The previous commit switched use of writefile() to binary mode so that
we could prevent a newline being added to a completely empty buffer.
Evidently, however, binary mode has side effects (see #567) so this
commit returns to non-binary mode - with a simpler fix for completely
empty files.
Unfortunately this implementation does not work for noeol files - see
the failing test - because writefile() does not take account of
'nofixeol' (unlike :write). This is suboptimal but acceptable because
noeol files are not be encountered often.
See #567.
An empty, unchanged file was considered to be changed (a new line
added).
This patch fixes s:write_buffer to use binary mode with `writefile` to
not append a newline always, and does so manually for non-empty buffers,
according to &endofline, &binary and &fixendofline.
This is taken out of Neomake, and tested there in
https://github.com/neomake/neomake/blob/091d148b/tests/utils.vader#L713-L759.
Until this commit the err callback closed the channel to prevent the close
callback from being invoked, because the close callback invoked the out
handler (which we don't want when the job was unsuccessful).
This mostly worked but if closing the channel took too long, which it
seemed it could when dealing with large files, the close callback could
be called in the meantime.
This commit removes the unreliable code. The parent commit ensured that
the close callback only calls the out handler when the job succeeded, so
we no longer need to try to prevent the close callback from being
called.
The diffing part of the plugin diffs two files which are written afresh
each time. When the same buffer was processed twice in quick
succession, the second process to write a file could write it before the
first process had finished reading it. This manifested as the "from"
file being read as empty, causing diff to report that all the lines in
the file had been added.
This commit adds a counter to disambiguate successive temporary files.
This prevents git-ls-files from escaping "unusual" characters in
pathnames. When this happened, gitgutter would feed the escaped name
back to git-show but git-show would not recognise it.
This commit makes git-ls-files use the -z option to output pathnames
verbatim. These pathnames also become null terminated so we have to
ensure we remove the null terminator too.
Closes#562.
Previously gitgutter could only diff the buffer against the index. This
change enables diffing against the working tree, which will be useful
for stashing.
Use a default with `getbufvar` to not get an empty string.
Fixes:
> Error detected while processing function <SNR>163_on_stderr_nvim[1]..<lambda>10[1]..gitgutter#utility#setbufvar:
> line 1:
> E712: Argument of get() must be a List or Dictionary
I can trigger it using https://github.com/AndrewRadev/linediff.vim when
closing the windows / the tab.
Likely triggered with/through Neomake.
In #303 there was a request to deactivate gitgutter when &diff is set,
although no reason was given initially. Then it was reported that
gitgutter showed signs, in the wrong place, during Fugitive's Gdiff.
However I cannot reproduce signs showing in Gdiff's index buffer. And
the signs in the working tree buffer are correct.
Furthermore the commit which deactivated gitgutter (85f6e136) had the
side-effect of breaking `:redir` and clearing Vim's startup message
(see #500).
Therefore since that change had no benefit, and a couple of drawbacks,
it makes sense to undo it.
Closes#500.
The previous implementation meant the temp file paths were treated as
regular expressions, which was vulnerable to problems with backslashes
etc.
See #494.
This a) logs the message to be seen in `:mess` and b) avoids the
hit-ENTER prompt (which is likely caused by the message not being
logged and Vim deciding that you need to see it then).
So a) fixes b).
An alternative would be to run `git` with no user config always, i.e.
setting `GIT_CONFIG_NOSYSTEM` and `HOME=/dev/null` (different for
Windows then) in its environment. Maybe there's an easier way / single
option for this..
This provides `g:gitgutter_hook_context` during the hook's execution and
removes the `:silent`, but uses `exists()` instead.
The bufnr might be necessary to know in the User autocommand, e.g. to
clear some cache.
Not using`:silent` is good practice in general to not hide (wanted)
output and errors etc.
