Type untyped JavaScript Maps, Sets and arrays from their uses (#2089)

packages/compiler/ambient/scriptc-overrides.d.ts

@@ -70,3 +70,24 @@ interface PromiseConstructor {
     reject: (reason: Error) => void;
   };
 }
+
+/* Carriers of the types the lowerer gives untyped JavaScript collections
+ * whose uses prove them (js-collection-types.ts). Nothing reads them at run
+ * time. */
+declare namespace __scriptcInferredCollections {
+  const map_string_string: Map<string, string>;
+  const map_string_number: Map<string, number>;
+  const map_string_boolean: Map<string, boolean>;
+  const map_number_string: Map<number, string>;
+  const map_number_number: Map<number, number>;
+  const map_number_boolean: Map<number, boolean>;
+  const map_boolean_string: Map<boolean, string>;
+  const map_boolean_number: Map<boolean, number>;
+  const map_boolean_boolean: Map<boolean, boolean>;
+  const set_string: Set<string>;
+  const set_number: Set<number>;
+  const set_boolean: Set<boolean>;
+  const array_string: string[];
+  const array_number: number[];
+  const array_boolean: boolean[];
+}

packages/compiler/src/frontend/lowering/js-collection-types.ts

@@ -0,0 +1,533 @@
+/* Types of untyped JavaScript collections that their uses prove.
+ *
+ * In a JavaScript file, `const m = new Map()`, `new Set()` and `[]` type as
+ * Map<any, any>, Set<any> and any[] to the checker, so every use lowers to
+ * the checked-dynamic collection operations. When every reference of such a
+ * `const` is a direct use of the collection, the value never leaves the
+ * binding. If its keys, values or elements are then each one primitive type
+ * wherever the program writes and looks them up, the binding lowers as the
+ * typed Map<K, V>, Set<K> or T[], as a JSDoc @type would make it:
+ *
+ * - Map: `m.get(k)`, `m.has(k)`, `m.delete(k)`, `m.clear()`, `m.size`, and
+ *   `m.set(k, v)` as a statement (its result is the map).
+ * - Set: `s.add(k)` as a statement, `s.has(k)`, `s.delete(k)`, `s.clear()`,
+ *   `s.size`.
+ * - Array: `a.push(...)`, `a.pop()`, `a.shift()`, `a.at(i)`, `a.includes(v)`,
+ *   `a.indexOf(v)`, `a.join(separator)`, `a.length`.
+ *
+ * Any other reference (passing, returning, storing, exporting, iterating,
+ * spreading, inspecting the collection, an optional call), a direct `eval`
+ * in the file, or any doubt about a type keeps the binding untyped.
+ *
+ * The types come from carrier declarations in scriptc-overrides.d.ts, which
+ * only the lowering program includes. */
+
+import * as ts from "../ts7/adapter.js";
+import { isJsSourceFile } from "../program.js";
+import { overridesDtsPath } from "../dts-paths.js";
+import type { Lowerer } from "./lowerer.js";
+
+type Primitive = "string" | "number" | "boolean";
+const PRIMITIVES: readonly Primitive[] = ["string", "number", "boolean"];
+
+type Collection = "map" | "set" | "array";
+
+interface Candidate {
+  readonly kind: Collection;
+  readonly declaration: ts.VariableDeclaration;
+  /** Every identifier that names the binding, its declaration's included. */
+  readonly names: ts.Identifier[];
+  /** Map keys or Set members written by set/add. */
+  readonly keyWrites: ts.Expression[];
+  /** Map values written by set. Array elements written by push. */
+  readonly valueWrites: ts.Expression[];
+  /** Arguments compared with keys or elements: get/has/delete, includes/indexOf. */
+  readonly lookups: ts.Expression[];
+  /** Array indexes passed to at(). */
+  readonly indexes: ts.Expression[];
+  /** Separators passed to join(). */
+  readonly separators: ts.Expression[];
+  key: Primitive | undefined;
+  value: Primitive | undefined;
+}
+
+const MAP_LOOKUPS = new Set(["get", "has", "delete"]);
+const SET_LOOKUPS = new Set(["has", "delete"]);
+const ARRAY_READS = new Set(["pop", "shift"]);
+const ARRAY_SEARCHES = new Set(["includes", "indexOf"]);
+/** The expression lowering refuses deeper nesting (SC1090). */
+const MAX_DEPTH = 200;
+
+/** The inferred type of the binding a `new Map()` or `new Set()` initializes,
+ * which the construction adopts. */
+export function initializedCollectionType(
+  lowerer: Lowerer,
+  expr: ts.NewExpression,
+): ts.Type | null {
+  const declaration = expr.parent;
+  return declaration !== undefined &&
+    ts.isVariableDeclaration(declaration) &&
+    declaration.initializer === expr &&
+    ts.isIdentifier(declaration.name)
+    ? inferredCollectionType(lowerer, declaration.name)
+    : null;
+}
+
+/** The inferred collection type of an identifier that names a proven
+ * binding, or null. Computed for every JavaScript file of the program on
+ * first use. */
+export function inferredCollectionType(lowerer: Lowerer, node: ts.Identifier): ts.Type | null {
+  const types = (lowerer.inferredCollectionTypes ??= inferProgram(lowerer));
+  return types.size === 0 ? null : (types.get(node) ?? null);
+}
+
+function inferProgram(lowerer: Lowerer): Map<ts.Node, ts.Type> {
+  const out = new Map<ts.Node, ts.Type>();
+  const carriers = carrierTypes(lowerer);
+  if (carriers === null) return out;
+  for (const file of lowerer.program.getSourceFiles()) {
+    if (file.isDeclarationFile || !isJsSourceFile(file)) continue;
+    for (const candidate of inferFile(lowerer, file)) {
+      const name =
+        candidate.kind === "map"
+          ? `map_${candidate.key}_${candidate.value}`
+          : candidate.kind === "set"
+            ? `set_${candidate.key}`
+            : `array_${candidate.value}`;
+      const type = carriers.get(name);
+      if (type === undefined) continue;
+      for (const id of candidate.names) out.set(id, type);
+    }
+  }
+  return out;
+}
+
+/** The carrier declarations' types, by name (`map_string_number`,
+ * `set_number`, `array_string`, ...), or null when the lowering program does
+ * not include them. */
+function carrierTypes(lowerer: Lowerer): Map<string, ts.Type> | null {
+  const file = lowerer.program.getSourceFile(overridesDtsPath());
+  if (file === undefined) return null;
+  const out = new Map<string, ts.Type>();
+  for (const statement of file.statements.nodes) {
+    if (
+      !ts.isModuleDeclaration(statement) ||
+      !ts.isIdentifier(statement.name) ||
+      statement.name.text !== "__scriptcInferredCollections" ||
+      statement.body === undefined ||
+      !ts.isModuleBlock(statement.body)
+    )
+      continue;
+    for (const member of statement.body.statements.nodes) {
+      if (!ts.isVariableStatement(member)) continue;
+      for (const declaration of member.declarationList.declarations.nodes)
+        if (ts.isIdentifier(declaration.name))
+          out.set(declaration.name.text, lowerer.checker.getTypeAtLocation(declaration.name));
+    }
+  }
+  return out.size > 0 ? out : null;
+}
+
+function inferFile(lowerer: Lowerer, file: ts.SourceFile): Candidate[] {
+  const candidates = new Map<ts.Symbol, Candidate>();
+  let directEval = false;
+  const declarations = (node: ts.Node): void => {
+    if (
+      ts.isCallExpression(node) &&
+      ts.isIdentifier(node.expression) &&
+      node.expression.text === "eval"
+    )
+      directEval = true;
+    if (ts.isVariableDeclaration(node)) {
+      const kind = candidateKind(lowerer, node);
+      const symbol = kind !== null ? lowerer.checker.getSymbolAtLocation(node.name) : undefined;
+      if (kind !== null && symbol !== undefined)
+        candidates.set(symbol, {
+          kind,
+          declaration: node,
+          names: [node.name as ts.Identifier],
+          keyWrites: [],
+          valueWrites: [],
+          lookups: [],
+          indexes: [],
+          separators: [],
+          key: undefined,
+          value: undefined,
+        });
+    }
+  };
+  ts.walkPreorder(file, declarations);
+  if (directEval || candidates.size === 0) return [];
+
+  const names = new Set(
+    [...candidates.values()].map((c) => (c.declaration.name as ts.Identifier).text),
+  );
+  const refused = new Set<ts.Symbol>();
+  const references = (node: ts.Node): void => {
+    if (ts.isIdentifier(node) && names.has(node.text)) {
+      const symbol = referencedSymbol(lowerer, node);
+      const candidate = symbol !== undefined ? candidates.get(symbol) : undefined;
+      if (symbol !== undefined && candidate !== undefined && node !== candidate.declaration.name) {
+        if (recordUse(candidate, node)) candidate.names.push(node);
+        else refused.add(symbol);
+      }
+    }
+  };
+  ts.walkPreorder(file, references);
+  for (const symbol of refused) candidates.delete(symbol);
+  const proven = [...candidates.values()].filter(
+    (c) => c.valueWrites.length > 0 || c.keyWrites.length > 0,
+  );
+  solve(lowerer, proven);
+  return proven.filter((c) =>
+    c.kind === "map" ? c.key && c.value : c.kind === "set" ? c.key : c.value,
+  );
+}
+
+/** The collection a `const` declaration starts empty, when the checker
+ * types it as an untyped Map, Set or array. */
+function candidateKind(lowerer: Lowerer, declaration: ts.VariableDeclaration): Collection | null {
+  if (!ts.isIdentifier(declaration.name) || declaration.initializer === undefined) return null;
+  const list = declaration.parent;
+  if (!ts.isVariableDeclarationList(list) || (list.flags & ts.NodeFlags.Const) === 0) return null;
+  const statement = list.parent;
+  if (
+    ts.isVariableStatement(statement) &&
+    ts.getModifiers(statement)?.some((m) => m.kind === ts.SyntaxKind.ExportKeyword)
+  )
+    return null;
+  const init = declaration.initializer;
+  let kind: Collection | null = null;
+  if (ts.isArrayLiteralExpression(init) && init.elements.nodes.length === 0) kind = "array";
+  else if (
+    ts.isNewExpression(init) &&
+    init.typeArguments === undefined &&
+    (init.arguments?.nodes.length ?? 0) === 0
+  ) {
+    if (lowerer.isStdlibGlobal(init.expression, "Map")) kind = "map";
+    else if (lowerer.isStdlibGlobal(init.expression, "Set")) kind = "set";
+  }
+  if (kind === null) return null;
+  // A JSDoc type, or anything else the checker already knows, stands.
+  const type = lowerer.checker.getTypeAtLocation(declaration.name);
+  const args =
+    kind === "array"
+      ? lowerer.checker.isArrayType(type)
+        ? lowerer.checker.getTypeArguments(type as ts.TypeReference)
+        : []
+      : lowerer.checker.getTypeArguments(type as ts.TypeReference);
+  const arity = kind === "map" ? 2 : 1;
+  if (args.length !== arity || !args.every((a) => (a.flags & ts.TypeFlags.Any) !== 0)) return null;
+  return kind;
+}
+
+function referencedSymbol(lowerer: Lowerer, id: ts.Identifier): ts.Symbol | undefined {
+  const parent = id.parent;
+  if (parent !== undefined && ts.isShorthandPropertyAssignment(parent) && parent.name === id)
+    return lowerer.checker.getShorthandAssignmentValueSymbol(parent);
+  const symbol = lowerer.checker.getSymbolAtLocation(id);
+  // `export { m }` names the export, an alias of the local binding.
+  if (
+    symbol !== undefined &&
+    parent !== undefined &&
+    ts.isExportSpecifier(parent) &&
+    parent.propertyName === undefined &&
+    (symbol.flags & ts.SymbolFlags.Alias) !== 0
+  )
+    return lowerer.checker.getAliasedSymbol(symbol);
+  return symbol;
+}
+
+/** Records a reference's use, or answers false when the reference is not
+ * one of the collection uses this inference proves. */
+function recordUse(candidate: Candidate, id: ts.Identifier): boolean {
+  const access = id.parent;
+  if (
+    access === undefined ||
+    !ts.isPropertyAccessExpression(access) ||
+    access.expression !== id ||
+    access.questionDotToken !== undefined
+  )
+    return false;
+  const name = access.name.text;
+  const call = access.parent;
+  if (call !== undefined && ts.isCallExpression(call) && call.expression === access) {
+    if (call.questionDotToken !== undefined) return false;
+    const args = call.arguments.nodes;
+    if (args.some(ts.isSpreadElement)) return false;
+    switch (candidate.kind) {
+      case "map":
+        if (MAP_LOOKUPS.has(name) && args.length === 1)
+          return (candidate.lookups.push(args[0]!), true);
+        if (name === "clear") return args.length === 0;
+        if (name === "set" && args.length === 2 && isStatement(call)) {
+          candidate.keyWrites.push(args[0]!);
+          candidate.valueWrites.push(args[1]!);
+          return true;
+        }
+        return false;
+      case "set":
+        if (SET_LOOKUPS.has(name) && args.length === 1)
+          return (candidate.lookups.push(args[0]!), true);
+        if (name === "clear") return args.length === 0;
+        if (name === "add" && args.length === 1 && isStatement(call))
+          return (candidate.keyWrites.push(args[0]!), true);
+        return false;
+      case "array":
+        if (name === "push") return (candidate.valueWrites.push(...args), true);
+        if (ARRAY_READS.has(name)) return args.length === 0;
+        if (name === "at" && args.length === 1) return (candidate.indexes.push(args[0]!), true);
+        if (ARRAY_SEARCHES.has(name) && args.length === 1)
+          return (candidate.lookups.push(args[0]!), true);
+        if (name === "join" && args.length <= 1) return (candidate.separators.push(...args), true);
+        return false;
+    }
+  }
+  return (candidate.kind === "array" ? name === "length" : name === "size") && isPlainRead(access);
+}
+
+function isStatement(call: ts.CallExpression): boolean {
+  let node: ts.Node = call;
+  while (node.parent !== undefined && ts.isParenthesizedExpression(node.parent)) node = node.parent;
+  return node.parent !== undefined && ts.isExpressionStatement(node.parent);
+}
+
+/** A read that neither writes nor deletes the property. */
+function isPlainRead(access: ts.PropertyAccessExpression): boolean {
+  let node: ts.Node = access;
+  while (node.parent !== undefined && ts.isParenthesizedExpression(node.parent)) node = node.parent;
+  const parent = node.parent;
+  if (parent === undefined) return false;
+  if (
+    ts.isBinaryExpression(parent) &&
+    parent.left === node &&
+    isAssignment(parent.operatorToken.kind)
+  )
+    return false;
+  if (
+    (ts.isPrefixUnaryExpression(parent) || ts.isPostfixUnaryExpression(parent)) &&
+    (parent.operator === ts.SyntaxKind.PlusPlusToken ||
+      parent.operator === ts.SyntaxKind.MinusMinusToken)
+  )
+    return false;
+  return !ts.isDeleteExpression(parent);
+}
+
+function isAssignment(kind: ts.SyntaxKind): boolean {
+  return kind >= ts.SyntaxKind.FirstAssignment && kind <= ts.SyntaxKind.LastAssignment;
+}
+
+/** Decides each candidate's key and value types: a fixed point in which a
+ * decided candidate's reads type the expressions built from them. */
+function solve(lowerer: Lowerer, candidates: Candidate[]): void {
+  const bySymbol = new Map<ts.Symbol, Candidate>();
+  for (const c of candidates) {
+    const symbol = lowerer.checker.getSymbolAtLocation(c.declaration.name);
+    if (symbol) bySymbol.set(symbol, c);
+  }
+  for (let changed = true; changed;) {
+    changed = false;
+    for (const c of candidates) {
+      if (c.kind !== "array" && c.key === undefined) {
+        const key = uniform(lowerer, bySymbol, c.keyWrites);
+        if (key !== null) {
+          c.key = key;
+          changed = true;
+        }
+      }
+      if (c.kind !== "set" && c.value === undefined) {
+        // Each candidate type is assumed for this collection's own reads.
+        const accepted = PRIMITIVES.filter((assumed) => {
+          c.value = assumed;
+          const found = uniform(lowerer, bySymbol, c.valueWrites);
+          c.value = undefined;
+          return found === assumed;
+        });
+        if (accepted.length === 1) {
+          c.value = accepted[0];
+          changed = true;
+        }
+      }
+    }
+  }
+  for (const c of candidates) {
+    const element = c.kind === "array" ? c.value : c.key;
+    // Lookups compare with the stored type; an index is a number.
+    const lookupsAgree =
+      element !== undefined &&
+      c.lookups.every((e) => primitiveOf(lowerer, bySymbol, e, new Set()) === element) &&
+      c.indexes.every((e) => primitiveOf(lowerer, bySymbol, e, new Set()) === "number") &&
+      c.separators.every((e) => primitiveOf(lowerer, bySymbol, e, new Set()) === "string");
+    if (!lookupsAgree) {
+      c.key = undefined;
+      c.value = undefined;
+    }
+  }
+}
+
+/** The one primitive type of every expression, or null. */
+function uniform(
+  lowerer: Lowerer,
+  bySymbol: Map<ts.Symbol, Candidate>,
+  exprs: readonly ts.Expression[],
+): Primitive | null {
+  let found: Primitive | null = null;
+  for (const e of exprs) {
+    const p = primitiveOf(lowerer, bySymbol, e, new Set());
+    if (p === null || (found !== null && p !== found)) return null;
+    found = p;
+  }
+  return found;
+}
+
+function checkerPrimitive(type: ts.Type): Primitive | null {
+  const flags = type.flags;
+  if (flags & (ts.TypeFlags.String | ts.TypeFlags.StringLiteral)) return "string";
+  if (flags & (ts.TypeFlags.Number | ts.TypeFlags.NumberLiteral)) return "number";
+  if (flags & (ts.TypeFlags.Boolean | ts.TypeFlags.BooleanLiteral)) return "boolean";
+  return null;
+}
+
+/** The primitive type `e` always evaluates to, or null. The checker answers
+ * where it knows; where it types `e` as `any`, the value's structure does:
+ * a read of a decided candidate, arithmetic and comparisons over such
+ * values, and constants defined from them. An expression nested deeper than
+ * the lowering's nesting limit answers null. */
+function primitiveOf(
+  lowerer: Lowerer,
+  bySymbol: Map<ts.Symbol, Candidate>,
+  e: ts.Expression,
+  visiting: Set<ts.Node>,
+  depth = 0,
+): Primitive | null {
+  if (depth > MAX_DEPTH) return null;
+  const type = lowerer.checker.getTypeAtLocation(e);
+  if ((type.flags & (ts.TypeFlags.Any | ts.TypeFlags.Unknown)) === 0) return checkerPrimitive(type);
+  const of = (x: ts.Expression): Primitive | null =>
+    primitiveOf(lowerer, bySymbol, x, visiting, depth + 1);
+  if (ts.isParenthesizedExpression(e)) return of(e.expression);
+  if (ts.isPrefixUnaryExpression(e)) {
+    if (e.operator === ts.SyntaxKind.ExclamationToken) return "boolean";
+    return e.operator !== ts.SyntaxKind.PlusPlusToken &&
+      e.operator !== ts.SyntaxKind.MinusMinusToken &&
+      of(e.operand) === "number"
+      ? "number"
+      : null;
+  }
+  if (ts.isConditionalExpression(e)) {
+    const yes = of(e.whenTrue);
+    return yes !== null && yes === of(e.whenFalse) ? yes : null;
+  }
+  if (ts.isBinaryExpression(e)) {
+    const op = e.operatorToken.kind;
+    switch (op) {
+      case ts.SyntaxKind.LessThanToken:
+      case ts.SyntaxKind.GreaterThanToken:
+      case ts.SyntaxKind.LessThanEqualsToken:
+      case ts.SyntaxKind.GreaterThanEqualsToken:
+      case ts.SyntaxKind.EqualsEqualsEqualsToken:
+      case ts.SyntaxKind.ExclamationEqualsEqualsToken:
+      case ts.SyntaxKind.EqualsEqualsToken:
+      case ts.SyntaxKind.ExclamationEqualsToken:
+        return "boolean";
+      case ts.SyntaxKind.QuestionQuestionToken:
+      case ts.SyntaxKind.BarBarToken: {
+        // `m.get(k) ?? d` and `a.pop() || d`: the stored type or a default of it.
+        const stored = storedRead(lowerer, bySymbol, e.left);
+        const fallback = of(e.right);
+        return stored !== null && stored === fallback ? stored : null;
+      }
+      case ts.SyntaxKind.PlusToken: {
+        const left = of(e.left);
+        const right = of(e.right);
+        if (left === "number" && right === "number") return "number";
+        return left !== null && right !== null && (left === "string" || right === "string")
+          ? "string"
+          : null;
+      }
+      case ts.SyntaxKind.MinusToken:
+      case ts.SyntaxKind.AsteriskToken:
+      case ts.SyntaxKind.SlashToken:
+      case ts.SyntaxKind.PercentToken:
+      case ts.SyntaxKind.AsteriskAsteriskToken:
+      case ts.SyntaxKind.BarToken:
+      case ts.SyntaxKind.AmpersandToken:
+      case ts.SyntaxKind.CaretToken:
+      case ts.SyntaxKind.LessThanLessThanToken:
+      case ts.SyntaxKind.GreaterThanGreaterThanToken:
+      case ts.SyntaxKind.GreaterThanGreaterThanGreaterThanToken:
+        return of(e.left) === "number" && of(e.right) === "number" ? "number" : null;
+      default:
+        return null;
+    }
+  }
+  if (ts.isCallExpression(e) || ts.isPropertyAccessExpression(e))
+    return candidateRead(lowerer, bySymbol, e);
+  // A constant defined from such an expression.
+  if (ts.isIdentifier(e) && !visiting.has(e)) {
+    const symbol = lowerer.checker.getSymbolAtLocation(e);
+    const declaration = symbol?.valueDeclaration;
+    if (
+      declaration !== undefined &&
+      ts.isVariableDeclaration(declaration) &&
+      declaration.initializer !== undefined &&
+      ts.isVariableDeclarationList(declaration.parent) &&
+      (declaration.parent.flags & ts.NodeFlags.Const) !== 0
+    ) {
+      visiting.add(e);
+      const p = primitiveOf(lowerer, bySymbol, declaration.initializer, visiting, depth + 1);
+      visiting.delete(e);
+      return p;
+    }
+  }
+  return null;
+}
+
+/** The candidate a use reads, when `e` is `c.member` or `c.member(...)` of a
+ * candidate `c`. */
+function candidateOf(
+  lowerer: Lowerer,
+  bySymbol: Map<ts.Symbol, Candidate>,
+  e: ts.Expression,
+): { candidate: Candidate; member: string } | null {
+  const access = ts.isCallExpression(e) ? e.expression : e;
+  if (!ts.isPropertyAccessExpression(access) || !ts.isIdentifier(access.expression)) return null;
+  const symbol = lowerer.checker.getSymbolAtLocation(access.expression);
+  const candidate = symbol !== undefined ? bySymbol.get(symbol) : undefined;
+  return candidate !== undefined ? { candidate, member: access.name.text } : null;
+}
+
+/** The type of a candidate's read that may answer undefined: `m.get(k)`,
+ * `a.pop()`, `a.shift()`, `a.at(i)`. */
+function storedRead(
+  lowerer: Lowerer,
+  bySymbol: Map<ts.Symbol, Candidate>,
+  e: ts.Expression,
+): Primitive | null {
+  while (ts.isParenthesizedExpression(e)) e = e.expression;
+  if (!ts.isCallExpression(e)) return null;
+  const read = candidateOf(lowerer, bySymbol, e);
+  if (read === null) return null;
+  const { candidate, member } = read;
+  if (candidate.kind === "map" && member === "get") return candidate.value ?? null;
+  if (candidate.kind === "array" && (ARRAY_READS.has(member) || member === "at"))
+    return candidate.value ?? null;
+  return null;
+}
+
+/** The type of a candidate's read that always answers a value: size,
+ * length, has, includes, indexOf, push. */
+function candidateRead(
+  lowerer: Lowerer,
+  bySymbol: Map<ts.Symbol, Candidate>,
+  e: ts.Expression,
+): Primitive | null {
+  const read = candidateOf(lowerer, bySymbol, e);
+  if (read === null) return null;
+  const { candidate, member } = read;
+  if (!ts.isCallExpression(e)) return member === "size" || member === "length" ? "number" : null;
+  if (member === "has" || member === "includes" || member === "delete") return "boolean";
+  if (member === "indexOf" || (candidate.kind === "array" && member === "push")) return "number";
+  if (candidate.kind === "array" && member === "join") return "string";
+  return null;
+}

packages/compiler/src/frontend/lowering/lower-classes.ts

@@ -116,6 +116,7 @@ import { PoisonError, dynFallbackType, jsFuncNameOf, newFnCtx, own } from "./low
 import { isAssertClassNew } from "./assert-instances.js";
 import { isCjsRequireGlobal } from "./lower-node-module.js";
 import { lowerArrayConstructor } from "./containers/array-construction.js";
+import { initializedCollectionType } from "./js-collection-types.js";
 import { lowerMapSeedNew, lowerSetSeedNew } from "./containers/collection-methods.js";
 import { lowerObjectConstructor } from "./containers/indexed-objects.js";
 import { knownBufEncoding, lowerBufferConstruct } from "./containers/bytes.js";
@@ -10699,7 +10700,11 @@ export function lowerNew(lowerer: Lowerer, expr: ts.NewExpression): IrExpr {
       // expression). The CONTEXTUAL type carries the annotation — adopt
       // it when it is a supported map. TS type arguments keep winning:
       // their expression type already maps.
-      if (mapped?.kind !== "map" || contextualSeed) {
+      const inferred = initializedCollectionType(lowerer, expr);
+      if (inferred) {
+        tsType = inferred;
+        mapped = lowerer.mapTypeOf(inferred);
+      } else if (mapped?.kind !== "map" || contextualSeed) {
         const ctx = lowerer.checker.getContextualType(expr);
         const ctxMapped = ctx ? lowerer.mapTypeOf(ctx) : null;
         if (ctx && ctxMapped?.kind === "map") {
@@ -10839,7 +10844,8 @@ export function lowerNew(lowerer: Lowerer, expr: ts.NewExpression): IrExpr {
     // `new Set<T>()`: construct + bulk add from supported values,
     // including live collection cursors. Duplicates collapse in insertion order.
     if (ts.isIdentifier(identifier) && lowerer.isStdlibGlobal(identifier, "Set")) {
-      const tsType = lowerer.typeOf(expr);
+      const inferred = initializedCollectionType(lowerer, expr);
+      const tsType = inferred ?? lowerer.typeOf(expr);
       let mapped = lowerer.mapTypeOf(tsType);
       const fieldType = assignedThisFieldType(lowerer, expr);
       if (mapped?.kind !== "set" && fieldType?.kind === "set") mapped = fieldType;

packages/compiler/src/frontend/lowering/lowerer.ts

@@ -156,6 +156,7 @@ import {
   trackClassFieldCreation,
 } from "./class-dynamic-dispatch.js";
 import { guardClassAttributeWrites } from "./class-attributes.js";
+import { inferredCollectionType } from "./js-collection-types.js";
 import { guardFrozenWrites, markSloppyRecordWrites } from "./frozen-records.js";
 import { applyHandleProperties } from "./handle-properties.js";
 import { installBufferValue } from "./containers/bytes.js";
@@ -2853,6 +2854,9 @@ export class Lowerer {
   /** Parameters read as `any` in the error-path copy of an error-first
    * callback. */
   readonly anySymbols = new Set<ts.Symbol>();
+  /** The collection types that uses prove for untyped JavaScript bindings,
+   * by every identifier naming one (js-collection-types.ts). */
+  inferredCollectionTypes: Map<ts.Node, ts.Type> | undefined = undefined;
   /** overload-choice's answers per call, binding and function. */
   readonly dynamicOverloadAnswers = new Map<ts.Node, boolean>();
   /** Error parameters of an error-first callback with a known value in
@@ -9075,6 +9079,10 @@ export class Lowerer {
   }
 
   typeOf(node: ts.Node): ts.Type {
+    if (ts.isIdentifier(node)) {
+      const inferred = inferredCollectionType(this, node);
+      if (inferred) return inferred;
+    }
     if (this.anySymbols.size > 0 && this.readsAnySymbol(node)) return this.checker.getAnyType();
     // A checked-dynamic argument that another overload could take at run
     // time leaves the resolved overload's return a guess, and JS bindings

packages/compiler/test/js-collection-types.test.ts

@@ -0,0 +1,64 @@
+import { mkdtempSync, readFileSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { expect, test } from "vitest";
+import { compile, deserializeModule, validateModule } from "../src/index.js";
+
+// Corpus 7985 exercises every refusal: only the collections whose uses
+// prove their types lower as typed collections; the rest stay untyped.
+test("untyped JavaScript collections take the types their uses prove, and only those", async () => {
+  const dir = mkdtempSync(join(tmpdir(), "scriptc-js-collection-types-"));
+  try {
+    const entry = join(
+      import.meta.dirname,
+      "../../../tests/corpus/7985-untyped-collection-types.mjs",
+    );
+    const output = join(dir, "main.ir.json");
+    const lowered = await compile(entry, { outputKind: "ir", outDir: dir, outPath: output });
+    if (!lowered.ok) throw new Error(JSON.stringify(lowered.diagnostics));
+    const module = deserializeModule(readFileSync(output, "utf8"));
+    expect(validateModule(module)).toEqual([]);
+    const types = new Map(module.globals.map((g) => [g.name, JSON.stringify(g.type)]));
+    const map = (key: string, value: string) =>
+      JSON.stringify({ kind: "map", key: { kind: key }, value: { kind: value } });
+    const set = (elem: string) => JSON.stringify({ kind: "set", elem: { kind: elem } });
+    const array = (elem: string) => JSON.stringify({ kind: "array", elem: { kind: elem } });
+    // Untyped collections hold checked-dynamic keys, values and elements.
+    const typed = (t: string) => !t.includes('"dyn"');
+    expect(Object.fromEntries([...types].filter(([, t]) => typed(t)))).toEqual({
+      counts: map("string", "f64"),
+      seen: set("f64"),
+      stack: array("f64"),
+      words: array("string"),
+      byNumber: map("f64", "string"),
+      nums: array("f64"),
+      flags: set("bool"),
+      cache: map("string", "f64"),
+      "%loaded": '{"kind":"bool"}',
+    });
+    for (const untyped of [
+      "strings",
+      "numbers",
+      "mixedKeys",
+      "mixedValues",
+      "mixedElements",
+      "passed",
+      "stored",
+      "iterated",
+      "eachSet",
+      "shown",
+      "chained",
+      "optional",
+      "indexed",
+      "truncated",
+      "spread",
+      "unread",
+      "exported",
+      "listed",
+      "renamed",
+    ])
+      expect(typed(types.get(untyped)!), untyped).toBe(false);
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});

packages/compiler/test/ts7/baselines/order-parity/tests/corpus/7985-untyped-collection-types.mjs.json

@@ -0,0 +1,6 @@
+{
+ "order": [
+  "<repo>/tests/corpus/7985-untyped-collection-types.mjs"
+ ],
+ "diags": []
+}

tests/corpus/7985-untyped-collection-types.mjs

@@ -0,0 +1,125 @@
+// Untyped Maps, Sets and arrays whose uses prove their key, value and
+// element types compile as typed collections; any other use keeps them
+// untyped. Both must answer as Node does, including SameValueZero keys,
+// lookups of another type, and every use that refuses the typing.
+
+// Typed: counters, membership sets and stacks used only through their
+// methods.
+const counts = new Map();
+const seen = new Set();
+const stack = [];
+const words = [];
+for (let i = 0; i < 40; i++) {
+  const key = "k" + (i % 7);
+  counts.set(key, (counts.get(key) ?? 0) + 1);
+  seen.add(i % 5);
+  stack.push(i * 2);
+  if (stack.length > 6) stack.shift();
+  words.push(key);
+}
+console.log(counts.size, counts.get("k3"), counts.get("missing"), counts.has("k6"), counts.delete("k0"), counts.size);
+console.log(seen.size, seen.has(4), seen.has(5), seen.delete(4), seen.has(4));
+console.log(stack.length, stack.at(-1), stack.at(0), stack.pop(), stack.includes(70), stack.indexOf(72), words.join("|").length);
+
+// SameValueZero keys and NaN elements on typed collections.
+const byNumber = new Map();
+byNumber.set(NaN, "nan");
+byNumber.set(-0, "zero");
+byNumber.set(1.5, "x");
+console.log(byNumber.get(NaN), byNumber.get(0), byNumber.get(-0), byNumber.size, byNumber.has(1.5));
+const nums = [];
+nums.push(NaN);
+nums.push(-0);
+console.log(nums.includes(NaN), nums.indexOf(NaN), nums.includes(0), nums.indexOf(0));
+const flags = new Set();
+flags.add(true);
+console.log(flags.has(true), flags.has(false), flags.size);
+
+// Captured by a closure that only uses it.
+const cache = new Map();
+const latest = () => cache.get("c") ?? -1;
+console.log(latest());
+for (let i = 0; i < 3; i++) cache.set("c", i * 10);
+console.log(latest(), cache.size);
+
+// Lookups of another type: the typing refuses, and the answers stay Node's.
+const strings = new Map();
+strings.set("1", "one");
+console.log(strings.get(1), strings.get("1"), strings.has(1));
+const numbers = [];
+numbers.push(1);
+console.log(numbers.includes("1"), numbers.indexOf("1"), numbers.at("0"));
+
+// Mixed key, value and element types.
+const mixedKeys = new Map();
+mixedKeys.set(1, "number");
+mixedKeys.set("1", "string");
+console.log(mixedKeys.size, mixedKeys.get(1), mixedKeys.get("1"));
+const mixedValues = new Map();
+mixedValues.set("a", 1);
+mixedValues.set("b", "two");
+console.log(mixedValues.get("a"), mixedValues.get("b"));
+const mixedElements = [];
+mixedElements.push(1);
+mixedElements.push("two");
+console.log(mixedElements.join(","), mixedElements.length);
+
+// Escapes and other uses that keep the collection untyped.
+const passed = new Map();
+passed.set("x", 1);
+function sizeOf(m) {
+  return m.size;
+}
+console.log(sizeOf(passed));
+function make() {
+  const made = new Set();
+  made.add("y");
+  return made;
+}
+console.log(make().has("y"));
+const stored = [];
+stored.push(3);
+const holder = { list: stored };
+holder.list.push(4);
+console.log(stored.length, holder.list.join("+"));
+const iterated = new Map();
+iterated.set("a", 1);
+iterated.set("b", 2);
+for (const [k, v] of iterated) console.log("entry", k, v);
+const eachSet = new Set();
+eachSet.add(3);
+eachSet.forEach((v) => console.log("member", v));
+const shown = new Map();
+shown.set("s", 1);
+console.log(shown);
+const chained = new Map();
+chained.set("a", 1).set("b", 2);
+console.log(chained.size);
+const optional = new Map();
+optional.set("o", 1);
+console.log(optional?.get("o"), optional.get?.("o"));
+const indexed = [];
+indexed.push(1);
+indexed[3] = 2;
+console.log(indexed.length, indexed.join(","));
+const truncated = [];
+truncated.push(1, 2, 3);
+truncated.length = 1;
+console.log(truncated.length, truncated.at(0));
+const spread = [];
+spread.push(5);
+console.log(Math.max(...spread));
+const unread = new Map();
+console.log(unread.size, unread.get("none"));
+export const exported = new Map();
+exported.set("e", 1);
+console.log(exported.get("e"));
+// Exported through an export list, under its own name or another: an
+// importer sees an untyped Map.
+const listed = new Map();
+listed.set("l", 1);
+console.log(listed.get("l"));
+const renamed = new Map();
+renamed.set("r", 1);
+console.log(renamed.get("r"));
+export { listed, renamed as alias };

tests/harness/corpus-costs/7985-untyped-collection-types.mjs.json

@@ -0,0 +1 @@
+{ "plain": 2400, "san": 2600 }