Previously, we were accidentally depending on the state of the display
layer when forcing it to update its index. This caused us to not index
enough content to cover the visibile area, which meant we weren't
querying enough lines to fill the screen in some situations.
This was causing problems in measurements because in that code path we
assume that text nodes are never empty. This commit also adds a test
verifying this invariant when a text decoration ending right after a
text tag is added.
This was leaving a measurement request in the map that was getting
picked up on the next frame. In some cases, the requested measurement
row was not present, causing an exception.
Some packages are interacting with this method assuming this behavior,
so this commit eliminates `screenPositionForPixelPositionSync` and
instead just performs the DOM update in `screenPositionForPixelPosition`
if it is needed.
Taking the initial measurement was setting the soft wrap column, which
was triggering a display layer reset, which was scheduling an update.
This update occurred at an unexpected time causing an exception.
I was hoping to strictly contain the layouts of highlights an lines
separately, since they are updated during different render phases.
Unfortunately, strict containment requires both divs to be positioned
absolutely. This in turn creates separate stacking contexts for lines
and highlights, which makes it impossible to render highlights in front
lines which themes sometimes need to do. For example,
atom-material-syntax pushes bracket matcher highlights to the front so
they are not obscured by the theme's solid black cursor line background.
/cc @as-cii. You should examine my work here and make sure I'm not
screwing something up with your line/block decoration update code.
Unfortunately Chromium does not trigger a `compositionstart` before
firing the text input event for the non accented character. Using `undo`
to remove such character from the history is risky because it could be
grouped with a previous change, thus making Atom undo too much.
With this commit we simply keep the behavior master exhibits as of
today. In the process of rewriting this code path, however, we fixed a
bug that occurred when opening the accented character menu while holding
another key, and improved test coverage as well by simulating the events
the browser triggers.
