Re: [PATCH 8/9] for-each-ref: add option to fully dereference tags

From: Victoria Dye <vdye@github.com>
Date: Nov 8, 2023, 01:13 UTC
Message-ID: <cf691b7c-288f-4cc9-a2ac-1a43972ae446@github.com>
In-Reply-To: <ZUoWWo7IEKsiSx-C@tanuki>

Patrick Steinhardt wrote:

Show 40 quoted lines

> On Tue, Nov 07, 2023 at 01:26:00AM +0000, Victoria Dye via GitGitGadget wrote:
>> From: Victoria Dye <vdye@github.com>
>>
>> Add a boolean flag '--full-deref' that, when enabled, fills '%(*fieldname)'
>> format fields using the fully peeled target of tag objects, rather than the
>> immediate target.
>>
>> In other builtins ('rev-parse', 'show-ref'), "dereferencing" tags typically
>> means peeling them down to their non-tag target. Unlike these commands,
>> 'for-each-ref' dereferences only one "level" of tags in '*' format fields
>> (like "%(*objectname)"). For most annotated tags, one level of dereferencing
>> is enough, since most tags point to commits or trees. However, nested tags
>> (annotated tags whose target is another annotated tag) dereferenced once
>> will point to their target tag, different a full peel to e.g. a commit.
>>
>> Currently, if a user wants to filter & format refs and include information
>> about the fully dereferenced tag, they can do so with something like
>> 'cat-file --batch-check':
>>
>>     git for-each-ref --format="%(objectname)^{} %(refname)" <pattern> |
>>         git cat-file --batch-check="%(objectname) %(rest)"
>>
>> But the combination of commands is inefficient. So, to improve the
>> efficiency of this use case, add a '--full-deref' option that causes
>> 'for-each-ref' to fully dereference tags when formatting with '*' fields.
> 
> I do wonder whether it would make sense to introduce this feature in the
> form of a separate field prefix, as you also mentioned in your cover
> letter. It would buy the user more flexibility, but the question is
> whether such flexibility would really ever be needed.
> 
> The only thing I could really think of where it might make sense is to
> distinguish tags that peel to a commit immediately from ones that don't.
> That feels rather esoteric to me and doesn't seem to be of much use. But
> regardless of whether or not we can see the usefulness now, if this
> wouldn't be significantly more complex I wonder whether it would make
> more sense to use a new field prefix instead anyway.
> 
> In any case, I think it would be helpful if this was discussed in the
> commit message.

I've been going back and forth on this, but I think a field specifier might be the way to go after all. Using a field specifier would inherently be more complex than the command line option (since the formatting code is a bit complicated), but that's not an insurmountable problem. The thing I kept getting caught up on was which symbol (or symbols?) to use to indicate a full object peel. I mentioned `**fieldname` in the cover letter, but that looks more like a double dereference than a recursive one.

I think `^{}fieldname` would be a good candidate, but it's *extremely* important (for the sake of avoiding user confusion/frustration) that it produces the same object & associated info as the standard revision parsing machinery [1]. One notable difference (it might be the only one) from `*fieldname` would be, if a ref points to a non-tag object, then that object's information would printed (rather than an empty string). But maybe that difference is what we'd want anyway, since it's a better one-for-one replacement of 'git for-each-ref | git cat-file --batch-check'.

I'll try implementing that for V2. If it doesn't work for some reason, though, I'll explain why in the commit message.

[1] https://git-scm.com/docs/git-rev-parse#Documentation/git-rev-parse.txt-emltrevgtemegemv0998em

> 
> Patrick
>