Unmount performance regression since 5.36.0

[x0k]

Describe the bug

I noticed a significant increase in unmout time (during component replacement) after switching to Svelte 5.36.0 (experimental.async disabled), from 55 ms to 8.5 seconds for the dev server and up to 3 seconds for the production build.

Perhaps this is related to #16618 but i don't use stores at all.

Reproduction

1. Open SJSF playground
2. Pick any other example (I chose single in my measurements)

Performance traces for versions 5.35.7 and 5.38.3 from the dev server:

• Trace-5.35.7.json

• Trace-5.38.3.json

Logs

System Info

System:
    OS: Linux 6.12 cpe:/o:nixos:nixos:25.05 25.05 (Warbler)
    CPU: (12) x64 Intel(R) Core(TM) i5-10400F CPU @ 2.90GHz
    Memory: 6.34 GB / 15.56 GB
    Container: Yes
    Shell: 5.2.37 - /run/current-system/sw/bin/bash
  Binaries:
    Node: 24.3.0 - /nix/store/7nfklihr3idvsd18aap75mqzbwxjjg0g-nodejs-24.3.0/bin/node
    npm: 11.4.2 - /nix/store/7nfklihr3idvsd18aap75mqzbwxjjg0g-nodejs-24.3.0/bin/npm
    pnpm: 10.12.4 - /nix/store/7qxbnlds7hd3ijgq1g5gmd32gis5cabx-pnpm-10.12.4/bin/pnpm
  Browsers:
    Brave Browser: 139.1.81.131
  npmPackages:
    svelte: catalog: => 5.38.1

Severity

blocking an upgrade

[x0k]

I think this may be related to the fact that each component of the form is calculated dynamically.

<script>
  const Component = $derived(getFormComponent(...));
</script>

<Component />

---

Additional codebase info:

• 100% of $effect's without teardown logic
• 0 uses of onMount and onDestroy

[dummdidumm]

Can you pin down which version exactly introduced the perf regression?

[x0k]

5.36.0

[dummdidumm]

Could you provide a Svelte playground that reproduces this? I don't know what your site does / what the code looks like to reproduce this.

[x0k]

I have created isolated example here and found out that removing double access in this derived reduces the unmounting time, triple access does not increase time.