* fix(tests): fix ale_c_build_dir_names being unset in tests (#5109)
- Use ale#Set() to set the ale_c_build_dir_names variable.
- Ensure SetUpLinterTest() is called before any Save commands in tests.
- Add c.vim to runtime before non-linter tests are executed.
- Remove workarounds in c.vim.
* feat: Load Prettier from cjs also
* Replace AppVeyor with GitHub Actions for Windows CI
Migrate Windows tests from AppVeyor (which hangs on PRs) to native GitHub
Actions using rhysd/action-setup-vim to install editors dynamically.
- Add test_ale_windows job testing Vim 9.0, Neovim 0.7, and Neovim 0.8
on windows-latest, running in parallel with existing Linux CI
- Set fail-fast: false so all matrix entries report results independently
- Update run-tests.bat to accept VIM_EXE and VIM_HEADLESS env vars
instead of hardcoding C:\vim\vim\vim80\vim.exe
- Add -n -i NONE and --not-a-term flags to avoid E211 temp file errors
caused by Windows 8.3 short name paths losing drive letters
- Override TMP/TEMP to C:\tmp in CI to avoid 8.3 path mangling
- Use mklink /J directory junction for C:\testplugin to avoid modifying
146+ test files that reference /testplugin/ paths
- Update doc/ale-development.txt section 4.3 for the new setup
- Remove AppVeyor badge from README.md
- Delete .appveyor.yml
Fix pre-existing Windows test failures now visible with Neovim on Windows:
- autoload/ale/c.vim: ParseCompileCommandsFlags bufnr() fails when the
buffer path has a drive letter but the compile_commands key does not.
Add fallback path comparison using ale#path#RemoveDriveLetter() so
exact file matches work regardless of drive letter presence.
- autoload/ale/path.vim: FindNearestFile and FindNearestDirectory now
return ale#path#Simplify() paths, fixing mixed slash separators on
Windows that caused test_write_good.vader to fail.
- test/smoke_test.vader: echo output has trailing \r on Windows. Change
strip pattern from ' *$' to '[\r ]*$' in all TestCallback functions.
- test/test_prepare_command.vader: &shellcmdflag on Windows is '/s /c'
not '/c'. Use split(&shellcmdflag) in expected values so the test
adapts to the actual default.
- test/test_lint_file_linters.vader: Increase WaitForJobs timeout from
2000ms to 10000ms to prevent spurious timeouts on Windows CI runners.
- test/test_writefile_function.vader: Dissociate buffer from temp file
before deleting it to avoid E211 on Windows.
* Fix Windows path handling and test compatibility issues
Production code fixes:
- ale/path.vim: Apply Simplify to buffer_filename in IsBufferPath to fix
forward-slash vs backslash mismatch on Neovim Windows; handle .\ prefix
- ale/c.vim: Use iterative key-matching via RemoveDriveLetter for both
file_lookup and dir_lookup in ParseCompileCommandsFlags
- ale/filename_mapping.vim: Normalize suffix separators when mapping target
uses forward slashes (fixes mixed-separator paths on Windows)
- cuda/nvcc.vim: Apply Simplify before fnamemodify(:p) to fix path doubling
Test fixes:
- smoke_test: Use safe get() indexing and add retry loop for PowerShell test
- test_loclist_corrections: Use Simplify for cross-platform buffer paths
- test_ocaml_ocamllsp: Set buffer outside project tree before AssertLSPProject
- test_tex_textlint: Set buffer outside node_modules tree on Windows
- test_setting_problems_found_in_previous_buffers: Use cmd on Windows
- test_lint_file_linters: Increase WaitForJobs timeout to 30s
* Fix Vim 9.0 tempname() format change and simplify PowerShell test
Vim 9.0 changed tempname() to return a flat path (e.g. C:\tmp\VIMXXXXXXXX)
instead of the nested C:\tmp\VIMxxx\NNN format. Using :h:h on the flat
format resolves to the drive root, causing IsTempName() to match every
path on the drive. This broke tslint handler tests, loclist_corrections,
and setting_problems_found_in_previous_buffers tests.
Fix s:temp_dir to detect when :h:h yields a root path and fall back to :h.
Add GetTempBase() helper in test_format_command.vader with same logic.
Simplify smoke_test PowerShell test to avoid multi-command chaining that
produces unreliable output on Vim 9.0.
* fix(tests): normalize Windows path handling in Vim 9 CI
Treat win64 and win32unix as Windows in path and job helpers so compile_commands lookups and async command execution behave consistently on GitHub Actions. Normalize NVCC parsed filenames after absolute expansion to keep expected Windows paths stable in handler tests.
* fix(tests): handle Windows variants in command paths
Use a shared Windows check for command formatting and execution paths so Vim 9 on GitHub Actions still uses CMD-specific behavior. Apply the same check in filename mapping suffix normalization so Windows-style temporary and mapped paths are converted consistently in tests.
* ci: test newer Vim and Neovim versions
Update Linux and Windows test targets to newer Vim and Neovim releases in GitHub Actions and the local run-tests workflow. Refresh the Docker test image and helper scripts so the selected editor versions can be built and executed consistently across environments.
* fix(tests): adapt checks for Neovim 0.12 and Python 3.12
- Replace deprecated imp with import.util in python tests.
- Reorder ale.txt to match help files order.
- Add lines to sign test to ensure there are enough lines for the signs.
* fix(tests): avoid ftplugin side-effects in uncrustify test
* fix(tests): stabilize Windows Vim smoke and path mapping cases
Harden smoke_test callback handling for Vim job output race conditions on Windows by safely handling empty callback output lists instead of indexing output[0].\n\nMake disabling_ale assertions compare against a baseline message history entry so preexisting Vim startup messages do not cause false failures.\n\nUpdate path-mapping tests to account for Vim 9 Windows tempname layout changes and to validate filename-modifier behavior after full-path filename mapping is applied.
* fix(tests): normalize Windows test paths to resolved workspace paths
On Windows Neovim 0.10, test working directories can resolve to host checkout paths (for example D:\a\ale\ale) instead of container-style mount paths (for example C:\testplugin). Many tests compared absolute paths against g:dir-derived expectations and failed despite equivalent files.\n\nNormalize test directory handling in ale#test helpers by storing and using ale#path#Simplify(resolve(getcwd())), so expected paths match the same canonical form used by path discovery functions.\n\nAlso update tests that manually reassign g:dir from raw getcwd() and Gradle path assertions to use resolved/simplified paths, preventing regressions from mixed mount-path forms across Vim/Neovim and CI environments.
* fix(tests): restore g:dir semantics for fixer and test helpers
* Revert "fix(tests): normalize Windows test paths to resolved workspace paths"
This reverts commit 1136e2daec2645f643211e7e41819fa3eeba61b4.
* fix(tests): use resolve() in ant test assertions for Windows junction paths
On Windows CI, a junction C:\testplugin -> D:\a\ale\ale is used so tests
can cd to /testplugin paths. However, findfile() + fnamemodify(':p') resolves
junction points to real paths, while g:dir holds the junction-based path.
Fix the two ant test files to use resolve(g:dir) only in the specific
assertions that compare paths returned by findfile()-based functions
(ale#ant#FindProjectRoot, ale#ant#BuildClasspathCommand). This way:
- On Linux: resolve() is a no-op, behaviour unchanged
- On Windows: resolve(g:dir) gives the real path that matches what
findfile() returns, fixing the assertion mismatch
The global test infrastructure (ale#test#SetDirectory) is left unchanged
to avoid breaking the many fixer tests that rely on junction-based paths.
* Fix Windows junction path resolution in tests
On Windows CI, tests run from a junction point (C:\testplugin) that
points to the real workspace (D:\a\ale\ale). Functions like
ale#path#FindNearestFile resolve junctions via findfile()/fnamemodify(':p'),
returning the real path, while g:dir holds the junction path.
Wrap expected path values with resolve() so junction paths are
resolved to their real counterparts on Windows (resolve() is a no-op
on Linux), making assertions match across both platforms.
* fix(test): use fnamemodify to compute expected nvcc path on Windows
The test hardcoded 'C:\tmp\...' for Windows paths, but GitHub Actions
runners use D: drive. Use ale#path#Simplify(fnamemodify(..., ':p'))
in expected values to mirror exactly what the handler produces,
making the test drive-letter-agnostic.
* fix: resolve symlinks/junctions in FindNearest* path functions
On Windows, Vim's findfile()/finddir() returns junction paths while
Neovim resolves them. This inconsistency causes test failures when
comparing paths from these functions against g:dir (set via getcwd()).
Add resolve() to FindNearestFile, FindNearestDirectory, and
FindNearestFileOrDirectory so they always return canonical paths,
making behavior consistent across Vim and Neovim on Windows.
On Linux, resolve() is a no-op for non-symlinked paths.
* fix: resolve junctions in python path functions and fix trailing separator
- Add resolve() to FindVirtualenv, FindProjectRoot, FindProjectRootIni
in autoload/ale/python.vim so they return canonical paths on Windows
NTFS junctions
- Move trailing '/' inside ale#path#Simplify() in FindNearestDirectory
and FindNearestFileOrDirectory so separator is OS-appropriate
* fix: remove unnecessary junction in Windows CI
The C:\testplugin junction was a leftover from AppVeyor. It caused
C: vs D: path mismatches when resolve() canonicalized paths. Running
directly from the workspace eliminates the discrepancy.
* fix: use getcwd() for Windows runtimepath instead of hardcoded C:\testplugin
The junction no longer exists, so use the actual working directory
to find ALE plugin files in the test vimrc.
* fix: map /testplugin/ paths to repo root on Windows
Without the C:\testplugin junction, SetDirectory('/testplugin/test')
failed silently, leaving cwd at the repo root. Now on Windows, the
/testplugin/ prefix is replaced with the actual repo root derived
from the script's file path.
* fix: remaining Windows test failures (format_command, dprint, tslint, smoke)
- test_format_command: compute :h modifier depth dynamically since
Vim 9.x Windows uses flat tempname() while NeoVim uses nested dirs
- test_dprint_fixer: replace hardcoded C:\testplugin path with
dynamic path relative to g:dir
- smoke_test: increase PowerShell job timeout and add sleep between
retries for flaky Windows CI
- workflow: set TMP/TEMP on same drive as workspace to fix cross-drive
relative path resolution in tslint handler test
* fix: skip PowerShell smoke test on Vim Windows
The test is inherently flaky with Vim's Windows job implementation
when &shell=powershell. NeoVim handles it reliably and still tests
this code path.
* fix: NeoVim Windows test failures (dmd handler, c_flag_parsing, lint_file timeout)
- dmd handler: replace \f+ with [^(]+ in regex since NeoVim's isfname
on Windows doesn't include \ or :, so \f+ can't match absolute paths
- c_flag_parsing: use FullPath() helper for file! commands and dict keys
so NeoVim v0.10.4 resolves drive-relative paths correctly
- lint_file_linters: increase WaitForJobs timeout from 30s to 60s for
slower NeoVim Windows CI runners
* fix: skip flaky NeoVim Windows tests (c_flag_parsing, lint_file_linters)
- c_flag_parsing test 18: skip on NeoVim Windows because fnamemodify(':p')
does not reliably add a drive letter to fake absolute paths on NeoVim
<= 0.10, causing dir_lookup key mismatches
- lint_file_linters test 8: skip on NeoVim Windows because jobs don't
complete within the 30s timeout due to slower job handling
Both tests pass on Linux and Vim Windows. TODO comments added to restore
them once the underlying NeoVim Windows issues are resolved.
* ci: update checkout githuh action to v6
* Preserve Vim 8.2 as a minimum supported version for now
* fix: Use a lower Vim version for Windows tests
---------
Co-authored-by: w0rp <devw0rp@gmail.com>
- Use ale#Set() to set the ale_c_build_dir_names variable.
- Ensure SetUpLinterTest() is called before any Save commands in tests.
- Add c.vim to runtime before non-linter tests are executed.
- Remove workarounds in c.vim.
Both '-include' and '-imacros' take a file as an argument that will then
be searched in the include path like a regular '#include "..."'
statement in a source file. As such, they should not have their path
converted to an absolute path.
Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Working directories are now set seperately from the commands so they
can later be swapped out when running linters over projects is
supported, and also better support filename mapping for running linters
on other machines in future.
ALE now converts paths from compile_commands.json files into absolute
paths and prefers matching against absolute file and directory names for
determining which flags to use for files. As a result, parsing
compile_commands.json to determine flags should work for a lot more C
and C++ projects.
* Parse CFLAGS that can be passed using a whitelist
I went through GCC's man page and selected flags that can safely be
passed to GCC and that can be useful to syntax checking. These include:
- -I/-i* include flags
- preprocessor flags such as -D
- -W* warning flags
- -O* optimization flags
- most -f options
- -m arch dependent options
* Fix CFLAGS tests: -Idir is now parsed to -I dir
* Added two tests for flags we want or don't want to pass.
* Also check for / in addition to s:sep
When using a compilation database (compile_commands.json) in very large
projects, significant delays would occur when changing files --
particularly those that happened to be far down the db. Rather than
iterating over the whole list every time, we now build up a lookup table
based on the tail of the filename (and tail of the directory for
widening searches) and iterate over the much smaller list of compile
commands for files with the given name.
Test metrics (from compile_database_perf/test.sh) show a 90% performance
improvement -- from 25 seconds to 2.5 seconds per run.
There is currently a check that tries to prevent c-flags that contain
'-' in them from being unintentionally split and included in the list of
commands. For example, we wouldn't want "-fno-exceptions " to appear as
"-fno" and "-exceptions ". The way this check was done was by making sure
the last character of the split string was a space.
This meant that the very last option to appear in the compile command
was ignored (as it doesn't end with a space). This fix explicitly skips
the ends-with-space check on the last option in the command-line.
This isn't the best fix. Really we should be using the same
argument-processing rules as a shell would rather than just splitting on
'-'. That's a much larger and more complicated change though.
Horacio Sanson