plato-232.md

July 31, 2026 · View on GitHub

Summary

Add discriminated sum types (tagged unions) with exhaustive match to Plato, targeting C#. Executes the RFC idea plato-077: this issue is the design + spec + test-corpus stage; the parser/AST front end (wave 2) and the stdlib migration (wave 3) are the implementation increments that follow. Removes the language's biggest partiality gap — the XxxKind + conditionally-valid-fields hand-encoding — and unblocks the stdlib Option/Result cleanup (plato-079).

Design (fixed)

  • Declaration: type gains an = case-list body as an alternative to the braced field body: type PathSegment2D = Move(EndPoint: Point2D) | ... | Close;. Payload-free cases (enums) omit parens: type FillRule = NonZero | EvenOdd;.
  • Match: an expression, exhaustive, positional binders in case-field declaration order, no default arm in v1. Errors name the sum type and the missing / unknown / duplicated cases.
  • Backend (C#): one readonly partial struct — int Kind tag (0-based, declaration order) + flattened Case_Field fields; private all-fields ctor; one static factory per case that zeroes the inactive fields (so struct equality over tag+fields is well-defined). match lowers during elaboration to a tag-conditional chain of existing TIR nodes — no new TIR node, downstream passes untouched. Generated code compiles on net8.0 default LangVersion (C# 12; no switch-expression exhaustiveness dependency).
  • Diagnostics: proposed CHK300–CHK307 (+ CHK320 for the non-C# writers rejecting sums). Full catalog with message texts in the design doc.
  • Generics: supported iff the monomorphizer specializes sums per type argument with no unifier change (expected yes); else v1 restricts with CHK306. option.plato decides.

Full spec: design doc. Survey (115 XxxKind types, ~11 true sums vs ~100 enums, flagship-5, 40-paths "after"): survey doc.

Scope

  • In (v1): declaration + match syntax; exhaustiveness/name checking; C# tagged-struct emission; monomorphic (and, if cheap, generic) sums.
  • Out (v1): GLSL/TS/Rust emission (CHK320 rejection); nested patterns; guards; default arm; recursive sum types (flat struct cannot contain itself); unique/affine interaction; per-case shared fields.
  • plato-077 — the RFC idea this executes.
  • plato-079 — stdlib partiality cleanup, hard-blocked on this.
  • plato-076 — Gratify kernel port; names this as its top language gap.
  • plato-230 — plato-src-v3, the survey's source corpus.

Done means

  • Design doc landed: declaration + match syntax, static/dynamic semantics, C# emission contract (hand-written PathSegment2D struct), diagnostics catalog.
  • Test corpus authored — plato-test-sum/ — 5 positive + 6 negative fixtures with the // EXPECT: <code> convention + README.
  • v3 survey: 115 XxxKind types classified (enum vs sum), flagship-5 recommended, 40-paths.plato "after" drafted.
  • Front end (wave 2): = case-list + match parse and type-check; exhaustiveness diagnostics CHK300–CHK306 (+CHK320) fire on the negative corpus; positive corpus compiles to C# on net8.0 default LangVersion. (Plato 1d3ed84 AST, 507de64 checking/lowering/emission; PlatoTests 126 → 142.)
  • Migration (wave 3): the flagship five migrated (PathSegment2D, Paint, MaskSource2D, ScalarFieldNode2D/3D, WindowFunction); lint plato-src-v3 0 parse/0 symbol errors (4584 → 4549), plato-src 193, PlatoTests 142/142. (Plato 37ed5e9.)

Closing note (2026-07-27)

Shipped: type X = Case(...fields) | Case | ...; declarations + exhaustive match expressions (positional binders, no default arm) — grammar (parakeet), AST, symbol resolution, checker (CHK300–CHK306, CHK320), match lowering during elaboration to a tag-conditional chain over existing TIR nodes (no new node), and C#-only tagged readonly partial struct emission (int Kind + flattened Case_Field fields + per-case factories + structural equality). Generics restricted in v1 (CHK306). Test corpus plato-test-sum/ + SumTypeCheckingTests (12) + SumTypeCodegenTests (4, in-proc Roslyn). Wave-3 migrated the five flagship plato-src-v3 kind-pattern types to real sums; 26-fields confirmed non-recursive (operands are Integer node indices) so no substitution was needed.

Commits: parakeet 12aad78 (grammar); Plato 1d3ed84 (wave-1 AST), 507de64 (wave-2 checking/lowering/emission), 37ed5e9 (wave-3 flagship migrations + v3 README + survey).

Remaining (follow-up, not this issue): the ~100 pure-enum XxxKind types — the kind-pattern sweep — plus the deferred features (bare constructors, generic sums, default arm, guards, nested patterns, recursive sums, GLSL/TS/Rust support). Tracked as plato-233.