Re: [QUESTION] how to diff one blob with nothing

From: ZheNing Hu <adlternative@gmail.com>
Date: Aug 5, 2023, 08:27 UTC
Message-ID: <CAOLTT8SBAb8bAV3uCPJ-q2btNeatMjmtRzDkRASffN5xTS1M0g@mail.gmail.com>
In-Reply-To: <xmqqmsz6ljk4.fsf@gitster.g>

Junio C Hamano <gitster@pobox.com> 于2023年8月5日周六 03:00写道：

Show 55 quoted lines

>
> Taylor Blau <me@ttaylorr.com> writes:
>
> > On Fri, Aug 04, 2023 at 10:28:53AM +0200, Christian Couder wrote:
> >> On Fri, Aug 4, 2023 at 6:42 AM ZheNing Hu <adlternative@gmail.com> wrote:
> >>
> >> > Actually, there is no need to support a default empty blob.
> >> > For example, with the command "git diff --no-index <file> /dev/null",
> >> > it can compare a file with /dev/null, but it can only compare <file>
> >> > and not <oid>.
> >> > Therefore, using commands like "git diff <oid> /dev/null",
> >> > "git diff --no-index <oid> /dev/null", or even "git diff <oid> --stdin"
> >> > could potentially solve this issue.
> >>
> >> Maybe it would be clearer to have a new option, called for example
> >> "--blob-vs-file", for that then. It could support both:
> >>
> >> $ git diff --blob-vs-file <blob> <file>
> >>
> >> and:
> >>
> >> $ git diff --blob-vs-file <file> <blob>
> >
> > Hmm. This feels like a case of trying to teach 'git diff' to do too
> > much.
>
> Worse yet, I do not quite get the original use case in the first
> place.  What is the series of diff output that result in comparing a
> random pair of blob object names going to be used for?
>
> The reply to <ZMKtcaN7xYaTtkcI@nand.local> says that the original
> use case was to express the evolution of a single path since its
> creation until its removal, but the thing is, a diff with an empty
> blob and a creation or a deletion event are expressed differently in
> the patch output, exactly because the patch has to be able to
> express "before this change, a file with zero byte content was
> there" and "before this change, there was nothing at this path"
> (vice versa for contents-removal vs deletion).
>
> For that reason, I have a hard time to find any merit in the earlier
> complaint that said "can be achieved by manually adding them, but it
> is not very compatible with the original logic", whatever the
> "original logic" refers to.  If creation needs to be recorded as
> creation and not as a change from an empty and existing blob, there
> has to be something that needs to be manually done to turn the
> latter (which is the only thing "diff" between two blobs or even a
> blob and a file can give) into the former *anyway*.  Whatever the
> thing that is looping over the history/evoluation of a single path
> needs to have a three-arm switch for each iteration to deal with
> creation, modification, and removal, and iterating over the contents
> of the files and prefixing "+" or "-" on each and every line would
> be the _easiest_ part of such a necessary tweak to turn "diff
> between an empty contents and something else" into "creation or
> deletion of a file."
>

Okay, let me clarify the background for using an empty blob diff. Essentially, it is a git web diff interface that requires real-time calculation of the diff between files across multiple versions and rendering them. Due to some reasons, the higher-level component did not provide multiple versions of commits but instead provided blob OIDs (Object IDs).

Therefore, I expected to generate the diff results directly using the
"git diff <oid> <oid>" command. (I only care about the patch part
in the diff and don't really care about the related information of OIDs
in the diff output.) Everything went smoothly except when a blob
is created or deleted, as there is no direct way to obtain
the diff for a blob using the "git diff <oid> <oid>" interface. Initially,
I intended to generate a patch by diffing an empty blob
 (e69de29bb2d1d6434b8b29ae775ad8c2e48c5391) with the blob ID.
However, unlike the empty tree (4b825dc642cb6eb9a060e54bf8d69288fbee4904),
the empty blob does not exist in the git repository by default.

I had to create an additional empty blob for the purpose of performing the diff, but this goes against the design of the web-based diff interface, which is intended to be read-only.

So, I might explore methods like "git diff <oid> /dev/null" or
 "git diff <oid> --stdin", they are read only, but it does not
currently exist...

Regardless, I hope that the empty blob diff can simulate the effect of an empty tree diff:

git diff 4b825dc642cb6eb9a060e54bf8d69288fbee4904

diff --git a/.cirrus.yml b/.cirrus.yml
new file mode 100644
index 0000000000..4860bebd32
--- /dev/null
+++ b/.cirrus.yml
@@ -0,0 +1,22 @@
+env:
+  CIRRUS_CLONE_DEPTH: 1
+
+freebsd_12_task:
+  env:
+    GIT_PROVE_OPTS: "--timer --jobs 10"
+    GIT_TEST_OPTS: "--no-chain-lint --no-bin-wrappers"
+    MAKEFLAGS: "-j4"
+    DEFAULT_TEST_TARGET: prove
+    DEVELOPER: 1
+  freebsd_instance:
+    image_family: freebsd-12-3
+    memory: 2G
+  install_script:
+    pkg install -y gettext gmake perl5
+  create_user_script:
+    - pw useradd git
+    - chown -R git:git .
+  build_script:
+    - su git -c gmake
+  test_script:
+    - su git -c 'gmake test'

Thanks,
ZheNing Hu