Re: [PATCH 3/4] cat-file --textconv/--filters: allow specifying the path separately

Johannes Schindelin [off-list ref] writes:

There are circumstances when it is relatively easy to figure out the
object name for a given path, but not the revision. For example, when

revision of the containing tree, I presume?

Let's simplify this dramatically, because we do not really need that
revision at all: all we care about is that we know the path. In the
scenario described above, we do know the path, and we just want to
specify it separately from the object name.

I wouldn't call it "simplifying dramatically".  It is just to
support two use cases that is different from the use case where you
want to use "<tree>:<path>".

We already have a precedent in "hash-object --path=<path>" for these
two different uses cases from the primary one.  That form can be
used when we know the contents but we do not know where the contents
came from.  It can also be used when we want to pretend that
contents came from a specific path, that may be different from the
file we are hashing.

Let's be consistent and (1) call it "--path", and (2) explain it as
a feature to allow you to tell the path separately or allow you to
pretend that the content is at a path different from reality.

After all, you would want to ignore <path2> part in this construct:

	git cat-file --filter --path=<path1> HEAD:<path2>

for the purpose of filtering, right?

The mode override is somewhat questionable.  Wouldn't you want to
reject

	git cat-file --filter --path=Makefile HEAD:RelNotes

when HEAD:RelNotes blob is known to be a symbolic link?  After all,
you would reject this

	git cat-file --filter --path=Makefile HEAD:t/

and it is quite similar, no?  I think this can be argued both ways,
but I have a feeling that obj_context.mode, if available, should be
honored with or without force_path.

I think I saw some code to ensure "when giving this option you need
that option in effect, too"; they should be tested here, too, no?

Hi Junio,

On Thu, 18 Aug 2016, Junio C Hamano wrote:

Johannes Schindelin [off-list ref] writes:

quoted

There are circumstances when it is relatively easy to figure out the
object name for a given path, but not the revision. For example, when

revision of the containing tree, I presume?

name of containing tree, actually.

quoted

Let's simplify this dramatically, because we do not really need that
revision at all: all we care about is that we know the path. In the
scenario described above, we do know the path, and we just want to
specify it separately from the object name.

I wouldn't call it "simplifying dramatically".  It is just to
support two use cases that is different from the use case where you
want to use "<tree>:<path>".

"this" was not the code in question. "this" is the use case. Just put my
shoes on for a moment. As described: you have a list of object names and
their path. Now you need to use the --textconv or --filters option, but
you do not have the commit name. What do you do? Concoct a script that
goes through `git rev-list --all -- <path>` in the hopes of finding a
revision soon? I hope after this little exercise you agree that
introducing the --path=<path> option simplifies this use case
dramatically.

We already have a precedent in "hash-object --path=<path>" for these
two different uses cases from the primary one.  That form can be
used when we know the contents but we do not know where the contents
came from.  It can also be used when we want to pretend that
contents came from a specific path, that may be different from the
file we are hashing.

Agreed. I was unaware of hash-object's existing option.

Let's be consistent and (1) call it "--path", and (2) explain it as
a feature to allow you to tell the path separately or allow you to
pretend that the content is at a path different from reality.

After all, you would want to ignore <path2> part in this construct:

	git cat-file --filter --path=<path1> HEAD:<path2>

for the purpose of filtering, right?

True, and I even make use of that in the test for the batch mode.

quoted

diff --git a/builtin/cat-file.c b/builtin/cat-file.c
index 0b74afa..5ff58b3 100644
--- a/builtin/cat-file.c
+++ b/builtin/cat-file.c

@@ -20,6 +20,8 @@ struct batch_options {
 	const char *format;
 };
 
+static const char *force_path;
+
 static int filter_object(const char *path, unsigned mode,
 			 const unsigned char *sha1,
 			 char **buf, unsigned long *size)

@@ -89,21 +91,24 @@ static int cat_one_file(int opt, const char *exp_type, const char *obj_name,
 		return !has_sha1_file(sha1);
 
 	case 'w':
-		if (!obj_context.path[0])
+		if (!force_path && !obj_context.path[0])
 			die("git cat-file --filters %s: <object> must be "
 			    "<sha1:path>", obj_name);
 
-		if (filter_object(obj_context.path, obj_context.mode,
+		if (filter_object(force_path ? force_path : obj_context.path,
+				  force_path ? 0100644 : obj_context.mode,
 				  sha1, &buf, &size))

The mode override is somewhat questionable.  Wouldn't you want to
reject

	git cat-file --filter --path=Makefile HEAD:RelNotes

when HEAD:RelNotes blob is known to be a symbolic link?  After all,
you would reject this

	git cat-file --filter --path=Makefile HEAD:t/

and it is quite similar, no?  I think this can be argued both ways,
but I have a feeling that obj_context.mode, if available, should be
honored with or without force_path.

I see your point. All I wanted to do was to enable

	git cat-file --filters --path=Makefile cafebabe

in which case obj_context.mode == S_IFINVALID. I fixed it (see interdiff
of the next iteration).

quoted

diff --git a/t/t8010-cat-file-filters.sh b/t/t8010-cat-file-filters.sh
index e466634..fd17159 100755
--- a/t/t8010-cat-file-filters.sh
+++ b/t/t8010-cat-file-filters.sh

@@ -31,4 +31,17 @@ test_expect_success 'cat-file --filters converts to worktree version' '
 	has_cr actual
 '
 
+test_expect_success 'cat-file --filters --use-path=<path> works' '
+	sha1=$(git rev-parse -q --verify HEAD:world.txt) &&
+	git cat-file --filters --use-path=world.txt $sha1 >actual &&
+	has_cr actual
+'
+
+test_expect_success 'cat-file --textconv --use-path=<path> works' '
+	sha1=$(git rev-parse -q --verify HEAD:world.txt) &&
+	test_config diff.txt.textconv "tr A-Za-z N-ZA-Mn-za-m <" &&
+	git cat-file --textconv --use-path=hello.txt $sha1 >rot13 &&
+	test uryyb = "$(cat rot13 | remove_cr)"
+'

I think I saw some code to ensure "when giving this option you need
that option in effect, too"; they should be tested here, too, no?

No, I would rather not test for that. These conditionals are purely for
any user's convenience, in case they specify an option that has no effect.
They are absolutely not essential for the function introduced in this
patch series.

Ciao,
Dscho