You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit c98b83b
Browse filesBrowse the repository at this point in the historyBrowse files
Copy file name to clipboardExpand all lines: scripts/size/scenarios.js
+1-1Lines changed: 1 addition & 1 deletion
Original file line number
Diff line number
Diff line change
@@ -2653,7 +2653,7 @@ module.exports = [
2653
2653
// Plain store arrays clone by slice() on the first draft write (2026-10-28, over next @ 17d4d1b5): 31.08 KB -> 31.10 KB, recorded minified 95,378 -> 95,467 B (+89 B), measured locally (linux) at 31,088 B brotli. `cloneRaw` took the descriptor path for every array (one descriptor per element, then defineProperties), so a push into a 10k-row store cost ~10.7 ms; a plain accessor-free array with no named keys now slices (~0.15 ms). Perf bug fix. Cap at measured + 10 B rounded up to 0.01 KB, never lowered; recorded minified = measured; CI re-bases.
2654
2654
// A list over a shallow store diffs its backing (#4345, 2026-10-28, over next @ 56a982cd): 31.10 KB -> 31.19 KB, recorded minified 95,467 -> 95,595 B (+128 B), measured locally (linux) at 31,173 B brotli. `mapArray` asks the store engine (types.ts `shallowView`, installed by store.ts's top level) for a shallow store's backing once per pass instead of reading every slot through the proxy: a one-row write on a 1k-row shallow store spent ~1.2 ms in per-slot traps. Apps with `<For>` and no stores pay the optional call only (+4 to +6 B minified, within every page's allowance). Perf bug fix; raise accepted by the maintainer (2026-10-28). Cap at measured + 10 B rounded up to 0.01 KB, never lowered; recorded minified = measured; CI re-bases.
2655
2655
// A guess's node re-asked from outside its action holds the frame with that action (D-34 / F-1, #4336, 2026-10-28, over next @ 17d4d1b5): 31.08 -> 31.36 KB, recorded minified 95,378 -> 96,028 B (+650 B over next), measured locally (linux) at 31,347 B brotli. This scenario is the only one that pays: it carries createOptimisticStore, which arms the core's guess-only paths (`guessing`, set by the first optimistic write, and lanes.ts's own copy of it, +16 B here, so its arms still fold when a page splits lanes.ts and scheduler.ts into separate chunks — the compiled live page, where the shared flag left them standing at +51 B; every other build folds them and ends 32-101 B smaller than next) — the outside-change join in `notifyStatus`, the held refresh in `recompute`, a lane pass taking another action's guess (`laneStage`), the `enterLane` repair of a held write read early — and the store's own: the derive's new question over a family's guesses stamped with the flush (`ask`), a frame still reading a guess at the seam joining the guess's action (`park`, so a keyed store's other keys never wait for it), reads of a target with no guess over that question suspending (`stands`), and a derive re-asking in another action's lane taking the guessed slots it writes (`_slotTake`). Measured alternatives: ask-time join (95,481 B) regressed the keyed store; a read-time join in `pullFamily` alone missed a render effect that does not re-run; deciding `stands` from the family's live guesses instead of a per-flush stamp (+24 B) changed reads in later flushes (#2951). Correctness; Size-Exception accepted by the maintainer (2026-10-28). Cap at measured + 10 B rounded up to 0.01 KB; recorded minified = measured; CI re-bases.
2656
-
// A displaced guess comes back with its taker's revert; a fallback's `latest()` does not pull a held swap ahead (#4336 / #4337, 2026-10-28, re-measured over next @ 1f558e50 after merging it): 31.19 -> 31.80 KB, recorded minified 95,595 -> 96,944 B (+1,361 B over next measured locally: +650 B is #4336's accepted exception above, +711 B this change), measured locally (linux) at 31,790 B brotli. Lanes (behind `guessed`, folded in every build without an optimistic write): a lane taking another action's guess keeps it aside (`take`; a store target's other guessed slots taken with it but not written, so the revert's re-ask is a question nothing stands in for), its readers re-derive with the lane (the split-effect form waits instead of showing `#1:` beside the guess), a correction that reverts marks its lane (`reverted`, a WeakSet: no new mangled field), and the landing of that lane's parent re-seats each guess still alive over the truth it lands (`endLanes`; a store slot on the node the slot has now, `_slotTake(n, true)`) in the same frame. Boundaries (+41 B in every build with a boundary, ≤ next except the lane pages: live +2 B, compiled live +9 B, live + router +2 B over next): the fallback read keeps the output's lane seat and verdict bits. Measured alternatives: restoring at the dissolve instead of the landing tore (`#1: dana | mine (sending)`); restoring without re-seating released store slots lost row 0 and the untracked read; reading the fallback with no observer broke #2701/#2949/#3540 pins; a `_reverted` field on the lane (-41 B here) renamed the mangled props and moved every other scenario's brotli by +1..+16 B; passing the revert to `supersede` as an argument (-26 B here) cost `+ isPending/latest` 5 B; an AND-mask for the fallback's flags (-5 B minified) cost the boundary apps up to +63 B brotli. Correctness; awaiting the maintainer's Size-Exception. Cap at measured + 10 B rounded up to 0.01 KB; recorded minified = measured; CI re-bases.
2656
+
// A displaced guess comes back with its taker's revert; a fallback's `latest()` does not pull a held swap ahead (#4336 / #4337, 2026-10-28, re-measured over next @ 1f558e50 after merging it): 31.19 -> 31.80 KB, recorded minified 95,595 -> 96,944 B (+1,361 B over next measured locally: +650 B is #4336's accepted exception above, +711 B this change), measured locally (linux) at 31,790 B brotli. Lanes (behind `guessed`, folded in every build without an optimistic write): a lane taking another action's guess keeps it aside (`take`; a store target's other guessed slots taken with it but not written, so the revert's re-ask is a question nothing stands in for), its readers re-derive with the lane (the split-effect form waits instead of showing `#1:` beside the guess), a correction that reverts marks its lane (`reverted`, a WeakSet: no new mangled field), and the landing of that lane's parent re-seats each guess still alive over the truth it lands (`endLanes`; a store slot on the node the slot has now, `_slotTake(n, true)`) in the same frame. Boundaries (+41 B in every build with a boundary, ≤ next except the lane pages: live +2 B, compiled live +9 B, live + router +2 B over next): the fallback read keeps the output's lane seat and verdict bits. Measured alternatives: restoring at the dissolve instead of the landing tore (`#1: dana | mine (sending)`); restoring without re-seating released store slots lost row 0 and the untracked read; reading the fallback with no observer broke #2701/#2949/#3540 pins; a `_reverted` field on the lane (-41 B here) renamed the mangled props and moved every other scenario's brotli by +1..+16 B; passing the revert to `supersede` as an argument (-26 B here) cost `+ isPending/latest` 5 B; an AND-mask for the fallback's flags (-5 B minified) cost the boundary apps up to +63 B brotli. Correctness; Size-Exception accepted by the maintainer 2026-10-28. Cap at measured + 10 B rounded up to 0.01 KB; recorded minified = measured; CI re-bases.
0 commit comments