Fixes an issue with CJK IMEs wherein a text input state update may be
sent to the framework that misleads the framework into assuming that IME
composing has ended.
As an example, when inputting Korean text, characters are built up keystroke by
keystroke until the point that either:
* the user presses space/enter to terminate composing and commit the
character, or;
* the user presses a key such that the character currently being
composed cannot be modified further, and the IME determines that the
user has begun composing the next character.
The following is an example sequence of events for the latter case:
1. User presses ㅂ. Begin compose event followed by change event
received with ㅂ. Embedder sends state update to framework.
2. User presses ㅏ. im_preedit_changed_cb with 바. Embedder sends state
update to framework.
3. User presses ㄴ. im_preedit_changed_cb with 반. Embedder sends state
update to framework.
4. User presses ㅏ. At this point, the current character being composed
(반) cannot be modified in a meaningful way, and the IME determines
that the user is typing 바 followed by 나. im_commit_cb received with
바, immediately followed by im_preedit_changed event with 나.
In step 4, we previously sent two events to the framework, one
immediately after the other:
* im_commit_cb triggers the text input model to commit the current
composing region to the string under edit. This causes the composing
region to collapse to an empty range.
* im_preedit_change_cb triggers the text input model to insert the new
composing character and set the composing region to that character.
Conceptually, this is an atomic operation. The fourth keystroke causes
the 반 character to be broken into two (바 and ㄴ) and the latter to be
modified to 나. From the user's point of view, as well as from the IME's
point of view, the user has NOT stopped composing, and the composing
region has simply moved on to the next character.
Flutter has no concept of whether the user is composing or not other
that whether a non-empty composing region exists. As such, sending a
state update after the commit event misleads the framework into
believing that composing has ended. This triggers a serious bug:
Text fields with input formatters applied do not perform input
formatting updates while composing is active; instead they wait until
composing has ended to apply any formatting. The previous behaviour
would thus trigger input formatters to be applied each time the user
input caused a new character to be input. This has the add-on negative
effect that once formatting has been applied, it sends an update back to
the embedder so that the native OS text input state can be updated.
However, since the commit event is immediately followed by a
preedit change, the state has changed in the meantime, and the embedder
is left processing an update (the intermediate state sent after the
commit) which is now out of date (i.e. missing the new state from the
change event).
The source of this bug is as follows:
* Commit event for a character/compose region is sent from the engine.
The engine TextInputModel still models its `composing` field as true.
An update is sent to the framework with the committed text and an
empty composing range such as (1, 1). Note that the engine previously
only sends a range of (-1, -1) when composing has ended, NOT just when
it has an empty composing region.
* Framework receives commit event and updates the text to match. The
framework does not model the system composing state; instead its
understanding of whether the user is composing or not is entirely
predicated on whether the composing region is empty or not. If it is,
it triggers input formatters, which in this case have no effect on the
text/selection. However, the framework consistently models empty
compose regions as (-1, -1) and resets the text editing value as such.
Because the framework triggered a change to the TextEditingValue, it
dutifully sends the update back to the engine.
* In the meantime, in parallel with the above step, the engine starts
processing the change event immediately following the commit, and
updates the text and composing region with the next character. This
change is promptly stomped on by the incoming framework update.
To avoid this, we have the engine consistently send empty compose
regions as (-1, -1) to the framework. After the input formatter is
applied on commit, the compose region is still (-1, -1) and there are
therefore no diffs, and the framework will not send an update back to
the engine and stomp on any new state on the engine side.
Longer-term, we really should add some form of versioning information to
the text edit protocol so as to detect and resolve conflicts rather than
relying entirely on not creating races in the first place.
This bug was revealed by flutter/flutter#90211
which applies an input formatter to single-line text fields in order to
suppress newlines.
Issue: flutter/flutter#97174
Helper functions first, then callbacks/implementations, then public functions.
Add function comments.
Use consistent function naming.
Remove unnecessary protoypes.
Make a function static that wasn't.
This PR fixes an assertion error when the physical-logical map changes during key synthesization. This error will lead to irregular key sequences, and might occur during an extremely rare edge case described in the unit test in this PR.
This is a reland of flutter/engine#29178 onto the main branch. The
original patch was landed on the master branch during the branch switch
from master to main.
* Web
* macos
* Linux
* Easier web impl
* doc and format
* Better linux impl
* Format
* Better impl mac
* Format
* Windows
* Format
* Apply suggestions from code review
Co-authored-by: Greg Spencer <gspencergoog@users.noreply.github.com>
Co-authored-by: Greg Spencer <gspencergoog@users.noreply.github.com>
