Back in March of 2021, as part of my maintenance work on Chrome/Chromium accessibility at Igalia, I inadvertently started a long, deep dive into the accessible name calculation code, that had me intermittently busy since then.
This is the third and last post in the series. We started with an introduction to the problem, and followed by describing how we fixed things in Chromium. In this post, we will describe some problems we detected in the accessible name computation spec, and the process to address them.
This post extends what was recently presented in a lightning talk at BlinkOn 17, in November 2022:
Hidden content and aria-labelledby
Authors may know they can use
aria-labelledby to name a node after another, hidden node. The attribute
aria-describedby works analogously for descriptions, but in this post we will refer to names for simplicity.
This is a simple example, where the produced name for the input is “foo”:
<input aria-labelledby="label"> <div id="label" class="hidden">foo</div>
This is because the spec says, in step 2B:
- if computing a name, and the current node has an aria-labelledby attribute that contains at least one valid IDREF, and the current node is not already part of an aria-labelledby traversal, process its IDREFs in the order they occur:
- or, if computing a description, and the current node has an aria-describedby attribute that contains at least one valid IDREF, and the current node is not already part of an aria-describedby traversal, process its IDREFs in the order they occur:
i. Set the accumulated text to the empty string.
ii. For each IDREF:
a. Set the current node to the node referenced by the IDREF.
b. Compute the text alternative of the current node beginning with step 2. Set the result to that text alternative.
c. Append the result, with a space, to the accumulated text.
iii. Return the accumulated text.
When processing each IDREF per 2B.ii.b, it will stumble upon this condition in 2A:
If the current node is hidden and is not directly referenced by aria-labelledby or aria-describedby, nor directly referenced by a native host language text alternative element (e.g. label in HTML) or attribute, return the empty string.
So, a hidden node referenced by
aria-describedby will not return the empty string. Instead, it will go on with the subsequent steps in the procedure.
Unfortunately, the spec did not provide clear answers to some corner cases: what happens if there are elements that are explicitly hidden inside the first hidden node? What about nodes that aren’t directly referenced? How deep are we expected to go inside a hidden tree to calculate its name? And what about
Consider another example:
<input id="test" aria-labelledby="t1"> <div id="t1" style="visibility:hidden"> <span aria-hidden="true">a</span> <span style="visibility:hidden">b</span> <span>c</span> d </div>
Given the markup above, WebKit generated the name “d”, Firefox said “abcd” and Chrome said “bcd”. In absence of clear answers, user agent behaviors differ.
There was an ongoing discussion in Github regarding this topic, where I shared my findings and concerns. Some conclusions that came out from it:
- It appears that the original intention of the spec was allowing the naming of nodes from hidden subtrees, if done explicitly, and to include only those children that were not explicitly hidden.
- It’s really hard to implement a check for explicitly hidden nodes inside a hidden subtree. For optimization purposes, that information is simplified away in early stages of stylesheet processing, and it wouldn’t be practical to recover or keep it.
- Given the current state of affairs and how long the current behavior has been in place, changing it radically would cause confusion among authors, backwards compatibility problems, etc.
The agreement was to change the spec to match the actual behavior in browsers, and clarify the points where implementations differed. But, before that…
A dead end: the innerText proposal
An approach we discussed which looked promising was to use the
innerText property for the specific case of producing a name from a hidden subtree. The spec would say something like: “if node is hidden and it’s the target of an aria-labelledby relation, return the innerText property”.
It’s easy to explain, and saves user agents the trouble of implementing the traversal. It went as far as to have a proposed wording and I coded a prototype for Chromium.
Unfortunately, innerText has a peculiar behavior with regard to hidden content: it will return empty if the node was hidden with the
The relevant spec says:
The innerText and outerText getter steps are:
1. If this is not being rendered or if the user agent is a non-CSS user agent, then return this’s descendant text content.
An element is being rendered if it has any associated CSS layout boxes, SVG layout boxes, or some equivalent in other styling languages.
An element with
visibility:hidden has an associated layout box, hence it’s considered “rendered” and follows the steps 2 and beyond:
- Let results be a new empty list.
- For each child node node of this:
i. Let current be the list resulting in running the rendered text collection steps with node. Each item in results will either be a string or a positive integer (a required line break count).
Where “rendered text collection steps” says:
- Let items be the result of running the rendered text collection steps with each child node of node in tree order, and then concatenating the results to a single list.
- If node’s computed value of ‘visibility’ is not ‘visible’, then return items.
The value of
hidden, hence it returns
items, which is empty because we had just started.
This behavior makes using innerText too big of a change to be acceptable for backwards compatibility reasons.
Clarifying the spec: behavior on naming after hidden subtrees
In the face of the challenges and preconditions described above, we agreed that the spec would document what most browsers are already doing in this case, which means exposing the entire, hidden subtree, down to the bottom. There was one detail that diverged: what about nodes with
aria-hidden="true"? Firefox included them in the name, but Chrome and WebKit didn’t.
Letting authors use
aria-hidden to really omit a node in a hidden subtree appeared to be a nice feature to have, so we first tried to codify that behavior into words. Unfortunately, the result became too convoluted: it added new steps, it created differences in what “hidden” means… We decided not to submit this proposal, but you can still take a look at it if you’re curious.
Instead, the proposal we submitted does not add or change many words in the spec text. The new text is:
“If the current node is hidden and is not part of an aria-labelledby or aria-describedby traversal, where the node directly referenced by that relation was hidden […], return the empty string.”
Instead, it adds a lot of clarifications and examples. Although it includes guidelines on what’s considered hidden, it does not make differences in the specific rules used to hide; as a result,
aria-hidden nodes are expected to be included in the name when naming after a hidden subtree.
Updating Chromium to match the spec:
aria-hidden in hidden subtrees
We finally settled on including
aria-hidden nodes when calculating a name or a description after a hidden subtree, via the
aria-describedby relations. Firefox already behaved like this but Chromium did not. It appeared to be an easy change, like removing only one condition somewhere in the code… The reality is never like that, though.
The change landed successfully, but it’s much more than removing a condition: we found out the removal had side effects in other parts of the code. Additionally, we had some corrections in Chromium to prevent authors to use
aria-hidden on focusable elements, but they weren’t explicit in the code; our change codifies them and adds tests. All this was addressed in the patch and, in general, I think we left things in a better shape than they were.
In the process to fix a couple of bugs about accessible name computation in Chrome, we detected undocumented cases in the spec and TODOs in the implementation. To address all of it, we ended up rewriting the traversal code, discussing and updating the spec and adding even more tests to document corner cases.
Many people were involved in the process, not only Googlers but also people from other browser vendors, independent agents, and members of the ARIA working group. I’d like to thank them all for their hard work building and maintaining accessibility on the Web.
But there is still a lot to do! There are some known bugs, and most likely some unknown ones. Some discussions around the spec still remain open and could require changes in browser behavior. Contributions are welcome!
Thanks for reading, and happy hacking!