git switch --force vs --discard-changes: docs don't match behavior

Hi,

https://github.com/git/git/blob/683c54c999c301c2cd6f715c411407c413b1d84e/Documentation/git-switch.adoc?plain=1#L115-L124 describes --force as "An alias for `--discard-changes`."

At least on git 2.47.2 though[1], it looks like --force will force more things than --discard-changes:

$ touch Makefile
$ git switch master
error: The following untracked working tree files would be overwritten 
by checkout:
         Makefile
Please move or remove them before you switch branches.
Aborting
$ git switch --discard-changes master
error: Untracked working tree file 'Makefile' would be overwritten by merge.
$ git switch --force master
branch 'master' set up to track 'origin/master'.
Switched to a new branch 'master'

Is this a bug in the code or documentation?

[1] I searched the mailing list to see if there were any relevant changes since then.

David Mandelberg <david@mandelberg.org> writes:

You'd need to explain this example a bit better. The reproduction must be done in a repository where the master branch has Makefile tracked, you are not on the master branch, and the commit you have checked out does not have Makefile tracked. IOW, this is not making Makefile stat-dirty, but is creating a new untracked file.

As I wasted a few minutes scratching my head before figuring out the above, I'm leaving a note for others.

Makes sense (note.  I do not use "git switch" myself).

I would guess that this "discard changes" takes the word "change" in a reasonbly strict sense (note. I do not use "git switch" myself, "git checkout" does not have this option, and I do not recall thinking about how "--discard-changes" should operate when the feature was added, so this is pretty much my first time to think about this issue). Untracked "Makefile", by definition, is not known by Git, so it cannot even tell if it was changed or not.

If "Makefile" is tracked in the current commit, this would have discarded changes from that modified file and allowed you to switch to the master branch.

And forcing would give us the same behaviour as "checkout -f".

I do not have a strong opinion either way. It may appear to some users that giving a finer grained control is a merit. Even when you are willing to throw away changes to already tracked content, getting stopped when you may lose a totally untracked thing might be nicer.

On the other hand, I suspect to many others this finer grained control does not give much value while adding more confusion.

I am obviously biased because I am accustomed not to have this distinction and accept "checkout -f" as a reasonable way to force switching to another branch discarding any and all local modifications including untracked new files that get in the way, though. But I do not feel strongly enough to say that the behaviour and the feature itself is misguided and we should rip it out. As long as that "finer grained control" is working in a consistent and explainable way, I'd actually vote for fixing the documentation to explain how "--discard-changes" is a bit milder than "--force'.

Thanks.

Junio C Hamano <gitster@pobox.com> writes:

It seems that this was designed from day one to work this way, or this was buggy from day one to work this way. Take your pick ;-)

When 3ec37ad1 (switch: add --discard-changes, 2019-03-29) introduced the --discard-changes option, it seems that it was intended to be truly the synonym, but the code change was done in a somewhat confusing way and added an extra bit (a new .discard_changes member was added, and is initialized by the value given to the .force member. Many existing conditionals that looked at the .force bit was changed to look at the .discard_changes bit, but not all of them.

It did not offer the protection against overwriting untracked files in builtin/checkout.c:reset_tree(). The .reset member was always set to 1 to allow it.

Later 480d3d6b (Change unpack_trees' 'reset' flag into an enum, 2021-09-27) added the "protect against overwriting untracked files" feature, and made it pay attention to the "force" option, not the other member. That is where this discrepancy was introduced. In retrospect, it might have been better if 3ec37ad1 removed .force and kept only "--force" at the UI layer, internally used the new .discard_changes everywhere. Or it would have been even better if the commit did not happen at all, and added "--discard-changes" only at the UI layer. Then the discrepancy would never have happened.

The author of 480d3d6b Cc'ed to ask for the original intention for reference, but that is merely to satisfy my curiosity. At this point, I think it is way too late for the original intention to sway the conclusion, as existing users of "switch" may be relying on the current behaviour. The best course of action may be to update the documentation to match the behaviour.

Thanks.

Op 2025-03-17 om 13:19 schreef Junio C Hamano:

Sorry about that. I had originally meant to copy/paste all the commands I ran, but the `git fetch` command produced a lot of output and I decided to only paste the ones that showed the bug, not the setup ones. Here are the commands I had run, starting in an empty directory:

git init git remote add origin https://github.com/git/git.git git fetch origin # a few status and show commands, but those hopefully aren't important

So I actually didn't have any commit checked out, but I think having a commit checked out that did not have a Makefile would have done the same thing.

Makes sense. For what it's worth, I don't have much of an opinion one way or another. I want some way to do what --force currently does, but I can always use `git checkout` for that if needed.