The text that is printed is generated when building assets, by analyzing LICENSE
and NOTICE files that comes with syntaxes and themes.
We take this opportunity to also add a NOTICE file as defined by Apache License 2.0.
We started to use lazycell because syntect already used it. But syntect has
changed to use once_cell. So we should also do that to prepare for using the
upcoming version of syntect.
By forwarding the task to find the `Plain Text` syntax to `assets`. Not only does
the code become simpler; we also get rid of a call to `self.get_syntax_set()`
which is beneficial to the long term goal of replacing `syntaxes.bin` with
`minimal_syntaxes.bin`.
Note that the use of `.expect()` is not a regression in error handling. It was
previously hidden in `.find_syntax_plain_text()`.
This significantly speeds up the startup time of bat, since only a single
linked SyntaxDefinition is loaded for each file. The size increase of the
binary is just ~400 kB.
In order for startup time to be improved, the --language arg must be used, and
it must match one of the following names:
"Plain Text", "ActionScript", "AppleScript", "Batch File", "NAnt Build File",
"C#", "C", "CSS", "D", "Diff", "Erlang", "Go", "Haskell", "JSON", "Java
Properties", "BibTeX", "LaTeX Log", "TeX", "Lisp", "Lua", "MATLAB", "Pascal",
"R", "Regular Expression", "Rust", "SQL", "Scala", "Tcl", "XML", "YAML", "Apache
Conf", "ARM Assembly", "Assembly (x86_64)", "CMakeCache", "Comma Separated
Values", "Cabal", "CoffeeScript", "CpuInfo", "Dart Analysis Output", "Dart",
"Dockerfile", "DotENV", "F#", "Friendly Interactive Shell (fish)", "Fortran
(Fixed Form)", "Fortran (Modern)", "Fortran Namelist", "fstab", "GLSL",
"GraphQL", "Groff/troff", "group", "hosts", "INI", "Jinja2", "jsonnet",
"Kotlin", "Less", "LLVM", "Lean", "MemInfo", "Nim", "Ninja", "Nix", "passwd",
"PowerShell", "Protocol Buffer (TEXT)", "Puppet", "Rego", "resolv", "Robot
Framework", "SML", "Strace", "Stylus", "Solidity", "Vyper", "Swift",
"SystemVerilog", "TOML", "Terraform", "TypeScript", "TypeScriptReact",
"Verilog", "VimL", "Zig", "gnuplot", "log", "requirements.txt", "Highlight
non-printables", "Private Key", "varlink"
Later commits will improve startup time for more code paths.
* fix some typos and misspellings
* CHANGELOG.md: Add Performance section (preliminary)
* Add a CHANGELOG.md entry for this PR
This will be needed to later support zero-copy deserialization of independent
syntax sets, but is interesting and useful on its own.
Instead of deferring serialization and deserialization to syntect, we implement it
ourselves in the same way, but make compression optional.
We can't use #[from] on Error::Msg(String) because String does not implement Error.
(Which it shouldn't; see e.g. https://internals.rust-lang.org/t/impl-error-for-string/8881.)
So we implement From manually for Error::Msg, since our current code was written
in that way for error-chain.
Move code to build assets to its own file. That results in better modularity and flexibility.
It also allows us to simplify HighlightingAssets a lot, since it will now always
be initialized with a SerializedSyntaxSet.
To improve startup performance, we will later load smaller `SyntaxSet`s instead
of one giant one. However, the current API assumes only one `SyntaxSet` is ever used,
and that that implicitly is the `SyntaxSet` from which returned `SyntaxReference`s
comes.
This change changes the API to reflect that `SyntaxSet` and `SyntaxReference`
are tightly coupled, and enables the use of several `SyntaxSet`.
Instead of 100 ms - 50 ms, startup takes 10 ms - 5 ms.
HighlightingAssets::get_syntax_set() is never called when e.g. piping the bat
output to a file (see Config::loop_through), so by loading the SyntaxSet only
when needed, we radically improve startup time when it is not needed.
They are just a way to get access to data embedded in the binary, so they don't
conceptually belong inside HighlightingAssets.
This has the nice side effect of getting HighlightingAssets::from_cache() and
::from_binary(), that are highly related, next to each other.
Or rather, introduce new versions of these methods and deprecate the old ones.
This is preparation to enable robust and user-friendly support for lazy-loading.
With lazy-loading, we don't know if the SyntaxSet is valid until after we try to
use it, so wherever we try to use it, we need to return a Result. See discussion
about panics in #1747.
Using BufReader makes sense for large files, but assets are never large enough
to require buffering. It is significantly faster to load the file contents in
one go, so let's do that instead.
Closes#1753
It already now reduces code duplication slightly, but will become even more
useful in the future when we add more complicated logic such as lazy-loading.
Since we only modify `pub(crate)` items, the stable bat-as-a-library API is not
affected.
This takes us one step closer to making SyntaxSet lazy-loaded, which in turn
takes us one step closer to solving #951.
Martin Nordholts
Bojan Đurđević
cbolgiano
Marcin Puc
Ville Skyttä
Keith Hall
Keith Hall
David Peter
David Peter