The engine can reach a Hanja reading back into the text the client
already holds, but only if the frontend hands that text over and can
take some of it away again. GTK 3 has both: retrieve-surrounding
brings the text around the cursor and delete-surrounding removes runes
before it, and a widget that answers neither leaves the text empty, so
nothing is ever reached into or taken from it.
The wire grows a control frame for the text, sent like the caret only
when it changes, and one byte in every response for the runes to take
back. That byte moves the length fields along, so the version goes to
2: an old daemon and a new module, either way round, fail the handshake
and the module falls through to GtkIMContextSimple rather than misread
a frame.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
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>
zwp_input_method_v2 over one connection, in one process: activate,
deactivate and content type applied at done, whose count is the serial a
commit must carry. The keyboard is grabbed at every activate and
released at every deactivate, because sway hands keys to the grab holder
without asking whether the input method is active, and a grab held while
inactive would take every key in the session. Keys come as evdev codes,
go through xkb and the Compose table the other frontends already share,
and reach the engine as an ipckeysym and a modifier mask; the grab's
modifiers go on to the virtual keyboard as well, since wlroots derives no
state from a virtual keyboard's own keys. A key the engine does not eat
goes back as a key, not as text, and whatever is still down is released
when the context goes. A password or a PIN purpose never reaches the
engine at all.
No popup yet: the X11 one still owns drawc, and with no Keycaret from us
it follows the pointer.
Checked by hand under headless sway 1.12 with a GTK3 entry driven by
wtype: ascii passes through, Ctrl+s selects Korean, rk shows the preedit
and Enter commits 가 and lets the Return on, Ctrl+c arrives with its
modifier, focus moving between two clients re-grabs and keeps typing, a
password entry types literally, a second strans says so and stops, and
WAYLAND_DISPLAY without a compositor falls back to XIM.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
An IBus client throws away a dead key its engine did not take — the GTK
module's own comment says so, and Qt does the same — so é and ü were
lost in every IBus application, while XIM composed them with a table of
its own. That table moves to compose.c, which both frontends now use;
xim.c is the shorter for it.
A finished sequence is text, not a key: the engine is asked to hand back
what it had pending, and the composed text follows it, so the composed
character can no longer land before the syllable typed before it.
popuplayout had two refit stages for a work area shorter than one
padded row — 40 pixels — and popupdraw re-checked the layout it had
just been handed, against an Imgh that bounded no buffer any more.
Pango and Cairo abort rather than return nil, so textinit is void and
the layout is never nil; the stride Cairo computes for RGB24 is w*4 by
definition; and textdraw set the layout's width and ellipsis twice.
dictlookup took a Lang for two reasons that were not its business: to
know whether to walk below the key (the emoji dictionary) and to drop a
candidate equal to the key except there. No dictionary lists its key
among its candidates — the rule was left from a first version that put
the reading first itself — so it goes, and the walk is dictprefix(),
called by the emoji search alone.
Ctrl+H took the pending syllable as the query, but the keys typed after
it started a composition of their own: gk, Ctrl+H, s showed 하ㄴ, and r,
Ctrl+H, k showed ㄱㅏ. transstr now composes from a pending text, so
the search goes on where the syllable left off. Escape ended the search
and dropped the syllable it had taken; it puts it back as pending text.
mkemoji wrote a row for every prefix of every alias, so that a query
matched as it was typed; the trie is a prefix index already, and with a
real emoji list those rows would be four times the aliases themselves.
Now emoji.dict has one row per alias, trienode() names a key's node, and
dictlookup walks the entries at and below it, the key's own first, up to
Maxkouho. Trie children are appended rather than pushed, so the walk
keeps the file's order and a bare digit still picks the superscript or
subscript it always did.
The hand-written symbol rows move to symbol.src; emoji.src is left to
the emoji. mkemoji reads both by default, or the files it is given.
ko_test and vi_test replayed keys with private copies of transstr's
loop, and test_util reimplemented the shown preedit; the engine exports
impre(Im*, Str*) and transstr() instead, so a change to composition
rules is one edit.
font.c exported three wrappers for one function, and the fit == -1 mode
existed only for the tests; textdraw() takes fit and colour. win.c
tested the empty picture in the thread and again in winshow, interned
two atoms by hand next to getatom(), kept a consumer-side drain that the
producer's already guarantees never finds anything, zeroed ten globals
before returning from the only thread that read them, and re-checked
sizes that resizebacking had just established. The page marker is laid
out once and drawn from the layout instead of being recomputed.
XIM text always went out through imdkit's COMPOUND_TEXT converter,
which already wraps every UTF-8 string in ESC%G; the UTF8_STRING
negotiation, the per-client encoding list, ximtext.c's fallback and its
wrapper never changed a byte on the wire. The spot location is honoured
for every style now (GTK's XIM module sends it with PreeditCallbacks)
and an unset focus window means the client window, as the spec says,
so those clients get the popup at the caret. readattrs/place kept four
transient fields to pass values between them; ximclose tore down state
right before die(); the OOM passthrough context was a third policy for
one small calloc where emalloc dies like everything else. Both frontends
now poll for engine-owner loss only while a preedit shows instead of
XIM waking on a pipe the engine had to know about; ibus stops waking
five times a second when idle. keymeaningful() is the engine's own
predicate. The standalone xim_test moved into the unit suite, so
xim/Makefile is gone.
The dictionary thread ran in imthread's own proc, so a lookup could
only start once the engine blocked, and it was a trie probe anyway; the
emoji and Hanja searches already called dictlookup directly. The
request/result channels, sequence numbers, staleness checks and the
second draw per Japanese key are gone; dictqjp fills the candidates
in place. Emit.dict and Lang.dictq only ever triggered lookups for
Vietnamese, which has no dictionary. dictlookup(Lang*, key, out, max)
returns the count. The Hanja lookup no longer pre-checks for a single
syllable; a reading either has an entry or it does not.
The hash map served a single exact-match lookup that the trie already
answers, at the price of a second container, a second file loader with
its own drift, and a Str-to-UTF-8 conversion on every chain probe. The
files are small (kanji.dict is 7.5k lines), so the trie holds both.
trieopen validates keys and each space-separated value word against Str
and reports path:line; trielookup takes the Str every caller holds and
treats a nil trie as an unloaded map; trienew/trieput exist for tests.
The overflow guards on growth, Trie.root (always 0) and the per-language
init loop written twice are gone.
