The symptom : when text is selected and the device is rotated,
sometimes the keyboard sets the word as being composed around
the start of the selection. Upon the next rotation this ends up
with the keyboard committing some text in place of the selection.
The cause : another bug in the framework with rotation >.>
The keyboard receives a call to startInput with a wrong cursor
position, namely one that does not represent a selection. The
keyboard sets a composition according to this wrong data. When
the keyboard is rotated again, it commits the text, which takes
the place of the selection.
The solution : actually when restarting input the keyboard
realizes that the cursor position is wrong. We cancel composition
at that time.
For robustness, this change also implements two other defensive
changes : upon call to onUpdateSelection, we actually realize
that the previous values were wrong, so we also fix it at that
time, and in addition, when rotating, we finishComposingText()
instead of commitText() which is less dangerous. Implementing
this later change also allows us to let less internal variables
from InputLogic escape to LatinIME, so it's also a good change
for design.
Bug: 14140799
Change-Id: Ib10de18e53e376ac1bbc8487e13d969828483346
The heuristic in RichInputConnection makes little sense
when textLength > mExpectedSelStart but we have
more than 1024 characters of text. If there are that many,
it's about 100% sure that 1024 is not the correct cursor
position. With no good guess, we'll just continue trusting
the app, even though we know it's lying : at least it will
make the problem visible to the app author.
Also, there have been a lot of confusion about initialSelStart
and initialSelEnd. The keyboard should log them so that
it helps us and editor authors debug more easily these
common problems.
Issue #65170 in AOSP and
Bug: 12772035
Change-Id: I6665a16c9f2832d33ee323f033bb38bcc092a3b4
When the cursor is moved by the user, the RichInputConnection
is told about it. However, to work around a framework bug, it
also looks at how many characters are in the buffer before the
cursor, and if that's more than the value it's been passed, it
deduces that's a framework bug and there are at least as many
characters as seen before the cursor, so it puts the expected
cursor position there.
When you move the cursor, TextView calls onUpdateSelection,
and when you move it fast, you'll get rapid-fire calls to
onUpdateSelection. This is fine, the RIC is equipped to
deal with that.
However, these calls take some time to make it to the IME. In
this instance, when the first call gets through and the IME
calls TextView (synchronously) for text before the cursor, the
cursor has already moved in the app, and TextView returns more
characters than the cursor position was declared to be in this
instance, so the RIC sets that as the expected cursor position.
Sure enough, a split second later, the second call to
onUpdateSelection arrives, with the new cursor position set
where the RIC had found it too early. The RIC takes that as an
"expected" cursor move, and the input does not get reset.
Luckily, we have a way out. As far as we know, the framework bug
only manifests itself upon rotation, which means we should only
have to adjust for it in onStartInputView. Doing it in
onUpdateSelection is too zealous (and probably too distrustful of
the app to send the correct cursor positions).
So we should just take care of the rotation case (by calling
tryFixLyingCursorPosition in onStartInputView) and remove the
compensating code in resetCachesUponCursorMoves.
Bug: 12982502
Change-Id: Ic3c1408a1ec45deaea63b01d98376a79ae567d77
During recorrection, the cursor position when calling
commitText is not necessarily at the end of the
composing text.
Besides, RichInputConnection assumes the cursor is
always after any composing text. This is not correct,
but in the practice, it seems all code paths work.
We should fix this in the future.
Bug: 13060691
Change-Id: I15f71fff62d36e80cf6e4a022c5e78af634b199d
This just mirrors what InputLogic#tryFixLyingCursorPosition
is doing. That method will go away in the next change.
Change-Id: Ifa2827dbc1f1d20e2c642d6f2d23514a01ed9203
This test was intended only for cases without a selection, and as
a safety net for cases where the app would pretend the cursor
is at N but we can get P chars from the editor where P > N.
When there is a selection, this is wrong. In the practice it works
because these values are not used in this case, but it's still wrong.
The case where P > N is arguable, but actually I see little reason
to trust the getTextBeforeCursor() method more than the
onUpdate selection method. Plus in the practice, I don't think
we are aware of any app with this bug, and it's probably not a
great idea to be too robust about this as it may encourage wrong
values sent to onUpdateSelection.
Change-Id: I42f2065d7aee668074e6b8e40b259da7e88e16e1
Don't use absolute cursor positions when making edits,
this leads to race conditions.
This is a bit ugly and will need to be fixed soon. Plans are
underway to clean this up.
Bug: 12390573
Change-Id: I69c09fc41b979880d0800c55a710e39373287cff
Don't use absolute cursor positions when making edits,
this leads to race conditions.
This is a bit ugly and will need to be fixed soon. Plans are
underway to clean this up.
Bug: 12390573
Change-Id: Ib42d4149343c642b1b5c1937b424e8afdbd4cc1f
This old method doesn't even re-read the old suggestions. It used to
recompute them without the coordinates.
Re-using the recorrection code, which is much more advanced, is
the right thing to do here.
Also, refining the test. It's no use trying to resume suggestion
if we don't have a suggestion strip, since we aren't going to
auto-correct anything anyway.
Not the motivation for this change, but this also fixes
Bug: 11620256
Change-Id: Id49efa32e293c49837c61fdc752c86bbac1d2c88
This returns the wrong string, but since it's used for getting the
previous word for bigrams, it only results in slightly worse
suggestions quality.
Bug: 11273655
Change-Id: I6ce5de2f76effc453ca691a654ab6bf17445b9e7
The PARAGRAPH type of span is dangerous, as concatenating
CharSequences that contain it may crash. We also don't use
other spans than SuggestionSpans, so we don't copy them.
Bug: 10622489
Change-Id: If4e44eca3cdc5bb02cf2e0c8c44ecd4bf27fae57
It's unclear what the concrete effects of this are, but they are not
very strong. This only happens in corner cases, when the input
connection is not active - while rotating, for example.
Change-Id: I1d22459a6e94a8ecccb53cfcbc2d301b1d502204
InputConnection#finishComposingText() should not change the position of the cursor,
so neither should it change its internal expectation of the cursor's position.
Change-Id: Ib3d39a5743cd1e8e356f438b04a5c30279430b2a
This change also eliminates a reference of
AudioAndHapticFeedbackManager from KeyboardSwitcher and MainKeyboard.
Bug: 6522943
Change-Id: Iac42ec8ff00c66deb76a660ffc07477923a58959
I548d899b introduced a new method to fix a sync miss between
the cursor position and the cached cursor position, but did not
take into account that it should also update the cached text
before and after the cursor in this case and that there was
already a method for doing this.
Change-Id: I31bd741893207c822827304e77791b1159774e1a