File renames aren't handled properly
dimonomid opened this issue · comments
What is the latest commit SHA in your installed vim-gitgutter?
f7b97666ae36c7b3f262f3190dbcd7033845d985
What vim/nvim version are you on?
VIM - Vi IMproved 9.0 (2022 Jun 28, compiled May 16 2023 16:59:51)
Included patches: 1-1561
First of all, thanks a lot for this plugin, it's a massive help overall.
One of the significant issues I'm having with it is that the file renames with small changes aren't handled properly:
- The quickfix list is broken (instead of using the new filename, it uses the old one with some
bsuffix) - If I manually navigate to the new file, it shows all lines as new, instead of showing the updated lines properly.
Here's a simple script to reproduce; the script will create a dir /tmp/rename-issue-demo, initialize a repository with a single 15-line file myfile.txt, commit it, and then update a couple lines in this file, rename it to myfile_renamed.txt, and stage all changes:
#!/bin/bash
# Make check if target dir exists
if [ -e /tmp/rename-issue-demo ]; then
if [ -d /tmp/rename-issue-demo ]; then
echo "/tmp/rename-issue-demo is an existing directory, gonna delete it"
rm -rf /tmp/rename-issue-demo
else
echo "/tmp/rename-issue-demo is not a directory, refusing to do anything" > /dev/stderr
exit 1
fi
fi
# Init a repo with a simple 15-line text file in it, and make the initial commit
mkdir /tmp/rename-issue-demo
cd /tmp/rename-issue-demo
git init
cat << EOF > myfile.txt
line01
line02
line03
line04
line05
line06
line07
line08
line09
line10
line11
line12
line13
line14
line15
EOF
git add myfile.txt
git commit -m 'Initial commit'
# Change a couple lines in the file, rename it, and stage all those changes
sed -i 's/line05/line05 updated/' myfile.txt
sed -i 's/line12/LINE12 Updated/' myfile.txt
mv myfile.txt myfile_renamed.txt
git add myfile.txt myfile_renamed.txtAfter that, git diff --staged properly shows file as renamed with some small changes:
diff --git a/myfile.txt b/myfile_renamed.txt
similarity index 75%
rename from myfile.txt
rename to myfile_renamed.txt
index 56957c0..4d03eb3 100644
--- a/myfile.txt
+++ b/myfile_renamed.txt
@@ -2,14 +2,14 @@ line01
line02
line03
line04
-line05
+line05 updated
line06
line07
line08
line09
line10
line11
-line12
+LINE12 Updated
line13
line14
line15However if I open the new myfile_renamed.txt in vim and do this:
:let g:gitgutter_diff_base = 'HEAD'
:GitGutterQuickFix
Then in the quickfix I see this:
myfile.txt b/m|5| -line05
myfile.txt b/m|12| -line12
As you see, it uses the old filename with some suffix b/m, and navigating to it obviously fails.
And if I do the :GitGutterEnable to see the diffs, then I see all lines as new ones:
So when doing reviews, for renamed files I have to review them elsewhere (not in vim with gitgutter), which is unfortunate. And otherwise the plugin mostly works great (again, thanks a lot for your work on it).
Do you think this issue with renamed files could be fixed?
Thanks for the detailed report.
GitGutter handles file renames if done within Vim, via :file or :saveas (for example, see 68f16eb). However that doesn't help when renames are done outside Vim.
Without thinking too hard I first tried adding the -M flag to the git-diff command:
diff --git i/autoload/gitgutter/diff.vim w/autoload/gitgutter/diff.vim
index 010a17b..c56afd9 100644
--- i/autoload/gitgutter/diff.vim
+++ w/autoload/gitgutter/diff.vim
@@ -137,7 +137,7 @@ function! gitgutter#diff#run_diff(bufnr, from, preserve_full_diff) abort
let cmd .= ' -c "diff.noprefix=false"'
let cmd .= ' -c "core.safecrlf=false"'
endif
- let cmd .= ' diff --no-ext-diff --no-color -U0 '.g:gitgutter_diff_args.' -- '.from_file.' '.buff_file
+ let cmd .= ' diff --no-ext-diff --no-color -M -U0 '.g:gitgutter_diff_args.' -- '.from_file.' '.buff_file
" Pipe git-diff output into grep.
if !a:preserve_full_diff && !empty(g:gitgutter_grep)Alas it didn't help.
GitGutter works by diffing the in-buffer file against a version of the file from the index (obtained via git show).
It looks a bit like this:
- Write buffer to temp file A
- Write base version of file to temp file B (
git show SHA:file > B) - Diff (
git diff --no-index -- A B)
The git-show command needs the filename, for which the code uses the buffer's filename.
I suppose instead we want to use the original filename – but I don't know how to get that in a neat way which could be used in the diff command.
It looks like the way to find renames is:
git diff --diff-filter=R --name-status SHA
And then parse the output for the original and current filenames.
I can't really see how to graft that into the plugin's current combined show and diff command. In which case it would need to be a separate, earlier step. But running an extra system call might add too much overhead.
Yeah after your initial explanation I was also looking at getting the file renames from git diff and also found the --name-status, which is exactly what we need here.
I can't really see how to graft that into the plugin's current combined show and diff command. In which case it would need to be a separate, earlier step.
I'm a bit confused wdym by not seeing how to graft that into the current design; I think we just need to make one more (earlier) step, as you also mentioned. So instead of the 3 steps you outlined above, we'll have 4:
- Write buffer to temp file A
- new step: Collect all file renames using
git diff --name-status .... - Write base version of file to temp file B (
git show SHA:file > B) (wherefilemight be either the same filename as we have now, or it can be older, depending on the renames collected earlier) - Diff (
git diff --no-index -- A B)
But running an extra system call might add too much overhead.
Tbh my experience with git diff in general (and also some brief experiments with git diff --name-status I've done just now) show that it should be plenty fast. But, to avoid worrying too much about the potential performance regression, it can be hidden behind a flag which is disabled by default, and then over time we might consider switching it to enabled by default instead (or not; doesn't matter much, since whoever needs it can just have it enabled explicitly)
Also btw, I can see that the diffing part can be addressed by this, but there is also the quickfix issue: right now it uses the old filename with a weird suffix. Is this part also fixable?
It's not the overhead of git-diff I'm thinking of, it's the overhead of Vim calling out to an external process and passing data to and fro across the process boundary. It used to be significant, though no doubt it's much better these days.
However I realised that we don't need to get the list of renames on every diff. We can get it once per diff base and cache it. So I think inserting a new step will be fine. If anyone complains about slowness, I'll make it opt-out.
The quickfix code is implemented independently of the diffing (I didn't have much time available when I wrote it).
It sets the prefixes to a/ and b/ so the prefixes have known lengths. However the prefixes shouldn't appear in the filename in the quickfix list. In your case the offsets seem to have gone awry – I suspect the m in b/m comes from b/myfile.txt being truncated in the wrong place.
I'll tackle the main diff function first and then the quickfix.
However I realised that we don't need to get the list of renames on every diff. We can get it once per diff base and cache it.
Just to double check though: if I e.g. set a diff base to HEAD~ or whatever, so the gitgutter will get the list of renames, and then the file gets renamed outside of vim without changing the diff base, ideally we do want gitgutter to pick that up, right? So while I agree it doesn't have to be done on every diff, but it probably should be done on some occasions even when the diff base did not change. Perhaps it also needs to be done every time we open a new file in vim (to check if this file is a new rename by any chance), or something like that (you know better for sure, just wanted to double check that this scenario is gonna be covered)
I'll tackle the main diff function first and then the quickfix.
Thanks!
Does this branch do what you need, at least as far as the signs go (I'll tackle quickfix later)? Can you get it to give you wrong signs?
I haven't written any automated tests yet but, testing manually, it works for me.
I tested it on a couple repos, and yeah the signs for renamed files are correct from what I can tell. Thanks a lot!
After the quickfix part is addressed as well, it'll be just perfect, as I can do the whole review right in the editor then.
Actually I think there's a new issue: if I change g:gitgutter_diff_base to a non-existing revision, I get those exceptions from obtain_file_renames:
(sorry I keep forgetting how can I copy-paste this error message from gvim, such a poor UX, so posting screenshot)
Stepping into this non-existing revision might be easier than you think: if I work on some particular repo, and set g:gitgutter_diff_base to a specific hash, and then briefly open a file from some other repo, and boom, it points to a non-existing revision now. I think on master it doesn't error out like that.
Thanks for pointing that out. I updated the branch to handle an invalid diff base.
Done in f7b9766...6a95f1b.
One more time, thanks a lot for doing this @airblade. I'm honestly impressed by how dedicated you are to supporting this project, not many open source projects are maintained like that. 🙇
Thank you!