[macOS] StyledText: an arrow key during Korean IME composition discards the composing syllable instead of committing it #3554

Activity

1. ethznn commented on Sep 2, 2026

   

   ethznn

   on Sep 2, 2026

   Author

   More actions

   Traced the event sequence from the Java side, and it narrows further than the report: the syllable is not dropped before it exists — it is composed into the text and then removed again.

   Listeners on SWT.ImeComposition, KeyDown, Verify, Modify and a CaretListener, typing 가 나 then beginning 다, then ←:

   Caret     -> offset=2  text="가나"    [U+AC00 U+B098]
   Caret     -> offset=3  text="가나ㄷ"  [U+AC00 U+B098 U+3137]
   KeyUp     keyCode=12599
   KeyUp     keyCode=12623
   Caret     -> offset=2  text="가나다"  [U+AC00 U+B098 U+B2E4]
   Caret     -> offset=1  text="가나"    [U+AC00 U+B098]
   KeyDown   keyCode=ARROW_LEFT  char=45796  stateMask=0
   KeyUp     keyCode=ARROW_LEFT
   

   Three things fall out of that:

   1. The preedit is carried in the widget's own text while composing — 가나ㄷ holds the compatibility jamo U+3137, and it resolves to 가나다 with U+B2E4 once the syllable completes. This is the inline IME path working as intended.

   2. The removal happens after the syllable is complete. 가나다 → 가나 is the composition being torn down, and it takes the finished character with it. So the loss is not a missed commit; it is a commit that gets reverted.

   3. The ARROW_LEFT KeyDown that follows carries char=45796, which is U+B2E4 — 다. The composed character is still in hand at the point the key is delivered; it has simply been taken out of the text by then.

   Reading IME.setMarkedText_selectedRange in Eclipse SWT/cocoa/.../IME.java, an empty marked string takes this path:

   if (text.length () == 0) {
       Shell s = parent.getShell ();
       s.keyInputHappened = true;
       startOffset = -1;
       resetStyles ();
   }

   The preceding COMPOSITION_CHANGED carries event.text = "", which is what clears the composition range in StyledText — including the completed syllable sitting in it. There is no branch that distinguishes "the input method cancelled this composition" from "the input method ended it because a caret command arrived", and the two want opposite outcomes: Esc should still discard, an arrow should commit.

   For contrast, Text on the same platform is backed by a native NSTextField, where AppKit performs that commit itself, which is why the same keystrokes keep 다 there.

   Snippet used for the trace, if it is useful:

   styled.addListener(SWT.ImeComposition, e -> log("ImeComposition detail=" + e.detail
           + " start=" + e.start + " end=" + e.end + " text=" + e.text));
   styled.addListener(SWT.KeyDown, e -> log("KeyDown keyCode=" + e.keyCode + " char=" + (int) e.character));
   styled.addListener(SWT.Modify, e -> log("Modify -> " + styled.getText()));
   styled.addCaretListener(e -> log("Caret -> offset=" + e.caretOffset + " text=" + styled.getText()));

   Happy to work up a patch if that is welcome — I have the reproduction on hand and can test any change against a real 2-Set Korean input source. Would want a steer first on where the distinction belongs, since IME is shared by every Canvas and not only StyledText.
2. tmssngr commented on Sep 11, 2026

   

   tmssngr

   on Sep 11, 2026

   Contributor

   More actions

   Thank you very much for the deep analysis. I would really appreciate a fix and it looks like you understand much more of the IME internals as I do (see #2412).

   Regarding Canvas or StyledText - I suggest to first fix it in StyledText as this is the most important keyboard driven Canvas-control right now.
3. ethznn commented on Sep 21, 2026

   

   ethznn

   on Sep 21, 2026

   Author

   More actions

   Thanks for the steer on StyledText first — and I have to walk back the delivery half of what I offered above.

   I'm not planning to sign the ECA — it asks for a legal name and a postal address, and I move often enough that keeping that current is upkeep I'd rather not take on — so the patch will have to come from someone else. Everything else I offered still stands: the reproduction is on hand, and I can run any candidate change against a real 2-Set Korean input source and report what it does. I did exactly that on #3307 earlier today.

   What the trace above leaves for whoever writes it:

   • The composed syllable is still in hand when the key is delivered — the ARROW_LEFT KeyDown carries char=45796, U+B2E4, 다 — so a fix has the character available and does not have to reconstruct it.
   • The clearing comes from a COMPOSITION_CHANGED whose event.text is empty, which IME.setMarkedText_selectedRange sends for an empty marked string. That same path serves a genuine cancel, so the distinction cannot be read from the event alone; it needs whatever the widget knows about the key that arrived with it.
   • Doing it in StyledText rather than IME keeps every other Canvas on the current behaviour, which is the conservative half of your suggestion and the reason I asked.

   If a change lands on a branch, point me at it and I'll report the measurements.