Re: [PATCH 4/6] [GSOC] ref-filter: add %(rest) atom and --rest option

From: Junio C Hamano <gitster@pobox.com>
Date: Jun 9, 2021, 07:00 UTC
Message-ID: <xmqq4ke7jzee.fsf@gitster.g>
In-Reply-To: <xmqqk0n4q1t6.fsf@gitster.g>

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

Show 5 quoted lines

> Is that what you want to add to give "cat-file --batch"?  Even in
> the context of "cat-file --batch", you can throw an object name for
> a blob to the command, but there is no path for the blob (a blob can
> appear at different places in different trees---think "rename), so I
> am not sure what benefit you are trying to derive from it.

I think I kind-of see what is going on here.  There is

    git cat-file blob --textconv --path="$path" "$blob_object_name"

that allows a blob to be fed to the command, pretend as if it appears at $path in a tree object and grab attribute for it, and show the blob contents converted using the textconv filter. If we were to mimic it by extending the format based substitutions, a design consistent with the behaviour is to teach --format=%(raw) to show the contents after applying the textconv filter instead of the raw blob contents.

And there is a corresponding

    git cat-file --batch --textconv

The "--path=$path" parameter is omitted when using --batch, as each object would sit at different path in the tree (so the input stream would be given as a run of "<blob> <path>" to give each item its own path).

So to answer my question in the previous message, yes, this is an attempt to support the "cat-file --textconv". So in the context of that command, something may need to be added. But I do not think it makes any sense to expose that to for-each-ref and friends, even if we were to share the internal machinery (after all, sharing of the internal machinery is a mere means to an end that is to make it easier to give the same syntax and same behaviour to end users and is not a goal itself; "because we use the same machinery, the users have to tolerate that irrelevant %(atoms) are accepted by the parser" is not making a good excuse for a sloppy implementation).

Having said all that, I somehow doubt that the "--batch=<format>" was designed to interact sensibly with the "--textconv" option. builtin/cat-file.c::expand_atom() does not know anything at all that the data could be modified from the raw contents of the blob, so --batch="%(contents) %(size)" --textconv, if existed, may show the conveted contents with size of blob before conversion, or something incoherent like that. And if your rewrite using the shared internal machinery results in a more coherent behaviour, that would be excellent. For example, we could imagine that the machinery, when textconv (or filter) is in use, would first grab the blob contents and run the requested conversion, and then work solely on that conveted contents when computing what to fill with %(raw:size) and other blob-related atoms.

Thanks.