Introduce DOM post-connection steps

[domfarolino]

This PR introduces a node's set of post-connection steps. For any given #concept-node-insert operation, these steps run for each inserted node synchronously after all DOM insertions are complete. This PR is meant to supersede the similar #732.

The goal here is to separate the following:

1. Script-observable but not-script-executing insertion side effects
   • This includes synchronously applying styles to the document, etc...
2. Script-executing insertion side effects
   • This includes creating an iframe's document and synchronously firing its load event

For any given call to #concept-node-insert, the above model allows us to keep all of (1) running synchronously after each node's insertion (as part of its insertion steps), while pushing all script-executing (or DOM tree-modifying or frame tree-modifying etc.) side effects to the new set of post-connection steps, which run synchronously during insertion, but after all nodes finish their insertion.

This essentially makes insertions "atomic" from the perspective of script, since script will not run until a given batch of DOM insertions are complete. This two-stage approach aligns the spec with a model most similar to Blink & Gecko's implementation, and fixes #808. This PR also helps out with whatwg/html#1127 and #575 (per #732 (comment)).

To accomplish, this we audited all insertion side effects on the web platform in https://docs.google.com/document/d/1Fu_pgSBziVIBG4MLjorpfkLTpPD6-XI3dTVrx4CZoqY/edit#heading=h.q06t2gg4vpw, and catalogued whether they have script-observable side-effects, whether they invoke script, whether we have tests for them, and how each implementation handles them. This gave us a list of present "insertion steps" that should move to the "post-connection steps", because they invoke script and therefore prevent insertions from being "atomic". This PR is powerless without counterpart changes to HTML, which will actually use the post-connection steps for all current insertion steps that invoke script or modify the frame tree. The follow-up HTML work is tracked, at the time of writing this:

• Use DOM's post-connection steps for <script> elements html#10188

• Meta element and DOM post-insertion steps html#10241

• <another incoming PR will be made for iframes>

• At least two implementers are interested (and none opposed):

  • ✅ Chrome (mostly implements this model; fails at least one "nested script ordering" test)
  • ✅ Firefox (mostly implements this model; fails a few tests that rely on iframe load event being fired synchronously, since I guess it fires this event asynchronously..)
  • 🟠 Safari: @annevk would you be able to provide a WebKit signal here or should I file a standards position issue?

• Tests are written and can be reviewed and commented upon at:

  • WPT results can be seen here
  • DOM: Add iframe insertion & removal steps WPTs web-platform-tests/wpt#44308
  • DOM: Added a few batch-insertion scenarios that are incompatible across engines web-platform-tests/wpt#44658
  • DOM: Add WPTs for removing steps web-platform-tests/wpt#44774
  • Add a few missing post-insertion-step tests  web-platform-tests/wpt#44906

• Implementation bugs are filed: N/A given Introduce DOM post-connection steps #1261 (comment).

• MDN issue is filed: …

• The top of this comment includes a clear commit message to use.

(See WHATWG Working Mode: Changes for more details.)

---

Two things need more discussion:

• Script child element mutation:
  • Use a queue of deferred steps in the node insertion algorithm #732 (review) and Use the new definition of the insertion steps html#4354 (comment) discuss Chromium & WebKit's model of not executing scripts when their children are modified except for insertion.
    • Regarding insertion only: All browsers are aligned on executing a script (that hasn't "already started") when a single child node is inserted. For insertion multiple child nodes atomically, see this WPT which Chrome & Firefox pass, and WebKit barely misses due to the post-connection steps¹. (Related: all browsers are consistent (see this test case) about not executing newly-inserted text nodes in a script that has "already started")
    • Regarding child removal/change: As a follow-up to Use a queue of deferred steps in the node insertion algorithm #732 (review) and Use the new definition of the insertion steps html#4354 (comment), I've investigated this a little more and found that actually all browsers follow the Chromium behavior here. See this WPT (from WPT: Script child removal does not trigger execution web-platform-tests/wpt#45085), where all browsers never trigger the execution of a script when a child node is removed. To accomplish this with the spec, we should probably just add an argument to the child changed steps, but that can be done in a follow-up PR to Use DOM's post-connection steps for <script> elements html#10188, since it is slightly orthogonal
  • The only oddity I can find with scripts in Chromium's model is that we don't follow this rule exactly, which mandates that when a script (that hasn't already started yet) gets a child <script> inserted into it, it must execute after that child script does. Chromium does not do that; the child mutation of the outer script (that inserts the inner <script>) triggers the execution of the outer script in its post-connection steps, and then after, the execution of the inner script during its post-connection steps. WebKit is aligned with Chrome on this and Firefox is the odd one out. Maybe we're OK aligning with Chrome/Firefox in this case?
• Dealing with "removal" steps
  • In general, removal is distinct from insertion, and I think has no bearing on the decisions we make in this PR and its corresponding HTML one. However, through some offline discussion, we didn't want to commit to this insertion model without having done some investigation into the removal side of things. I've started this in this doc, but see the summary here. Basically there are only two places during Node removal where browsers ever synchronously execute script: (1) pagehide which all browsers do on same-origin iframes, and (2) blur events (only Chrome does this).
  • Chrome is committed to experimenting with not firing in either of these scenarios. Along with the deprecation of mutation events, this should mark the end of all synchronous script execution on Node removal

---

Preview | Diff

Footnotes

1. The only reason WebKit fails this is because due to the lack of post-connection steps, the script executes after the first text node is appended, so by the time the second text node gets attached, the script is already started ↩

[domfarolino]

PTAL @annevk! I'll prepare an HTML PR and commit message for this, as well as file the appropriate browser bugs. But I'd love your thoughts on the meat of this so far, since it should be good to once those particulars are done.

[domfarolino]

Just a quick note for reviewers that have already taken a look: I've just edited the OP to elaborate on the fact that all browsers never execute a script as a result of removing a child node from it. This was previously pointed out in #732 (review) and whatwg/html#4354 (comment) as a maybe-Chromium-only thing — although it appears at the time, at least Chromium & WebKit behaved this way, and as of now, all browsers agree on this. Above, I've elaborated on what follow-up work can be done to DOM/HTML to align the spec with implementations. (Like I said, I don't think it needs to be handled immediately in this PR or in whatwg/html#10188.)

[domfarolino]

FWIW, even though much of the prior discussion about this and #808 makes it seem like WebKit doesn't implement the post-insertion steps that this PR introduces, the only reason that Blink implements this correctly is because it inherited the code from WebKit!

See https://github.com/WebKit/WebKit/blob/main/Source/WebCore/dom/ContainerNodeAlgorithms.cpp#L54 in WebKit, where we track which elements have post-insertion steps to call later, as well as the and the equivalent code in Blink https://source.chromium.org/chromium/chromium/src/+/main:third_party/blink/renderer/core/dom/container_node.cc;l=1095-1096;drc=54b5f6dcc0e51751aec186771b4012ef663b538b.

Also see https://github.com/WebKit/WebKit/blob/main/Source/WebCore/dom/ContainerNode.cpp#L302-L307 in WebKit, where we run each eligible element's post-insertion steps after the children change steps, and the equivalent code in Blink https://source.chromium.org/chromium/chromium/src/+/main:third_party/blink/renderer/core/dom/container_node.cc;l=1071-1077;drc=9ab453c8e7d336f430c50488e8638ca802b69d80.

For script execution, WebKit even requests that post-insertion steps are used, I just don't really understand why WebKit fails tests like this then: https://wpt.fyi/results/dom/nodes/insertion-removing-steps/Node-appendChild-script-and-button-from-div.tentative.html?label=experimental&label=master&aligned. Someone like @rniwa probably knows though.

[rniwa]

Also see https://github.com/WebKit/WebKit/blob/main/Source/WebCore/dom/ContainerNode.cpp#L302-L307 in WebKit, where we run each eligible element's post-insertion steps after the children change steps, and the equivalent code in Blink https://source.chromium.org/chromium/chromium/src/+/main:third_party/blink/renderer/core/dom/container_node.cc;l=1071-1077;drc=9ab453c8e7d336f430c50488e8638ca802b69d80.

For script execution, WebKit even requests that post-insertion steps are used, I just don't really understand why WebKit fails tests like this then: https://wpt.fyi/results/dom/nodes/insertion-removing-steps/Node-appendChild-script-and-button-from-div.tentative.html?label=experimental&label=master&aligned. Someone like @rniwa probably knows though.

It's because in WebKit, updating of element.form also happens in the post insertion steps.

[domfarolino]

It's because in WebKit, updating of element.form also happens in the post insertion steps.

Good to know. Do you think WebKit would be willing to change that? It's not a decision that needs to block this PR, since this PR just gives us the primitives available for HTML to do whatever we decide. But so far it seems 2/3 impls agree on that specific case.

[rniwa]

It's because in WebKit, updating of element.form also happens in the post insertion steps.

Good to know. Do you think WebKit would be willing to change that? It's not a decision that needs to block this PR, since this PR just gives us the primitives available for HTML to do whatever we decide. But so far it seems 2/3 impls agree on that specific case.

Unlikely. There are some design & performance constraints about this.

[domfarolino]

Unlikely. There are some design & performance constraints about this.

I'd appreciate some elaboration on this, just to understand why other browsers seem to be OK not doing this. However, it is a low priority discussion relative to this PR, since this PR doesn't change anything WRT that scenario specifically. But if you'd be willing to review this PR that would be great.

[domfarolino]

Gentle ping for reviews (and for a response to #1261 (comment)).

[annevk]

I plan on merging this next week Wednesday assuming there's no further comments.