Stopgap solution for a crash.
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.
Cherry-pick of I6ce5de2f
Bug: 11273655
Change-Id: I17fb6d74f18fb31bd8f8518f80456d74ae30a2c3
diff --git a/java/src/com/android/inputmethod/latin/RichInputConnection.java b/java/src/com/android/inputmethod/latin/RichInputConnection.java
index e43cab5..c212f9c 100644
--- a/java/src/com/android/inputmethod/latin/RichInputConnection.java
+++ b/java/src/com/android/inputmethod/latin/RichInputConnection.java
@@ -294,7 +294,14 @@
// the text field, then we can return the cached version right away.
if (cachedLength >= n || cachedLength >= mExpectedCursorPosition) {
final StringBuilder s = new StringBuilder(mCommittedTextBeforeComposingText);
- s.append(mComposingText);
+ // We call #toString() here to create a temporary object.
+ // In some situations, this method is called on a worker thread, and it's possible
+ // the main thread touches the contents of mComposingText while this worker thread
+ // is suspended, because mComposingText is a StringBuilder. This may lead to crashes,
+ // so we call #toString() on it. That will result in the return value being strictly
+ // speaking wrong, but since this is used for basing bigram probability off, and
+ // it's only going to matter for one getSuggestions call, it's fine in the practice.
+ s.append(mComposingText.toString());
if (s.length() > n) {
s.delete(0, s.length() - n);
}