Re: [PATCH 1/2] blob.c: remove buffer & size arguments to parse_blob_buffer()

On Sat, Apr 10 2021, Ævar Arnfjörð Bjarmason wrote:

On Fri, Apr 09 2021, Jeff King wrote:

quoted

On Fri, Apr 09, 2021 at 10:07:27AM +0200, Ævar Arnfjörð Bjarmason wrote:

quoted

As noted in the comment introduced in 837d395a5c0 (Replace
parse_blob() with an explanatory comment, 2010-01-18) the old
parse_blob() function and the current parse_blob_buffer() exist merely
to provide consistency in the API.

We're not going to parse blobs like we "parse" commits, trees or
tags. So let's not have the parse_blob_buffer() take arguments that
pretends that we do. Its only use is to set the "parsed" flag.

See bd2c39f58f9 ([PATCH] don't load and decompress objects twice with
parse_object(), 2005-05-06) for the introduction of parse_blob_buffer().

OK. Calling it parse_blob_buffer() is a little silly since it doesn't
even take a buffer anymore. But I guess parse_blob() might imply that it
actually loads the contents from disk to check them (which the other
parse_foo() functions do), so that's not a good name.

So this might be the least bad thing. Given that there are only two
callers, just setting blob->object.parsed might not be unreasonable,
either. But I don't think it's worth spending too much time on.

quoted

@@ -266,7 +266,7 @@ struct object *parse_object(struct repository *r, const struct object_id *oid)
 			error(_("hash mismatch %s"), oid_to_hex(oid));
 			return NULL;
 		}
-		parse_blob_buffer(lookup_blob(r, oid), NULL, 0);
+		parse_blob_buffer(lookup_blob(r, oid));
 		return lookup_object(r, oid);

Not new in your patch, but I wondered if this could cause a segfault
when lookup_blob() returns NULL. I _think_ the answer is "no". We'd hit
this code path when either:

  - lookup_object() returns an object with type OBJ_BLOB, in which case
    lookup_blob() would return that same object

  - lookup_object() returned NULL, in which case lookup_blob() will call
    it again, get NULL again, and then auto-create the blob and return
    it

So I think it is OK. But there are a bunch of duplicate hash lookups in
this code. It would be clearer and more efficient as:

diff --git a/object.c b/object.c
index 2c32691dc4..2dfa038f13 100644
--- a/object.c
+++ b/object.c

@@ -262,12 +262,14 @@ struct object *parse_object(struct repository *r, const struct object_id *oid)
 	if ((obj && obj->type == OBJ_BLOB && repo_has_object_file(r, oid)) ||
 	    (!obj && repo_has_object_file(r, oid) &&
 	     oid_object_info(r, oid, NULL) == OBJ_BLOB)) {
+		if (!obj)
+			obj = create_object(r, oid, alloc_blob_node(r));
 		if (check_object_signature(r, repl, NULL, 0, NULL) < 0) {
 			error(_("hash mismatch %s"), oid_to_hex(oid));
 			return NULL;
 		}
-		parse_blob_buffer(lookup_blob(r, oid), NULL, 0);
-		return lookup_object(r, oid);
+		parse_blob_buffer(obj, NULL, 0);
+		return obj;
 	}
 
 	buffer = repo_read_object_file(r, oid, &type, &size);

but I doubt the efficiency matters much in practice. Those hash lookups
will be lost in the noise of computing the hash of the blob contents.

I was trying to keep the changes smaller, but what about just doing this?:

Sent a bit too soon...:

Well, aside from this segfault think-o introduced while
experimenting. Needs to be:

@@ -262,12 +261,16 @@ struct object *parse_object(struct repository *r, const struct object_id *oid)
 	if ((obj && obj->type == OBJ_BLOB && repo_has_object_file(r, oid)) ||
 	    (!obj && repo_has_object_file(r, oid) &&
 	     oid_object_info(r, oid, NULL) == OBJ_BLOB)) {
+		if (!obj) {
+			struct blob *blob = create_blob(r, oid);
+			obj = &blob->object;
+		}
 		if (check_object_signature(r, repl, NULL, 0, NULL) < 0) {
 			error(_("hash mismatch %s"), oid_to_hex(oid));
 			return NULL;
 		}
-		parse_blob_buffer(lookup_blob(r, oid), NULL, 0);
-		return lookup_object(r, oid);
+		obj->parsed = 1;
+		return obj;
 	}
 
 	buffer = repo_read_object_file(r, oid, &type, &size);