Hojun-Cho
650f00812d
One xkb_compose_state served the whole session, so a dead key left half typed in one window was completed by the first key in the next. Two IBus contexts through the real processkey(), before this: want length 0, got length 2; byte 0: want <end>, got 0xc3 (text) want 0, got 1 (req.op) Context A pressed dead_acute; context B pressed e, got é, and went down the composed-text branch instead of sending a key. Only COMPOSING is sticky -- xkbcommon starts over by itself after COMPOSED and CANCELLED, which the table in compose_test already pins -- and nothing here ever called xkb_compose_state_reset. That static was also fed from three procs, two of them live at once, with no lock, which xkbcommon forbids. Reaching it needs two focused windows, so it cannot be made to fail on demand and has no test: it goes because the sharing goes, not because anything guards it. So a state per frontend, each starting over when the key comes from another context. All are made in composeinit(), on threadmain, before proccreate, because xkb_compose_state_new refs the table and that ref is a plain increment -- b160: mov (%rdi),%edx b16a: add $0x1,%edx b170: mov %edx,(%rdi) -- so making a state from a shared table on two procs would race in place of the feed. Made before the procs exist they can still share the one table. An owner address that gets reused says so with composedrop: ibus hands out a contexts[] slot again, xim can malloc an Ic at a freed one, and wl's single context, which serves every client in turn, drops at every activate and deactivate. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
strans
strans is a small, single-user input method for Korean, Japanese, English,
emoji and symbols, with Vietnamese Telex as a compatibility mode. One
engine serves four frontends: zwp_input_method_v2 on Wayland — sway and
any other compositor that offers it — IBus, which GTK 4 and Qt use, XIM,
and a GTK 3 module.
Modes
| Key | Mode |
|---|---|
Ctrl+S |
Korean Hangul (2-beolsik) |
Ctrl+N |
Japanese Hiragana, converting to Kanji |
Ctrl+K |
Japanese Katakana |
Ctrl+T |
English |
Ctrl+V |
Vietnamese Telex |
Ctrl+E |
Emoji and symbol search |
Ctrl+H |
One-shot Hanja search |
Only a plain Ctrl chord is a strans key: with Shift, Alt or Super
held it commits what is pending and goes to the application, so
Ctrl+Shift+V still pastes. The mode switched to — 한, あ, ア, ă,
A — shows in the popup until the next key. A Korean keyboard's 한/영 and
한자 keys stand for Ctrl+S/Ctrl+T and Ctrl+H; a Japanese one's
半角/全角, ひらがな/カタカナ, 変換 and 無変換 for the kana modes, Space
and English.
Composing
Japanese composes a whole reading and then offers its Kanji with none
chosen. Space and Tab step through the candidates and wrap (Shift
reverses), the reading in Katakana last, so a word the dictionary lacks
converts with one Space; Up/Down and PageUp/PageDown move without
wrapping. Enter commits the chosen candidate, or the reading, as 0 and
typing on always do. Once a candidate is chosen 1-9 take that row of
the page shown; before that the digits type. Backspace deletes the last
kana shown, and romaji that made no kana stays as typed until it is mended;
Esc cancels the reading, and on a chosen candidate both go back to it.
The dictionary keys verbs and adjectives by stem and okurigana, so kaku
offers かく's readings and then 書く and its kin — typing the okurigana
with Shift, as SKK does, kaKu, asks for that split first. Katakana
mode does not convert.
Korean composes one syllable at a time. Enter, Tab, Esc and any key
that is not a jamo commit the syllable and go on to the application, so
Esc still leaves insert mode. Two lone consonants join into the compound
final they make (rt → ㄳ), and a vowel typed before its consonant is
reordered under it (kr → 가), as in libhangul.
A search takes the keys until Enter, Space or 1-9 picks a result or
Esc cancels, then returns to the previous mode. Emoji matches the typed
keys, and what they spell in the current mode, against a prefix of every
emoji's CLDR name and keywords in English, Korean and Japanese, and against
ASCII aliases such as -> and <=. Ctrl+H takes the syllable being
composed as its query and composes on from it, converting a word as well as
a syllable — 한자 gives 漢字, 대한민국 gives 大韓民國 — and Esc gives the
syllable back. Text already committed belongs to the application and
cannot be converted.
Dead keys and Compose sequences are composed by strans itself for the
Wayland, XIM and IBus frontends, from XCOMPOSEFILE or the locale; the
GTK 3 module leaves them to GtkIMContextSimple.
Preedit and candidates
| Frontend | Preedit | Candidates |
|---|---|---|
| Wayland input method | inline in the client | popup |
| GTK 3 and IBus | inline in the client | popup |
| XIM PreeditCallbacks | inline through XIM callbacks | popup |
| XIM PreeditPosition, PreeditNothing | popup | popup |
On Wayland the popup is a surface the compositor places at the text cursor,
flipping it above the line when there is no room below. Elsewhere it is an
X11 window, placed from the XIM spot or from the GTK caret, and XIM text
travels as COMPOUND_TEXT. Either way GDK_SCALE sizes it for HiDPI
displays.
Build and test
git submodule update --init
make docker-image
make docker-build
That leaves strans, the daemon, and gtk/im-strans.so, the GTK 3 module.
Tests come in three tiers:
make check # generated-map validation and unit tests
make check-live # one IBus, GTK, XIM and IPC daemon smoke each
make check-stress # randomized, capacity, collision, failure and restart
make docker-check runs all three in the container, and
make docker-valgrind the unit suite under Valgrind; rerun
make docker-image after changing Dockerfile. UNITARGS=hangul filters
the unit suite alone, not the other tiers.
A native build wants a C toolchain, Make, pkg-config and Plan 9 port, plus
development files for D-Bus, XCB and xcb-imdkit, Wayland and
wayland-scanner, xkbcommon, Pango and Cairo, and GTK 3; the tests also
want Xvfb, Python 3 and fonts covering Latin, CJK and emoji.
Dockerfile names the exact packages.
Run
./run.sh # restart the daemon in the background
./strans map # or run it in the foreground
./strans DIR reads hira.map, kata.map, telex.map, kanji.dict,
emoji.dict and hanja.dict from DIR at runtime; nothing but the GTK
module is installed. A service supervisor wants the session's
XDG_RUNTIME_DIR and its WAYLAND_DISPLAY or DISPLAY: the GTK module
finds the daemon at $XDG_RUNTIME_DIR/strans.sock, else
/tmp/strans.UID, and IBus clients through the address file libibus
expects under ~/.config/ibus/bus/, which strans writes as ibus-daemon
would.
strans shows one popup per session, so it picks a frontend at startup: a
compositor offering zwp_input_method_v2 gets the Wayland frontend, and
then neither XIM nor the X11 popup runs; every other session gets XIM. The
IBus endpoint and the GTK 3 module are served in both.
On Wayland, then, set none of the variables below. GTK, Qt and Firefox
speak text-input-v3 themselves — GTK 4 binds it with GTK_IM_MODULE unset
— and setting these is what pushes them off the path that works: such a
client still composes inline but shows no candidates, because the popup
belongs to the Wayland frontend. Chromium is beyond help either way, since
it asks for text-input-v1, which wlroots does not implement.
export XMODIFIERS=@im=strans # XIM
doas make -C gtk install # GTK 3 module
export GTK_IM_MODULE=strans
GLFW_IM_MODULE=ibus kitty # IBus client example
strans provides its own IBus endpoint, so ibus-daemon and fcitx are not
needed. Building the GTK module wants GTK development files; installing
one built elsewhere does not. The install target takes GTK_MODULE_DIR
when set, else asks GTK where its modules live; doas make -C gtk uninstall
removes it. With neither WAYLAND_DISPLAY nor DISPLAY the daemon still
serves IBus and its socket, but nothing draws a popup.
Dictionary data
After changing an input source, regenerate and verify:
python3 map/mktelex.py >map/telex.map
map/mkemoji >map/emoji.dict
map/mkhanja >map/hanja.dict
make verify-map
map/README says where the emoji, Hanja and Japanese data
comes from.
Benchmark
make bench builds bench/bench. With the daemon stopped, ./bench.sh
starts one, warms it up and records the workload with Linux perf into
bench/perf.data.
Licensing
Third-party notices are in LICENSES. The repository declares
no license for the strans source as a whole.