The fonts we embed in Zed binary (Zed Sans & Zed Mono) weigh about 30Mb in total and we are cloning them several times during startup and loading of embedded assets (once explicitly in Zed and then under the hood in font-kit). Moreover, after loading we have at least 2 copies of each font in our program; one in .rdata and the other on the heap for use by font-kit.
This commit does away with that distinction (we're no longer allocating the font data) and slightly relaxes the interface of `TextSystem::add_fonts` by expecting one to pass `Cow<[u8]>` instead of `Arc<Vec<u8>>`. Additionally, `AssetSource::get` now returns `Cow<'static, [u8]>` instead of `Cow<'self, [u8]>`; all existing implementations conform with that change.
Note that this optimization takes effect only in Release builds, as the library we use for asset embedding - rust-embed - embeds the assets only in Release mode and in Dev builds it simply loads data from disk. Thus it returns `Cow<[u8]>` in it's interface. Therefore, we still copy that memory around in Dev builds, but that's not really an issue.
This patch makes no assumptions about the build profile we're running under, that's just an intrinsic property of rust-embed.
Tl;dr: this should shave off about 30Mb of memory usage and a fair chunk (~30ms) of startup time.
Release Notes:
- Improved startup time and memory usage.
Something is stomping over the timestamp of generated scene.h right
after the build script finishes:
`Dirty gpui v0.1.0 (/Users/someonetoignore/work/zed/zed/crates/gpui):
the file target/debug/build/gpui-ca04eedfe8d0e13c/out/scene.h has
changed (1705922004.637000680s, 1s after last build at
1705922003.507431315s)`
^ That's an output of `-v` flag added to either `cargo run` or `cargo
build`.
This comes up when you do `cargo build` followed by `cargo run` or
another `cargo build`; the artifact is not getting reused, even though
it should be possible?
Release Notes:
- N/A
Was moved by mistake in: ebbe52e6b0
This was throwing errors everywhere on Linux because anyhow is needed by
the code and is not macOS dependent.
Release Notes:
- N/A
Simplify Zed's collaboration system by:
- Only allowing member management on root channels.
- Disallowing moving sub-channels between different roots.
- Disallowing public channels nested under private channels.
This should make the mental model easier to understand, and makes it
clearer
who has what access. It is also significantly simpler to implement, and
so
hopefully more performant and less buggy.
Still TODO:
- [x] Update collab_ui to match.
- [x] Fix channel buffer tests.
Release Notes:
- Simplified channel membership management.
One of the complaints of users on our first Hack call was that the error
messages you got when channel joining failed were not great.
This aims to fix that specific case, and lay the groundwork for future
improvements.
It adds two new methods to anyhow::Error
* `.error_code()` which returns a value from zed.proto (or
ErrorCode::Internal if the error has no specific tag)
* `.error_tag("key")` which returns the value of the tag (or None).
To construct errors with these fields set, you can use a builder API
based on the ErrorCode type:
* `Err(ErrorCode::Forbidden.anyhow())`
* `Err(ErrorCode::Forbidden.message("cannot join channel").into())` - to
add any context you want in the logs
* `Err(ErrorCode::WrongReleaseChannel.tag("required", "stable").into())`
- to add structured metadata to help the client handle the error better.
Release Notes:
- Improved error messaging when channel joining fails.
Sometimes Cocoa calls app delegate methods (notably the display link)
while we're calling Cocoa methods. This causes a deadlock unless we
are careful to run cocao methods while we're not holding our internal
locks
Since our last measurements, startup time in Zed had regressed quite
significantly. This was due to several issues:
- We were loading IBMPlex, which we're not really using in the UI.
- Images were being parsed in the foreground, thus blocking the main
thread
- Language models (for semantic index and assistant) were being loaded
in the foreground, thus blocking the main thread
- Interaction with the keychain was blocking the main thread
In addition to these, with this pull request we will now deserialize the
items for a pane in parallel, as opposed to doing so sequentially.
All combined, when running the zed binary directly this brings startup
time from ~350ms to ~210ms on my machine.
Release Notes:
- Improved startup time.
We recently doubled the size of the `ELEMENT_ARENA` after someone ran
into a panic due to the arena running out of space.
This adds some logging so that we can hopefully develop a better
understanding of when the element area's allocation is elevated.
Release Notes:
- N/A
This fixes the performance problem we saw in
https://github.com/zed-industries/community/issues/2405.
In a trace we could see that if a font is missing we'd constantly look
it up and never cache that it's missing.
This changes that and does cache the font-is-missing result.
Drawback is that one would need to restart Zed after installing a
missing font that was configured in settings. That seems acceptable for
now, though.
Release Notes:
- Improved rendering performance when configured fonts are missing on
the system.
We recently doubled the size of the `ELEMENT_ARENA` after someone ran
into a panic due to the arena running out of space.
This adds some logging so that we can hopefully develop a better
understanding of when the element area's allocation is elevated.
Co-authored-by: Antonio <antonio@zed.dev>
This fixes the performance problem we saw in https://github.com/zed-industries/community/issues/2405.
In a trace we could see that if a font is missing we'd constantly look
it up and never cache that it's missing.
This changes that and does cache the font-is-missing result.
Drawback is that one would need to restart Zed after installing a
missing font that was configured in settings. That seems acceptable for
now, though.
Co-authored-by: Antonio <antonio@zed.dev>
This fixes a panic we just saw on nightly. It's the first time we
see this panic, so I think it's pretty safe to assume having twice as
much capacity will fix it.
This PR does some cleanup of the `gpui` docs, making sure we're linking
things appropriately and following good Rust documentation style.
Release Notes:
- N/A
This PR adds a barebones README to the `gpui` crate, just so folks have
somewhere to land on when we link them here from
[gpui.rs](https://www.gpui.rs/).
We can flesh this out as we go.
Release Notes:
- N/A
This isn't my favorite idea of a fix, but it does work for now, and it
seems likely the terminal will need to configure other aspects of action
dispatch in the future.
In the future we should explore making it possible to do this via the
keymap, either by making disabling bindings more robust; or by having a
way to indicate immediate mode per binding.
Release Notes:
- Fixed a bug where cmd-k in terminal took 1s
This isn't my favorite idea of a fix, but it does work for now, and it
seems likely the terminal will need to configure other aspects of action
dispatch in the future.
In the future we should explore making it possible to do this via the
keymap, either by making disabling bindings more robust; or by having a
way to indicate immediate mode per binding.
This optimizes rendering time by saving computation of the layer_id and
comparison when inserting it into the `BTreeMaps`.
Co-authored-by: Antonio <antonio@zed.dev>
I was unable to run the collab tests locally because I would run out of
file descriptors.
From some digging it turned out that tokio allocates a new file
descriptor to do work on the CurrentThread using KQUEUE.
We create a new tokio Runtime with each database connection, and these
database connections were being retained by the Client, which is
retained by the Context.
Cleaning up our leaked contexts (and an unrelated retain cycle in the
UserStore) fixes the problem (though does make me
wonder if a different approach might be preferrable).
Release Notes:
- N/A
I was unable to run the collab tests locally because I would run out of
file descriptors.
From some digging it turned out that tokio allocates a new file
descriptor to do work on the CurrentThread using KQUEUE.
We create a new tokio Runtime with each database connection, and these
database connections were being retained by the Client, which is
retained by the Context.
Cleaning up our leaked contexts fixes the problem (though does make me
wonder if a different approach might be preferrable).
