commit | 28bf572ae39115c939e44e5f6203325ef67658ff | [log] [tgz] |
---|---|---|
author | Morten Stenshorne <mstensho@chromium.org> | Wed Nov 02 20:11:02 2022 |
committer | Blink WPT Bot <blink-w3c-test-autoroller@chromium.org> | Wed Nov 02 20:18:16 2022 |
tree | cfd3c2a7bc6ab0c8bc66112855300f318eab9087 | |
parent | 7a78a321019d8f4455f75ec539776e3af06622bc [diff] |
Invalidate layout from changed spanner up to containing block. We need to mark all ancestors between the spanner and its containing block for layout, because they may contain references to the spanner. Just like for out-of-flow positioned elements, there may be references to column spanners from such ancestors. More precisely, there may be a descendant break token for the spanner. So if the spanner is removed, they will point to a dead object. Make sure to mark the ancestors for layout, so that we never hit the cache. It's very hard to hit the cache in front of a spanner, since we force column balancing, which very easily makes us miss the cache (and nuke it). So it was tricky to reproduce. However, it was possible if we had a multicol on a fragmentable table caption (or somewhere inside it), because of a cache miss while side-effects were disabled (which prevented the missed entry from being nuked). Grid layout also sometimes disables side-effects, so it might have been possible there as well. The table layout code initially lays out captions with side-effects disabled, so a stale entry in LayoutBox is kept even if the cache is missed, and then hit when laying out for "real" afterwards (via NGTableLayoutAlgorithm::GenerateFragment()). The entry (fragment) would have an outgoing break token tree containing a dead spanner. We always need a balancing pass in front of a spanner. If the parent of a spanner used to fragment, we'll typically miss the cache and nuke it when performing layout after a DOM / style change. However, with side-effects disabled, this doesn't happen. We just miss the cache, but keep the old result lying around (which should be okay to do). When we get to actual layout with a known fragmentainer block-size, there's nothing that prevents us from hitting the cache. There's a TODO in ComputeCaptionFragments() suggesting that we remove the disable-side-effects scope, if we instead stow away enough caption info in a break token. That's a good idea, solving this bug that way would just hide it. There may be other ways of reproducing it, and I think what the table code does is valid. The test case needs to trigger a break in a parallel flow before a spanner, since that's probably the only way of creating a break token for the spanner. This happens in the IsColumnSpanAll() when laying out children in NGBlockLayoutAlgorithm::Layout(). Bug: 1379635 Change-Id: Ic82462876d0c8b5dbf38ae21be88fff428fbb7fb Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/3998595 Commit-Queue: Morten Stenshorne <mstensho@chromium.org> Reviewed-by: Ian Kilpatrick <ikilpatrick@chromium.org> Cr-Commit-Position: refs/heads/main@{#1066634}
The web-platform-tests Project is a cross-browser test suite for the Web-platform stack. Writing tests in a way that allows them to be run in all browsers gives browser projects confidence that they are shipping software that is compatible with other implementations, and that later implementations will be compatible with their implementations. This in turn gives Web authors/developers confidence that they can actually rely on the Web platform to deliver on the promise of working across browsers and devices without needing extra layers of abstraction to paper over the gaps left by specification editors and implementors.
The most important sources of information and activity are:
wpt:matrix.org
matrix channel; includes participants located around the world, but busiest during the European working day.If you'd like clarification about anything, don't hesitate to ask in the chat room or on the mailing list.
Clone or otherwise get https://github.com/web-platform-tests/wpt.
Note: because of the frequent creation and deletion of branches in this repo, it is recommended to “prune” stale branches when fetching updates, i.e. use git pull --prune
(or git fetch -p && git merge
).
See the documentation website and in particular the system setup for running tests locally.
The wpt
command provides a frontend to a variety of tools for working with and running web-platform-tests. Some of the most useful commands are:
wpt serve
- For starting the wpt http serverwpt run
- For running tests in a browserwpt lint
- For running the lint against all testswpt manifest
- For updating or generating a MANIFEST.json
test manifestwpt install
- For installing the latest release of a browser or webdriver server on the local machine.wpt serve-wave
- For starting the wpt http server and the WAVE test runner. For more details on how to use the WAVE test runner see the documentation.On Windows wpt
commands must be prefixed with python
or the path to the python binary (if python
is not in your %PATH%
).
python wpt [command]
Alternatively, you may also use Bash on Ubuntu on Windows in the Windows 10 Anniversary Update build, then access your windows partition from there to launch wpt
commands.
Please make sure git and your text editor do not automatically convert line endings, as it will cause lint errors. For git, please set git config core.autocrlf false
in your working tree.
The master branch is automatically synced to wpt.live and w3c-test.org.
Save the Web, Write Some Tests!
Absolutely everyone is welcome to contribute to test development. No test is too small or too simple, especially if it corresponds to something for which you've noted an interoperability bug in a browser.
The way to contribute is just as usual:
git checkout -b topic
../wpt lint
as described above.If you spot an issue with a test and are not comfortable providing a pull request per above to fix it, please file a new issue. Thank you!