@scriptc/compiler 0.0.0 → 0.0.2

package/ambient/scriptc-overrides.d.ts

@@ -0,0 +1,95 @@
+/* scriptc's divergence and precision overrides — interface merges that
+ * re-type a handful of standard-library members to scriptc's documented
+ * runtime behavior. Merged interfaces resolve overloads later-declaration-
+ * first, so these win at call sites that match them; call forms only the
+ * lib signatures accept still typecheck and are fenced by the lowerer.
+ *
+ * This file is part of the LOWERING program only. Preflight's second-chance
+ * pass (checkPreflight) rebuilds the program WITHOUT it — the project's own
+ * type world — so a program that typechecks under its own tsc is never
+ * rejected over an override-manufactured error; the affected sites meet the
+ * lowerer's honest fences instead (SC1100's checked-cast family for
+ * JSON.parse results, the executor fences for Promise shapes). The
+ * primitives every world needs (comptime, setTimeout, __island_eval) live
+ * in scriptc.d.ts, which ships to BOTH programs. */
+
+/* `pop(): T`, not the lib's `T | undefined`: popping an empty array traps
+ * at runtime — a documented divergence (SEMANTICS.md) this override keeps
+ * typecheckable. Check `.length` first, as real code must. */
+interface Array<T> {
+  pop(): T;
+}
+
+/* `at(): string`, not the lib's `string | undefined`: undefined is
+ * unrepresentable, so an out-of-range at() THROWS a catchable TypeError at
+ * the validated island exit instead of returning undefined (documented
+ * divergence, SEMANTICS.md). */
+interface String {
+  at(index: number): string;
+}
+
+/* Object.values/entries with the actual field-value union, not the lib's
+ * `any` (the lib's precise overloads only cover index-signature and
+ * ArrayLike sources). A PRECISION override, not a divergence: the lowering
+ * emits the field list statically and the values are exactly T[keyof T] —
+ * without this the result would be `any`-typed and drag every downstream
+ * use into the island. */
+interface ObjectConstructor {
+  values<T extends object>(o: T): Array<T[keyof T]>;
+  entries<T extends object>(o: T): Array<[string, T[keyof T]]>;
+}
+
+/* `parse` returns `unknown`, not the lib's `any` — the dynamic boundary. A
+ * checked cast (`JSON.parse(s) as Config`) validates the value against the
+ * target type at runtime and THROWS on mismatch — the mechanism that makes
+ * trusting TS types sound at data boundaries. The lib's any-returning
+ * overload stays reachable only through the reviver form, which the
+ * lowerer rejects (as it does stringify's replacer/space parameters).
+ *
+ * `stringify` takes `unknown`, not the lib's `any`: the same honest
+ * surface, but the parameter must not CONTEXTUALLY TYPE literal arguments
+ * as `any` (an any-typed slot is the island boundary; `JSON.stringify([1,
+ * 2])` must keep building a static array and serializing it
+ * type-directedly). */
+interface JSON {
+  parse(text: string): unknown;
+  stringify(value: unknown): string;
+}
+
+/* The supported Promise construction shape: an executor whose resolve takes
+ * the plain value. Preferred over the lib signature at inference time, so
+ * `new Promise<T>((resolve) => ...)` types resolve as `(value: T) => void`
+ * — no PromiseLike union. The reject parameter is a real closure rejecting
+ * the promise; its reason is pinned to Error (the lib says `reason?: any`)
+ * because rejection payloads ride the same representation as thrown values
+ * and the idiomatic reason IS an Error — reject(nonError) and bare
+ * reject() typecheck only in the project world and meet the lowerer's
+ * honest coercion fences instead. First settle wins, exactly JS: a reject
+ * after resolve (or a second reject) is a no-op, and an executor throw
+ * after any settle is swallowed. Executors that use the lib signature's
+ * other extras (resolving with a thenable) still typecheck and fail
+ * lowering on their types. */
+interface PromiseConstructor {
+  new <T>(
+    executor: (resolve: (value: T) => void, reject: (reason: Error) => void) => void,
+  ): Promise<T>;
+  /* ES2024's Promise.withResolvers, in the executor's own scriptc shape:
+   * resolve takes the plain value (no PromiseLike union; the conditional
+   * collapses Promise<void>'s resolve to () => void so the record's
+   * field types map exactly), reject is the executor's Error-pinned
+   * rejection closure. The lib's own declaration would otherwise be
+   * unrepresentable field-by-field. */
+  withResolvers<T>(): {
+    promise: Promise<T>;
+    // `0 extends 1 & T` detects any (untyped JS: Promise.withResolvers()
+    // infers it) — its settled value rides the dyn arm like unknown's,
+    // so resolve takes one unknown; a naked conditional over any would
+    // otherwise union BOTH branches, which no field can map.
+    resolve: 0 extends 1 & T
+      ? (value: unknown) => void
+      : [T] extends [void]
+        ? () => void
+        : (value: T) => void;
+    reject: (reason: Error) => void;
+  };
+}

package/ambient/scriptc.d.ts

@@ -0,0 +1,74 @@
+/* scriptc's shipped declarations — the ALWAYS-SHIPPED CORE. Programs
+ * compile against the REAL TypeScript standard library (lib.es2023, no
+ * DOM, no @types): the checker sees the full standard surface, and the
+ * LOWERER is the scope fence — any reached use of standard-library surface
+ * without a lowering is a SC2020 diagnostic at its use site. This file
+ * declares only what the es2023 lib does not:
+ *
+ * 1. scriptc's own primitives (comptime, __island_eval).
+ * 2. setTimeout/clearTimeout and the Timeout handle (Node/DOM territory the
+ *    lib files don't cover; plain declarations, so with @types/node present
+ *    they MERGE as overloads instead of colliding — the Timeout handle maps
+ *    to the numeric timer id either way, and .unref()/.ref()/.hasRef() are
+ *    loop-liveness bookkeeping).
+ *
+ * This file ships to EVERY program scriptc builds — the lowering program
+ * and preflight's project-world second-chance program alike (a project
+ * using comptime must typecheck in both). The divergence/precision
+ * overrides (JSON.parse(): unknown, pop(): T, the Promise executor shape,
+ * ...) live in scriptc-overrides.d.ts, which joins the LOWERING program only.
+ *
+ * console, process, and the "node:fs" module live in scriptc-node-fallback.d.ts,
+ * shipped ONLY when the target project has no @types/node — with it, the
+ * project's real Node types are the type surface and the fallback stands
+ * down (its `declare const`/`declare module` forms would collide). The
+ * lowering tables recognize the same members either way, by name +
+ * provenance. */
+
+/* The timer handle — setTimeout's return, mapped to the numeric timer id.
+ * unref() drops it from the event loop's keep-alive set (the process may
+ * exit with the timer still armed — it never fires then, exactly Node);
+ * ref() restores it; hasRef() reports the state. The methods return the
+ * handle for chaining. Under @types/node this MERGES with NodeJS.Timeout
+ * (both map to the same numeric handle). */
+interface Timeout {
+  ref(): Timeout;
+  unref(): Timeout;
+  hasRef(): boolean;
+  /* Re-arms the timer to fire at now + the original delay (Node's
+   * Timeout.refresh). Chaining like ref/unref. */
+  refresh(): Timeout;
+}
+/* The callback is invoked with NO arguments; the second overload admits a
+ * callback declared with one parameter for the sleep idiom —
+ * `setTimeout(resolve, ms)` with Promise<unknown>'s resolve — where the
+ * zero-argument invocation delivers undefined (@types/node's generic
+ * signature admits the same shape and MERGES with these). Promise<void>'s
+ * resolve keeps matching the first overload (void params accept absence). */
+declare function setTimeout(callback: () => void, ms?: number): Timeout;
+declare function setTimeout(callback: (value?: unknown) => void, ms?: number): Timeout;
+/* The trailing-argument form: Node passes the extras to the callback
+ * (`setTimeout(cb, 0, 'foo')` fires cb('foo')). The callback slot accepts
+ * any function shape (never[] rest — parameter-contravariance's bottom);
+ * the delivered call rides the checked-dynamic boundary, so each argument
+ * is validated against the callback's real signature at fire time. */
+declare function setTimeout(callback: (...args: never[]) => void, ms?: number, ...args: unknown[]): Timeout;
+declare function clearTimeout(handle?: Timeout | number | null | undefined): void;
+
+/* Compile-time evaluation. The callback must be an inline arrow/function
+ * expression with NO references to outer bindings (it is extracted
+ * source-textually and executed in an isolated node:vm context inside the
+ * COMPILER's Node process); the returned value is baked into the binary as a
+ * literal — numbers (finite), strings, booleans, arrays, and records,
+ * nested. What type-checks runs with real JavaScript semantics, so the
+ * baked result is exactly what Node would compute at runtime (the
+ * differential harness verifies this with a `comptime = (f) => f()` shim). */
+declare function comptime<T>(compute: () => T): T;
+
+/* INTERNAL TESTING HOOK — not part of the supported surface, deliberately
+ * undocumented. Evaluates `code` in the embedded dynamic-island engine and
+ * returns String(result); island exceptions arrive as catchable errors.
+ * Requires --dynamic (a clean diagnostic otherwise). This exists to prove
+ * the engine embed end-to-end in tests; user-facing constructs will lower
+ * to the island through their own paths, not through this. */
+declare function __island_eval(code: string): string;

package/dist/backend/cc.d.ts

@@ -0,0 +1,169 @@
+export interface CcOptions {
+    /** Path of the generated (or hand-written) program TU: a .c file, or the
+     * LLVM backend's .ll — clang compiles IR text natively on the same
+     * command line, so both ride this one seat. */
+    cPath: string;
+    /** Path of the native executable to produce. */
+    outPath: string;
+    /** Build with ASan + the runtime RC audit (test/debug lane). */
+    sanitize?: boolean;
+    /** Embed the dynamic-island engine (--dynamic): compiles scr_island.c,
+     * defines SCR_DYNAMIC, and links the cached libqjs.a. Off = the static
+     * default, byte-identical to builds predating the flag. */
+    dynamic?: boolean;
+    /** The program contains a regex construct (index.ts detects it on the
+     * IR): compiles scr_regex.c and links the vendored libregexp — as cached
+     * standalone objects in static builds, from the engine archive under
+     * --dynamic (one libregexp per binary; its host hooks want the island's
+     * JSContext there). Off = regex-free: the command line is exactly the
+     * historical one, so regex-free binaries cannot change by a byte. */
+    regex?: boolean;
+    /** The embedded npm graph references fetch (index.ts detects it on the
+     * IR): compiles the NATIVE fetch bridge (scr_fetch.c over scr_net +
+     * scr_tls + scr_http's client parser + zlib — the socket units join
+     * the link implicitly, no libcurl anywhere), which builds for every
+     * target the socket units reach: hosts, linux cross, win32 cross. Only
+     * meaningful under `dynamic`; fetch-free builds keep their exact link
+     * line. SCRIPTC_FETCH_CURL=1 selects the retired curl reference instead
+     * (scr_fetch_curl.c + system libcurl on hosts / the generated soname
+     * stub on linux cross targets — ensureCurlStub), kept compilable for
+     * one release as the flip's reference. */
+    fetch?: boolean;
+    /** The embedded npm graph imports node:http or node:https (index.ts
+     * detects it on the IR): compiles the island's http/https client
+     * bridge (scr_net_island.c) and implies the socket units into the
+     * link, exactly like fetch. The emitted main calls
+     * scr_net_island_install on the same predicate (native-fetch builds
+     * also register it from scr_fetch_install). */
+    netIsland?: boolean;
+    /** The program uses zlib (index.ts detects zlib.* libCalls on the IR):
+     * compiles scr_zlib.c — the regex/curl gating precedent, so zlib-free
+     * binaries keep their exact link line. Host builds link the SYSTEM libz
+     * (macOS ships it), byte-identical to the historical line; cross targets
+     * compile the vendored zlib per target instead (ensureZlibObjects — zig
+     * has no libz in its sysroots). Compressed bytes may differ between the
+     * system and vendored libraries, which is why the corpus only ever
+     * compares round-trips and fixed-blob inflation, never raw deflate
+     * output. */
+    zlib?: boolean;
+    /** The program uses node:assert (index.ts detects assert.* libCalls on
+     * the IR): compiles scr_assert.c — the zlib gating precedent, so
+     * assert-free binaries keep their exact size class. scr_regex.c calls
+     * the assert throw/inspect helpers (assert.match lives there) and
+     * scr_symbol.c calls the equality message assemblers (assert.eqSym
+     * lives there), so the regex and symbol switches also pull this
+     * file. */
+    assert?: boolean;
+    /** The program uses util.inspect/format (index.ts detects insp.*
+     * libCalls on the IR): compiles scr_inspect.c — the assert gating
+     * precedent, so inspect-free binaries keep their exact size class. */
+    inspect?: boolean;
+    /** The program dispatches prototype methods on dyn receivers (index.ts
+     * detects dynInvoke nodes / dyn.defineProps libCalls on the IR):
+     * compiles scr_dyn_invoke.c — the assert gating precedent, so
+     * dispatch-free binaries keep their exact size class. */
+    dynInvoke?: boolean;
+    /** The program uses the diagnostics_channel surface (index.ts detects
+     * dc.* libCalls on the IR): compiles scr_dc.c — pure data structure
+     * over the DOM (no loop hooks, no install), cross-compiles everywhere.
+     * Channel-free binaries keep their exact size class. */
+    dc?: boolean;
+    /** The program uses the checked-dynamic async surfaces
+     * (moduleUsesDynAsync on the IR, or the dynInvoke/dc gates — their TUs
+     * call into this one): compiles scr_async_dyn.c — DOM-promise
+     * reactions, AsyncLocalStorage, the unhandledRejection/warning
+     * process events. */
+    dynAsync?: boolean;
+    /** The program uses the process-events surface (signal/exit listeners,
+     * stdin events, for-await over stdin — moduleUsesProcessEvents on the
+     * IR): compiles scr_events.c into the binary. Event-free binaries keep
+     * their exact link line and size class. */
+    events?: boolean;
+    /** The program uses the node:events EventEmitter surface
+     * (moduleUsesEmitter on the IR): compiles scr_events_emitter.c into the
+     * binary — the events gating precedent, but pure data structure (no
+     * loop hooks, no install), so it cross-compiles everywhere win32
+     * included. Emitter-free binaries keep their exact link line. */
+    emitter?: boolean;
+    /** The program uses ES Symbol values (moduleUsesSymbol on the IR):
+     * compiles scr_symbol.c into the binary — the emitter gating precedent:
+     * pure data structure (no loop hooks, no install; the Symbol.for
+     * registry initializes lazily), so it cross-compiles everywhere.
+     * Symbol-free binaries keep their exact link line. */
+    symbol?: boolean;
+    /** The program uses the URLSearchParams surface (moduleUsesSearchParams
+     * on the IR): compiles scr_url_params.c into the binary — the symbol
+     * gating precedent: pure data structure (no loop hooks, no install),
+     * cross-compiles everywhere. sp-free binaries keep their exact link
+     * line (scr_url.c never references the unit). */
+    searchParams?: boolean;
+    /** The program uses the node:stream class surface (moduleUsesStream on
+     * the IR): compiles scr_stream.c into the binary — always alongside
+     * scr_events_emitter.c, which moduleUsesEmitter answers true for
+     * whenever this does (the stream classes root at the emitter). Pure
+     * data structure plus the loop's deferred-tick hook — no poller, so it
+     * cross-compiles everywhere win32 included. */
+    stream?: boolean;
+    /** The program uses the node:net surface (moduleUsesNet on the IR):
+     * compiles scr_net.c into the binary — the events gating precedent, so
+     * net-free binaries keep their exact link line. */
+    net?: boolean;
+    /** The program uses the node:http server surface (moduleUsesHttpServer
+     * on the IR): compiles scr_http.c — always alongside scr_net.c, which
+     * moduleUsesNet answers true for whenever this does. */
+    http?: boolean;
+    /** The program uses the REAL node:http2 surface (moduleUsesHttp2 on
+     * the IR): compiles scr_http2.c — always alongside scr_net.c, which
+     * moduleUsesNet answers true for whenever this does. */
+    http2?: boolean;
+    /** The program uses the node:dgram or node:dns surface (moduleUsesDgram
+     * on the IR): compiles scr_dgram.c into the binary — the net gating
+     * precedent, so dgram-free binaries keep their exact link line. */
+    dgram?: boolean;
+    /** The program uses fs.watch (moduleUsesFsWatch on the IR): compiles
+     * scr_watch.c into the binary — the net gating precedent, so watch-free
+     * binaries keep their exact link line. */
+    watch?: boolean;
+    /** The program uses node:test (moduleUsesNodeTest on the IR): compiles
+     * scr_test.c into the binary — the net gating precedent, so test-free
+     * binaries keep their exact link line. */
+    nodeTest?: boolean;
+    /** The program uses the node:tls or node:https surface (moduleUsesTls on
+     * the IR): compiles scr_tls.c and links the vendored mbedTLS archive
+     * (built lazily like the engine archive, cached per flavor). Always
+     * alongside scr_net.c and scr_http.c, which moduleUsesNet /
+     * moduleUsesHttpServer answer true for whenever this does. TLS-free
+     * binaries keep their exact link line and never compile mbedTLS. */
+    tls?: boolean;
+}
+export declare function runtimeSrcDir(): string;
+export interface CcDriver {
+    /** The compiler argv prefix: ["clang"] (default) or ["zig", "cc"]. */
+    argv: string[];
+    /** The SCRIPTC_TARGET triple, or null for a host-native build. */
+    target: string | null;
+    /** Extra compile args the target demands; ALWAYS [] on the default path. */
+    targetArgs: string[];
+}
+export declare function resolveCc(env?: NodeJS.ProcessEnv): CcDriver;
+/** The OS the produced binary runs on: the triple's OS under SCRIPTC_TARGET,
+ * the host's otherwise — so platform-conditional link flags follow the
+ * TARGET, not the machine running the compiler. Exported for compile()/
+ * analyze(): the FRONTEND consults it too (path.sep / os.EOL literals and
+ * the path-module binding follow the target — a win32 triple compiles
+ * Node-on-Windows semantics, path.win32 backing the bare module). */
+export declare function targetPlatform(driver: CcDriver): string;
+/** Compiles one C program together with the runtime sources.
+ * Without a cache dir (the production default), the runtime (a dozen small
+ * files) is recompiled on every build in one historical clang invocation —
+ * no cached-archive staleness bugs. --dynamic additionally compiles
+ * scr_island.c under SCR_DYNAMIC and links the cached engine archive (built
+ * lazily, see above); regex-using programs additionally compile scr_regex.c
+ * and link libregexp (the cached objects, or the archive's own copy under
+ * --dynamic). Without either, the command line is exactly the historical
+ * one — regex-free static builds must stay byte-identical.
+ *
+ * With SCRIPTC_CACHE_DIR set (the test lanes; see the cache block above),
+ * unchanged programs skip clang via the binary cache, and misses link the
+ * program's own TU against cached per-flavor runtime objects. */
+export declare function compileC(opts: CcOptions): Promise<void>;