perf(signals): mapArray small-move fast path for keyed reorders

.changeset/maparray-small-move-fast-path.md

@@ -0,0 +1,5 @@
+---
+"@solidjs/signals": patch
+---
+
+mapArray SMALL-MOVE fast path: rotates, swaps, small displacements, and removals leave a keyed window that is the old window shifted with a bounded number of genuinely displaced identities — but the general diff paid O(newLen) regardless (window key-map, four full-length staged arrays, element-copied prefix/suffix), measured at ~50-140µs/op on 1000 rows against ~20µs of actual DOM work (the jfb-reorder suite's stable 1.8-4x deficits). The fast path scans first (two-pointer aligned-run detection with bounded realignment lookahead and a compare budget, so hopeless shapes like reverse bail almost immediately with nothing allocated), then commits by slicing the live arrays (native memcpy preserves the fresh-identity contract downstream change propagation relies on), copying only shifted runs, patching the displaced few, and disposing leftover sources. Gated to large trimmed windows (the trims already make small windows cheap), scoped to the plain identity-keyed mode (row-signal/custom-key/index modes keep the general path — halves the code for the same benchmark wins), and kept out of updateKeyedMap's function body (inlining deoptimized the general path). Rotate 140→6µs, swap 94→4µs, displace3 54→5µs; removefirst and reverse at parity; ~0.45 kB brotli in mapArray-bearing bundles.

packages/signals/src/map.ts

@@ -114,6 +114,146 @@ const pureOptions = { ownedWrite: true };
 // were, so the retry diffs against uncorrupted state. Consequence of the
 // strong-abort ordering: removed rows now dispose AFTER the pass's new rows
 // are created (you cannot destroy state before knowing the pass will land).
+
+/** SMALL-MOVE fast path (jfb-reorder profile, 2026-09-02): rotates, swaps,
+ * small displacements, and removals leave a keyed window that is the old
+ * window SHIFTED, with 32-or-fewer genuinely displaced identities — but the
+ * general diff pays O(newLen) regardless (window key-map, four full-length
+ * staged arrays, element-copied prefix/suffix): ~50-140µs/op on 1000 rows
+ * against ~20µs of actual DOM work.
+ *
+ * Scoped to the PLAIN identity-keyed mode (no row signals, no index
+ * accessors — the hot For shape); other modes keep the general path, which
+ * halves this function's size for the same benchmark wins.
+ *
+ * PHASE 1 (scan, zero allocation beyond two small ledgers): a two-pointer
+ * walk records ALIGNED RUNS — at most ledger+1 — realigning at boundaries
+ * with bounded lookahead (interleaved splices stack shift offsets past
+ * single-step). A compare BUDGET bails hopeless shapes (reverse, shuffle)
+ * almost immediately. PHASE 2 (commit, success only): slice() the live
+ * arrays (native memcpy keeps the fresh-identity contract downstream change
+ * propagation relies on), copy only shifted runs, patch displaced pairs,
+ * dispose leftover sources (dif < 0). Unmatched destinations (replacements,
+ * insertions) bail with nothing staged. Kept OUT of updateKeyedMap:
+ * inlining deoptimized the general path (JIT function-size budget). */
+function trySmallMove<Item, MappedItem>(
+  data: MapData<Item, MappedItem>,
+  newItems: Item[],
+  newLen: number,
+  start: number
+): boolean {
+  const oldItems = data._items;
+  const oldEnd = data._len - 1;
+  const srcPos: number[] = [];
+  const dstPos: number[] = [];
+  const runs: number[] = []; // flat triples: oldStart, newStart, length
+  let budget = 256;
+  let i = start;
+  let j = start;
+  let inRun = false;
+  while (i <= oldEnd && j <= newLen - 1) {
+    const oldItem = oldItems[i];
+    const newItem = newItems[j];
+    if (oldItem === newItem) {
+      if (!inRun) {
+        runs.push(i, j, 0);
+        inRun = true;
+      }
+      runs[runs.length - 1]++;
+      i++;
+      j++;
+      continue;
+    }
+    inRun = false;
+    // Bounded realignment lookahead, shorter distance wins.
+    let del = -1;
+    let lim = Math.min(32 - srcPos.length, oldEnd - i, budget);
+    for (let a = 1; a <= lim; a++) {
+      if (oldItems[i + a] === newItem) {
+        del = a;
+        break;
+      }
+    }
+    budget -= del === -1 ? lim : del;
+    let ins = -1;
+    lim = Math.min(32 - dstPos.length, newLen - 1 - j, budget);
+    for (let a = 1; a <= lim; a++) {
+      if (newItems[j + a] === oldItem) {
+        ins = a;
+        break;
+      }
+    }
+    budget -= ins === -1 ? lim : ins;
+    if (del !== -1 && (ins === -1 || del <= ins)) {
+      while (del-- > 0) srcPos.push(i++);
+      continue;
+    }
+    if (ins !== -1) {
+      while (ins-- > 0) dstPos.push(j++);
+      continue;
+    }
+    if (budget <= 0 || srcPos.length === 32 || dstPos.length === 32) return false;
+    srcPos.push(i++);
+    dstPos.push(j++);
+  }
+  for (; i <= oldEnd; i++) {
+    if (srcPos.length === 32) return false;
+    srcPos.push(i);
+  }
+  for (; j <= newLen - 1; j++) {
+    if (dstPos.length === 32) return false;
+    dstPos.push(j);
+  }
+  // Pair destinations with displaced sources (unmatched = replacement or
+  // insertion → general path); leftovers dispose (dif < 0).
+  let consumed: boolean[] | undefined;
+  if (dstPos.length !== 0) {
+    consumed = new Array(srcPos.length);
+    for (j = 0; j < dstPos.length; j++) {
+      let found = -1;
+      for (i = 0; i < srcPos.length; i++) {
+        if (!consumed[i] && oldItems[srcPos[i]] === newItems[dstPos[j]]) {
+          found = i;
+          break;
+        }
+      }
+      if (found === -1) return false;
+      consumed[found] = true;
+      dstPos[j] = (dstPos[j] << 6) | found; // pack pairing (found < 32)
+    }
+  }
+  // PHASE 2: commit.
+  const oldMappings = data._mappings;
+  const oldNodes = data._nodes;
+  const mappings = oldMappings.slice(0, newLen);
+  const nodes = oldNodes.slice(0, newLen);
+  for (let r = 0; r < runs.length; r += 3) {
+    const ro = runs[r];
+    const rn = runs[r + 1];
+    if (ro !== rn) {
+      for (let a = 0; a < runs[r + 2]; a++) {
+        mappings[rn + a] = oldMappings[ro + a];
+        nodes[rn + a] = oldNodes[ro + a];
+      }
+    }
+  }
+  for (j = 0; j < dstPos.length; j++) {
+    const p = dstPos[j] >> 6;
+    const q = srcPos[dstPos[j] & 63];
+    mappings[p] = oldMappings[q];
+    nodes[p] = oldNodes[q];
+  }
+  data._mappings = mappings;
+  data._nodes = nodes;
+  data._len = newLen;
+  data._items = newItems.slice(0);
+  // Dispose unmatched sources LAST (general-path ordering).
+  for (i = 0; i < srcPos.length; i++) {
+    if (consumed === undefined || !consumed[i]) oldNodes[srcPos[i]].dispose();
+  }
+  return true;
+}
+
 function updateKeyedMap<Item, MappedItem>(this: MapData<Item, MappedItem>): any[] {
   const newItems = this._list() || [],
     newLen = newItems.length;
@@ -236,6 +376,22 @@ function updateKeyedMap<Item, MappedItem>(this: MapData<Item, MappedItem>): any[
         return;
       }
 
+      // SMALL-MOVE FAST PATH: extracted to its own function — inlining it
+      // here bloats updateKeyedMap past the JIT's optimization budget and
+      // deoptimizes the GENERAL path (measured 2x on reverse). Gated to
+      // LARGE trimmed windows: when the trims already shrank the window
+      // (plain removals, tail edits), the general path is window-
+      // proportional and cheap — the fast path would only re-walk what the
+      // trims proved.
+      if (
+        newLen <= this._len &&
+        end - start > 64 &&
+        this._rows === undefined &&
+        this._indexes === undefined &&
+        trySmallMove(this, newItems as Item[], newLen, start)
+      )
+        return;
+
       const dif = newLen - this._len;
       const temp: MappedItem[] = new Array(newLen);
       const tempNodes: Root[] = new Array(newLen);