Problem:
Undo "u" is wrong after cascading an insert that adds a line (`A{<CR>`, `o`, …).
The cascade replays the insert as a per-cursor `edit()`, where `u_save`
stores a line multiple times into the undo block at different line
counts; then `getbot` computes conflicting entries that undo each other.
Solution:
`u_save` the whole affected line-range as one entry at session start,
the same approach used by Vim for linewise ops (`op_shift`, `ex_sort`).
The "enclosing" `u_save` wins and the invalid entries are overridden.
Only join paths for matches and entries needed for traversal. This
avoids string processing for rejected files and improves selective
searches across the Neovim tree by about 19-25%.
AI-assisted
Problem: Generating helptags builds and queries a full vimdoc syntax
tree. This is much slower than the previous scanner and also happens
before discovering that the tags file cannot be written.
Solution: Extract tags with an LPeg grammar, reading each file once and
skipping example blocks and ordinary text. Open the output first so
unwritable directories can return immediately.
This restores legacy handling of tags inside inline code spans. On an
arm64 Mac, the 5.36 MB runtime corpus takes 24 ms instead of 604 ms,
with identical generated tags.
AI-assisted
vim-patch:015c84ce5 runtime(doc): add missing usr_52 entry to toc
vim-patch:59bd74ed4 translation: include vim.pot in the repository
vim-patch:5d552d652 translation: ignore vim.pot creation date, regenerate it, rm allfiles
vim-patch:3c8d32e4f translation: comment out deletion of *.pot file in Makefiles
vim-patch:238b0b058 tests(commondumps): Mark and fold lines in screendump views
vim-patch:cb65fe17e tests(commondumps): Make mark-line-related optimisations
vim-patch:e543abbf8 tests(commondumps): Use character counts when marking columns
vim-patch:e17d10c2d CI: Bump github/codeql-action
vim-patch:9.2.1039: MS-Windows: Unix domain socket channels fail
vim-patch:9.2.1050: tests: test_suspend() fails on Solaris
vim-patch:9.2.1053: libvterm: characters can get a different width in a terminal window
vim-patch:46ff48ba77bbf167a646f84053e96342357446b3
vim-patch:9.2.1043: matchlist() allocates empty strings for unmatched submatches
vim-patch:9.2.1044: Vim script execution is slower than necessary
vim-patch:9.2.1051: getdigits() overflow behaviour is not portable
vim-patch:9.2.1061: computing the displayed width of a line is slower than necessary
Problem:
InsertCharPre is triggered for stuffed text (".", i_CTRL-R, i_CTRL-A).
But that text was already transformed when it was typed: the `v:char`
result is appended to the redo buffer literally (`redo_append_lit()`).
So "." transforms it again, e.g. an autopair handler turns "()" into
"())".
Solution:
Skip InsertCharPre for stuffed text.
Precedent: `vgetorpeek()` disables abbreviations for stuffed text for
the same reason (it is post-expansion).
Macros are unaffected.
Problem:
Resetting w_scbind_pos to the displayed position discards pending
scroll distance. This breaks scrollbind with asymmetric folds and
leaves persistent offsets after tall virtual-line blocks. Wrapped
lines with smoothscroll also show increased follower lag.
Solution:
Revert the original change and its test. This restores the previous
behavior, including the original tall-virt_lines oscillation, while
a replacement is reviewed separately.
This reverts commit b781e95a24.
AI-assisted
Enable checks for unused code, deprecated calls, return types, and
annotations.
Fix the warnings and replace the deprecated highlight function. Keep
exceptions where old APIs are still needed or the checker gets it wrong.
AI-assisted
Problem: Autocmd callbacks currently have to rely on curwin, which may
already have changed by the time they run. This is especially noticeable
for WinNew, WinClosed, WinResized and WinScrolled.
Solution: Track the window in the autocmd context and expose it as ev.win.
Pass it explicitly for the window events that need it.
Problem:
When a `:!` command runs, file timestamps are checked and W12 or
`FileChangedShell` is triggered. But not when a `:terminal` job exits.
Only `enter_buffer()` checks, so whether you get the W12 warning depends
on how you happen to return to the file: `:term` works, `:tabnew | term`
does not.
echo foo > testfile
nvim --clean testfile
" Edit the buffer, don't save it.
:!mv testfile testfile2
:tabnew | term sh
touch testfile2 && mv testfile2 testfile
exit
'autoread' is defeated because `:!mv` deletes the file, so
`should_watch()` fails and the watcher stops. `ensure_watcher()` only
re-runs on `BufReadPost`/`BufWritePost`/`OptionSet`, so nothing watches
the file when the terminal moves it back.
Solution:
Set `need_check_timestamps` on terminal-job exit, same as `:!cmd`
(`os_shell`).
Problem:
When vim.lsp.rpc.connect is used for an LSP and the connection fails
(due to ECONNREFUSED, for example), the LSP client is left
uninitialized. Subsequent calls to vim.lsp.start will reuse the
uninitialized client, but no new connection will be attempted.
Solution:
When vim.lsp.rpc.connect encounters a connection error, terminate the
client. This allows the user to do something in response to the failed
connection in the on_exit callback, and also makes it so subsequent
attempts to start the LSP will try to establish a new connection rather
than using the previous uninitialized client.
Co-authored-by: Jason Stenftenagel <jasonstenftenagel@gmail.com>
Problem:
In follow-mode, a motion that fails at the primary cursor (e.g. "j" at
EOB) still cascades to other cursors, so they move while the primary
does not, which causes them to be deduplicated...
Also affects Visual sessions: "Vjd" at EOB deletes 1 line at the primary
but (potentially) 2 lines at every multicursor.
Solution:
If a motion "beeps", don't replay it.
"3w" near the end still replays (it moved), as does "0" at column
0 (does not beep). This feels like a usable and intuitive compromise.
Problem:
"Press any key" modal prompt is shown when `:quit`
from unsaved buffer, even if ui2 is enabled (which
should never show hit-enter prompt).
Solution:
Skip `wait_return` for `ext_messages` UIs.
Problem:
Servers cannot request refreshing stale LSP folding ranges after a
project-wide change.
Solution:
Advertise `workspace.foldingRange.refreshSupport` and re-request folding
ranges for active buffers when handling `workspace/foldingRange/refresh`.
AI-assisted
Problem:
- With `clipboard=unnamed[plus]`, per-cursor yanks are joined into '"'
on exit, but not the clipboard.
- Multicursor `"+yy` calls the clipboard provider per-cursor, behaving
as "last wins".
Solution:
- Never cascade to/from the clipboard provider.
- Implicit clipboard (`clipboard=unnamed[plus]`) uses the cursor-local
unnamed register, so the cascade needs no `start_batch_changes()`. On
exit, the clipboard gets the same joined result as the `"` register.
- Operations that reference the explicit clipboard registers `"+` / `"*`
while multicursor is active, use the cached (primary) clipboard value.
Problem: Watcher failures are reported inconsistently, and callers
cannot detect when a watcher is no longer usable.
Solution: Route startup and event failures and unexpected inotifywait
exits through an optional on_error callback. Log to nvim-watch.log by
default. Ignore disappearing child directories and exits caused by
cancellation. Let LSP log failures and notify once per message, using
INFO for missing roots and ERROR otherwise, without interrupting
registration.
AI-assisted
Problem: In Virtual Replace mode, backspacing over a character that
replaced several multi-byte characters deletes the padding
that follows it, so the text after the cursor loses its
alignment.
Solution: In replace_do_bs() advance by the length of the character at
the current offset instead of always measuring the first
restored character (Volodymyr Chernetskyi).
After the original characters are restored, replace_do_bs() adds up
their screen width so it knows how much of the alignment padding to drop
again. The loop advanced "i" by mb_ptr2len(p) - 1, which always returns
the length of the *first* restored character rather than the length of
the character at the offset being looked at.
One backspace can restore more than one character, because a wide
character may have replaced several narrow ones. When those characters
do not all have the same byte length the index lands in the middle of a
character: chartabsize() then measures a trailing byte, counts it as an
unprintable <xx> worth four cells, and the inflated width makes the
following loop delete padding spaces that should have been kept.
call setline(1, 'aé xyz')
call feedkeys("gR\u4e00\<BS>\e", 'xt')
leaves "aéxyz" instead of restoring the original line. Going the other
way round ('éa') happens to work, since there the first character is the
longer one.
This has been wrong since the loop was added in Vim 7.0.
closes: vim/vim#212115d934b1bdb
Co-authored-by: Volodymyr Chernetskyi <19735328+chernetskyi@users.noreply.github.com>
Problem: vim_strbyte() can be improved
Solution: Replace the byte-by-byte scan in vim_strbyte() with strchr(),
preserving its byte-range and NUL behavior (Yasuhiro Matsumoto).
closes: vim/vim#21240df4ef959d5
Co-authored-by: Yasuhiro Matsumoto <mattn.jp@gmail.com>
Problem:
A Visual-mode mapping that creates cursors (`xmap I Q0i`) replays its
commands only if cursors already existed when it started.
Solution:
Always collect `composite`. Consumers are still decided per frame
(`CmdFrame.consumers`) and at emit/cascade.
- call-non-callable
- missing-fields
- redefined-label
- redefined-local
- undefined-field
- unreachable-code
Promote redefined-local and unreachable-code to warnings.
Fix DOS line-ending handling for multiline semantic tokens, and support
intersection types in the help parser.
AI-assisted
Problem: matchfuzzypos() can be improved
Solution: Compute the uppercase character once per character in
has_match(), instead of repeating the conversion while
scanning each candidate (Yasuhiro Matsumoto).
closes: vim/vim#21241c0c3fbd967
Co-authored-by: Yasuhiro Matsumoto <mattn.jp@gmail.com>
Problem: CursorLineFold and CursorLineSign are only applied when
'cursorlineopt' contains "number" or is "both"
(Evgeni Chasnovski).
Solution: Apply those groups whenever 'cursorline' is set, independent
of 'cursorlineopt' (Vrushali Zampalkar).
fixes: vim/vim#20480closes: vim/vim#211121f8d3c44c5
Co-authored-by: Vrushali-Z <zampalkarvasu@gmail.com>
Problem: Netrw raises E46 when changing directory fails.
Solution: Remove unused assignments to the read-only directory argument
and cover both error-handling branches with a regression test
(Qiming zhao).
When :lcd fails with E472, s:NetrwLcd() assigns to a:newdir instead of
returning -1. The fallback assignment also refers to dirname, which is
undefined on this path. Neither assignment can update the caller's
directory variable. Keep the existing option restoration and failure
return while removing the invalid assignments.
Supported by AI.
closes: vim/vim#212262f72e4c72f
Co-authored-by: Qiming zhao <chemzqm@gmail.com>
Problem: With 'regexpengine' set to 1 a case-insensitive match against
a literal string fails when the string starts with a
multi-byte character that is longer than a character following
it, so the two regexp engines disagree (after v9.1.0645).
Solution: In cstrncmp() advance by the length of the character at the
current position instead of always measuring the first
character of "s1" (Volodymyr Chernetskyi).
cstrncmp() walks "s1" to find how many characters make up "*n" bytes, so
that it can measure out the same number of characters in "s2". The loop
decremented the remaining byte count by mb_ptr2len(s1), which always
returns the length of the *first* character, rather than the length of
the character at the current position "p".
When the first character is longer than a later one the byte count runs
out too early, the character count comes up short, and MB_STRNICMP2() is
handed a length for "s2" that is too small, so the comparison fails. For
example matching "\cüber" against "Überraschung": "über" is five bytes,
but each iteration subtracts two (the length of "ü"), so the loop runs
three times instead of four.
:set regexpengine=1
echo matchstr('Überraschung', '\cüber')
returns an empty string, while 'regexpengine' set to 2 correctly returns
"Über". The default value of 0 uses the NFA engine and is unaffected.
related: vim/vim#14756
closes: vim/vim#212124e5ac0d68a
This is a bit more boilerplate because of painful impedance mismatch (is
augroup names the root source of truth? is ids the source of truth? who
knows? what does truth even mean?), but show an error reflecting what
the caller actually tried (they passed in an invalid id, not a null
string)
context: vim_snprintf provides a "portable" replacement for a system
vsnprintf(). However we rely on system vsnprintf() in places. This
is quite arbitrary. I want to use system vsnprintf() only, and
reduce the machinery in string.c for typval printf only, which will
delet a lot of duplicated code.
The biggest hurdle here is not "portably" (we only support sane c
runtimes) but the following discrepancy: standard libc printf()
considers NULL args to %s to be undefined behavior strictly
while vim_snprintf() _defacto_ allows this by replacement to "[NULL]"
IMO, this should be seen as "graceful" error handling (nicer than crashing
on the user), not a valid means to have the string "[NULL]" intentionally
be shoved in the users face. more robust code should explicitly
check for NULL and replace it with a _appropriate_ fallback for
the situation, depending on what a NULL string actually means in the
context (unnamed buffer? anonymous namespace? global augroup?)
soo, this PR provides the most _gently_ incremental nudge towards
considering this case again as an error one could possible imagine.
If sanitizers are complied in, this minor incursion gets printed
into the sanitizers' log as a "report-and-continue" error with a
traceback. This doesn't annoy the end user or "stop the world"
in the test suite, but the CI will see the reported error and
fail the build, just like other non-fatal CI errors.
Problem:
A server that supports dynamic registration must not also declare
completionProvider at initialize, so server_capabilities.completionProvider
is nil for anything reading it directly.
Solution:
Do not advertise it by default. A client that follows a registration can
declare it itself.
Problem:
Scrolling a 'scrollbind' window whose peer has a virt_lines block
taller than the window makes the peer jump back and forth.
Solution:
After scrollup()/scrolldown(), sync w_scbind_pos to the actual
get_vtopline() when the target was not clamped to buffer bounds.
Ref: esmuellert/codediff.nvim#519
AI-assisted
Problem:
Type hierarchy (#28388, #28467) and `workspace/executeCommand` (#11607) are
implemented, but `textDocument.typeHierarchy` and `workspace.executeCommand`
are missing from `make_client_capabilities()`.
Solution:
Declare both capabilities.
AI-assisted
"vim.pot" file and tasks are N/A.
"clean" task is N/A after migrating to CMake.
N/A patches:
- 59bd74ed4c9ab366182c93bdc430b186729abbad
- 5d552d652b0197063565ab937d30f92a9ed28545
- 3c8d32e4fc17b4b38b17d7c039ee6203d1b01078
Problem:
The eol_offset used to decode multiline semantic tokens is computed as
vim.bo.fileformat[bufnr] == 'dos' and 2 or 1
which indexes the option value rather than the buffer, so it is never
"dos" and the offset is always 1. A token that crosses a line boundary
in a 'fileformat' of "dos" is then decoded one code unit short per line
ending, and the highlight extends past the end of the token.
Solution:
Read the option with the buffer-scoped form, vim.bo[bufnr].fileformat.
The existing multiline test covers both formats now. Its token length of
82 counts the line endings it spans, so the same token reaches four
characters less far once the buffer is "dos".
AI-assisted
Problem:
vim.lsp.util.show_document() can fail to load the target buffer, most
visibly with E325 when another process owns the file's swapfile and the
user declines to open it. The Vim error escapes show_document() and
surfaces inside whichever LSP handler called it, instead of being
reported as a failure to show the document.
Solution:
Load the buffer up front with vim.fn.bufload(), before any of that
state is touched, and report the error instead of raising.
AI-assisted
Problem: LuaLS struggles with the generics used in Nvim's runtime,
requiring broad diagnostic suppressions. Indexing is also slow.
Solution: Use EmmyLua for type checks in the build and CI. It offers
more sophisticated type checking, substantially better support for
generics, and much better flow analysis.
Correct the affected annotations. Use `@internal`, supported directly by
EmmyLua, instead of `@nodoc` for shared internal declarations, and
support it in the help parser.
AI-assisted
Problem: K on generic and nullable type annotations can jump to the
wrong help tag. Generated See references repeat the type names.
Solution: Strip generic arguments and nullable suffixes when resolving
help tags. Remove the redundant generated See references.
AI-assisted
Problem:
dynamicRegistration is declared false, and completionProvider is read
from server_capabilities only.
Solution:
Declare it, and prefer registerOptions per field.
Problem:
The LSP client excludes paths such as Git objects and installed
dependencies from file watching. However, watchdirs() still traverses
these subtrees before filtering which directories receive watchers,
paying the filesystem and Lua traversal cost for excluded content.
Solution:
Prune excluded subdirectories during the initial walk, and document
that excluding a directory also excludes its descendants.
Using the LSP client's existing **/node_modules/*/** exclusion on a
synthetic macOS tree with 100 installed packages and 4,000 files:
Before After
Directory scans 503 103
Entries inspected 4502 202
Directory watchers 103 103
The measurements use real filesystem calls and watcher handles, with
identical watcher paths for this exclusion.
Ref: #23291
AI-assisted
Problem:
msgpack#strptime() finds a timestamp by bisecting strftime() output, and
brackets the search with the extreme UTC offsets. The lower bound
subtracts 12 hours, but -12:00 is the westernmost offset, which produces
the *largest* timestamp for a given local time. The easternmost offset
is +14:00, so in any zone east of UTC+12 the search can start above the
target and the function throws:
internal-start-string:Internal error: start > string
With TZ=GMT-14, four of the five timestamps in msgpack_spec.lua fail,
as do three tests in shada_spec.lua, which reaches the same function
through shada#strings_to_sd().
Solution:
Subtract 14 hours instead. The upper bound is already generous enough at
+14, and widening a bisection bracket downwards cannot change the result
in zones that worked before.
The test disables the python3 provider, because msgpack#strptime() only
uses the Vimscript implementation changed here when no provider answers,
and otherwise hands the work to datetime.strptime().
AI-assisted
