every exit; normal exits replay defer, error exits errdefer+defer. Return value evaluates into a hidden temp BEFORE the replay (Zig order). Deviation: registration is lexical, fn-scoped.
• #3 optionals ?T — auto-registered UClass ("?i64") with __has/__val; null, assignment, if (x) |v| capture (fresh local + prepended field read, the Rust match-bind trick), orelse (AN_TERNARY), .? unwrap. Locals only.
• #3 error unions !T — errno-style global __zig_err slot with a caller-clears convention; error names auto-register (codes from 1); return error.X (defers replay first, slot set after), try f() (statement or whole var-init; propagates via errdefer+defer replay + exit), f() catch v / catch |e| { }, error.X as a comparable value. Scalar payloads ride the normal return register.
• #4 slices (minimal) — auto-registered UClass ("[]i64") with __ptr/__len; a[lo..hi] (AN_ADDR over AN_INDEX for the base address), s[i] read/write via raw i64 pointer math + AN_DEREF (the C frontend's decay trick), s.len. Initializer-position slicing of fixed arrays only.

Test test/test_zig_advanced.zig (switch/defer/errdefer/optionals/error propagation/slices in one program) wired into make test. Gate: all three Zig tests green, testmgr quick 11/11, self-host fixedpoint byte-identical.

Probe verdict: still no shared-internals bug — AN_TERNARY, AN_DEREF on computed addresses, AN_ADDR over AN_INDEX, deep nested AN_SEQ/AN_IF chains from desugaring all worked as documented. One self-inflicted lesson worth recording: the paramless-recursion pitfall (bare ZParseStatement reads the Result alias instead of recursing) bit AGAIN inside the defer dispatch — third frontend it has bitten; see frank2-paramless-name-semantics.

Sub-ticket #6 landed too (2026-07-09, type-param subset)

After the user pointed out Pascal generics were mostly parser-side, #6 followed the same route: fn f(comptime T: type, ...) monomorphized by token-buffer substitution (rparser's RSpecializeGenericFn ported as ZSpecializeGenericFn — Zig is easier: the type arrives explicitly as the first call argument, no inference). Mangled f$i64 specializations, cached; comptime param dropped from the copy; mid-stream jmp bracketing reused; two Zig-specific hazards handled (caller's defer stash and CurProc/FrameSize saved around the nested compile). Explicitly NOT a comptime VM: no comptime values, control flow, or builtins.

What remains is exactly the non-frontend work (parked, would upscale to Track A only on its own merits per experimental/README.md):

• a real comptime VM (#6's full form) — values/control-flow/builtins.
• record-ABI shapes — optional/struct/slice params and returns.
• std breadth (#8, lib/zrtl) — Track B, on demand. No lib/zrtl exists on disk (2026-09-12); nothing has been started here, so do not go looking.
• real tySlice / tagged-union primitives — shared with the Rust tickets ([[feature-rust-borrowed-slice-type]], [[feature-rust-match-enum-payload]]).

Sub-ticket 2 landed (2026-07-07, Track Z — user-directed scaffolding pass)

zig-structs-and-pointers done, purely frontend-side as the gap map predicted (zparser.inc + test + Makefile only; zero shared-internals edits):

• Structs: top-level const Name = struct { f: T, ... }; (scalar fields) registered via the same AddUClass/AddUField prescan rparser.inc uses; var s: Name = undefined;, field read/write s.f, struct-literal init var s = Name{ .f = expr, ... }; lowered to per-field AN_ASSIGNs (the Rust no-construct-node shape).
• Pointers: *T / *const T to scalars — &x (AN_ADDR, plain vars only), p.* read + assignment target (AN_DEREF), pointer locals (annotated or inferred from &x via the AN_ADDR operand's symbol) and pointer fn params (8-byte scalar in the same SysV register slots; pointee recorded through LastTypePointer* → AllocParam).
• Arrays: var a: [N]T = undefined; (AllocArray, 0-based), a[i] read/write (AN_INDEX), a.len folded to an AN_INT_LIT constant.
• Statement-level postfix lvalues (s.f = e; / a[i] = e; / p.* = e;) parse as expression-then-= (plain = only; compound stays plain-variable-only). Postfix is single-level by construction (fields are scalar, so chains cannot arise).

Test test/test_zig_structs.zig wired into make test. Probe verdict: again no shared-internals bug — record/pointer/array lowering worked first try; self-host fixedpoint byte-identical. Cross-target matrix offloaded to Track T per the watcher protocol.

Skeleton landed (2026-07-06, Track Z)

Sub-ticket #1 (zig-frontend-skeleton) is done, mirroring the Ada-skeleton precedent exactly (additive only: compiler/zlexer.inc + compiler/zparser.inc new files, isZig flag + .zig dispatch in compiler.pas, one-line isZig var in defs.inc — no new AST nodes, no new IR, no backend work). zparser reuses rparser.inc's node helpers (RSeqAppend/RMakeIdent/RBinOp/RWiden), so the two C-family skeletons share plumbing.

Landed subset: top-level [pub] fn (≤4 scalar params, i8–u64/usize/ isize/bool/void), calls + recursion, var/const with type inference and = undefined, assignment + compound (+= etc), if/else-if/else, while with : (continue-expression), range for (lo..hi) |i| (exclusive hi, per Zig), break/continue, and/or, _ = expr; discard, const std = @import("std"); header swallowed, and std.debug.print("fmt", .{args}) lowered onto AN_WRITE/AN_WRITELN with {}-placeholder splitting (segments point straight into the fmt token's TokChars — no copies). Test: test/test_zig_skeleton.zig wired into make test next to the Ada one.

Probe verdict: no shared-internals bug found this pass — shared pipeline (AN_* → IR → x86-64) handled the Zig shapes first try. Both bugs hit while building were frontend-local: (1) Zig integer / initially mapped to tkSlash = Pascal REAL division → double bits reinterpreted as int64 garbage; fixed by lowering / to tkDiv (trunc div) since all skeleton types are integers — note rparser.inc still maps Rust / to tkSlash, same latent hazard, its tests just never divide ([[feature-rust-frontend]], worth a one-line note there); (2) a { past } } comment-brace typo in my own code.

Known deviations (accepted, documented in zparser.inc header): continue inside while ... : (expr) skips the continue-expression (Zig runs it); const mutability not enforced; no overflow safety; {{ }} fmt escapes unhandled; x86-64 only.

• Opened: 2026-07-03 (feature request — filed next to [[feature-rust-frontend]])
• Priority: unranked — scoping ticket, not a greenlit build

Reframed under the esoteric-frontend-probe category (2026-07-05)

Parked as a full-language effort (below), but a skeleton-only pass (lexer/ parser for a trivial subset, lowering onto existing IR, no comptime engine) is back in scope — not to make Zig usable, but as a bug-probe against shared internals. See [[feature-esoteric-frontend-probes]] for the category rule.

Why parked (2026-07-05, user decision — rationale corrected 2026-07-05)

Original log entry claimed comptime's recursive/Turing-complete evaluation "undermines determinism" and conflicts with the byte-identical fixedpoint self-host gate. That claim was wrong and is withdrawn. Recursion and Turing-completeness threaten termination (an unbounded comptime loop — hence Zig's own branch-quota limiter), not determinism: a pure interpreter with no access to wall-clock/randomness/filesystem/mutable global state gives the same output for the same input regardless of recursion depth. And the fixedpoint gate is about the PXX compiler itself (written in Pascal) reproducing byte-identically across bootstrap stages — a Zig frontend's comptime interpreter only runs while compiling Zig source, never during the compiler's own self-compile, so it cannot threaten a gate it is never part of.

Actual reason to park: pure engineering scope, not architectural risk. comptime is pervasive (not a corner feature — no separate generics syntax; generics, @TypeOf, conditional compilation, and most of std all lean on it), and the builtin surface is large (@import, @sizeOf, @ptrCast, @field, …, each comptime-typed) with no "ignore and link" escape hatch the way C headers give other frontends. See "Notes on scale" below — that section already had the honest framing; the determinism paragraph was an unnecessary and incorrect addition on top of it.

Decision: scope an Erlang frontend first — see [[feature-erlang-frontend-scoping]], on its own merits (different engineering domain: runtime/scheduler work vs. a compile-time interpreter), not because it is "more deterministic" — Erlang's actor-model preemptive scheduling is itself a classically nondeterministic runtime execution model (message arrival order, process interleaving), so determinism was never a valid axis to rank these two on. Revisit Zig if a comptime-style engine becomes independently justified by something else needing it.

Original scoping (2026-07-03, pre-park)

Motivation

Add a Zig-syntax frontend (5th, after Pascal / Nil-Python / C / Rust-planned) lowering to the existing shared IR/backends, same shape as the C frontend's v80 bring-up (clexer.inc/cparser.inc/cpreproc.inc → shared IR). Zig is C-style / systems: manual memory (explicit allocators, no GC), structs, pointers, plain integers, C-like control flow — so its type theory maps onto the current statically-typed IR directly, much like C, with no ownership/move/trait system to model (simpler than Rust in that dimension). But "simpler type theory" does NOT mean trivial: reaching a subset that compiles real Zig is likely non-trivial because comptime is pervasive and the builtin surface is large — see "Notes on scale" below. It is the more tractable of the two new-frontend requests (JS parked as architecturally out-of-scope, see [[feature-js-frontend-parked]]), not an easy one.

Explicit non-goals (scope cuts up front, like the C/Rust precedent)

• No comptime metaprogramming engine initially. comptime is Zig's compile-time-evaluation + generics substrate — a full interpreter over the AST. v1 handles only trivially-const comptime (constant expressions, comptime-known array sizes) as ordinary const-eval; generic functions (fn f(comptime T: type, ...)) come later via the existing monomorphization engine (GenericFuncs, defs.inc), not a general comptime VM.
• No async/await ecosystem. Zig's colorblind async is evolving upstream and is not a v1 target; if pursued, desugar onto PXX's stackful coroutine runtime (coroutine_emit.inc) like the Rust plan — correct for our own code, not wire-compatible with any Zig async runtime.
• No build.zig / package manager / build system. Compile single files / a fixed file set, as the C frontend does — not the Zig build graph.
• No full std. A thin lib/zigrtl subset (basic std.debug.print, std.mem, allocators-as-thin-wrappers-over-our-heap) grown on demand, the way lib/crtl grew for C. Not std.io/net/fs breadth up front.
• Compiling arbitrary upstream Zig packages is out of scope for v1 — same rule as the Rust umbrella: hand-port a call surface if needed, don't let one dependency drag the ticket into full-Zig scope.

What already exists to reuse (confirmed reusable machinery)

• C-frontend bring-up pattern — clexer/cparser/cpreproc → shared AST (AN_*) → shared IR. Zig's zlexer/zparser follow the identical shape; the whole IR/backend/ABI/ELF stack is unchanged.
• Structs / pointers / fixed & dynamic arrays / integers — already first- class in the IR; Zig structs, [N]T, []T slices, *T map onto them.
• Error unions (E!T) and optionals (?T) — a tagged value (payload + error/null tag). tyVariant (defs.inc) is the 16-byte proto; the generalized tagged-union work planned for Rust ([[feature-rust-match-enum-payload]]) is the SAME primitive — share it.
• defer / errdefer — scope-exit execution; the scope-cleanup machinery (IRLowerCleanupToDepth, class dtor insertion) is the precedent to reuse, same family as the Rust drop work ([[feature-rust-drop-move-tracking]]).
• Slices []T / []const u8 — non-owning ptr+len view; identical need to Rust's [[feature-rust-borrowed-slice-type]] — share the type.
• Monomorphization (GenericFuncs/specialization, defs.inc) — backs comptime-generic functions when that phase lands.
• Local type inference (tyAuto + EnableAutoVar, parser.inc ~7731, 10250) — PXX already infers a binding's type from its initializer for inline var (var x := 5+3 in a statement block, for var x in coll). Verified empirically: var x := 5+3 → Integer, var s := 'hi' → string. This is the SAME local-inference-from-initializer mechanism as Zig's const x = expr / var x = expr, so Zig's inferred bindings reuse the existing expression typer, not new machinery. Caveats: ours is currently gated on EnableAutoVar and scoped to inline var (global declaration-section var still needs an explicit type); Zig infers pervasively on every const/var. So the local inference is a genuine reuse win — but note the boundary vs. Zig's comptime-types below, which is the part we do NOT share.

Much of the Rust RTL/type groundwork (tagged unions, slices, drop/defer) is directly shared with Zig — sequencing the two together would amortize it.

AST/IR gap map — "pure frontend" vs "needs shared internals" (2026-07-04)

Verified against the current node/type inventory (defs.inc) + empirical checks.

Lowers onto EXISTING AST/IR — a skeleton subset is pure lexer+parser, zero Track-A internals change:

Needs NEW shared machinery (Track A internals — NOT lexer+parser), and each overlaps an already-planned Rust ticket:

Bottom line: skeleton (#1) = pure frontend, shippable, zero internals risk. A useful subset = skeleton + the 3 shared additions (tagged-union, slice, optional) — which ARE the Rust Track-A tickets. Zig and Rust share their hard parts; building either advances the other. The "auto typing" is on the free side; the tagged union is the real cost.

Sub-tickets (split when work starts — don't flood the board yet)

Track A (compiler internals — shared AST/IR/symtab/backends):

1. zig-frontend-skeleton — zlexer.inc/zparser.inc, entry point (.zig dispatch in compiler.pas), minimal subset: fn, pub, integers, var/ const, if/while/for/switch, blocks, return, basic operators. Gates everything else. Mirrors C-frontend skeleton scope.
2. zig-structs-and-pointers — struct, field access, *T/*const T, &x, .* deref, [N]T arrays.
3. zig-optionals-and-error-unions — ?T, E!T, orelse, catch, try, if (opt) |x| capture, unreachable. Depends on the generalized tagged-union primitive (shared with [[feature-rust-match-enum-payload]]).
4. zig-slices — []T/[]const T ptr+len views, slicing a[lo..hi], .len/.ptr. Shared with [[feature-rust-borrowed-slice-type]].
5. zig-defer-errdefer — scope-exit + error-path-exit execution over the existing cleanup machinery.
6. zig-comptime-generics — fn f(comptime T: type, ...) via monomorphization; comptime-known sizes. NOT a general comptime VM.
7. zig-switch-and-tagged-enum — switch on enums/tagged unions with payload capture; enum/union(enum).

Track B (lib/zigrtl):

8. zig-rtl-core — allocators as thin wrappers over the pxx heap (std.heap.page_allocator-shape), std.debug.print / std.log wired to existing write machinery, std.mem basics (copy, eql, span).

Notes on scale — Zig is NOT an easy source (2026-07-03, user flag)

Earlier framing ("smaller lift than Rust") is only half true and needs tempering. Zig has no ownership/move/trait system, so its type theory is simpler than Rust's — but reaching a practically useful subset (one that compiles real Zig) is likely NON-trivial, for reasons specific to Zig:

• comptime is pervasive, not a corner feature. Zig has no separate generics syntax — generics, type construction, conditional compilation, and much of std are all comptime. Our "no comptime VM" cut (a reasonable v1 scope, like C/Rust's cuts) therefore excludes a LARGER fraction of real Zig than "no macro_rules!" excludes of real Rust. Everyday fn f(comptime T: type), @This(), @TypeOf, @hasField etc. all lean on it. Impact: "compiles real programs" is gated on comptime-generics much earlier than the Rust equivalent.
• Huge builtin surface. @import, @sizeOf, @intCast, @ptrCast, @field, @memcpy, … (hundreds). Many are comptime-typed. Each needs frontend handling; there is no "ignore and link" fallback like C headers.
• Error unions !T + optionals ?T are threaded through everything — idiomatic Zig, so the tagged-union + try/catch/orelse propagation must be solid EARLY (sub-ticket #3), not a late add.
• defer/errdefer ordering + error-path cleanup — correctness-sensitive (wrong = skipped cleanup / wrong error path, a runtime bug, not a compile error). Land with tests first.
• Pre-1.0 moving target. Zig syntax/semantics still shift between releases; pin a target version.
• Allocator-passing idiom. Every allocation takes an allocator param; std leans on it. Mappable to our heap, but pervasive.

Net: comparable-or-harder than the C frontend's multi-session bring-up to reach "compiles real Zig", despite the simpler type theory — the cost is comptime + builtins, not the type system. #1 gates everything; #3 (error unions/optionals) and #6 (comptime-generics) are the real gates on usefulness. Sequence with the Rust tagged-union/slice/drop work to share the primitives.

Log

• 2026-07-03 — umbrella opened from a feature request; filed alongside the Rust frontend family. Zig chosen over JS (parked — see [[feature-js-frontend-parked]]) because it is C-style and maps directly onto the existing IR. No code yet; skeleton (#1) branches off master under Track A per the C-frontend precedent when work starts.

5/6-param internal calls fixed (2026-07-16, Track Z / A)

Sibling-sweep of the Rust 5/6-param spill fix: ZParseTopLevelFn had the same bespoke case i of 0..3 param-register spill the Rust frontend did (the code comment literally said "byte-for-byte the sequence RParseTopLevelFn emits") — emitting no modrm byte for param index 4/5, so mov [rbp+off], r8 came out as 48 89 <off32> and SIGILL'd on the 5th param. Both frontends now share REmitParamRegSpill (computes REX.R for r8/r9); the Zig param cap is raised from 4 to the register-convention 6. test/test_zig_manyparams.zig (add5/add6 + 5-param recursion). Regressions green: all three prior Zig tests, quick tier, self-host byte-identical. Lesson reinforced: a bug copied byte-for-byte across sibling frontends must be fixed (or factored) across all of them — see the sweep-sibling-dispatch-branches rule.

Slice params + array literals + a real-load chess perft (2026-07-16, Track Z)

Two frontend enablers landed (zparser.inc only; self-host byte-identical):

• []T slice parameters — fn f(s: []T, ...): the 16-byte __ptr/__len record passed by address (IsRef), so s[i] rw and s.len work through the pointer, mirroring the Rust &[T] param path. Added to the signature prescan (register slice class, mark tyRecord param IsRef) and the body parser (AllocParam with the slice recid + isRef). test/test_zig_slice_params.zig.
• Array literals — anonymous dot-brace and typed [N]T{ ... } forms, lowered to one AN_INDEX store per element (the struct-literal no-construct shape). Both annotation-present and RHS-typed (const t = [4]i64{...}) forms. LANDMINE re-hit: a .{ inside a { } .inc comment desynced pxx's own brace-nesting comment lexer and produced an "unexpected character" ~94k lines downstream in the self-compile — keep literal braces out of .inc comments (see nested-comment-brace-selfhost-landmine).

Then the payoff — test/test_zig_chess_perft.zig, a full-legality chess perft compiled from Zig (the esoteric-frontend-probe's point: exercise the shared pipeline under real load). Slice params carry the board + move list, 5-param recursion drives the tree, array literals hold the offset tables, deep nested control flow runs movegen/make/unmake. Node counts are reference-exact: startpos perft(4)=197281, Kiwipete perft(3)=97862. Probe verdict: no shared-internals bug — the one bug found this pass was the frontend-local 5/6-param spill (fixed above, shared with Rust). Regressions green: all Zig tests, quick tier, self-host byte-identical.

Re-measurement 2026-09-12 (frankZ, Track Z) — the completion claim is stale

Prompted by the stale-hazard rule: an unranked experimental umbrella asserting completion is a warning that decays like a lock. Measured at 356d4d4fa with a freshly converged HEAD binary (808076de24be), CWD at the repo root. That binary is byte-identical to the one pin v408 ships (last.sha256 = 808076de24be..., pinned 2026-09-12 from tree 14934e60d; I built at 356d4d4fa, and only docs landed between — determinism, not the same tree), so this census describes exactly the compiler that is now $(PXX_STABLE). This used the HEAD-built compiler, NOT $(PXX_STABLE), so it is independent of pin age — and every gap below is a PARSE error, raised before any builtin or RTL is reached, which no pin could produce.

The existing suite is GREEN and that is not the same claim. All six test/test_zig_*.zig compile, run, and match their Makefile expected strings byte-for-byte, chess perft included. The frontend has not rotted. But the suite is written in the subset that works — it uses for (0..N) |i| in all four loop sites and never for (arr) |v|, and declares arrays only as [5]i64 = undefined. That is the passing-arrangement population, so it certifies the skeleton rather than probing it.

28 isolated single-construct probes, one file each (isolation on purpose: a combined program reports a first failure and hides the walls behind it; the scaffolding-only control compiles, so the failures are the constructs). 6 pass — bool, usize, pub fn, string literal + {s}, while (c) : (i += 1), and the control. 22 fail. Three first-round failures were CONFOUNDED and were re-probed before attribution: for (a) |v| died on the [_]i64{} literal, not on the loop. Disambiguated, [3]i64{1,2,3} compiles (the ticket claims exactly that and is correct); [_]T{} does not.

Grouped by MECHANISM, not by probe count — four gates produce most of it:

1. The struct body is a flat token walk requiring ident : ident (zparser.inc:1611-1618). One shape constraint refuses methods, const Self = @This(), default field values, and every non-scalar field type (arrays, ?T, nested structs) — 7 probes, one cause. Methods fail regardless of position, so it is not an ordering bug.
2. Top level accepts only fn/pub fn (zparser.inc:2111-2112) — no top-level const, no enum, no union(enum), no error{} set, no test block — 5 probes.
3. for unconditionally parses a range — for (arr) |v| reaches the shared range parser and errors expected '..' with no Zig: prefix, which is the tell that it never reached a frontend check.
4. The type vocabulary is integers, bool, void and declared structs — the diagnostic says so in those words. No floats at all: f64 is unknown type.

Singles: switch ranges 0...3 (the ticket does say "No ranges"), comptime, labeled blk: { break :blk v }, unreachable, multiline strings, []const u8 parameters, and pub fn main() !void (refused by name).

What this changes for ranking. Of the gaps above, items 1, 2, 3, [_]T{}, labeled blocks and multiline strings need NO shared machinery — they are parser work in zparser.inc, which is Track Z's own file. The ticket's standing position, that what remains is Track A's comptime VM and real tySlice/tagged-union primitives, is true only of floats, ranges and comptime. The single highest-value item is (1): struct methods are how essentially all real Zig is organised, and the shape constraint that refuses them also refuses six other things.

Nothing was fixed here — this is a measurement, and the frontend stays low-prio experimental. lib/zrtl, which the header names as a Track Z file, does not exist on disk at all.