| Commit message (Collapse) | Author | Age |
|---|
| ... | |
| |
|
|
|
| |
There's some bounds checks we do that panic if the condition is ever
true.
|
| |
|
|
|
|
|
| |
Using the vte crate allows removal of the ansi parser state machine and
enables us to just be concerned with actions described in the protocol.
In addition to making alacritty simpler, this also improves correctness
and performance.
|
| |
|
|
|
| |
Allegedly x11 wants you to wait for a refresh event before using the
context, but that doesn't arrive on all platforms (notably macOS).
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is only like the third time I've made this change. The issue of
having a blank screen at startup is due to x11 event loop + glX
interactions. Not sure what the problem is specifically, but
glXMakecurrent was blocking until the x11 event loop advanced.
The input and rendering are able to live on the same thread while still
removing unnecessary renders due to the
glutin::WindowProxy::wakeup_event_loop() method. The PtyReader just
kicks the event loop when there's something to do; otherwise, the event
loop just waits for something to happen and _doesn't_ draw in free run
mode.
|
| |
|
|
|
|
|
|
|
|
| |
Currently has a bug where screen is blank at startup. That aside,
Alacritty uses basically 0 CPU now. The input thread is still separate
from the render thread, but, given the ability to wake the event loop,
it may be possible to merge them again. I'm not sure if that's actually
desirable.
Performance is seemingly unchanged.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
To minimize rendering, the input must be handled in a separate thread.
To see, why, consider the optimal rendering solution: renders are only
necessary when the pty has data that changes the terminal state, OR
there is a window event which changes the graphics state. When not
drawing, the render thread is to remain parked at a condition variable,
and it's not possible to handle input while parked! Thus, we need a
separate thread.
In addition to adding the separate thread, each subsystem thread is now
spawned in a separate function to (hopefully) improve readability.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch adds support for rendering italic fonts and bold fonts.
The `font` crate has a couple of new paradigms to support this: font
keys and glyph keys. `FontKey` is a lightweight (4 byte) identifier for
a font loaded out of the rasterizer. This replaces `FontDesc` for
rasterizing glyphs from a loaded font. `FontDesc` had the problem that
it contained two strings, and the glyph cache needs to store a copy of
the font key for every loaded glyph. `GlyphKey` is now passed to the
glyph rasterization method instead of a simple `char`. `GlyphKey`
contains information including font, size, and the character.
The rasterizer APIs do not define what happens when loading the same
font from a `FontDesc` more than once. It is assumed that the
application will track the resulting `FontKey` instead of asking the
font to be loaded multiple times.
|
| | |
|
| |
|
|
|
|
| |
This is experimental on a separate branch of Glutin. It's intended to
fix the problem of certain key events not being delivered on alt-tab and
breaking the modifier state tracking.
|
| |
|
|
|
|
|
| |
Previous version of serde no longer worked; cargo packages were updated
as a result. `Zero` and `One` traits were deprecated. Use of those was
removed. The `Step` trait gained a lot more methods, and the index::$ty
implementations were updated.
|
| |
|
|
|
| |
The cursor now matches perfectly cell widths and actually shows the
character underneath.
|
| |
|
|
|
| |
terminal.resize was not being called. Additionally, it was being called
too much on other platforms (resize events were completely coalesced).
|
| |
|
|
|
|
| |
The default characters sent for this were incorrect. Delete now sends
xterm-compatible escapes (dch1, kdch1), and Backspace sends the delete
code 0x7f.
|
| |
|
|
|
|
|
| |
The extra render thread just resulted in extra complexity. There were
also some startup bugs not resolved with that architecture. Finally,
there was no noticeable performance boost from having the additional
thread.
|
| |
|
|
|
|
|
|
| |
The grid and term modules already rely on the index types, and ansi is
about to be updated with strongly typed APIs. Since Cursor, Line, and
Column are fundamental to the code in several modules, namespacing them
under one of them seems less correct than a module that stands by
itself.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The Grid no longer knows about a `Cell` and is instead generic. The
`Cell` type is coupled to the `term` module already, and it's been moved
there to reflect the strong relationship.
Grid APIs previously accepted `usize` for many arguments. If the caller
intended rows to be columns, but the function accepted them in reverse,
there would be no compiler error. Now there is, and this should prevent
such bugs from entering the code.
The Grid internals grew significantly to accomodate the strongly typed
APIs. There is now a `grid::index` module which defines Cursor, Line,
and Column. The Grid APIs are all based on these types now. Indexing for
Ranges proved to be somewhat awkward. A new range had to be constructed
in the implementation. If the optimizer can't figure out what's going on
in that case, the ranges may not be a zero-cost abstraction.
|
| |
|
|
|
|
| |
It's a generic impl of `input::Notify` for `Write` types; as such, it
seems completely reasonable to include in the input module. Moving it
also serves to declutter main.
|
| |
|
|
|
|
|
|
|
|
|
| |
The pty read thread now runs the parser and directly updates the
terminal in the same thread. This obviates the need for a channel which
sends every char read from the pty; this is a huge performance boon.
Synchronization between the updater and the renderer is now achieved
with a PriorityMutex. Previously, an atomic bool was (poorly) used to
request the lock on terminal. The PriorityMutex is dead simple to use,
and it _Just Works_.
|
| |
|
|
|
|
|
|
|
|
|
| |
The upcoming Utf8Chars iterator was vendored from a libstd PR. The
iterator works on BufRead types which is critical for improving
performance. A small modification was made where the number of unused
bytes is included with Utf8CharsError::IncompleteUtf8.
The pty reader thread was updated to use this new type. Next steps will
be moving the parsing there and either sending parse results in batches
or updating the terminal directly from that thread.
|
| |
|
|
|
|
|
|
|
| |
Configuration may now be specified in either `$HOME/.alacritty.yml` or
`$HOME/.config/alacritty.yml`. See `alacritty.yml` in the repository
root for an example.
When a configuration file cannot be located, a default configuration is
used.
|
| | |
|
| |
|
|
|
|
|
| |
The resize callback is the only way to perform live resizing on macOS. A
callback is provided, and a static variable is used to provide a Sender
to that function so that resize events may be processed in the usual
way.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
The resize event is received from glutin on the update thread, but the
renderer needs to be informed as well for updating the viewport and
projection matrix. This is achieved with an mpsc::channel.
To support resizing, the grid now offers methods for growing and
shrinking, and there are several implementations available for
clear_region based on different Range* types.
Core resize logic is all in Term::resize. It attempts to keep as much
context as possible when shinking the window. When growing, it's
basically just adding rows.
|
| |
|
|
|
|
|
| |
This moves more logic out of main() and prepares the Term type to handle
resizing. By providing all size data to Term, it is now possible to
implement a resize function there which handles all resizing logic save
for the rendering subsystem.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The input/pty processing loop previously consumed input on a channel and
hit the rendering when the queue became empty. This had a couple of
problems
1. It was possible to be overwhelmed with input and not give the
renderer an opportunity to update the screen. This gave the
appearance of locking up.
2. Multiple frames could be rendered for a single vblank (redundant
work)
3. Open loop rendering would inevitably have buffer swapping out of sync
with vblanks and visual tearing would occur.
This commit enables vsync on the glutin window. The rendering was all
moved onto a separate thread from input/pty processing to support vsync.
However, the rendering thread must be 100% synchronized with the updater
thread. There's now a Mutex on the Term, and an atomic bool to ask the
input/pty processing to yield to the renderer.
One aspect of this feature that hasn't been worked out is how to limit
the frame rate. Currently, it just free runs at the screen refresh rate.
The initial attempt here included the input/pty processor holding a lock
while waiting for input. This *almost* worked, but there was a (not
uncommon) edge case where the terminal state was "dirty" but the
renderer was not ready to draw.
Instead of blocking on the refresh issue, it's being punted to after an
MVP is released. The overhead of drawing at 60Hz was profiled to be ~5%
CPU usage, and this is deemed acceptable for an MVP.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
There's a number of keys/combinations that should emit escape sequences
to the PTY when triggered. This commit adds a framework to support that.
The input::Processor is a type which tracks state of modifier keys. When
special keys (like arrow, function) are detected, the processor pulls up
a list of candidate escapes to send, and picks the first one based on
terminal mode and active modifier keys. The input::Processor is generic
over the thing receiving the escape sequences, the input::Notify type.
Included is a wrapper for `&mut io::Write` which implements
input::Notify and is currently used to connect the processor to the PTY
stream.
This added handling of the APP_CURSOR mode which changes affects input
processing.
|
| |
|
|
|
|
|
| |
The sense of set_mode and unset_mode was inverted. The
TextCursor/ShowCursor mode depended on the incorrect behavior, and that
was fixed as well. TextCursor was renamed to ShowCursor to be perfectly
clear on the intent.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Alacritty now runs on macOS using CoreText for font rendering.
The font rendering subsystems were moved into a separate crate called
`font`. The font crate provides a unified (albeit limited) API which
wraps CoreText on macOS and FreeType/FontConfig on other platforms. The
unified API differed slightly from what the original Rasterizer for
freetype implemented, and it was updated accordingly.
The cell separation properties (sep_x and sep_y) are now premultiplied
into the cell width and height. They were previously passed through as
uniforms to the shaders; removing them prevents a lot of redundant work.
`libc` has some differences between Linux and macOS. `__errno_location`
is not available on macOS, and the `errno` crate was brought in to
provide a cross-platform API for dealing with errno.
Differences in `openpty` were handled by implementing a macOS specific
version. It would be worth investigating a way to unify the
implementations at some point.
A type mismatch with TIOCSCTTY was resolved with a cast.
Differences in libc::passwd struct fields were resolved by using
std::mem::uninitialized instead of zeroing the struct ourselves. This
has the benefit of being much cleaner.
The thread setup had to be changed to support both macOS and Linux.
macOS requires that events from the window be handled on the main
thread. Failure to do so will prevent the glutin window from even
showing up! For this reason, the renderer and parser were moved to their
own thread, and the input is received on the main thread. This is
essentially reverse the setup prior to this commit. Renderer
initialization (and thus font cache initialization) had to be moved to
the rendering thread as well since there's no way to make_context(null)
with glx on Linux. Trying to just call make_context a second time on the
rendering thread had resulted in a panic!.
|
| |
|
|
|
|
|
|
|
|
|
| |
Of note are the `ansi` and `grid` modules becoming public. There are
several bits of unused code in each of these. In the case of `grid`, the
unused parts are generally useful, like some indexing implementations.
In ansi, there are pieces that will be used once the parser is more
complete. In any case, these modules are fairly generic and mostly
usable outside of Alacritty.
Unused cargo packages were also removed.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The main thing preventing this system being event driven in the past was
input from the keyboard had to be polled separately from pty activity.
This commit adds a thread for the window event iterator and sends them
on the same channel as pty characters.
With that in place, the render loop looks like
- Block on 1 available input
- Get all remaining available input that won't cause blocking
- Render
Which means that rendering is only performed on state changes. This
obsoleted the need for a `dirty` flag in the Term struct.
|
| |
|
|
|
|
| |
Calling ::std::process::exit() from the SIGCHLD handler would sometimes
deadlock some OpenGL internal shutdown procedure. To resolve this, a
flag was added that can be checked with `process_should_exit`.
|
| |
|
|
| |
When the flag is unset, the cursor is not rendered.
|
| |
|
|
|
|
|
|
| |
This is achieved by setting a `dirty` flag when the terminal receives
an event that causes visible state to change. The implementation is
pretty much crap because most methods know about the flag.
Figure out something better later.
|
| |
|
|
| |
The main loop is now exitted if the char sender hangs up.
|
| | |
|
| |
|
|
|
|
| |
They're still unhandled, but they won't silently pass by anymore.
TODO
|
| | |
|
| |
|
|
|
|
|
|
|
| |
The glyph cache was previously initialized with a list of glyphs from
INIT_LIST, and never updated again. This meant that code points not
included in that list were not displayed. Now, the glyph cache has
gained the ability to load new glyphs at render time.
This seems to have lightly decreased performance for some reason.
|
| |
|
|
|
|
|
|
|
|
| |
Per-instanced data was previously stored in uniforms. This required
several OpenGL calls to upload all of the data, and it was more complex
to prepare (several vecs vs one).
Additionally, drawing APIs are now accessible through a `RenderApi`
(obtained through `QuadRenderer::with_api`) which enables some RAII
patterns. Specifically, checks for batch flushing are handled in Drop.
|
| |
|
|
|
| |
Draw calls are now batched for performance. Render times on git log at
the default size are now ~200usec.
|
| |
|
|
|
|
|
|
| |
This moves some logic that was previously being done per-character into
the vertex shader. At this time, we've traded CPU computation for
additional gl::Uniform2f calls. This is only a marginal improvement.
However, this patch positions the renderer well for instanced drawing,
and that will be a huge performance win.
|
| |
|
|
|
|
|
| |
Recompiling the entire program whenever a shader changes is slow, and it
can interrupt flow. Shader reloads are essentially instantaneous now. If
the new shader fails to compile, no state is changed; the previous
program continues to be used.
|
| |
|
|
|
|
| |
This dramatically reduces the number of BindTexture calls needed when
rendering the grid. Draw times for a moderately full terminal of the
default size are ~1ms with this patch.
|
| |
|
|
|
|
|
|
| |
This moves the rendering logic to draw the grid, to draw strings, and to
draw the cursor into the renderere module. In addition to being an
organizational improvement, this also allowed for some optimizations
managing OpenGL state. Render times for a moderate screen of text
dropped from ~10ms to ~4ms.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch introduces basic support for terminal emulation. Basic means
commands that don't use paging and are not full screen applications like
vim or tmux. Some paging applications are working properly, such as as
`git log`. Other pagers work reasonably well as long as the help menu is
not accessed.
There is now a central Rgb color type which is shared by the renderer,
terminal emulation, and the pty parser.
The parser no longer owns a Handler. Instead, a mutable reference to a
Handler is provided whenever advancing the parser. This resolved some
potential ownership issues (eg parser owning the `Term` type would've
been unworkable).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is the initial terminal stream parsing implementation for
Alacritty. There are currently several TODOs, FIXMEs, and unimplemented!
things scattered about still, but what's here is good enough to
correctly parse my zsh startup.
The `Parser` implementation is largely based on the suck-less _simple
terminal_ parser. Because this is Rust and Rust has a fantastic type
system, some improvements are possible. First, `Parser` is a struct, and
its data is stored internally instead of statically. Second, there's no
terminal updates hard-coded into the parser. Instead, `Parser` is
generic over a `Handler` type which has methods for all of the actions
supported by the parser. Because Parser is generic, it should be
possible (with proper inlining) to have equivalent performance to the
hard-coded version.
In addition to using _simple terminal_ as a reference, there's a doc in
Alacritty's repository `docs/ansicode.txt`, a summary of the ANSI
terminal protocol, which has been referenced extensively.
There's probably a large number escapes we don't handle, and that's ok.
There's a lot that aren't necessary for everyday terminal usage. If you
feel like something that's not supported should be, feel free to add it.
Please try not to become overzealous and adding support for sequences
only used by folks trapped in 1988.
|
| |
|
|
|
|
|
|
| |
Opens a pty, forks a child process, and execs the shell defined in
user's /etc/passwd file. Bytes from the pty are currently just written
to Alacritty's stdout as a sanity check that things are hooked up.
Thanks to `st` for some guidance on setting this up.
|
| |
|
|
| |
Optimization is impossible without measurement!
|
