Tested using cross-compiling from linux:
zig build -Dcross=true -Dtarget=x86_64-windows nvim_bin
Note: not fully functional without a runtime, which still has to be
fuddled with manually
Macos and windows builds require a recent zig 0.15+dev version
As this zig master branch is currently too much in flux, we can't make
our CI depend on zig master.
Revisit CI after zig 0.15 release or at least feature freeze.
Problem:
We temporarily disabled linux arm ci because of stability issues with
the runner. #32339 Since then, the hardware was changed, so we can try
re-enabling ARM linux CI. https://github.com/actions/partner-runner-images/issues/47#issuecomment-2678170225
Solution:
- re-enable arm linux ci. reverts 8e4b77134a
- also use arm image for these jobs, where arm seems to run much faster:
- `lint` (step: `clang-tidy`)
- master: 1m5s
- this pr (linux ARM): 37s
- `clang-analyzer` (step: `cmake --build ...`)
- master: 10m
- this pr (linux ARM) 5m 55s
- `with-external-deps` (step: `Build`)
- master: 26s
- this pr (linux ARM): 21s
NEW BUILD SYSTEM!
This is a MVP implementation which supports building the "nvim" binary,
including cross-compilation for some targets.
As an example, you can build a aarch64-macos binary from
an x86-64-linux-gnu host, or vice versa
Add CI target for build.zig currently for functionaltests on linux
x86_64 only
Follow up items:
- praxis for version and dependency bumping
- windows 💀
- full integration of libintl and gettext (or a desicion not to)
- update help and API metadata files
- installation into a $PREFIX
- more tests and linters
Enable translations in the GCC release job, as it's the fastest.
Although there aren't many tests for translation behavior, this still
allows ensuring that Nvim isn't broken when translations are enabled.
This automatically downloads and uses the correct luals binary for the
currently used system. `make luals` will run luals on all lua files in
`runtime`.
We download lua-language-server manually instead of relying on
contributors downloading it on their own (like with stylua) as
lua-language-server is updated frequently which may cause unnecessary
friction. Therefore, we download a pinned version of luals which we then
can manually bump when needed. This can be re-evaluated if luals becomes
more stable in the future.
Currently this is not run when using `make lint` since cmake style "file
caching" doesn't seem possible at the moment. This is because checking a
single file doesn't seem to work.
Work on https://github.com/neovim/neovim/issues/24563.
Problem: Some CI jobs are redundant: `RelWithDebInfo` is already tested
on Linux-Arm64; `MinSizeRel` and Ninja Multi Config are not sufficiently
relevant in practice to spend CI cycles on.
Solution: Remove `build-types` job.
Tests should not rely on being run inside a git clone, so the Linux
builds cover this use case. The macOS builds will continue running with
the .git directory so there's still unix-ish coverage within a git
clone.
This ensures that no tests fail due to differences between release and debug builds.
The release build-type check is now unnecessary, too, so remove it.
Problem: Adding support for modern Nvim features (reflow, OSC 8, full
utf8/emoji support) requires coupling libvterm to Nvim internals
(e.g., utf8proc).
Solution: Vendor libvterm at v0.3.3.
Auto-merging is a useful feature by github, but it requires required
checks which requires a few adjustments. The primary change is that
required checks can't use `paths` or `paths-ignore` as that risks not
running the job, and required checks must always be run.
A workaround for this is to introduce a dummy workflow which is used for
every path not used in the real workflow. That way the required job is
"always" run as far as github is concerned. The workaround is unweildly
so it's only useful to do it for costly workflows where the potential
benefits are big. If not it's better to simply remove any `paths` or
`paths-ignore` from a workflow instead.
Add more filters so that LuaJIT can parse headers on macOS 14.
The system headers use a style of enum introduced in C++11 (and allowed
as an extension in C by clang) of the form:
enum Name : Type {
The system headers also use bitfields in the mach_vm_range_recipe* types:
struct Foo { int bar : 32; }
Neither of these constructs can be parsed by LuaJIT, so filter the lines
out. Neither of these declarations are used by neovim's unittests.
There is a (now closed) issue about bitfields for LuaJIT:
https://github.com/LuaJIT/LuaJIT/issues/951Fixes#26145.
Python 3.12+ throws an error if you try to install a package in an
externally managed environment. Using `--break-system-packages` is not
recommended for personal use, but for CI it should be fine and is
probably the most straightforward solution.
`github.ref` is now defined for both pull requests and pushes, meaning
that it can be used to simplify the concurrency group.
`cancel-in-progress` is set to true only if the trigger is a pull
request, as we don't want master runs to cancel each other out.
Problem: No test coverage on ARM.
Solution: Add `macos-14` tests, which now run on M1. Skip unit tests as these don't work on M1, see #26145. Also test universal build on M1.
Note: `macos-14` will be `macos-latest` in Q2 2024, so we'll want to switch these to keep Intel and unittest coverage on macos (while GH still offers Intel runners).
Explicitly set the build type for both deps and Nvim. They are already
explicitly set on Windows to RelWithDebInfo. Now also explicitly set
them to Debug on POSIX.
Python3 provider tests suddenly became extremely flaky on macos for
unknown reasons. For some reason, installing python with the
setup-python action over using the default python fixes the flakiness.
Use this workaround for the time being to unblock CI while we figure out
the root cause.
- silence false warnings on MSVC
- merge `clang-tidy` cmake target into `lintc` and remove the
corresponding make target
- use cmake's built-in endianness detection
This will run the three test suites (unit, functional and old) in
parallel, meaning that neovim is built for each test and run separately.
This has a slight increase in total CI usage, but it allows rerunning
only the specific test suite that failed for flaky tests, which will
save some time.
Ideally we'd remove any drawbacks by building neovim once and reusing it
for each test suite, but that is not currently possible due to poor
upload/download speeds of the upload-artifact and download-artifact
actions. This has been addressed in
https://github.com/actions/toolkit/pull/1488, but will only be made
available in upload-artifact@v4 and download-artifact@v4.
This adds the checks in https://neovim.io/doc/reports/clang/ when using
clang-tidy. The strategy is to enable all clang-analyzer checks, and
disable only the checks for the warnings that exist currently. This
allows us to eliminate each warning type without blocking ongoing work,
but also without adding bugs for already eliminated warnings.
The plan is to eventually eliminate https://neovim.io/doc/reports/clang/
by completely integrating it into the clang-tidy check.
Also add make and cmake targets `clang-analyzer` to run this check.
This reverts commit e71c7898ca.
Triggering jobs on users own fork turned out to be not that useful, and
only necessary in rare moments. It's easier to adjust the CI scripts if
the users wants CI results before creating a pull request. It also
reduces the complexity of the CI code.
This partially reverts commit 7d0479c558.
The job has been particularly unstable when used with docker on
cirrus-ci, which is especially bad as it's meant to be a non-flaky and
simple test.
This will abort if lint programs are not found, and is meant primarily
for the lint job in CI. Supersedes the REQUIRED argument in
add_glob_target as it's a superior replacement by being a built-in
solution.
