This is done in order to help prevent us from loading the same
font face over and over again under separate keys. We still incur
the performance hit of doing the fontconfig search each new glyph,
but that's unavoidable without more extensive refactoring.
Loading a glyph from the cache is a very hot operation in the renderer.
The original implementation would first check if a glyph was loaded and
then call `get()` which would have to search a second time. This showed
up as a very slow point in profiles.
This patch addresses glyph cache access in two ways: by using a faster
hasher optimized for small keys (fnv), and by using the entry API for
fetching a cached glyph. The `fnv` hasher is faster than the default and
is very efficient for small keys. Using the entry API on the HashMap
means only 1 lookup instead of two. The entry API has a downside where
the key needs to get cloned on fetches.
Reducing the GlyphKey width to 64-bits helps in both areas. Copying an
8-byte wide type is very cheap and thus limits downside of the entry
API. The small width also helps with the hasher performance.
Over all, this patch reduced typical render times by several hundred
microseconds on a 2013 MacBook Pro with a full screen terminal full of
text.
Makes thin stroke rendering for darwin configurable by a new toplevel
key under `font:` in the config file. Defaults to false, has no impact
on non macos.
This work started because we wanted to be able to simply say "monospace"
on Linux and have it give us some sort of font. The config format for
fonts changed to accomodate this new paradigm. As a result, italic and
bold can have different families from the normal (roman) face.
The fontconfig based font resolution probably works a lot better than
the CoreText version at this point. With CoreText, we simply iterate
over fonts and check it they match the requested properties. What's
worse is that the CoreText version requires a valid family. With
fontconfig, it will just provide the closest matching thing and use it
(unless a specific style is requested).
There's now a proper wrapper in place for working with the FontConfig
library. This should help significantly with error handling with font
loading; at least, the FontConfig code shouldn't panic. The FreeType
rasterizer still needs to be updated to handle missing fonts, and a more
sensible default font should be specified.
Ref tests use a recording of the terminal protocol and a serialization
of the grid state to check that the parsing and action handling systems
produce the correct result. Ref tests may be recorded by running
alacritty with `--ref-test` and closing the terminal by using the window
"X" button. At that point, the recording is fully written to disk, and a
serialization of important state is recorded. Those files should be
moved to an appropriate folder in the `tests/ref/` tree, and the
`ref_test!` macro invocation should be updated accordingly.
A couple of changes were necessary to make this work:
* Ref tests shouldn't create a pty; the pty was refactored out of the
`Term` type.
* Repeatable lines/cols were needed; on startup, the terminal is resized
* by default to 80x24 though that may be changed by passing
`--dimensions w h`.
* Calculating window size based on desired rows/columns and font metrics
required making load_font callable multiple times.
* Refactor types into library crate so they may be imported in an
integration test.
* A whole bunch of types needed symmetric serialization and
deserialization. Mostly this was just adding derives, but the custom
deserialization of Rgb had to change to a deserialize_with function.
This initially adds one ref test as a sanity check, and more will be
added in subsequent commits. This initial ref tests just starts the
terminal and runs `ll`.
The FreeType font Rasterizer API is updated to match the CoreText
Rasterizer. This enabled bold/italic fonts since the rest of the
codebase has already been updated.
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 adds a bunch of APIs to CGContext (and supporting types) that
aren't actually necessary to turn on subpixel rendering. The key for
subpixel rendering were the options passed to bitmap_context_create().
Specifically, kCGImageAlphaPremultipliedFirst | kCGBitmapByteOrder32Host
are necessary to enable it.
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!.
