GitLab

Original CL: http://crrev.com/c/2134848
Revert: http://crrev.com/c/2138415

Landed at the same time as another change that affected test
expectations, just needed to be rebaselined and landed again.

This is an issue with AXObjects that are ignored and not
included in the accessibility tree. When we fire an event
on such a node, the serializer owned by
RenderAccessibilityImpl first walks up to find a common
ancestor that is in the tree, and then serializes that
node, even if it was already serialized. It breaks the
mechanism that avoids re-serializing nodes that were
already just serialized.

In extreme circumstances, this led to a bug where the
size of the AXTreeUpdate for a large table grew to 200MB
for one particular website, even though the resulting
AXTree had only around 2MB worth of data.

The fix is simple: when a node changes, only fire a
notification on the nearest ancestor that's in the tree.
That way the serializer can avoid redundantly processing
the same node more than once.

Covered by a new test and a DCHECK that asserts that we
should never get an AXTreeUpdate with more nodes than the
size of the resulting tree. Without the new fix, the test
fails with an AXTreeUpdate containing 1000 nodes with a
tree of size 10.

Bug: 1066880
Tbr: Adam Ettenberger <Adam.Ettenberger@microsoft.com>
Tbr: Aaron Leventhal <aleventhal@chromium.org>
Relnotes: N/A
Change-Id: I485e8d98d8fe9638706e26ceff642664a3fa6e3e
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/2140033Reviewed-by: Dominic Mazzoni <dmazzoni@chromium.org>
Reviewed-by: Adam Ettenberger <Adam.Ettenberger@microsoft.com>
Commit-Queue: Dominic Mazzoni <dmazzoni@chromium.org>
Cr-Commit-Position: refs/heads/master@{#757221}