Confused by "skipped updates" when rebasing, restoring or abandoning commits.

[binarybana]

I found myself with a few expected changes and many unwanted changes in my current commit.

❯ jj st
Working copy changes:
M python/tutorials/10-block-scaled-matmul.py  <--- expected change
M third_party/nvidia/include/cublas_instance.h  <--- expected change
M third_party/nvidia/include/cublas_types.h.      <--- expected change
M third_party/nvidia/triton_nvidia.cc                    <--- expected change
D third_party/proton/Dialect/include/Analysis/ScopeIdAllocation.h   <--- Start of unwanted changes 
D third_party/proton/Dialect/include/Conversion/CMakeLists.txt
... (many more deletions)

No problem, I'll just blow those away with a restore for that fileset,

❯ jj restore -f @- -t @ 'glob:third_party/proton/**'
Working copy  (@) now at: xurykvqu e0c1410e Claude adding block scaled floating cuBLAS baselines
Parent commit (@-)      : rypswprz 90c2ebbe DONOTMERGE
Added 73 files, modified 0 files, removed 16 files
Warning: 16 of those updates were skipped because there were conflicting changes in the working copy.
Hint: Inspect the changes compared to the intended target with `jj diff --from e0c1410ec00a`.
Discard the conflicting changes with `jj restore --from e0c1410ec00a`.

Hmm.. that's odd. And when I look at jj st some of the Deletions are still there, but some have now moved to renames:

❯ jj st
Working copy changes:
... (expected changes)
D third_party/proton/Dialect/CMakeLists.txt
... (more deletions)
R third_party/proton/{Dialect => dialect}/include/Analysis/ScopeIdAllocation.h
.... (more renames)

Okay, let me try spliting on the fileset (since everything is still in that one other directory) and abandoning:

❯ jj split 'glob:third_party/proton/**' -m "proton stuff"
Selected changes : xurykvqu 829980f3 proton stuff
Remaining changes: pyyqmzwk 6bcd0874 Claude adding block scaled floating cuBLAS baselines
Working copy  (@) now at: pyyqmzwk 6bcd0874 Claude adding block scaled floating cuBLAS baselines
Parent commit (@-)      : xurykvqu 829980f3 proton stuff

❯ jj st
Working copy changes:
M python/tutorials/10-block-scaled-matmul.py
M third_party/nvidia/include/cublas_instance.h
M third_party/nvidia/include/cublas_types.h
M third_party/nvidia/triton_nvidia.cc
Working copy  (@) : pyyqmzwk 6bcd0874 Claude adding block scaled floating cuBLAS baselines
Parent commit (@-): xurykvqu 829980f3 proton stuff

Perfect! Now to just blow away the stuff I don't want,

❯ jj abandon -r @-
Abandoned 1 commits:
  xurykvqu 829980f3 proton stuff
Rebased 1 descendant commits onto parents of abandoned commits
Working copy  (@) now at: pyyqmzwk 625d7bd7 Claude adding block scaled floating cuBLAS baselines
Parent commit (@-)      : rypswprz 90c2ebbe DONOTMERGE
Added 73 files, modified 0 files, removed 57 files
Warning: 57 of those updates were skipped because there were conflicting changes in the working copy.
Hint: Inspect the changes compared to the intended target with `jj diff --from 625d7bd7f384`.
Discard the conflicting changes with `jj restore --from 625d7bd7f384`.

❯ jj st
Working copy changes:
M python/tutorials/10-block-scaled-matmul.py
M third_party/nvidia/include/cublas_instance.h
M third_party/nvidia/include/cublas_types.h
M third_party/nvidia/triton_nvidia.cc
D third_party/proton/Dialect/include/Analysis/ScopeIdAllocation.h
D third_party/proton/Dialect/include/Conversion/CMakeLists.txt
D third_party/proton/Dialect/include/Conversion/ProtonGPUToLLVM/CMakeLists.txt
...

🤯

My first guess was something different in .gitignore between @ and @- but I didn't find anything there. Also, no other scripts/watchers/processes that should be watching this directory and changing things behind jj's back.

I'm also not sure how to create a reproducer for this. Any suggestions on what to look at/try?

❯ jj version
jj 0.34.0-22900c9a9ba362efa442fed2dd4e6e1d5c22cc7a

On a mac with a local filesystem.

[jennings]

I can't see exactly what happened, but this seems like a pretty good clue to me:

R third_party/proton/{Dialect => dialect}/include/Analysis/ScopeIdAllocation.h

This rename is just changing the casing of the Dialect directory.

The git file system is case-sensitive, but your APFS file system is almost certainly case-insensitive.

This causes weird things to happen in both git and jj. For example, if your working copy is supposed to have two different files checked out:

third_party/proton/dialect/something
third_party/proton/Dialect/something_else

The directory on disk can only have one of the two casings, so one of them must lose and be placed in the "wrong" directory.

If you use jj file list third_party/proton, do you see both casings listed?

[binarybana]

Yes, that directory casing was the issue. jj file list third_party/proton only had the lower case dialect which made me think that maybe the git index was "wanting" it to be upper case.

And sure enough, doing a mv third_party/proton/dialect third_party/proton/Dialect immediately resolved the issues.

Wow, that was confusing since I wasn't touching those files either, they just got picked up through a rebase and hoisted into my change. Is there any git or jj config to prevent this? Or should I file an issue to help warn the user what might be happening in these cases?

[binarybana]

[jennings]

What about the directory on disk?