chcp parsing is fragile because of the system locale. There's no convenient way to parse out the codepage value without regex just by relying on cmd.exe builtins and default binaries in PATH.
Vim can be used to parse chcp output but it requires an additional `system` per `s:system` and `chcp` can change within the same console so caching the value won't work on the terminal.
Powershell supports regex but it has a long startup even with `-NoProfile` so running it when `&shell` is not powershell slows down `:PlugInstall` more.
* output of chcp was not parsed correctly
On Windows, when wrapping a batch command with the function s:wrap_cmds(),
when calling 'chcp' to get the current code page, the code assumed that
the output was in the format "active code page: XXX" (where XXX is the
code page), whereas the actual output is localized (for instance, in
French, the output would be: "page de code active: XXX").
The parsing of the output relied on that, and this failed for a
message different from "active code page" (i.e., English).
This patch changes the parsing to split the output of chcp on the colon
instead of spaces. Assuming that the output is always in the format
"<localized message>: XXX", regardless of the locale, hopefully, this is
a bit more robust.
Excerpt from `:h shell-powershell`:
To use powershell (on Windows):
set shell=powershell shellquote=( shellpipe=\| shellxquote=
set shellcmdflag=-NoLogo\ -NoProfile\ -ExecutionPolicy\ RemoteSigned\ -Command
set shellredir=\|\ Out-File\ -Encoding\ UTF8
Previously, `:PlugDiff` would show every new commit to a plugin's git
repo. This makes sense for the general case, but makes less sense when a
plugin lives in a subdirectory of the repo (and is configured with the
'rtp' option). This makes it difficult to determine which commits relate
to the plugin and which are unrelated.
This changes `:PlugDiff` to filter out any commits outside of the 'rtp'
folder.
Some consequences:
* This does not change the `:PlugUpdate` UI. This means `:PlugUpdate`
may pull down non-plugin commits, display that it has updated the
plugin, and then `:PlugDiff` will show no updates (since such commits
fall out of the 'rtp' path).
* It also means there's no UI to revert non-plugin updates, as they
don't show up in `:PlugDiff`.
Git v2.8 introduced support for parallel fetching of submodules. This
can result in a substantial performance boost when performing a "git
submodule update" in a repository with many submodules, potentially
further magnified when updating submodules recursively.
This commit introduces a check on the Git version to see if it supports
parallel submodule fetching. If it does, we add the '--jobs=N' argument
to our call to "git submodule update", where N is g:plug_threads or the
vim-plug default of 16.
This allows you to override the default behavior of `o' or `<cr>'
binding in PlugDiff window.
e.g.
" Move to preview window
autocmd! FileType vim-plug nmap <buffer> o <plug>(plug-preview)<c-w>P
Close#749Close#768
`writefile()` always output LF without CR each lines.
But batch file on Windows needs CR and LF, at end of lines.
And if the path of home directory contains non-ASCII
characters like Japanese username (e.g. `C:\Users\太郎`),
batch file without CR can't be executed correctly.
Close#635Close#668Close#539
Use a temporary batchfile for :!, system(), and jobs and run it in cmd.exe.
This bypasses Vim/Neovim issues in Windows and reduces the need to set more options.
Also, s:shellesc_cmd works in a batchfile only.
Set shellredir for system() in Windows
$SHELL sets the default value of 'shell' (see :h 'shell').
This affects shellredir but cmd.exe requires '>%s 2>&1'.
When loading 'deoplete.nvim' for the 2nd time during InsertEnter
manually, the `s:dobufread` (or `s:lod` itself) prevents it to work
properly - likely because the plugin gets resourced.
Maybe there could be a way to force this (and reload plugins always),
but by default it seems to make sense to skip already loaded plugins.
Checking out a branch or tag may not succeed if there is a path entry
with the same name. This commit appends -- to git checkout commands to
clarify that we're not referring to a file or a directory.
Close#602
Installer is not synchronous when started on VimEnter since
has('vim_starting') returns 0 in that case. We needed a way to make the
installer synchronous.
Related: #574
- Same output format across different installers
- Avoid unnecessary construction of result string
- Fix multi-line error logging when exit_cb is called before out_cb
Having VisualHostKey enabled for ssh can cause lines containing a single
"^M" character to appear in the git output, which causes E684 to be
thrown if not filtered out.
Now `:PlugInstall!` and `:PlugUpdate!` forces to update submodules.
`:PlugInstall!` now also checks out to new specified branch, tag, or
commit.
Also added tests for changing branch, tag, or commit.
This commit extracts duplicate logic out of three installers.
Pros.
- Better maintainability
- Easier to add/extend the features
- Fixes a bug when 'commit' option is used, submodules are updated
before the designated commit is checked out
Cons.
- The whole process takes slightly longer due to lost parallelism after pull
- Especially, submodule updates are not parallelized
- However, this shouldn't matter much in practice as there are few
plugins that heavily rely on submodules
