@solidjs/signals 2.0.0-rc.1 → 2.0.0-rc.10

package/dist/prod/store/utils.js

@@ -12,44 +12,873 @@ import "../core/effect.js";
 
 import { createMemo } from "../signals.js";
 
-import { ownEnumerableKeys, $PROXY } from "./store.js";
+import { $PROXY, $RECORD, $TARGET, ownEnumerableKeys } from "./store.js";
 
 function trueFn() {
     return true;
 }
 
-const propTraps = {
-    get(e, r, t) {
-        if (r === $PROXY) return t;
-        return e.get(r);
+/** @internal What a source ENTRY is, decided once when the view is built
+ * (`merge()` learns it while flattening; `omit()` from its argument) and
+ * carried beside the entry — `MergeView.kinds[i]`, `OmitView.kind` — so no
+ * read has to ask. Asking is the cost: any brand check on a Proxy is a trap
+ * (`instanceof` is a `getPrototypeOf` trap, ~20 ns on a store — as much as
+ * the read itself), and a merge over a store did two per read. */ const SOURCE_PLAIN = 0;
+
+ // a plain object: own keys fixed, data is data
+const SOURCE_OMIT = 1;
+
+ // an `OmitView` record (only as a merge entry)
+const SOURCE_PROXY = 2;
+
+ // a store or foreign proxy: everything is a trap
+const SOURCE_MEMO = 3;
+
+ // a function source (merge's memo): swaps objects
+const SOURCE_MERGE = 4;
+
+ // a `MergeView` record (only as an omit's source)
+const EMPTY = Object.freeze({});
+
+// The object behind a LEAF entry (any kind but OMIT): a memo is read
+// (tracked, as merge's own reads are) and a nullish result has no keys.
+function leafOf(e, n) {
+    return n === SOURCE_MEMO ? (e = e()) == null ? EMPTY : e : e;
+}
+
+/** @internal The record behind an `omit()` proxy: `source` with `hidden`
+ * keys removed. It is the proxy's TARGET, so the shared handler reads it as
+ * plain fields — no per-instance closures — and it is what props consumers
+ * walk directly (`merge`, `spread`, `ssrElement`): a view never materializes
+ * a copy, and a consumer that knows the record never goes through its traps
+ * (a `getOwnPropertyDescriptor` trap per key allocates a descriptor and a
+ * getter, so enumerating a proxy costs more than the copy it was avoiding).
+ * `hidden` is a key list or a predicate (`omit(props, k => k[0] === "$")`).
+ *
+ * An omit over a `merge()` holds the merge's RECORD (`MergeView`, kind
+ * `SOURCE_MERGE`) — never its proxy, so no read hops through a trap — and is
+ * one record however many leaves the merge has. A consumer walks it as ONE
+ * filtered entry (`sourceKeys` / `sourceGet` recurse into the merge's
+ * sources by function call), and a later `merge()` over it carries the
+ * record as one entry instead of copying its leaves: on a component chain of
+ * defaults + omit + spread (`merge(omit(merge(omit(props))))`) the layers
+ * nest as records, each a few fields, where a flatten to leaf views built a
+ * view and a combined key list per leaf per layer — the largest allocation
+ * of a Kobalte-shaped render. A merge leaf may be merge's memo for a
+ * function source; it is resolved on access. */ class OmitView {
+    source;
+    kind;
+    hidden;
+    /** see `MergeView.table` */
+    table=0;
+    /** see `tableOwnKeys` / `tableDescriptor` */
+    keys=undefined;
+    descs=undefined;
+    constructor(e, 
+    /** of `source` — PLAIN, PROXY (a store or a foreign proxy), MEMO, or
+     * MERGE (a `MergeView` record); never OMIT, a view over a view folds
+     * into one. */
+    n, t) {
+        this.source = e;
+        this.kind = n;
+        this.hidden = t;
+    }
+}
+
+function isHidden(e, n) {
+    const t = e.hidden;
+    return typeof t === "function" ? t(n) : t.includes(n);
+}
+
+// Both filters as one. Two key lists stay a key list (one `includes`, no
+// closure); a predicate on either side needs a closure. An omit over a
+// merge builds one combined list per leaf, per component layer, so the
+// copy's form matters in every tier: `concat` runs the species/spreadable
+// protocol (2–3× the cost of a copy once optimized), a hand loop is 2–4×
+// `concat` in the interpreter and baseline tiers (a bytecode per element
+// against one builtin), and a presized `new Array(n)` is holey, which takes
+// `includes` off its fast path. `slice` + `push` of the (short) second list
+// is within a third of the best form in every tier, and packed.
+function combineHidden(e, n) {
+    if (typeof e !== "function" && typeof n !== "function") {
+        const t = e.slice();
+        for (let e = 0; e < n.length; e++) t.push(n[e]);
+        return t;
+    }
+    return t => (typeof e === "function" ? e(t) : e.includes(t)) || (typeof n === "function" ? n(t) : n.includes(t));
+}
+
+// The object a view filters (see `leafOf`).
+function viewSource(e) {
+    return leafOf(e.source, e.kind);
+}
+
+// The record behind a `$PROXY`-marked object that is one of OUR views, else
+// undefined for a store or a foreign proxy. ONE read: the view traps answer
+// `$RECORD` first thing and a store's `get` trap answers it `undefined` on
+// its symbol fast path, so the question never reaches a store's generic
+// read path (firewall gate, tracked key read), which is what an unknown
+// symbol would take — and a foreign proxy forwards it to a target that has
+// no such key. Asking `$TARGET` first and then each view kind was two to
+// four trap hops per source at every `merge()`, `omit()` and `ssrElement`.
+// Call only after `$PROXY in o` is known true.
+function recordOf(e) {
+    return e[$RECORD];
+}
+
+/** @internal The `OmitView` behind an `omit()` proxy, or undefined. */ function omitView(e) {
+    if (e == null || !($PROXY in e)) return undefined;
+    const n = e[$RECORD];
+    return n instanceof OmitView ? n : undefined;
+}
+
+// A props SOURCE ENTRY is a plain object, a proxy (store, merge, omit — the
+// last two are normally unwrapped first: `mergeView` / `omitView`), a memo,
+// or an `OmitView` record, and travels with its kind. These three answer for
+// an entry what `Object.keys` / `in` / `[]` answer for an object, so every
+// consumer walks entries with one code path, an `OmitView` is filtered
+// rather than materialized, and nothing is asked of a proxy but the read.
+// Own keys of a leaf: a proxy answers through ONE `ownKeys` trap (a store's
+// keeps the key set tracked; `Object.keys` on a proxy would add a descriptor
+// trap per key); a plain object its enumerable string keys. What a memo
+// holds is only known once read.
+function leafKeys(e, n) {
+    if (n === SOURCE_PLAIN) return Object.keys(e);
+    if (n === SOURCE_PROXY || e[$PROXY] === e) return Reflect.ownKeys(e);
+    return Object.keys(e);
+}
+
+/** @internal Own string keys of a source entry — every consumer skips
+ * symbols itself. */ function sourceKeys(e, n) {
+    if (n === SOURCE_OMIT) {
+        if (e.kind === SOURCE_MERGE) return mergeKeysOf(e.source, false, e);
+        const n = leafKeys(viewSource(e), e.kind);
+        const t = [];
+        for (let r = 0; r < n.length; r++) if (!isHidden(e, n[r])) t.push(n[r]);
+        return t;
+    }
+    return leafKeys(leafOf(e, n), n);
+}
+
+/** @internal `key in entry`. */ function sourceHas(e, n, t) {
+    if (n === SOURCE_OMIT) {
+        if (isHidden(e, t)) return false;
+        return e.kind === SOURCE_MERGE ? mergeHas(e.source, t) : t in viewSource(e);
+    }
+    return t in leafOf(e, n);
+}
+
+/** @internal `entry[key]` — the source's getter runs once, here. */ function sourceGet(e, n, t) {
+    if (n === SOURCE_OMIT) {
+        if (isHidden(e, t)) return undefined;
+        return e.kind === SOURCE_MERGE ? mergeGet(e.source, t) : viewSource(e)[t];
+    }
+    return leafOf(e, n)[t];
+}
+
+// An entry whose own key set is fixed: a plain object, or a view over one —
+// an omit of a plain object, or of a merge whose entries all are. Not a
+// store (its key set is a tracked signal), not a merge memo source (it
+// swaps whole objects), and not any proxy that declares itself with
+// `$PROXY in s` — a frames slot proxy answers `has` for every key and lists
+// none, so only the `in` walk is right for it.
+function entryHasStaticKeys(e, n) {
+    if (n === SOURCE_PLAIN) return true;
+    if (n !== SOURCE_OMIT) return false;
+    if (e.kind === SOURCE_PLAIN) return true;
+    return e.kind === SOURCE_MERGE && mergeHasStaticKeys(e.source);
+}
+
+function mergeHasStaticKeys(e) {
+    const n = e.sources, t = e.kinds;
+    for (let e = 0; e < n.length; e++) if (!entryHasStaticKeys(n[e], t[e])) return false;
+    return true;
+}
+
+/** @internal Whether the own key set of a props object cannot change
+ * reactively: a plain object, or a merge/omit view over plain objects only.
+ * A consumer may then decide from `Object.getOwnPropertyDescriptor` once —
+ * "no `children` key" or "a data `children`" holds for the object's lifetime,
+ * so no tracking scope is needed for it (#3388). For a store, or a view with
+ * a store or memo leaf, keys can appear later and the reactive path is the
+ * only correct one. */ function hasStaticKeys(e) {
+    if (!($PROXY in e)) return true;
+    const n = recordOf(e);
+    if (n instanceof MergeView) return mergeHasStaticKeys(n);
+    return n !== undefined && entryHasStaticKeys(n, SOURCE_OMIT);
+}
+
+/**
+ * Whether `o[key]` can never change for the lifetime of `o`: the key is a
+ * data property of a plain object, or is absent from an object whose key set
+ * is fixed. A getter, a key on a store, a memo-backed `merge()` source, or
+ * any key of an object whose keys can appear later (a store) is not static.
+ *
+ * Looks through `merge()`/`omit()` views to the leaf that owns the key. Any
+ * object will do, but props are the case it exists for: the compiler encodes
+ * a literal at the call site (`as="button"`) as a data property and an
+ * expression (`as={isLink() ? "a" : "button"}`) as a getter, so a component
+ * library reads the caller's own static/dynamic classification of a prop at
+ * runtime — identically on server and client, the compiler emitting the same
+ * own DESCRIPTORS on both (the object behind them may differ: the server
+ * builds props as a plain-prototype instance with shared getters) — and can
+ * take a no-computation path for the literal:
+ *
+ * ```tsx
+ * const Tag = dynamic(() => props.as, { static: isStatic(props, "as") });
+ * ```
+ *
+ * One descriptor lookup; no read of the value, nothing tracked.
+ */ function isStatic(e, n) {
+    if ($PROXY in e) {
+        // A store answers its descriptor trap with a value; through a view the
+        // descriptor is truthful (see `sourceDescriptor`). A foreign proxy is
+        // opaque: nothing about it is known to be fixed.
+        if (viewOf(e) === undefined) return false;
+        const t = Reflect.getOwnPropertyDescriptor(e, n);
+        return t === undefined ? hasStaticKeys(e) : t.get === undefined;
+    }
+    const t = Reflect.getOwnPropertyDescriptor(e, n);
+    return t === undefined || t.get === undefined && t.set === undefined;
+}
+
+function accessorDescriptor(e, n = true) {
+    return {
+        configurable: true,
+        enumerable: n,
+        get: e,
+        set: trueFn
+    };
+}
+
+/** The descriptor a consumer should see for `key` on an entry —
+ * the view proxies answer `getOwnPropertyDescriptor` with it, so it tells the
+ * truth through any depth of merge/omit layers.
+ *
+ * A DATA descriptor means "nothing reactive can hide behind this value": the
+ * key is a data property of a plain-object leaf — the compiler's own
+ * encoding of a static attribute. Everything else is an accessor: a getter
+ * on a leaf, a key on a store proxy (its "data" is a signal), or a key on a
+ * merge memo source (the whole object is reactive). That is what lets a
+ * consumer skip a reactive node for a static prop at the bottom of a
+ * component chain, and it is why the store case must NOT forward the store's
+ * own descriptor, which reports a value.
+ *
+ * `configurable: true` always — the target has no such property, and the
+ * Proxy invariants forbid reporting a non-configurable one. */
+// `present` says the caller has already established `key in s` (a trap's
+// shadowing walk did), so a store is not asked a second time.
+function sourceDescriptor(e, n, t, r = false) {
+    if (n === SOURCE_OMIT) {
+        if (isHidden(e, t)) return undefined;
+        return sourceDescriptor(e.source, e.kind, t, r);
+    }
+    // An omit's merge record: the descriptor of the last entry that has the
+    // key, as the merge proxy's own trap answers (see `mergeDescriptor`).
+        if (n === SOURCE_MERGE) return mergeDescriptor(e, t);
+    // A memo source (`merge(() => …)`) is reactive wholesale: whatever shape
+    // the memo's current object has, the key is an accessor here.
+        if (n === SOURCE_MEMO) {
+        return r || t in leafOf(e, n) ? accessorDescriptor(() => leafOf(e, n)[t]) : undefined;
+    }
+    if (n === SOURCE_PROXY) {
+        // Another view (an omit's source may be a merge proxy) already answers
+        // truthfully. A store's reported "data" is a signal, and a foreign proxy
+        // (frames slot props) has no own descriptors: for both, existence is
+        // `in` and the kind is accessor — one trap, and never the store's
+        // descriptor path.
+        if (recordOf(e) !== undefined) return Reflect.getOwnPropertyDescriptor(e, t);
+        return r || t in e ? accessorDescriptor(() => e[t]) : undefined;
+    }
+    const i = Reflect.getOwnPropertyDescriptor(e, t);
+    if (i === undefined) return undefined;
+    if (i.get !== undefined || i.set !== undefined) return accessorDescriptor(() => e[t], i.enumerable);
+    // The proxy target has no such key, so the descriptor must be configurable;
+    // Reflect's is a fresh object, so a configurable one is handed out as is.
+        if (i.configurable) return i;
+    return {
+        configurable: true,
+        enumerable: i.enumerable,
+        writable: true,
+        value: i.value
+    };
+}
+
+// Own ENUMERABLE keys, symbols included, of an entry — the user-facing key
+// set (`Object.keys(merged)`), where enumerability matters (#2769).
+function sourceEnumerableKeys(e, n) {
+    {
+        if (e.kind === SOURCE_MERGE) return mergeEnumerableKeys(e.source, e);
+        const n = ownEnumerableKeys(viewSource(e));
+        const t = [];
+        for (let r = 0; r < n.length; r++) if (!isHidden(e, n[r])) t.push(n[r]);
+        return t;
+    }
+}
+
+// The target of a merge() proxy: the flattened sources, read by one shared
+// handler — like OmitView, no per-instance closures. `sources` is what
+// `mergeSources` answers.
+/** @internal */ class MergeView {
+    sources;
+    kinds;
+    /** key → the plain leaf that owns it (later sources win), built by an
+     * enumeration or once the reads have paid for it (see `resolvedTable`)
+     * when every leaf has static keys; `null` when one doesn't. Until then
+     * the slot counts the per-key trap reads so far. One slot rather than a
+     * counter field of its own: a view is built per source per component
+     * layer, and each field initializer is a measurable share of a
+     * constructor that small in the lower JIT tiers. */
+    table=0;
+    /** see `tableOwnKeys` / `tableDescriptor` */
+    keys=undefined;
+    descs=undefined;
+    constructor(e, 
+    /** `kinds[i]` is what `sources[i]` is (see `SourceKind`). */
+    n) {
+        this.sources = e;
+        this.kinds = n;
+    }
+}
+
+/** @internal The `MergeView` behind a `merge()` proxy — its flattened
+ * `sources` with their `kinds` — or undefined. */ function mergeView(e) {
+    if (e == null || !($PROXY in e)) return undefined;
+    const n = e[$RECORD];
+    return n instanceof MergeView ? n : undefined;
+}
+
+/** @internal The record behind a merge() or omit() proxy — a `MergeView`
+ * (flattened `sources` with their `kinds`) or an `OmitView` — or undefined
+ * for anything else (a plain object, a store, a foreign proxy). Two fast
+ * traps on a proxy (`has`, then `$RECORD`), none on a plain object. */ function viewOf(e) {
+    return e != null && $PROXY in e ? e[$RECORD] : undefined;
+}
+
+/** @internal The resolved key table of a merge/omit view — every own key of
+ * the view mapped to the plain object that owns it, in merged order (a key
+ * at the position of the last source that carries it, see `tableSet`) — or
+ * undefined when it has none: a leaf is a store or a memo source, whose
+ * keys can change, or the object is not a view at all.
+ *
+ * This is the flat object the eager copy used to build, made lazily and
+ * without copying: one pass over the leaves' own keys, then every
+ * `get`/`has`/descriptor is one lookup plus one read of the owning leaf, and
+ * a consumer (`spread` rerunning its effect) walks the table instead of
+ * re-deriving shadowing from the leaves each time. Own keys only, as the
+ * copy's were: a plain source's key set is fixed once merged (keys added to
+ * it later are not seen — the copy didn't see them either).
+ *
+ * It is built by an ENUMERATION — the `ownKeys` trap, or a consumer asking
+ * for it here — or once per-key reads have paid for it (`READS_FOR_TABLE`),
+ * not on the first read. A per-key read has a direct answer (a walk of the
+ * sources, last to first, one `in` each) whose cost is the source count,
+ * while the table's is every key of every leaf, so the walk wins until a
+ * view has been read about as many times as it has keys. On the server it
+ * never is: a component reads its props a few times, the element enumerates
+ * them once through its own source walk, and the view is gone — building on
+ * first read there cost a component chain a table per layer (profiled on
+ * the Kobalte-shaped chain: a third of SSR time in the table code and its
+ * garbage). On the
+ * client a view read on every reactive rerun crosses the threshold in its
+ * first few updates and is one lookup per read from then on, as before.
+ * Once built — by a `spread`, `Object.keys`, `{...props}`, or the count —
+ * every trap uses it. */ function resolvedTable(e) {
+    if (e == null || !($PROXY in e)) return undefined;
+    const n = recordOf(e);
+    if (n instanceof OmitView) return omitTable(n);
+    return n === undefined ? undefined : mergeTable(n);
+}
+
+function mergeTable(e) {
+    let n = e.table;
+    if (typeof n !== "object") {
+        // a read count: not decided yet
+        const t = e.sources, r = e.kinds;
+        for (let n = 0; n < t.length; n++) {
+            if (!entryHasStaticKeys(t[n], r[n])) {
+                e.table = null;
+                return undefined;
+            }
+        }
+        n = new Map;
+        collectTable(n, e, undefined);
+        e.table = n;
+    }
+    return n === null ? undefined : n;
+}
+
+// One pass over a merge record's leaves into `table`, through the nested
+// omit-over-merge entries — the filters enclosing the current leaf are the
+// `filters` stack — so a component chain builds ONE table at the view that
+// asked, not one per layer. A key an entry hides that an EARLIER entry owned
+// must stay: a filter applies to its own entry's contribution, not to the
+// merge, which is exactly what the stack expresses. Every entry has static
+// keys (the caller checked), so a leaf's own keys are the truth.
+function collectTable(e, n, t) {
+    const r = n.sources, i = n.kinds;
+    for (let n = 0; n < r.length; n++) {
+        const f = r[n];
+        if (i[n] === SOURCE_OMIT) {
+            if (f.kind === SOURCE_MERGE) {
+                if (t === undefined) t = [ f ]; else t.push(f);
+                collectTable(e, f.source, t);
+                t.pop();
+                continue;
+            }
+            const n = f.source;
+            const r = Reflect.ownKeys(n);
+            for (let i = 0; i < r.length; i++) {
+                const u = r[i];
+                if (!isHidden(f, u) && !hiddenByAny(t, u)) tableSet(e, u, n);
+            }
+        } else {
+            const n = Reflect.ownKeys(f);
+            for (let r = 0; r < n.length; r++) {
+                const i = n[r];
+                if (!hiddenByAny(t, i)) tableSet(e, i, f);
+            }
+        }
+    }
+}
+
+function hiddenByAny(e, n) {
+    if (e !== undefined) for (let t = e.length - 1; t >= 0; t--) if (isHidden(e[t], n)) return true;
+    return false;
+}
+
+// Key order is the merged one — every key at the position of the LAST source
+// that carries it — the order `ssrElement`'s array form serializes in and the
+// eager copy enumerated in, so a spread through a view and a spread over the
+// sources emit the same attribute order.
+function tableSet(e, n, t) {
+    if (e.has(n)) e.delete(n);
+    e.set(n, t);
+}
+
+// An omit view's table: its source's (a merge's table, or a plain object's
+// own keys) minus the hidden keys. Cached on the record.
+function omitTable(e) {
+    let n = e.table;
+    if (typeof n !== "object") {
+        const t = e.source;
+        if (e.kind === SOURCE_MERGE) {
+            // One pass over the merge's leaves with this filter on the stack —
+            // the merge record builds no table of its own for it.
+            if (!mergeHasStaticKeys(t)) {
+                e.table = null;
+                return undefined;
+            }
+            n = new Map;
+            collectTable(n, t, [ e ]);
+        } else if (e.kind === SOURCE_PLAIN) {
+            n = new Map;
+            const r = Reflect.ownKeys(t);
+            for (let i = 0; i < r.length; i++) {
+                const f = r[i];
+                if (!isHidden(e, f)) n.set(f, t);
+            }
+        } else {
+            // a store or a memo: keys can change, no table
+            e.table = null;
+            return undefined;
+        }
+        e.table = n;
+    }
+    return n === null ? undefined : n;
+}
+
+// The user-facing key set of a resolved table: its keys that are enumerable
+// on the leaf that owns them (`Object.keys(merged)`, #2769). Fixed, like the
+// table, so it is built once per view: an `ownKeys` trap may hand back the
+// same array every time (the engine copies it).
+const propertyIsEnumerable = Object.prototype.propertyIsEnumerable;
+
+function tableOwnKeys(e, n) {
+    let t = e.keys;
+    if (t === undefined) {
+        t = e.keys = [];
+        for (const [e, r] of n) if (propertyIsEnumerable.call(r, e)) t.push(e);
+    }
+    return t;
+}
+
+// The descriptor for a table key — its owning leaf's, truthful (see
+// `sourceDescriptor`) — with the shape cached per key so an enumeration
+// (`for…in`, `Object.keys`, `{...props}`: a descriptor trap per key, on
+// every pass) does not re-read the leaf's descriptor and re-allocate a
+// getter each time. An accessor reads live, so its descriptor is reused as
+// is; a data descriptor is rebuilt with the current value.
+function tableDescriptor(e, n, t) {
+    const r = n.get(t);
+    if (r === undefined) return undefined;
+    let i = e.descs;
+    if (i === undefined) i = e.descs = new Map;
+    let f = i.get(t);
+    if (f === undefined) {
+        f = sourceDescriptor(r, SOURCE_PLAIN, t);
+        if (f === undefined) return undefined;
+        i.set(t, f);
+        return f;
+    }
+    if (f.get !== undefined) return f;
+    return {
+        configurable: true,
+        enumerable: f.enumerable,
+        writable: f.writable,
+        value: r[t]
+    };
+}
+
+// Per-key trap reads a view takes before building its table, from the
+// break-even: a build is ~60 ns per key (an `ownKeys` share, a `has`, a
+// `set`), a walk ~20 ns per source (an `in`, a hidden-list check), and a
+// leaf carries about five keys — so the table has paid for itself after
+// ~15 reads. Measured on the Kobalte-shaped chain: a walk is 2× a lookup at
+// depth 1 and up to 10× for a first-source key at depth 7 (a walk through
+// seven nested layers, a hidden-list check at each), so a view read on
+// every update wants the table; a view read a handful of times (every
+// server-side view) never wants it.
+const READS_FOR_TABLE = 16;
+
+// The table for a per-key trap: the one a view HAS (built by an enumeration
+// or an earlier read, see `resolvedTable`), or the one this read pays for,
+// or none — a walk answers. One per view type, so a trap pays no type check.
+// A settled slot is an object (the Map, or `null`); a number is the count.
+function mergeReadTable(e) {
+    const n = e.table;
+    if (typeof n === "object") return n === null ? undefined : n;
+    if (n + 1 < READS_FOR_TABLE) {
+        e.table = n + 1;
+        return undefined;
+    }
+    return mergeTable(e);
+}
+
+// Only for an omit over a merge (`kind === SOURCE_MERGE`; the caller checks,
+// inline — a call is not free in every tier): an omit over one object reads
+// it directly — a list check and a property read, nothing a table would
+// shorten.
+function omitReadTable(e) {
+    const n = e.table;
+    if (typeof n === "object") return n === null ? undefined : n;
+    if (n + 1 < READS_FOR_TABLE) {
+        e.table = n + 1;
+        return undefined;
+    }
+    return omitTable(e);
+}
+
+// The table a record HAS — built already by an enumeration or a trap's read
+// count — or undefined. What a nested walk asks: a record reached through an
+// outer view's entry counts no reads of its own (the outer view decides for
+// the whole tree, and its table build then builds the inner ones), so the
+// inner merges of a component chain build nothing on the server where the
+// leaves are read a few times each.
+function tableOf(e) {
+    const n = e.table;
+    return typeof n === "object" && n !== null ? n : undefined;
+}
+
+// "no entry has the key" — distinct from an entry that holds `undefined`.
+const MISSING = Symbol();
+
+// The read: the value of the last entry that has the key, or MISSING. ONE
+// walk — a nested omit-over-merge entry answers presence and value together,
+// so a chain of layers is walked once per read, not once per layer per
+// level.
+function mergeLookup(e, n) {
+    const t = tableOf(e);
+    if (t !== undefined) {
+        const e = t.get(n);
+        return e === undefined ? MISSING : e[n];
+    }
+    const r = e.sources, i = e.kinds;
+    for (let e = r.length - 1; e >= 0; e--) {
+        const t = i[e];
+        // The common leaf first, read in place: a component's props view is a
+        // few plain objects, read once per key on the server.
+                if (t === SOURCE_PLAIN) {
+            const t = r[e];
+            if (n in t) return t[n];
+            continue;
+        }
+        if (t === SOURCE_OMIT) {
+            const t = r[e];
+            if (isHidden(t, n)) continue;
+            if (t.kind === SOURCE_MERGE) {
+                const e = mergeLookup(t.source, n);
+                if (e !== MISSING) return e;
+                continue;
+            }
+            const i = viewSource(t);
+            if (n in i) return i[n];
+        } else {
+            const i = leafOf(r[e], t);
+            if (n in i) return i[n];
+        }
+    }
+    return MISSING;
+}
+
+function mergeGet(e, n) {
+    const t = mergeLookup(e, n);
+    return t === MISSING ? undefined : t;
+}
+
+// `key in merge`, on the record.
+function mergeHas(e, n) {
+    const t = tableOf(e);
+    if (t !== undefined) return t.has(n);
+    const r = e.sources, i = e.kinds;
+    for (let e = r.length - 1; e >= 0; e--) if (sourceHas(r[e], i[e], n)) return true;
+    return false;
+}
+
+// The proxy's `getOwnPropertyDescriptor`, on the record.
+function mergeDescriptor(e, n) {
+    const t = tableOf(e);
+    if (t !== undefined) return tableDescriptor(e, t, n);
+    const r = e.sources, i = e.kinds;
+    for (let t = r.length - 1; t >= 0; t--) {
+        if (!sourceHas(r[t], i[t], n)) continue;
+        // `in` also answers for inherited keys, which have no own descriptor.
+                return sourceDescriptor(r[t], i[t], n, true) ?? accessorDescriptor(() => mergeGet(e, n));
+    }
+    return undefined;
+}
+
+// Own keys of a merge record in merged order — every key at the position
+// of the LAST entry that carries it, the order the table keeps and
+// `ssrElement` serializes in. `enumerable` selects the user-facing set
+// (`Object.keys`, #2769) over every own string key (a consumer's walk);
+// `filter` is the omit this record is read through, applied as the keys
+// are gathered so an omit over a merge builds ONE list per layer. A list
+// with `indexOf` rather than a Set: a props object has a dozen keys, and a
+// Set's hash store was 2 KB per row on the Kobalte-shaped chain.
+function mergeKeysOf(e, n, t) {
+    const r = [];
+    collectKeys(e, t === undefined ? undefined : [ t ], n, r, null);
+    return r;
+}
+
+// One pass over a merge record's leaves — through its nested omit-over-merge
+// entries, `filters` the omits enclosing the current leaf (see
+// `collectTable`) — appending each leaf's keys to `keys` in merged order
+// (a key already listed moves to the end: later wins) and, when `owners` is
+// given, the object that owns the key at the same index. A consumer that
+// reads every key once (`ssrElement`) then reads `owners[i][keys[i]]`: no
+// `in` walk per key, no table. A memo leaf is resolved once here.
+function collectKeys(e, n, t, r, i) {
+    const f = e.sources, u = e.kinds;
+    for (let e = 0; e < f.length; e++) {
+        let o = f[e], c = u[e];
+        let s;
+        if (c === SOURCE_OMIT) {
+            if (o.kind === SOURCE_MERGE) {
+                if (n === undefined) n = [ o ]; else n.push(o);
+                collectKeys(o.source, n, t, r, i);
+                n.pop();
+                continue;
+            }
+            s = o;
+            c = o.kind;
+            o = o.source;
+        }
+        o = leafOf(o, c);
+        const d = t ? ownEnumerableKeys(o) : leafKeys(o, c);
+        for (let e = 0; e < d.length; e++) {
+            const t = d[e];
+            if (s !== undefined && isHidden(s, t)) continue;
+            if (hiddenByAny(n, t)) continue;
+            addKey(r, i, t, o);
+        }
+    }
+}
+
+// Append `key` owned by `owner`, moving an earlier listing to the end: later
+// wins, and the position is the last owner's (the merged order).
+function addKey(e, n, t, r) {
+    const i = e.indexOf(t);
+    if (i !== -1) {
+        e.splice(i, 1);
+        if (n !== null) n.splice(i, 1);
+    }
+    e.push(t);
+    if (n !== null) n.push(r);
+}
+
+/** @internal Every own string key of a props SOURCE — a plain object, a
+ * store or foreign proxy, or a merge/omit view — appended to `keys` in
+ * merged order with the object that owns each at the same index of
+ * `owners`: a key already listed (by this source or an earlier one) moves
+ * to the end, so several sources collected in turn give the order and the
+ * winners a merge of them would. A consumer that reads each key once
+ * (`ssrElement`) then reads `owners[i][keys[i]]` — the owner's getter runs
+ * there, once — and asks nothing else of a view: no table, no `in` walk per
+ * key through the merge/omit layers, no key list per leaf. One pass,
+ * however deep the layers nest. Symbols are listed; the consumer skips
+ * them. */ function sourceOwners(e, n, t) {
+    if ($PROXY in e) {
+        const r = viewOf(e);
+        if (r === undefined) {
+            // a store: one `ownKeys` trap, reads through `[]`
+            const r = Reflect.ownKeys(e);
+            for (let i = 0; i < r.length; i++) addKey(n, t, r[i], e);
+            return;
+        }
+        if (r instanceof OmitView) {
+            if (r.kind === SOURCE_MERGE) return collectKeys(r.source, [ r ], false, n, t);
+            const e = viewSource(r);
+            const i = leafKeys(e, r.kind);
+            for (let f = 0; f < i.length; f++) if (!isHidden(r, i[f])) addKey(n, t, i[f], e);
+            return;
+        }
+        return collectKeys(r, undefined, false, n, t);
+    }
+    const r = Object.keys(e);
+    for (let i = 0; i < r.length; i++) addKey(n, t, r[i], e);
+}
+
+// The user-facing key set of a merge record, read through `filter` if given.
+function mergeEnumerableKeys(e, n) {
+    const t = mergeTable(e);
+    if (t === undefined) return mergeKeysOf(e, true, n);
+    const r = tableOwnKeys(e, t);
+    if (n === undefined) return r;
+    const i = [];
+    for (let e = 0; e < r.length; e++) if (!isHidden(n, r[e])) i.push(r[e]);
+    return i;
+}
+
+const mergeTraps = {
+    get(e, n, t) {
+        // The private keys are symbols; a string read (every prop) skips the
+        // compares. `$TARGET` is answered so a store check never walks the
+        // sources (a leaf that is a store would answer it).
+        if (typeof n === "symbol") {
+            if (n === $PROXY) return t;
+            if (n === $RECORD) return e;
+            if (n === $TARGET) return undefined;
+        }
+        // A trap read counts toward the table (see `mergeReadTable`); the walk
+        // itself is the record's. `mergeReadTable` and the plain walk of
+        // `mergeLookup` are inlined here: a trap is entered from the runtime, so
+        // nothing below it is inlined for it, and a component's props view is a
+        // few plain objects read once per key on the server — the walk is the
+        // whole read. A leaf that is not plain hands the walk to `mergeGet`,
+        // which starts over (a plain leaf walked twice is two `in` checks).
+                const r = e.table;
+        let i;
+        if (typeof r !== "object") {
+            if (r + 1 < READS_FOR_TABLE) {
+                e.table = r + 1;
+                const t = e.sources, i = e.kinds;
+                for (let r = t.length - 1; r >= 0; r--) {
+                    if (i[r] !== SOURCE_PLAIN) return mergeGet(e, n);
+                    // Read first, `in` only to tell a missing key from one holding
+                    // undefined: the last source is the one that usually has the key.
+                                        const f = t[r][n];
+                    if (f !== undefined || n in t[r]) return f;
+                }
+                return undefined;
+            }
+            i = mergeTable(e);
+            if (i === undefined) return mergeGet(e, n);
+        } else if (r === null) return mergeGet(e, n); else i = r;
+        const f = i.get(n);
+        return f === undefined ? undefined : f[n];
     },
-    has(e, r) {
-        if (r === $PROXY) return true;
-        return e.has(r);
+    has(e, n) {
+        if (n === $PROXY) return true;
+        if (n === $TARGET || n === $RECORD) return false;
+        const t = mergeReadTable(e);
+        if (t !== undefined) return t.has(n);
+        return mergeHas(e, n);
     },
     set: trueFn,
     deleteProperty: trueFn,
-    getOwnPropertyDescriptor(e, r) {
-        return {
-            configurable: true,
-            enumerable: true,
-            get() {
-                return e.get(r);
-            },
-            set: trueFn,
-            deleteProperty: trueFn
-        };
+    getOwnPropertyDescriptor(e, n) {
+        if (n === $PROXY || n === $TARGET || n === $RECORD) return undefined;
+        const t = mergeReadTable(e);
+        if (t !== undefined) return tableDescriptor(e, t, n);
+        return mergeDescriptor(e, n);
     },
     ownKeys(e) {
-        return e.keys();
+        return mergeEnumerableKeys(e);
     }
 };
 
-function resolveSource(e) {
-    return !(e = typeof e === "function" ? e() : e) ? {} : e;
-}
+// An omit view reads its source directly — a hidden-key check and one
+// property read; over a merge record, the merge's own walk by function call,
+// never a trap. Once an enumeration or the read count has built its table
+// (over a merge with plain leaves only: the merge's, filtered) every trap
+// answers from that instead.
+const omitTraps = {
+    get(e, n, t) {
+        if (n === $PROXY) return t;
+        // `$RECORD` is THIS view — never the underlying merge's record, which is
+        // unfiltered: handing a re-merge the merge's own sources would leak the
+        // omitted keys (#3014). A consumer walks the omit record as ONE filtered
+        // entry (`recordOf` tells the two records apart by class).
+                if (n === $RECORD) return e;
+        if (n === $TARGET) return undefined;
+        if (e.kind === SOURCE_MERGE) {
+            const t = omitReadTable(e);
+            if (t !== undefined) {
+                const e = t.get(n);
+                return e === undefined ? undefined : e[n];
+            }
+            if (isHidden(e, n)) return undefined;
+            return mergeGet(e.source, n);
+        }
+        if (isHidden(e, n)) return undefined;
+        return viewSource(e)[n];
+    },
+    has(e, n) {
+        if (n === $PROXY) return true;
+        if (n === $TARGET || n === $RECORD) return false;
+        if (e.kind === SOURCE_MERGE) {
+            const t = omitReadTable(e);
+            if (t !== undefined) return t.has(n);
+            if (isHidden(e, n)) return false;
+            return mergeHas(e.source, n);
+        }
+        if (isHidden(e, n)) return false;
+        return n in viewSource(e);
+    },
+    set: trueFn,
+    deleteProperty: trueFn,
+    getOwnPropertyDescriptor(e, n) {
+        if (n === $PROXY || n === $TARGET || n === $RECORD) return undefined;
+        if (e.kind === SOURCE_MERGE) {
+            const t = omitReadTable(e);
+            if (t !== undefined) return tableDescriptor(e, t, n);
+        }
+        return sourceDescriptor(e, SOURCE_OMIT, n);
+    },
+    ownKeys(e) {
+        if (e.kind === SOURCE_MERGE) {
+            const n = omitTable(e);
+            if (n !== undefined) return tableOwnKeys(e, n);
+            // No table (a store or memo leaf): the merge's own key set, filtered.
+                        return sourceEnumerableKeys(e);
+        }
+        const n = Reflect.ownKeys(viewSource(e));
+        const t = [];
+        for (let r = 0; r < n.length; r++) if (!isHidden(e, n[r])) t.push(n[r]);
+        return t;
+    }
+};
 
-const $SOURCES = Symbol(0);
+/** @internal The flattened sources behind a `merge()` proxy, or undefined.
+ * A merge's writes are no-ops, so its sources are the whole truth. A COPY of
+ * a merge (`{...merged}`, a descriptor copy) is a plain object that carries
+ * no sources — `ownKeys` never lists $RECORD — so what is on the copy is
+ * the truth there and every consumer reads it directly (#3384). */ function mergeSources(e) {
+    const n = mergeView(e);
+    return n === undefined ? undefined : n.sources;
+}
 
 /**
  * Merges multiple props-like objects into a single proxy that *preserves
@@ -59,9 +888,23 @@ const $SOURCES = Symbol(0);
  * Function arguments are treated as memo-backed sources — useful for passing
  * derived defaults whose computation should track reactively.
  *
+ * The result is a live VIEW of its sources, never a copy: creating it costs
+ * nothing per key, every read goes to the source that owns the key (a getter
+ * runs there, a data property is read live), and writing to it is a no-op.
+ * A single non-function source is returned as is. To own a mutable object,
+ * copy it: `{ ...merged }` snapshots the current values.
+ *
  * Use this in component bodies to merge defaults / overrides without losing
  * Solid's per-property tracking.
  *
+ * Reading props, here and anywhere: a prop's getter is defined only for a
+ * read through its own object (`props.x`, `{ ...props }`, `Reflect.get`,
+ * these views). Forwarding its descriptor onto another object and reading it
+ * there is not supported — the compiler's server-side props keep their state
+ * on the instance, so the getter needs its object as receiver. A copy that
+ * must stay live defines its own getter that reads through the source, as
+ * the no-Proxy paths of merge() and omit() do.
+ *
  * @example
  * ```tsx
  * function Button(_props: { label: string; type?: string; disabled?: boolean }) {
@@ -72,130 +915,154 @@ const $SOURCES = Symbol(0);
  * ```
  */ function merge(...e) {
     if (e.length === 1 && typeof e[0] !== "function") return e[0];
-    let r = false;
-    const t = [];
-    for (let n = 0; n < e.length; n++) {
-        const o = e[n];
-        r = r || !!o && $PROXY in o;
-        const s = !!o && o[$SOURCES];
-        if (s) {
-            for (let e = 0; e < s.length; e++) t.push(s[e]);
-        } else t.push(typeof o === "function" ? (r = true, createMemo(o)) : o);
-    }
-    if (SUPPORTS_PROXY && r) {
-        return new Proxy({
-            get(e) {
-                if (e === $SOURCES) return t;
-                for (let r = t.length - 1; r >= 0; r--) {
-                    const n = resolveSource(t[r]);
-                    if (e in n) return n[e];
-                }
-            },
-            has(e) {
-                for (let r = t.length - 1; r >= 0; r--) {
-                    if (e in resolveSource(t[r])) return true;
-                }
-                return false;
-            },
-            keys() {
-                const e = new Set;
-                for (let r = 0; r < t.length; r++) {
-                    const n = ownEnumerableKeys(resolveSource(t[r]));
-                    for (let r = 0; r < n.length; r++) e.add(n[r]);
+    // Sized to the argument count up front: a view is built per source per
+    // component layer, and growing two empty arrays by `push` was a third of
+    // its construction. A nested view's sources may run past the count (the
+    // array grows), a falsy source leaves it short (trimmed below).
+        const n = new Array(e.length);
+    const t = new Array(e.length);
+    let r = 0;
+    // The one non-falsy source, if there is exactly one: it IS the merge.
+        let i = undefined;
+    let f = 0;
+    for (let u = 0; u < e.length; u++) {
+        const o = e[u];
+        if (!o) continue;
+        f++;
+        i = o;
+        if (typeof o === "function") {
+            n[r] = createMemo(o);
+            t[r++] = SOURCE_MEMO;
+            continue;
+        }
+        if ($PROXY in o) {
+            // A store (or a foreign proxy) is a leaf as it is. A merge() proxy is
+            // flattened through: its writes are no-ops, so its sources are exactly
+            // what it reads. An omit() proxy joins as its view record — ONE entry,
+            // its filter travelling with it, whether it is over a plain object or
+            // a whole merge (never the merge's own sources, which would leak the
+            // omitted keys, #3014). A consumer's walk recurses into the record.
+            const e = recordOf(o);
+            if (e instanceof MergeView) {
+                for (let i = 0; i < e.sources.length; i++) {
+                    n[r] = e.sources[i];
+                    t[r++] = e.kinds[i];
                 }
-                return [ ...e ];
+            } else if (e !== undefined) {
+                n[r] = e;
+                t[r++] = SOURCE_OMIT;
+            } else {
+                n[r] = o;
+                t[r++] = SOURCE_PROXY;
             }
-        }, propTraps);
+            continue;
+        }
+        n[r] = o;
+        t[r++] = SOURCE_PLAIN;
+    }
+    if (r !== n.length) {
+        n.length = r;
+        t.length = r;
+    }
+    if (SUPPORTS_PROXY) {
+        if (f === 1 && typeof i !== "function") return i;
+        // Always a view, never a copy. Building a plain object here costs a
+        // descriptor read, a bound getter and a defineProperty per key per
+        // layer, and component libraries stack several layers per element
+        // (defaults → omit → call-site statics → …), so the copies dominated
+        // their render cost while every consumer that matters — `spread`,
+        // `ssrElement`, a nested merge — reads the flattened sources directly
+        // anyway (#3448). The view is O(1) to create and reads through to the
+        // sources, so a data property on a source is read live, like a getter.
+        // Writes to the result are no-ops (a consumer that needs its own object
+        // copies: `{...merged}`, which the traps answer truthfully). Copies of
+        // the result never carry $RECORD (#3384): `ownKeys` answers only the
+        // sources' keys.
+                return new Proxy(new MergeView(n, t), mergeTraps);
     }
-    const n = Object.create(null);
+    // No Proxy: an eager descriptor copy, semantics as close to the view as a
+    // plain object allows (getters stay live; data properties are snapshots).
+        const u = Object.create(null);
     let o = false;
-    let s = t.length - 1;
-    for (let e = s; e >= 0; e--) {
-        const r = t[e];
-        if (!r) {
-            e === s && s--;
+    let c = n.length - 1;
+    for (let e = c; e >= 0; e--) {
+        const t = n[e];
+        if (!t) {
+            e === c && c--;
             continue;
         }
-        const u = Object.getOwnPropertyNames(r);
-        for (let t = u.length - 1; t >= 0; t--) {
-            const c = u[t];
-            if (c === "__proto__" || c === "constructor") continue;
-            if (!n[c]) {
-                o = o || e !== s;
-                const t = Object.getOwnPropertyDescriptor(r, c);
-                n[c] = t.get ? {
+        const r = Object.getOwnPropertyNames(t);
+        for (let n = r.length - 1; n >= 0; n--) {
+            const i = r[n];
+            if (i === "__proto__" || i === "constructor") continue;
+            if (!u[i]) {
+                o = o || e !== c;
+                const n = Object.getOwnPropertyDescriptor(t, i);
+                u[i] = n.get ? {
                     enumerable: true,
                     configurable: true,
-                    get: t.get.bind(r)
-                } : t;
+                    get: n.get.bind(t)
+                } : n;
             }
         }
     }
-    if (!o) return t[s];
-    const u = {};
-    const c = Object.keys(n);
-    for (let e = c.length - 1; e >= 0; e--) {
-        const r = c[e], t = n[r];
-        if (t.get) Object.defineProperty(u, r, t); else u[r] = t.value;
+    if (!o) return n[c];
+    const s = {};
+    const d = Object.keys(u);
+    for (let e = d.length - 1; e >= 0; e--) {
+        const n = d[e], t = u[n];
+        if (t.get) Object.defineProperty(s, n, t); else s[n] = t.value;
     }
-    u[$SOURCES] = t;
-    return u;
+    return s;
 }
 
-/**
- * Returns a reactive proxy of `props` with the listed keys hidden. Tracking
- * on the remaining keys is preserved.
- *
- * Use it to forward "rest" props to a child element while pulling out the
- * keys your component handles itself — the equivalent of `splitProps(p, ["a","b"])[1]`.
- *
- * @example
- * ```tsx
- * function Input(props: { label: string; value: string; onInput: (v: string) => void } & JSX.HTMLAttributes<HTMLInputElement>) {
- *   const rest = omit(props, "label", "value", "onInput");
- *
- *   return (
- *     <label>
- *       {props.label}
- *       <input
- *         {...rest}
- *         value={props.value}
- *         onInput={e => props.onInput(e.currentTarget.value)}
- *       />
- *     </label>
- *   );
- * }
- * ```
- */ function omit(e, ...r) {
-    if (SUPPORTS_PROXY && $PROXY in e) {
-        return new Proxy({
-            get(t) {
-                // $SOURCES must not tunnel through the filter: merge() flattens
-                // whatever answers it, so forwarding would hand a re-merge the
-                // UNFILTERED sources of an underlying merge proxy and the omitted
-                // keys leak back in (#3014 — the SSR element-spread path re-merges
-                // static attributes with the rest object). Opaque here: merge
-                // composes omit proxies through their traps instead.
-                return t === $SOURCES || r.includes(t) ? undefined : e[t];
-            },
-            has(t) {
-                return t !== $SOURCES && !r.includes(t) && t in e;
-            },
-            keys() {
-                return ownEnumerableKeys(e).filter(e => !r.includes(e));
+function omit(e, ...n) {
+    let t = n.length === 1 && typeof n[0] === "function" ? n[0] : n;
+    if (SUPPORTS_PROXY) {
+        // A view over a view folds: one record, both filters, the original
+        // source — so a consumer walks the real object however deep the omits go.
+        // Over a merge() proxy the source is the merge's RECORD (see OmitView):
+        // one record whatever the leaf count, read by function call.
+        let n = e;
+        let r = SOURCE_PLAIN;
+        if (typeof e === "function") r = SOURCE_MEMO; else if ($PROXY in e) {
+            r = SOURCE_PROXY;
+            const i = recordOf(e);
+            if (i instanceof OmitView) {
+                n = i.source;
+                r = i.kind;
+                t = combineHidden(i.hidden, t);
+            } else if (i !== undefined) {
+                n = i;
+                r = SOURCE_MERGE;
             }
-        }, propTraps);
+        }
+        return new Proxy(new OmitView(n, r, t), omitTraps);
     }
-    const t = {};
-    const n = Object.getOwnPropertyNames(e);
-    const o = r.length > 4 && n.length > r.length ? new Set(r) : undefined;
-    for (const s of n) {
-        if (o ? !o.has(s) : !r.includes(s)) {
-            const r = Object.getOwnPropertyDescriptor(e, s);
-            !r.get && !r.set && r.enumerable && r.writable && r.configurable ? t[s] = r.value : Object.defineProperty(t, s, r);
+    const r = {};
+    const i = Object.getOwnPropertyNames(e);
+    const f = typeof t === "function" ? t : t.length > 4 && i.length > t.length ? (e => n => e.has(n))(new Set(t)) : e => t.includes(e);
+    for (const n of i) {
+        if (!f(n)) {
+            const t = Object.getOwnPropertyDescriptor(e, n);
+            if (!t.get && !t.set && t.enumerable && t.writable && t.configurable) {
+                r[n] = t.value;
+            } else if (t.get || t.set) {
+                // An accessor is re-homed with its source as receiver, never copied
+                // as-is: a props getter is only defined for a read THROUGH its own
+                // object (the compiler's server props keep their state on the
+                // instance, so a forwarded descriptor read on the copy throws). Same
+                // rule as merge()'s copy path above.
+                Object.defineProperty(r, n, {
+                    enumerable: t.enumerable,
+                    configurable: true,
+                    get: t.get && t.get.bind(e),
+                    set: t.set && t.set.bind(e)
+                });
+            } else Object.defineProperty(r, n, t);
         }
     }
-    return t;
+    return r;
 }
 
-export { merge, omit };
+export { MergeView, OmitView, SOURCE_MEMO, SOURCE_MERGE, SOURCE_OMIT, SOURCE_PLAIN, SOURCE_PROXY, hasStaticKeys, isStatic, merge, mergeSources, mergeView, omit, omitView, resolvedTable, sourceGet, sourceHas, sourceKeys, sourceOwners, viewOf };