@@ -10362,3 +10362,79 @@ a scheduler design question, not a tweak. A Vapor-style text cache would
1036210362change behaviour. Local only: the branches `perf/write-path-constant`,
1036310363`perf/insert-text-in-place` and `perf/normalize-split` with its stash, and
1036410364the fork builds `dist-pw-*`, `dist-ip-*`, `dist-ns-*` and `dist-scan`.
10365+
10366+ ### 18 (Vapor lane, follow-up): one template clone path — `importNode` or adopt-once, priced on the board (2026-10-25)
10367+
10368+ `template(html, flag)` clones through `cloneNode(true)` from the node parsed
10369+ into a `<template>`'s inert document, and switches to
10370+ `document.importNode` when the compiler sets flag 1: custom elements, `is`
10371+ attributes, or a lazy-loading `<img>`/`<iframe>`
10372+ (`template_subtree_is_import_node`, `packages/compiler/src/dom/element.rs`).
10373+ `importNode` creates the clone in the document, so a custom element upgrades
10374+ at clone time, before Solid's walk sets its properties. A `cloneNode` copy
10375+ from the inert document upgrades only when it is inserted. The question was
10376+ whether one path could replace the special case, and what that costs on the
10377+ Octane board.
10378+
10379+ **No board fixture emits flag 1.** A scan of every built Solid fixture bundle
10380+ found no flagged template, no `loading=` attribute and no custom-element
10381+ tag. Dropping the flag in favour of plain `cloneNode` changes nothing on the
10382+ board, but it reopens the upgrade-timing problem the flag exists for. The
10383+ measurable choice is a uniform path that is also correct for custom
10384+ elements:
10385+
10386+ - **import** (`perf/template-import`, 3d4f9ffa9): every template clones
10387+ through `document.importNode`.
10388+ - **adopt** (`perf/template-adopt`, d0c7633bb): `create` adopts the parsed
10389+ node into the document once (`document.adoptNode`), then every clone uses
10390+ `cloneNode(true)`. The clone is born in the document, so nothing is adopted
10391+ on insertion.
10392+
10393+ Both pass every web suite that passes on `next` in that worktree. Four
10394+ `UNCOMMITTED_INPUT_OVERWRITTEN` dev tests and one #4240 server test fail
10395+ there on `next` too.
10396+
10397+ **Setup.** `next` @ 4ea2f4986 and both probes, two rounds interleaved per
10398+ round (next, import, adopt), 11 client suites chosen for template mounts.
10399+ Results are in `results-tpl-{next,import,adopt}-r{1,2}-oct25`, the log in
10400+ `.logs/template-clone-2026-10-25.log`. Each figure is Solid's own per-op
10401+ score against `next` from the same round, as a geomean over ops and then
10402+ over rounds. list-clear has no Solid target, which leaves 10 suites.
10403+
10404+ | suite | import | adopt | `next` r2 / r1 |
10405+ | --- | ---: | ---: | ---: |
10406+ | js-framework | +4.7% | +2.6% | −2.8% |
10407+ | todomvc | +7.6% | +6.8% | +0.3% |
10408+ | chat-stream | +11.5% | +9.8% | +10.2% (noise) |
10409+ | svg-dashboard | +0.8% | +0.3% | 0.0% |
10410+ | uibench | +0.5% | +1.4% | −0.2% |
10411+ | spa-navigation | +9.6% | +6.4% | +2.0% |
10412+ | application-composition | −0.8% | −2.5% | +0.8% |
10413+ | scaling-curves | −5.4% | −5.4% | −0.5% |
10414+ | effectful-list | +5.4% | +6.0% | +0.4% |
10415+ | memo-wall | −0.9% | +1.3% | −0.1% |
10416+ | **geomean, 10 suites** | **+3.2%** | **+2.6%** | \|noise\| 1.7% |
10417+
10418+ Ops that moved more than 5% the same way in both rounds, for both probes:
10419+ the creation ops. js-framework `run` +13–14%, `runlots` +16–18%, `add`
10420+ +11%, `replace` +8% (import). spa-navigation `nav_nested`, `nav_deep` and
10421+ `nav_mount` +9–16%. todomvc toggles +24–26%. uibench `render` cases +5–8%.
10422+ uibench's deep-tree `removeAll` is +33–34% on both, which suggests the cost
10423+ does not stop at the clone: nodes born in the document are slower to tear
10424+ down as well. js-framework `select` (−13–14%) and todomvc `clearCompleted`
10425+ (−8%) improved on both probes. Neither clones a template, so they probably
10426+ reflect a different heap shape, not the path itself.
10427+
10428+ **Board impact.** Holding the 15 untouched suites at zero, the change to the
10429+ 25-suite board geomean is +1.3% for import and +1.0% for adopt. Without
10430+ chat-stream, whose own round-to-round spread is 10%, it is +0.8% and +0.7%.
10431+ The cost lands on creation ops, 10–17% on the most clone-heavy.
10432+
10433+ **Verdict.** Keep the special case. `cloneNode` from the template's inert
10434+ document, with adoption on insertion, is the fastest path in Chromium. It is
10435+ faster than creating each clone in the document, by `importNode` or from a
10436+ pre-adopted node: adopt-once beats `importNode` by about half a percent and
10437+ still loses to `next`. The flag costs only templates that need it, and none
10438+ of the board's do. Removing it is worth doing only if custom-element upgrade
10439+ timing is dropped as a guarantee, which is a semantics call, not a speed
10440+ call. The probe branches stay local.
0 commit comments