perf(build): avoid quadratic link scan in the preload helper

[antur84]

Hey

I found this little opportunity while looking at browser performance traces originally, I've run this patch in prod with measured (inline in my patch) improvement metrics across the population. Sharing in case it makes sense for everyone.

Before

The __vitePreload runtime helper checked whether each dependency was already preloaded by scanning every <link> in the document, backwards, once per unseen dependency. For a call with L links and U unseen dependencies that is O(U * (L + U)), since the helper appends a link for every miss. A large code-split app whose dynamic imports carry long dependency lists pays it on the main thread at boot.

After

Collect the existing link hrefs into two Sets (any rel for scripts, stylesheet only for CSS) and dedupe in O(1) per dependency, so a call costs O(L + U). The sets are built on the first dependency not yet in seen, so a call whose dependencies are all seen does no link work, as before. seen already guarantees a dependency reaches the dedupe at most once, so links the helper creates are not written back into the sets.

Behaviour is unchanged: same links, same order, same CSS load promises and rejection, same vite:preloadError path, same nonce handling.

Performence measurement

A large code-split app ships this change as a patch currently. The helper times its own main-thread cost per editor boot, and on a small random sample of page loads it also replays the removed per-dependency link scan read-only, so the same boot records what the stock scan would have cost on that device. Traffic across desktop and mobile browsers; about 900 <link>s in the document after boot.

Column	Meaning
Calls	Median __vitePreload calls per boot; times are summed over them
Stock p50 / p90	Main-thread ms the removed link scan would have cost on these boots, median / 90th percentile. Scan only, without even having to do all the work, so a lower bound
PR p50 / p90	Main-thread ms this PR's helper spent on the same boots, everything included (resolution, snapshot, link insertion)
Δ	Stock minus PR, in ms saved per boot; conservative, since the PR column includes work stock also did outside the scan

Boot	Calls	Stock p50	PR p50	Δ p50	Stock p90	PR p90	Δ p90
Full page load	36	144	102	${\color{green}-42}$	580	418	${\color{green}-162}$
In-app navigation	9	90	36	${\color{green}-54}$	472	202	${\color{green}-270}$

[sapphi-red]

Thanks!