* rework positioning/rendering, enables softwrap/virtual text
This commit is a large rework of the core text positioning and
rendering code in helix to remove the assumption that on-screen
columns/lines correspond to text columns/lines.
A generic `DocFormatter` is introduced that positions graphemes on
and is used both for rendering and for movements/scrolling.
Both virtual text support (inline, grapheme overlay and multi-line)
and a capable softwrap implementation is included.
fix picker highlight
cleanup doc formatter, use word bondaries for wrapping
make visual vertical movement a seperate commnad
estimate line gutter width to improve performance
cache cursor position
cleanup and optimize doc formatter
cleanup documentation
fix typos
Co-authored-by: Daniel Hines <d4hines@gmail.com>
update documentation
fix panic in last_visual_line funciton
improve soft-wrap documentation
add extend_visual_line_up/down commands
fix non-visual vertical movement
streamline virtual text highlighting, add softwrap indicator
fix cursor position if softwrap is disabled
improve documentation of text_annotations module
avoid crashes if view anchor is out of bounds
fix: consider horizontal offset when traslation char_idx -> vpos
improve default configuration
fix: mixed up horizontal and vertical offset
reset view position after config reload
apply suggestions from review
disabled softwrap for very small screens to avoid endless spin
fix wrap_indicator setting
fix bar cursor disappearring on the EOF character
add keybinding for linewise vertical movement
fix: inconsistent gutter highlights
improve virtual text API
make scope idx lookup more ergonomic
allow overlapping overlays
correctly track char_pos for virtual text
adjust configuration
deprecate old position fucntions
fix infinite loop in highlight lookup
fix gutter style
fix formatting
document max-line-width interaction with softwrap
change wrap-indicator example to use empty string
fix: rare panic when view is in invalid state (bis)
* Apply suggestions from code review
Co-authored-by: Michael Davis <mcarsondavis@gmail.com>
* improve documentation for positoning functions
* simplify tests
* fix documentation of Grapheme::width
* Apply suggestions from code review
Co-authored-by: Michael Davis <mcarsondavis@gmail.com>
* add explicit drop invocation
* Add explicit MoveFn type alias
* add docuntation to Editor::cursor_cache
* fix a few typos
* explain use of allow(deprecated)
* make gj and gk extend in select mode
* remove unneded debug and TODO
* mark tab_width_at #[inline]
* add fast-path to move_vertically_visual in case softwrap is disabled
* rename first_line to first_visual_line
* simplify duplicate if/else
---------
Co-authored-by: Michael Davis <mcarsondavis@gmail.com>
The new version of the `toml` crate is based on `toml_edit` and does
not support zero copy deserialization anymore. So we need to deserialize
`String` instead of `&str` in the keympa
increment/decrement (C-a/C-x) had some buggy behavior where selections
could be offset incorrectly or the editor could panic with some edits
that changed the number of characters in a number or date. These stemmed
from the automatic jumping behavior which attempted to find the next
date or integer to increment. The jumping behavior also complicated the
code quite a bit and made the behavior somewhat difficult to predict
when using many cursors.
This change removes the automatic jumping behavior and only increments
or decrements when the full text in a range of a selection is a number
or date. This simplifies the code and fixes the panics and buggy
behaviors from changing the number of characters.
After changes in #5239, the loaded configuration wasn't stored,
resulting in a success message even if the instance kept the previous
configuration values.
Commit 1b89d3e535 introduced a regression
where opening a new file would no longer work, because attempting to
canonicalize its path would lead to a "No such file or directory"
error. Fall back to opening a new file when encountering an error to
fix this case.
This commit addresses issue 5193, where the author
requested that the name of the binary needed is printed along
with the rest of the health information.
This commit adds a format! macro which formats in the name of the
binary and then it will be printed along with the rest of the
debug information. The value in cmd is referenced to the call
to which, and then consumed upon the call to format!
This roughly matches the behavior of the diagnostic picker: when
jumping to a diagnostic with `[d`/`]d`/`[D`/`]D`, the range of the
diagnostic is selected instead of the start point.
A language server might send None as the response to workspace symbols.
We should treat this as the empty Vec rather than the server sending
an error status. This fixes the interaction with gopls which uses
None to mean no matching symbols.
If the new results shown by the picker select a file that hasn't been
previewed before, the idle timeout would not trigger highlighting on
that file. With this change, we reset the idle timeout and allow that
file to be highlighted on the next idle timeout event.
This change uses the idle-timeout event to trigger fetching new results
in the DynamicPicker, so idle-timeout becomes a sort of debounce. This
prevents querying the language server overly aggressively.
Most language servers limit the number of workspace symbols that
are returned with an empty query even though all symbols are
supposed to be returned, according to the spec (for perfomance
reasons). This patch adds a workspace symbol picker based on a
dynamic picker that allows re-requesting the symbols on every
keypress (i.e. when the picker query text changes). The old behavior
has been completely replaced, and I have only tested with
rust-analyzer so far.
The error messages for a theme that failed to be deserialized (or
otherwise failed to load) were covered up by the context/with_context
calls:
* The log message for a bad theme configured in config.toml would only
say "Failed to deserilaize theme"
* Selecting a bad theme via :theme would show "Theme does not exist"
With these changes, we let the TOML deserializer errors bubble up, so
the error messages can now say the line number of a duplicated
key - and that key's name - when a theme fails to load because of a
duplicated key.
Providing a theme which does not exist to :theme still gives a helpful
error message: "No such file or directory."
* Reset mode when changing buffers
This is similar to the change in
e4c9d4082a: reset the editor to normal
mode when changing buffers. Usually the editor is already in normal
mode but it's possible to setup insert-mode keybindings that change
buffers.
* Move normal mode entering code to Editor
This should be called internally in the Editor when changing documents
(Editor::switch) or changing focuses (Editor::focus).
* add command and keybding to jump to next/prev hunk
* add textobject for change
* Update helix-vcs/src/diff.rs
Co-authored-by: Michael Davis <mcarsondavis@gmail.com>
* select entire hunk instead of first char
* fix selection range
Co-authored-by: Michael Davis <mcarsondavis@gmail.com>
* Add `View::ensure_cursor_in_view_center` to adjust view after searching and jumping
Also `offset_coodrs_to_in_view` was refactored to reduce duplicated position calculations.
* Fix a wrong offset calculation in `offset_coords_to_in_view_center`
It ignored `scrolloff` if `centering` is false.
Completion edits - either basic `insert_text` strings or structured
`text_edit`s - are assumed by the LSP spec to apply to the current
cursor (or at least the trigger point). We can use the range (if any)
and text given by the Language Server to create a transaction that
changes all ranges in the current selection though, allowing auto-
complete to affect multiple cursors.
* Change default TS object bindings
Changes 'match inside/around' bindings for:
- type definition from `c` to `t`
- comments from `o` to `c`
- tests from `t` to `T`
Also changes those for the `]` / `[` bindings.
* Update docs for changed keybinds
Co-authored-by: Michael Davis <mcarsondavis@gmail.com>
* Add a undo/redo split test case for crossing branches
* history: Switch up/down transaction chaining order
The old code tends to work in practice because, usually, either up_txns
or down_txns are empty. When both have contents though, we can run into
a panic trying to compose them all since they will disagree on the
length of the text. This fixes the panic test case in the parent
commit.
* Show (git) diff signs in gutter (#3890)
Avoid string allocation when git diffing
Incrementally diff using changesets
refactor diffs to be provider indepndent and improve git implementation
remove dependency on zlib-ng
switch to asynchronus diffing with similar
Update helix-vcs/Cargo.toml
fix toml formatting
Co-authored-by: Ivan Tham <pickfire@riseup.net>
fix typo in documentation
use ropey reexpors from helix-core
fix crash when creating new file
remove useless use if io::Cursor
fix spelling mistakes
implement suggested improvement to repository loading
improve git test isolation
remove lefover comments
Co-authored-by: univerz <univerz@fu-solution.com>
fixed spelling mistake
minor cosmetic changes
fix: set self.differ to None if decoding the diff_base fails
fixup formatting
Co-authored-by: Ivan Tham <pickfire@riseup.net>
reload diff_base when file is reloaded from disk
switch to imara-diff
Fixup formatting
Co-authored-by: Blaž Hrastnik <blaz@mxxn.io>
Redraw buffer whenever a diff is updated.
Only store hunks instead of changes for individual lines to easily allow
jumping between them
Update to latest gitoxide version
Change default diff gutter position
Only update gutter after timeout
* update diff gutter synchronously, with a timeout
* Apply suggestions from code review
Co-authored-by: Blaž Hrastnik <blaz@mxxn.io>
Co-authored-by: Michael Davis <mcarsondavis@gmail.com>
* address review comments and ensure lock is always aquired
* remove configuration for redraw timeout
Co-authored-by: Blaž Hrastnik <blaz@mxxn.io>
Co-authored-by: Michael Davis <mcarsondavis@gmail.com>
This matches the insert-mode behavior for Vim and Kakoune: if the
current line is empty except for whitespace, `<ret>` should insert a
line ending at the beginning of the line, moving any indentation to the
next line.
The 'revisions' field on History can't be treated as linear: each
Revision in the revisions Vec has a parent link and an optional child
link. We can follow those to unroll the recent history.
When using undo/redo, the history revision can be decremented. In that
case we should apply the inversions since the given revision in
History::changes_since. This prevents panics with jumplist operations
when a session uses undo/redo to move the jumplist selection outside
of the document.
This case panics since undo/redo call View::apply and here, the edit
that moves the jumplist selection out-of-bounds is not yet applied when
View::apply is called in undo/redo. View::apply should only be called
by the EditorView now.
* Add a test case for updating jumplists across windows
* Apply transactions to all views on history changes
This ensures that jumplist selections follow changes in documents, even
when there are multiple views (for example a split where both windows
edit the same document).
* Leave TODOs for cleaning up View::apply
* Use Iterator::reduce to compose history transactions
Co-authored-by: Blaž Hrastnik <blaz@mxxn.io>
Co-authored-by: Blaž Hrastnik <blaz@mxxn.io>
Language Servers may signal that they do not support a method in
the initialization result (server capabilities). We can check these
when making LSP requests and hint in the status line when a method
is not supported by the server. This can also prevent crashes in
servers which assume that clients do not send requests for methods
which are disabled in the server capabilities.
There is an existing pattern the LSP client module where a method
returns `Option<impl Future<Output = Result<_>>>` with `None` signaling
no support in the server. This change extends this pattern to the rest
of the client functions. And we log an error to the statusline for
manually triggered LSP calls which return `None`.
Previously, jumplists could grow unchecked. Every transaction is
applied to jumplist selections to ensure that they are up to date
and within document bounds, so this would cause every edit to become
more expensive as jumplist lengths increased throughout a session.
Setting a maximum number of entries limits the cost.
Vim and Neovim limit their jumplists:
* b298fe6cba/src/structs.h (L141)
* e8cc489acc/src/nvim/mark_defs.h (L57)
Notably, Kakoune does not. In Kakoune, changes are applied to jumplist
entries lazily as you hit `<C-o>`/`<C-i>` though, so Kakoune doesn't
have the same growing cost concerns. Kakoune also does not have a
concept of a View which limits the cost further.
Vim and Neovim limit to 100. This seems unreasonably high to me so I've
set this to 30 to start. We can increase if this is problematically
low.
d6323b7cbc changed the behavior of paste
to select the newly inserted text. This is preferrable in normal mode
because it's useful to be able to act on the new text. This behavior
is worse for insert or select mode though:
* In insert mode, the cursor ends up on the last character of the newly
selected text, so further typing inserts text before the last
character.
* In select mode, the current selection is replaced with the new text
selection which doesn't extend the current selection. With this
change, the selection is extended to include the new text.
This aligns the behavior more closely with Kakoune, but it's
coincidental instead of intentional: Kakoune doesn't implement
bracketed paste (AFAIK) which causes this behavior in insert mode,
and Kakoune doesn't have a select mode.
Previously, commands such as `r<tab>` (replace with tab) or `t<tab>`
(select till tab) had no effect. This is because `KeyCode::Tab` needs
special treatment (like `KeyCode::Enter`).
This change handles a language server exiting. This was a UX sore-spot:
if a language server crashed, Helix did not recognize the exit and
continued to send requests to it. All requests would timeout since they
would not receive responses. This would also hold-up Helix closing
itself down since it would try to gracefully shutdown the server which
is implemented in the LSP spec as a request.
We could attempt to automatically restart the language server on crash.
I left this for future work since that change will need to be slightly
complicated: it will need to cover the case of a language server
repeatedly crashing.
d7d0d5ffb7 resolves completion items on
the idle-timeout event. The `Completion::resolve_completion_item`
function blocks on the LSP request though, which blocks the compositor
and in turn blocks the event loop. So until the language server returns
the resolved completion item, Helix is unable to respond to keypresses
or other LSP messages.
This is typically ok since the resolution request is fast but for some
language servers this can be problematic, and ideally we shouldn't be
blocking like this anyways.
When receiving a `completionItem/resolve` request, the Volar server
sends a `workspace/configuration` request to Helix and blocks itself
on the response, leading to a deadlock. Eventually the resolve request
times out within Helix but Helix is locked up and unresponsive in that
window.
This change resolves the completion item without blocking the
compositor.
PR #4134 switched the autocomplete menu from alphabetical to fuzzy
sorting. This commit removes the still existing filtering by prefix and
should enable full fuzzy sorting of the autocomplete menu.
closes#3084, #1807
Co-authored-by: Blaž Hrastnik <blaz@mxxn.io>
This fixes an edge case for completing shellwords. With a file
"a b.txt" in the current directory, the sequence `:open a\<tab>`
will result in the prompt containing `:open aa\ b.txt`. This is
because the length of the input which is trimmed when replacing with
completion is calculated on the part of the input which is parsed by
shellwords and then escaped (in a separate operation), which is lossy.
In this case it loses the trailing backslash.
The fix provided here refactors shellwords to track both the _words_
(shellwords with quotes and escapes resolved) and the _parts_ (chunks
of the input which turned into each word, with separating whitespace
removed). When calculating how much of the input to delete when
replacing with the completion item, we now use the length of the last
part.
This also allows us to eliminate the duplicate work done in the
`ends_with_whitespace` check.
The text within the command palette used a custom format to display
the keybinding for a command. This change switches to the key sequence
format that we use for pending keys and macros.
* init
* cargo fmt
* optimisation of the scrollbar render both for Menu and Popup. Toggling off scrollbar for Popup<Menu>, since Menu has its own
* rendering scroll track
* removed unnecessary cast
* improve memory allocation
* small correction
d6323b7cbc introduced a regression for
shell commands like `|`, `!`, and `<A-!>` which caused the new
selections to be incorrect. This caused a panic when piping (`|`)
would cause the new range to extend past the document end.
The paste version of this bug was fixed in
48a3965ab4.
This change also inherits the direction of the new range from the old
range and adds integration tests to ensure that the behavior isn't
broken in the future.
* dynamically resize line number gutter width
* removing digits lower-bound, permitting spacer
* removing max line num char limit; adding notes; qualified successors; notes
* updating tests to use new line number width when testing views
* linenr width based on document line count
* using min width of 2 so line numbers relative is useful
* lint rolling; removing unnecessary type parameter lifetime
* merge change resolution
* reformat code
* rename row_styler to style; add int_log resource
* adding spacer to gutters default; updating book config entry
* adding view.inner_height(), swap for loop for iterator
* reverting change of current! to view! now that doc is not needed
If `a\ b.txt` were a local file, `:o a\ <tab>` would fill the prompt
with `:o aa\ b.txt` because the replacement range was calculated using
the shellwords-parsed part. Escaping the part before calculating its
length fixes this edge-case.
This changes the completion items to be rendered with shellword
escaping, so a file `a b.txt` is rendered as `a\ b.txt` which matches
how it should be inputted.
8584b38cfb switched to shellwords for
completion in command-mode. This changes the conditions for choosing
whether to complete the command or use the command's completer.
This change processes the input as shellwords up-front and uses
shellword logic about whitespace to determine whether the command
or argument should be completed.
* Fix range offsets in multi-selection paste
d6323b7cbc introduced a regression with
multi-selection paste where pasting would not adjust the ranges
correctly. To fix it, we need to track the total number of characters
inserted in each changed selection and use that offset to slide each
new range forwards.
* Inherit selection directions on paste
* Add an integration-test for multi-selection pasting
The sequence "_y"_p panics because the blackhole register contains an
empty values vec. This causes a panic when pasting since it unwraps
a `slice::last`.
This follows changes in Kakoune to the same effects:
* p/<space>p: 266d1c37d0
* !/<A-!>: 85b78dda2e
Selecting the new data inserted by shell or pasting is often more
useful than retaining a selection of the pre-paste/insert content.
* Clamp highlighting range to be within document
This fixes a panic possible when two vsplits of the same document
exist and enough lines are deleted from the document so that one of
the windows focuses past the end of the document.
* Ensure cursor is in view on window change
If two windows are editing the same document, one may delete enough of
the document so that the other window is pointing at a blank page (past
the document end). In this change we ensure that the cursor is within
view whenever we switch to a new window (for example with `<C-w>w`).
* Update helix-term/src/ui/editor.rs
Co-authored-by: Blaž Hrastnik <blaz@mxxn.io>
Co-authored-by: Blaž Hrastnik <blaz@mxxn.io>
When backward-deleting a character, if this character and the following
character form a Pair, we want to delete both. However, there is a bug
that deletes both characters also if both characters are closers of some
Pair.
This commit fixes that by adding an additional check that the deleted
character should be an opener in a Pair.
Closes https://github.com/helix-editor/helix/issues/4544.
Most commands that accept an argument show their current value if no
argument is specified. The `:theme` command previously displayed an
error message in the status bar if not provided with an argument:
```
Theme name not provided
```
It now shows the current theme name in the status bar if no argument is
specified.
Signed-off-by: James O. D. Hunt <jamesodhunt@gmail.com>
Signed-off-by: James O. D. Hunt <jamesodhunt@gmail.com>
This bug occurs on `shell_insert_output` and `shell_append_output`
commands.
The previous implementation would create a child process using the Rust
stdlib's `Command` builder. However, when nothing should be piped in
from the editor, the default value for `stdin` would be used. According
to the Rust stdlib documentation that is `Stdio::inherit` which will
make the child process inherit the parent process' stdin. This would
cause the terminal to freeze.
This change will set the child process' stdin to `Stdio::null` whenever
it doesn't pipe it. In the `if` statement where this change was made
there was an extra condition for windows that I am not sure if would
require some special treatment.
This is mostly for the sake of the diagnostics pickers: without
rendering the diagnostic styles, it's hard to tell where the entries
in the picker are pointing to.
Some language servers may not send the `documentation` field if it
is expensive to compute. Clients can request the missing field with
a completionItem/resolve request.
In this change we use the idle-timeout event to ensure that the current
completion item is resolved.
This complicates the code a little but it often divides by two the number of allocations done by
the functions. LSP labels especially can easily be called dozens of time in a single menu popup,
when listing references for example.
When we do auto formatting, the code that takes the LSP's response and applies
the changes to the document are just getting the currently focused view and
giving that to the function, basically always assuming that the document that
we're applying the change to is in focus, and not in a background view.
This is usually fine for a single view, even if it's a buffer in the
background, because it's still the same view and the selection will get updated
accordingly for when you switch back to it. But it's obviously a problem for
when there are multiple views, because if you don't have the target document in
focus, it will ask the document to update the wrong view, hence the crash.
The problem with this is picking which view to apply any selection change to.
In the absence of any more data points on the views themselves, we simply pick
the first view associated with the document we are saving.
When force quitting, we need to block on the pending writes to ensure
that write commands succeed before exiting, and also to avoid a crash
when all the views are gone before the auto format call returns from
the LS.
* Autosave all when the terminal loses focus
* Correct comment on focus config
Co-authored-by: Blaž Hrastnik <blaz@mxxn.io>
* Need a block_try_flush_writes in all quit_all paths
Co-authored-by: Blaž Hrastnik <blaz@mxxn.io>
* Fix test::print for Unicode
The print function was not generating correct translations when
the input has Unicode (non-ASCII) in it. This is due to its use of
String::len, which gives the length in bytes, not chars.
* Fix multi-code point auto pairs
The current code for auto pairs is counting offsets by summing the
length of the open and closing chars with char::len_utf8. Unfortunately,
this gives back bytes, and the offset needs to be in chars.
Additionally, it was discovered that there was a preexisting bug where
the selection was not computed correctly in the case that the cursor
was:
1. a single grapheme in width
2. this grapheme was more than one char
3. the direction of the cursor is backwards
4. a secondary range
In this case, the offset was not being added into the anchor. This was
fixed.
* migrate auto pairs tests to integration
* review comments
This change removes language server configuration from the default
languages.toml config for integration tests. No integration-tests
currently depend on the availability of a language server but if any
future test needs to, it may provide a language server configuration
by passing an override into the `test_syntax_conf` helper.
Language-servers in integration tests cause false-positive failures
when running integration tests in GitHub Actions CI. The Windows
runner appears to have `clangd` installed and all OS runners have
the `R` binary installed but not the `R` language server package.
If a test file created by `tempfile::NamedTempFile` happens to have a
file extension of `r`, the test will most likely fail because the
R language server will fail to start and will become a broken pipe,
meaning that it will fail to shutdown within the timeout, causing a
false-positive failure. This happens surprisingly often in practice.
Language servers (especially rust-analyzer) also emit unnecessary
log output when initializing, which this change silences.
`helix_view::apply_transaction` closes over `Document::apply` and
`View::apply` to ensure that jumplist entries are updated when a
document changes from a transaction. `Document::apply` shouldn't
be called directly - this helper function should be used instead.
If a document is written with a new path, currently, in the event that
the write fails, the document still gets its path changed. This fixes
it so that the path is not updated unless the write succeeds.
The way that document writes are handled are by submitting them to the
async job pool, which are all executed opportunistically out of order. It
was discovered that this can lead to write inconsistencies when there
are multiple writes to the same file in quick succession.
This seeks to fix this problem by removing document writes from the
general pool of jobs and into its own specialized event. Now when a
user submits a write with one of the write commands, a request is simply
queued up in a new mpsc channel that each Document makes to handle its own
writes. This way, if multiple writes are submitted on the same document,
they are executed in order, while still allowing concurrent writes for
different documents.
Instead of repeatedly checking if it is in_bounds, calculate the
max_indent beforehand and just loop. I added a debug_assert to "prove"
that it never tries drawing out of bounds.
Better performance, and otherwise very long lines with lots of tabs
will wrap around the u16 and come back on the other side, messing up
the beginning skip_levels.
Also changes workspace diagnostic picker bindings to <space>D and
changes the debug menu keybind to <space>g, the previous diagnostic
picker keybind. This brings the diagnostic picker bindings more in
line with the jump to next/previous diagnostic bindings which are
currently on ]d and [d.
