The whole-row memcmp settled an unchanged row, but on a changed row the cell
loop then read again the prefix memcmp had already compared, and the trailing
trim walked unchanged cells one by one. Chunks of eight cells now find the
first and last changed cell, so no cell is compared twice and only the span
between them is scanned cell by cell. Output is byte-identical.
Measured on frames captured from vif at 60 fps (median of 15, one core):
vanguard 120x40 73.7 -> 66.6 us a frame (-10%), vanguard 240x66
158.8 -> 131.8 (-17%), main with the gun firing 40.0 -> 34.1 (-15%), main
idle 20.3 -> 14.0 (-31%). Dropping the row compare instead costs +15% to
+120% on the same frames.
* mouse: the wheel's four directions and the side buttons
parseSGRMouse read a report's low two bits alone, so a wheel tilted left or
right (66, 67) reported as WheelDown, and a mouse's back and forward buttons
(128, 129) as a left or middle press. It now decodes by bank: bits 0-1 number
a button within the main three and none, the wheel's four directions (bit 6),
or buttons 8 to 11 (bit 7). WheelLeft and WheelRight join MouseButton after
Forward, so every existing value keeps its number; Back and Forward are
reported for buttons 8 and 9; buttons 10 and 11, and a report with both bank
bits set, are dropped as KeyNone. A wheel step is a press with no release in
every direction.
* tui: ScrollThumb and ScrollDrag
ScrollThumb is the thumb ScrollBarStyled draws, its first row and height,
exported so an application hit-tests the bar it drew. ScrollDrag drives a bar
by the pointer, rows counted from the track's top: a press on the thumb takes
hold where it lands, a press on the track moves the thumb's middle there and
takes hold, and each drag keeps the held row under the pointer, clamped at the
track's ends, at the nearest offset when there are fewer offsets than rows.
Each call returns the offset to show.
A table of SGR reports covers each bank and drops the buttons past 9; a
property test over every track of 3 to 10 rows and content up to 30 checks
each press and drag against the thumb ScrollThumb draws.
* linux console source: releases from the console's own keyboard, typed by its keymap
SetKeyReport with KeySourceConsole, on a Linux virtual console the game runs
on, switches the keyboard to raw keycodes (K_MEDIUMRAW) while releases are
reported. The kernel allows the switch on the caller's controlling tty, so no
root is needed; a console that refuses is not asked again and the protocol is
next. Going back to presses, Fini and EmergencyReset switch it back.
Each keycode is typed as the console types it cooked, after
drivers/tty/vt/keyboard.c, by the console's own tables read at the switch
(KDGKBENT, KDGKBSENT, KDGKBDIACRUC, KDGKBMETA): modifiers and their left and
right groups, Caps and Num Lock with their lights kept in step (KDSKBLED),
sticky and locked modifiers, Unicode entries, dead keys and compose, the
keypad by Num Lock, function strings, Meta as ESC prefix or bit 7, console
switching by Alt+Fn and Alt+arrows to consoles in use, and from kernel 7.1
cursor and CSI keys with their modifiers. Translation starts from the
console's modifiers and locks (TIOCLINUX 6, KDGKBLED). Each key is held by
its keycode; a key typing several keys taps all but its last. Switching to
another console ends every hold Lost. Mouse reports are off while the
keyboard is raw, since the console types them into the same input, and come
back after it.
The guard: before the first switch the package runs the binary again, named
console-guard, holding the console and the read end of a pipe only the game
writes. If the pipe ends while the keyboard is raw (kill -9, the OOM killer, a
Go fatal error), the guard switches the keyboard back, restores the termios
the game found and shows the cursor, without a reset that would clear the
game's last words. A game leading its console's session takes the console
with it as it dies; its guard then leads a session of its own and opens the
console again once the kernel has hung it up, to restore it.
Checked in QEMU on a real VT, kernels 6.8 and 7.2, the game as uid 1000:
the same key strokes read identically raw and cooked (Shift released first,
Ctrl, Alt, AltGr, Caps and Num Lock, the keypad, function keys, modified
cursor keys on 7.2; dead grave and tilde, compose, a euro sign and Cyrillic
from a loaded keymap), Alt+F2 away ended the hold Lost and Alt+F1 came back,
and the keyboard came back to K_UNICODE with ICANON and ECHO after Fini, a
panic with EmergencyReset, kill -9, and kill -9 of a game leading its session
(left raw without the guard's own session).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B3zCVjspRDuEuMrDTumVhp
* freebsd console source: releases from vt(4)'s keyboard, typed by its keymap
On FreeBSD's vt(4) console the keyboard switches to keycodes (K_CODE),
which vt allows any process with the console open, no root needed. Each
keycode types what vt types for it cooked: the keyboard's keymap and
accent map (GIO_KEYMAP, GIO_DEADKEYMAP) run through a port of the
kernel's genkbd_keyaction, kbdmux's left Alt keypad codes, and
vt_processkey's xterm sequences, ESC for Alt and console switching. The
locks go back to the keyboard and its lights with KDSKBSTATE. Scroll Lock
is inert while raw; with the console's history mode on (Scroll Lock lit)
the keyboard stays as it is and the ask reports presses.
Shared with the Linux source now: the restore EmergencyReset runs, the
guard, which opens again a console revoked or hung up with its session's
leader, and what a key asks of the console. A console that can't go raw
just now (Scroll Lock on, its guard gone, output not drained) is tried
again at the next ask rather than never; a refusal at prepare still
holds for good.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B3zCVjspRDuEuMrDTumVhp
* example-keys: show each key event, for trying key releases
A viewer for the key events a terminal or console reports: F2 asks for
releases from the console or the terminal's protocol, or gives them
back, and the report each ask came to shows above the events.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B3zCVjspRDuEuMrDTumVhp
---------
Co-authored-by: Claude <noreply@anthropic.com>
* key model for releases: KeyAction, SetKeyReport, focus and key report events, hold pairing
A key event now carries its Action: KeyPress, KeyRepeat or KeyRelease. Every
terminal still reports presses only, so Action stays KeyPress and every event
is what v0.4.3 sent, until a game asks with SetKeyReport. The ask names the
sources releases may come from (KeySourceConsole, KeySourceProtocol); an
EventKeyReport answers each ask with the level reached and its source. No
source exists yet, so every ask answers presses and nothing is written.
KeyHolds pairs a release with its press by physical key, so a release carries
what its press carried, and turns a source's lost reports into Lost releases.
It is exported for hosts that read native key events (KeySourceHost).
The reader sends keys through the level in force: at presses a repeat is a
press and a release is dropped; at releases a key with no physical key behind
it is a tap. Focus reports end every hold, Lost, before EventFocus; they are
read only once a source turns them on. Mouse events gain a Touch bit for hosts.
Terminal gains SetKeyReport, so this is v0.5.0.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B3zCVjspRDuEuMrDTumVhp
* protocol source: releases from the terminal's key event protocol, Alacritty's
SetKeyReport with KeySourceProtocol asks the terminal whether it has the key
event protocol (CSI ? u, fenced by a DA1 request), pushes its flags (all five,
CSI > 31 u) and asks what they came to before reporting releases. A terminal
must confirm event types, every key as an escape code and the text keys type;
one that answers DA1 alone, takes the push without them, or answers nothing
within two seconds stays at presses, and is not asked again. Detection goes by
these answers only, never by TERM or the terminal's name.
While the flags are pushed, key reports are decoded into the events a legacy
terminal sends for the same keys, so a game can keep the protocol on for its
whole run: Shift goes into the character, Ctrl with a character into its
control key, Alt stays a modifier; Super, the locks, lone modifiers, media keys
and F13 and up report nothing. Ctrl with a character that has no control key
keeps ModCtrl. Releases pair with their presses through KeyHolds; a key the
terminal names none for (text from an input method) is a tap.
Focus reports come on with the protocol and end every hold, Lost, as focus
leaves. Going back to presses releases what is held, pops the flags and stops
focus reports, decoding until the terminal confirms the pop. Fini pops the
flags before the alternate screen goes, and EmergencyReset when they are
pushed. Fini now stops the input reader without holding the terminal's lock,
since the reader writes its requests through it.
Checked against Alacritty 0.13.2 on Xvfb with xdotool: 58 key combinations
read identically with the protocol on and off (Ctrl+1 keeps ModCtrl), every
press got its release, Shift released first still released 'A', focus loss
and a release swallowed by Alacritty's Vi mode ended holds Lost. Inside tmux
3.4 the ask came to presses.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B3zCVjspRDuEuMrDTumVhp
---------
Co-authored-by: Claude <noreply@anthropic.com>
Lixen
lixen