Files
sx/current/CHECKPOINT-REIFY.md
agra ae27cffe9d plan(reify): F1 findings + lock the zero-compiler-reify end state
Record the verified pass-order / define-timing / parse / dispatch findings
from F1 investigation, and make explicit that the floor work MUST delete
reifyType + the E :: reify decl hook + findReturnReifyCall (reify lives only
in meta.sx). Removal can't precede the floor, so they land together; suite
never left red across a session boundary.
2026-06-16 20:15:21 +03:00

14 KiB
Raw Blame History

CHECKPOINT-REIFY — comptime type_info / reify (async-first foundation, step 3)

Companion to PLAN-REIFY.md. Update after every step (one step at a time, per the cadence rule).

Last completed step

Phase 3.1 (green) — RecvResult/TryResult done. Added RecvResult($T) and TryResult($T) to meta.sx as type-fns over reify (a value-or-closed, and a value-or-empty-or-closed enum). They needed NO new compiler machinery — they're reify-of-a-literal in a type-fn body, i.e. exactly the Phase 1 path — so the channel result types are pure sx library code. examples/0617 green (both construct + match, incl. payload-less .closed/.empty). Full suite green (673 examples). Cadence: 3.0 xfail (undefined → RED) → 3.1 green.

Phase 3.1 is the last GREEN step. Phase 3.2 was attempted + reverted; the stream now pivots to the declare/define RE-ARCHITECTURE (see the section below + the F1F5 table in PLAN-REIFY). type_info NOT done (old Phase 2.2, orthogonal).

(prior) Phase 2.1 (green) — field_type done. field_type($T, i) -> Type is

implemented over the type table (fieldTypeOf in lower/generic.zig, re-exported on Lowering): struct field / tagged-union + union variant payload (.void for a tagless variant) / tuple element / array + vector element; OOB and memberless types poison to .unresolved with a loud diagnostic (never a silent default). It folds at lower time, so it composes inside type_eq / type_name / any type-arg slot. examples/0616 green (struct fields name+type, type_eq fold, tagged-union payloads incl. quit → void). Full suite green (672 examples, 447 unit). Cadence: 2.0 xfail (empty marker, RED) → 2.1 green (this commit).

type_info is NOT done — it still bails loudly in call.zig:tryLowerReflectionCall. It builds a full TypeInfo value (inverse of reify) and is the larger Phase 2.2 step (see Next step).

(prior) Phase 1 (type-fn → reify identity) — COMPLETE. A type-fn body that returns

reify(...) now mints the enum under the instantiation's name: instantiateTypeFunction (lower/generic.zig) detects a reify-returning body (findReturnReifyCall) and routes it to reifyType(<mangled-or-alias name>, reify_call) with the type-arg bindings active (so payload = T resolves to the bound arg) — placed BEFORE the general case, which would otherwise route the reify call to the inline-position loud bail. Registering under the mangled name lets the fn's top-of-instantiation cache return the SAME TypeId on the second instantiation → Box(i64) at two independent sites is ONE type (Contract 1). examples/0615 green (build()consume() cross-site, plus b : Box(i64) = .none); full suite green (671 examples, 447 unit). Cadence: 1.0 xfail (empty marker, RED) → 1.1 green (this commit).

(prior) Phase 0.2 (green) — Phase 0 COMPLETE. Implemented reify(.enum(...)):

Lowering.reifyType (in lower/nominal.zig) reads the flat-enum TypeInfo literal off the AST, synthesizes an ast.EnumDecl, and feeds it through the SAME type_bridge.buildEnumInfo path source enums use → the minted type is byte-identical to a hand-written enum { value: i64; closed; } and flows through enum codegen unmodified (Contract 2 confirmed — a source enum exhibits the exact same construct/match behavior, verified by probe). Wired at the E :: reify(...) const-decl hook in lower/decl.zig. examples/0614 green (value 3 / closed, exit 0); snapshots captured; full suite green (670 examples, 447 unit). Unsupported reify shapes bail loudly via reifyBail (never a silent default).

Also landed (user-directed, separate commit feat(parser):): reserved keyword as a member name after . (.enum, case .enum:, x.enum) — so the reify example reads reify(.enum(...)) without a backtick. readme updated.

Two earlier Phase-0 commits

  • 0.1 (xfail). Added examples/0614 + empty .exit marker → RED.
  • 0.0 (lock). Added the comptime type-metaprogramming surface as the on-demand module library/modules/std/meta.sx (NOT the prelude — see decision below): EnumVariant/EnumInfo/TypeInfo data types + bodyless #builtin decls reify / type_info / field_type. Each builtin bails LOUDLY when reached unimplemented (no silent default). Unit test src/parser.test.zig (registered in src/root.zig) locks that the decls parse. zig build test green (447/447 unit, 669/669 examples).

Current state

  • modules/std/meta.sx declares the surface; the variant uses the backtick raw escape `enum (reads as the keyword, not a mangled enum_).
  • Loud bails wired (unimplemented → diagnostic, never a silent type):
    • reify(...) in a :: type-alias position → decl.zig (the .call const-decl branch) emits "reify is not yet implemented (REIFY Phase 0.2)" and poisons the alias to .unresolved. This is also where Phase 0.2 will hook the real construction.
    • reify / field_type in any other type position → generic.zig:resolveTypeCallWithBindings (defense-in-depth).
    • type_info(...) in expression position → call.zig:tryLowerReflectionCall.
  • No interpreter-side construction yet — reify mints nothing.

Decision (0.0)

Meta lives in modules/std/meta.sx, not the prelude (core.sx). Declaring the data types in the always-loaded prelude interns them into every module's type table and shifts every .ir snapshot (broke 37 examples in a trial). An on-demand import keeps the prelude clean; reify users #import "modules/std/meta.sx". (User-directed.)

RE-ARCHITECTURE in progress (user-directed, 2026-06-16)

declare/define are to be the ONLY compiler primitives; reify / make_enum / RecvResult / TryResult move to sx in meta.sx. reify becomes literally { h := declare(); define(h, info); return h; }. The AST-walking reifyType and all syntactic reify recognition (decl.zig E :: reify hook, generic.zig findReturnReifyCall) get DELETED, replaced by generic comptime evaluation of a Type-returning expression. Full F1F5 phase table + the resolved naming/identity design is in PLAN-REIFY.md → "RE-ARCHITECTURE".

**A Phase 3.2 attempt (make_enum + an eval-decode reader, reifyEnumFromVariantsValue

  • reifyValueFn) was implemented this session, then REVERTED** (git reset --hard 9306ad5) per this decision — the reader grew the compiler instead of shrinking it. The eval-decode decode logic is salvage for define (see commit 4d6c463 in reflog if needed); the syntactic routing is discarded.

Next step

F1 — declare()/define() primitives + a directly-minted RECURSIVE enum.

  • declare() -> Type: intern an empty forward nominal slot (reserveShadowEnumSlot shape, nominal.zig:98), anonymous, return a first-class Type.
  • define(handle, info): decode the TypeInfo value (reuse the reverted decode: interp Value → names + payload Type-tags → tagged_union matching buildEnumInfo), fill via updatePreservingKey.
  • X :: declare() decl hook + scan-time define(...) side-effect (must complete before codegen needs the layout — same scan-time window the old reifyType used).
  • Example: List :: declare(); define(List, .enum(.{ … payload = *List … })) — cons/nil list, construct + match. Self-ref works because the handle predates the body. Then F2 (reify→sx + general comptime-Type-decl eval, delete reifyType), F3 (type-fn body eval, delete findReturnReifyCall, re-green 0615/0617), F4 (make_enum in sx), F5 (validation). type_info (old 2.2) stays orthogonal/pending.

END STATE IS NON-NEGOTIABLE: ZERO reify in the compiler. reifyType, the E :: reify(...) decl hook (decl.zig ~660), and findReturnReifyCall routing (generic.zig) must ALL be deleted by the end of the floor work — reify lives only in meta.sx. The compiler keeps only declare/define (+ the generic comptime Type-expr evaluation that runs them). Removal can't precede the floor (the reify examples would have no path), so the floor + removal land together; the suite must never be left red across a session boundary.

F1 investigation findings (verified 2026-06-16 — don't re-derive)

  • Pass order (decl.zig:lowerRoot): Pass 1 scanDecls (line 441 — registers all types; the old reify hook mints here) → Pass 1f UnknownTypeChecker → Pass 2 lowerMainAndComptime (1320 — force-lowers main via lazyLowerFunction, which transitively lowers bodies; also creates #run comptime fns). #run side-effects EXECUTE only at emit time (emit_llvm.zig:runComptimeSideEffects).
  • define timing: must run AFTER scanDecls (needs all types) but BEFORE body lowering (match/construct resolve variant indices at lower-time, Pass 2). Neither existing hook fits → add a new Pass 1g (after scanDecls, before lowerMainAndComptime) that executes recorded declare/define pairs via the interp. Self-ref works because declare binds the name in Pass 1, so *List in the define body resolves.
  • Parse constraint: a bare top-level define(List, ...) does NOT parse (expected '::' / ':=' / ':' after identifier). So define is invoked either via a recognized top-level form or #run define(...) whose comptime fn Pass 1g executes early (instead of waiting for emit). Decide the surface in F1.
  • Builtin dispatch: resolveBuiltin (call.zig:1297) is a tiny name→BuiltinId table; reify/type_info/field_type are NOT there — they're recognized syntactically (decl hook / tryLowerReflectionCall at call.zig:1679). declare/ define need their own dispatch (likely comptime interp builtins, since they mutate the type table at comptime).
  • Salvage: the reverted eval-decode (reifyEnumFromVariantsValue + decodeVariantElements, commit 4d6c463 in reflog) IS define's decode core.

(superseded) earlier Next step — type_info value construction Reflect a type into a TypeInfo

value — the inverse of reify. Two sub-pieces, both non-trivial: (a) widen the meta.sx TypeInfo data model beyond `enum (struct / tuple variants); (b) build the value at comptime — a []EnumVariant-style slice of structs holding strings (name) + Type tags (payload), populated from the type table. This is comptime value-CONSTRUCTION (allocating slice/struct/string/type-tag values in the interpreter), materially larger than 2.1's fold-to-a-TypeId. Currently bails loudly in call.zig:tryLowerReflectionCall. Best done as its own session for context room.

NOTE (round-tripping): reifyType still reads a LITERAL TypeInfo off the AST (fine for the inline-literal + type-fn-over-literal cases of Phases 01). Once 2.2 produces a COMPUTED TypeInfo, feeding it back into reify needs the reader generalized (or interp evaluation of the TypeInfo value) — handle that when 2.2 enables round-tripping.

Alternatively jump to Phase 3 (make_enum + RecvResult/TryResult sx lib over reify) — that only needs reify (have it) + type-fns (have them), not type_info.

SELF-REFERENCE = Phase 4, API DECIDED (user-directed): explicit declare()define(h, info) (the declaration-vs-definition split; NOT a reify_rec((self)=>…) closure). declare() returns a forward nominal Type handle (named from the :: LHS) usable freely in any later TypeInfo*List, []List, and across types for MUTUAL recursion the one-self closure couldn't express; define(handle, info) fills the body. reify(info) stays as the one-shot sugar. Maps onto existing machinery: declare = empty nominal slot (reserveShadowEnumSlot-style), define = buildEnumInfo

  • updatePreservingKey, struct-stub→tagged_union re-key already handled (internNamedTypeDecl's adoptsForwardStructStub). Loud invariants: never-defined declare = hard error; by-value self-inclusion rejected (infinite size). Full write-up in PLAN-REIFY Phase 4.

Known issues

None yet.

Log

  • 3.1 (green) — RecvResult/TryResult done. Added as type-fns over reify in meta.sx (no new machinery — Phase 1 reify-of-literal-in-type-fn); 0617 green. make_enum (3.2) pending on the generalized reify reader.
  • 3.0 (xfail). 0617 + empty marker → RED (RecvResult undefined).
  • 2.1 (green) — field_type done. fieldTypeOf over the type table (struct/tagged-union/union/tuple/array/vector; OOB+memberless = loud poison); folds at lower time, composes in type_eq/type_name; 0616 green. type_info still pending (Phase 2.2 — builds a TypeInfo value).
  • 2.0 (xfail). 0616 + empty .exit marker → RED (field_type bailed).
  • 1.1 (green) — Phase 1 done. Type-fn body return reify(...) routes through reifyType under the instantiation name (findReturnReifyCall + mangled-name registration); Box(i64) at two sites = one type (Contract 1); 0615 green.
  • 1.0 (xfail). 0615 + empty .exit marker → RED (reify bailed in type-fn body).
  • 0.2 (green) — Phase 0 done. reifyType mints a flat enum from a literal TypeInfo via the shared buildEnumInfo path; 0614 green; Contract 2 confirmed (reify'd enum == source enum, same construct/match). Payload-less variant idiom in the example is c : E = .closed; (enum-literal target), same as a source enum.
  • parser (user-directed). Keyword as member name after . — see feat(parser).
  • 0.1 (xfail). 0614 + empty .exit marker → RED.
  • 0.0 (lock). Meta surface in modules/std/meta.sx (data types + 3 bodyless #builtin decls), loud bails at all three reach points, src/parser.test.zig parse-lock. Two user-directed refinements folded in: variant uses `enum raw escape; surface moved out of the prelude into its own module to avoid type-table / .ir-snapshot churn. zig build test green.
  • Stream carved. Selected as the first async-first foundation: reify gates both channel result types (RecvResult($T)) and race's synthesized union, is fully validated (3 reviewers), and is a self-contained compiler/type-system feature testable in isolation (06xx comptime). Generic-enum syntax dropped in its favor.