Pool the trackers of tracking frames

[NullVoxPopuli-ai-agent]

A tracking frame no longer makes a Tracker, a Set and an array. In a benchmark of createCache graphs, one update takes about half the time: the weighted mean is 0.6x of main.

This is the first part of #21650, and it merges alone. Only packages/@glimmer/validator changes.

case	main	this PR
propagate: 100 chains x 100 deep	724.68 µs	337.41 µs (0.5x)
kairo: diamond	428 ns	212 ns (0.5x)
rows: 1000 rows, write all	118.18 µs	67.92 µs (0.6x)
create: 1000 computeds, read each	41.11 µs	23.68 µs (0.6x)
weighted geometric mean	1.0x	0.6x

What changes

• beginTrackFrame takes the tracker for its depth from a pool. Frames are strictly nested, so one tracker for each depth is enough.
• A tracker keeps its tags in an array, not in a Set.
• A tag keeps the index at which a tracker took it last, in a new field slot. A tracker has the tag if its entry at that index is the tag. A tag that the frame consumes again costs one comparison.
• resetTracking() clears the pooled trackers.

What changes for callers

• No function changes its signature.
• Each tag has one more field, slot.

One edge case, and why the tracker uses an index

• A tag that a nested frame takes at another index, between two consumptions of the outer frame, is in the outer frame two times. The combined tag has the same revision. If the nested frame takes the tag at the same index, the outer frame has the tag one time.
• A stale index cannot hide a tag, because the tracker compares the entry at that index with the tag.
• A frame counter was the first design. It leaves the small-integer range of V8 after about one billion frames, and the benchmark was then 1.1 to 1.3 times slower.
• An untrack frame takes a depth but no tracker, so the pool can have holes. resetTracking() skips them.
• In CPU profiles of three graphs on main, the tracker, its Set and beginTrackFrame took 27% to 45% of the samples.

All 20 cases

case	main	this PR
propagate: 1 chains x 1 deep	124 ns	63 ns (0.5x)
propagate: 10 chains x 10 deep	5.51 µs	2.89 µs (0.5x)
propagate: 100 chains x 100 deep	724.68 µs	337.41 µs (0.5x)
propagate: 1 chains x 1000 deep	57.18 µs	31.34 µs (0.5x)
propagate: 1000 chains x 1 deep	103.01 µs	51.98 µs (0.5x)
kairo: avoidable propagation	382 ns	166 ns (0.4x)
kairo: broad propagation	7.35 µs	3.81 µs (0.5x)
kairo: deep propagation	3.09 µs	1.45 µs (0.5x)
kairo: diamond	428 ns	212 ns (0.5x)
kairo: mux	20.23 µs	12.90 µs (0.6x)
kairo: repeated observers	365 ns	173 ns (0.5x)
kairo: triangle	755 ns	386 ns (0.5x)
kairo: unstable	574 ns	374 ns (0.7x)
rows: 1000 rows, write 1	7.19 µs	7.03 µs (1.0x)
rows: 1000 rows, write all	118.18 µs	67.92 µs (0.6x)
batch: 10 writes, 1 output	667 ns	368 ns (0.6x)
avoidable: write the same value	6 ns	6 ns (1.0x)
create: 1000 signals	17.84 µs	18.60 µs (1.0x)
create: 1000 computeds, read each	41.11 µs	23.68 µs (0.6x)
create: 1000 outputs	50.30 µs	35.34 µs (0.7x)
weighted geometric mean	1.0x	0.6x

The benchmark is https://github.com/NullVoxPopuli-ai-agent/ember-reactivity-bench. One measurement is the writes of one frame, then one flush that brings every output up to date. Each case of each build runs in its own process, pinned to one core. The numbers are the median of 6 mirrored rounds on Node 24.20.

The run has main, each of the five branches, and #21650 as columns: table and method.

Tests

• The full suite passes locally: 9500 pass, 18 skipped, 0 failed.
• type-check:internals, ESLint and Prettier pass.
• New tests: a frame that consumes one tag many times, and a tag that a nested frame consumes between two consumptions.
• A new test for nested frames that take one tag at the same index.
• New tests for resetTracking(): with an open frame, and after a frame that began inside untrack frames.

The parts of #21650

#21650 has all five changes in one branch. It stays open as the reference for the combined numbers. The five branches merge with no conflict, and their merge is the same code as the head of #21650.

PR	change	reactivity benchmark, mean against main
#21663 (this PR)	Pool the trackers of tracking frames	0.6x
#21664	Reuse the combined tag of a tracking frame. Needs #21663	0.6x, with the pool
#21665	Index loop over the subtags of a combined tag	1.0x
#21666	Make the functions of a TrackedValue on their first use	0.9x, create is 0.5x
#21667	Store 0 first in the value field of a TrackedValue	1.0x, removes one deopt
#21650	all five	0.6x

Rendering was measured for the combined change only: pnpm bench shows 2.1% less script time, in this comment on #21650.

🤖 Generated with Claude Code

[NullVoxPopuli]

ci on a real app green here: NullVoxPopuli/limber#2288

(lots of unconventional reactivity timings, kinda)