Re: [PATCH v3] Support ent:relative_path

From: Junio C Hamano <junkio@cox.net>
Date: May 5, 2007, 03:38 UTC
Message-ID: <7vwszolz26.fsf@assigned-by-dhcp.cox.net>
In-Reply-To: <Pine.LNX.4.64.0705050324580.4015@racer.site>

Johannes Schindelin <Johannes.Schindelin@gmx.de> writes:

Show 5 quoted lines

> Think of the "todo" branch, for example. And this is not even far fetched. 
> In many repositories I have to have separate related, but non-branch 
> HEADs. For example, I track some projects with custom scripts. These 
> scripts do _not_ belong into that branch. However, they are related, so I 
> store them in the same repository.

I said I won't talk about it until 1.5.2 final, but you forced me into it.

> Plus, my example of the bare-repository still has not been answered by 
> _anyone_ in favour of changing the current behaviour.

I am not sure if there is anything to answer on this one.

When you are talking to a bare-repository, your place relative to the root of the working tree is clearly undefined, and it is natural that "<tree-ish>:./<path>", "<tree-ish>:/<path>" and "<tree-ish>:<path>" cannot mean anything but relative to the root of the tree-ish. "<tree-ish>:../<path>" is obviously nonsense. So it does not matter if unadorned <path> is relative to cwd or root in this case.

Your other example, however, gives a much better illustration. An unrelated 'todo' branch where your cwd does not have any relevance to the contents of that tree-ish, or worse yet, a commit from git-gui project in git.git repository, where taking your cwd into account has an actively wrong effect, demonstrates why we would need a way to say "By this path, I mean from the top, I do not want you to take it as relative to where I am".

One way to ensure that is to keep the current "it is a path from the top" behaviour, and extended it with "... unless it begins with ./". Doing this forever however penalizes the case where you want to use relative paths by requiring ./ at the beginning.

Another is to do the usual POSIXy path interpretation and "unless it begins with /, it is taken as relative to where you are". This penalizes the 'todo' and git-gui commit use case because the user explicitly needs to say "where I am does not matter" by prefixing the path with '/', and also necessitates a change to the syntax for looking backwards for a commit with that message, because the existing syntax to look for a string clashes with it [*1*].

Both have merits and demerits. If we did not have any existing code and users, the latter is clearly what we would have done, as it is more consistent. The path handling feels more natural (in line with the way we expect paths to be handled on POSIX systems), the "look backwards" search feels more natural ( you use '/' for forward search, '?' for backwards).

I also suspect the latter is more often convenient. When working on a flat project, it does not matter if the default is relative to cwd or to the root. But if your project is deep, and if you somehow do "git show" more often than "git diff" (I don't, but different people may do so for different reasons), it would start to hurt if you always have to say "./".

It is however clearly a bigger change to existing users. Correcting earlier mistakes is painful, so it certainly is tempting to take the approach that the path is always absolute and require "./" for relative. I agree it is an easier change, but I am not convinced yet that it is the right design in the longer term.

[Footnote]

*1* This is only true for looking for a path in the index case, as <commit>:/<string> does not seem to work. I think this is a bug in the current code -- shouldn't it limit the search to commits that are reachable from that named one?