Mismatched HEAD default behavior from git log

> It appears the way --stdin processes input discards nonexistent
> commits before the machinery that decides whether you provided any
> revs or not runs, and so if every --stdin rev is discarded then you
> get the default HEAD. If you provide them via the command line,
> though, then it seems like they're discarded later and you don't get a
> default.
>
> I'm not sure whether this is intentional or not (certainly I don't see
> it anywhere in the git log documentation for --ignore-missing or
> --stdin), but it results in a behavior mismatch that's impossible to
> reconcile without requiring extra git processes. I can't always
> provide HEAD since, if multiple revs are supplied, if any revs exist
> then HEAD would not be included regardless of whether the revs were
> supplied via the command line or --stdin.

> Bryan Turner <bturner@atlassian.com> writes:
> 
> > It appears the way --stdin processes input discards nonexistent
> > commits before the machinery that decides whether you provided any
> > revs or not runs, and so if every --stdin rev is discarded then you
> > get the default HEAD. If you provide them via the command line,
> > though, then it seems like they're discarded later and you don't get a
> > default.
> >
> > I'm not sure whether this is intentional or not (certainly I don't see
> > it anywhere in the git log documentation for --ignore-missing or
> > --stdin), but it results in a behavior mismatch that's impossible to
> > reconcile without requiring extra git processes. I can't always
> > provide HEAD since, if multiple revs are supplied, if any revs exist
> > then HEAD would not be included regardless of whether the revs were
> > supplied via the command line or --stdin.
> 
> As the intent for adding the "--stdin" option to any subcommand has
> always been "we may need to feed many many things, that may bust the
> command line length limit, hence we let you feed these things from
> the standard input, but otherwise there should be no change in
> behaviour or semantics", when the behaviour of command line and
> "--stdin" differ, it is a bug in the latter.

>> As the intent for adding the "--stdin" option to any subcommand has
>> always been "we may need to feed many many things, that may bust the
>> command line length limit, hence we let you feed these things from
>> the standard input, but otherwise there should be no change in
>> behaviour or semantics", when the behaviour of command line and
>> "--stdin" differ, it is a bug in the latter.
>
> Agreed. It also helps in this case that the command-line behavior is
> sensible and the --stdin one is not. :)
>
> I think the solution is probably something like:

> diff --git a/revision.c b/revision.c
> index 96630e3186..f5bbefa091 100644
> --- a/revision.c
> +++ b/revision.c
> @@ -2099,12 +2099,13 @@ static void read_pathspec_from_stdin(struct strbuf *sb,
>  		strvec_push(prune, sb->buf);
>  }
>  
> -static void read_revisions_from_stdin(struct rev_info *revs,
> -				      struct strvec *prune)
> +static int read_revisions_from_stdin(struct rev_info *revs,
> +				     struct strvec *prune)
>  {
>  	struct strbuf sb;
>  	int seen_dashdash = 0;
>  	int save_warning;
> +	int got_rev_arg = 0;
>  
>  	save_warning = warn_on_object_refname_ambiguity;
>  	warn_on_object_refname_ambiguity = 0;
> @@ -2124,12 +2125,14 @@ static void read_revisions_from_stdin(struct rev_info *revs,
>  		if (handle_revision_arg(sb.buf, revs, 0,
>  					REVARG_CANNOT_BE_FILENAME))
>  			die("bad revision '%s'", sb.buf);
> +		got_rev_arg = 1;
>  	}
>  	if (seen_dashdash)
>  		read_pathspec_from_stdin(&sb, prune);
>  
>  	strbuf_release(&sb);
>  	warn_on_object_refname_ambiguity = save_warning;
> +	return got_rev_arg;
>  }
>  
>  static void add_grep(struct rev_info *revs, const char *ptn, enum grep_pat_token what)
> @@ -2754,7 +2757,8 @@ int setup_revisions(int argc, const char **argv, struct rev_info *revs, struct s
>  				}
>  				if (revs->read_from_stdin++)
>  					die("--stdin given twice?");
> -				read_revisions_from_stdin(revs, &prune_data);
> +				if (read_revisions_from_stdin(revs, &prune_data))
> +					got_rev_arg = 1;
>  				continue;
>  			}
>  
>
> Possibly it would make sense to push that flag into rev_info, though,
> and let handle_revision_arg() set it. That would fix this bug and
> prevent similar ones in other code paths (though we're not likely to get
> revisions from anywhere else, I suppose).
>
> -Peff

> Jeff King <peff@peff.net> writes:
> 
> >> As the intent for adding the "--stdin" option to any subcommand has
> >> always been "we may need to feed many many things, that may bust the
> >> command line length limit, hence we let you feed these things from
> >> the standard input, but otherwise there should be no change in
> >> behaviour or semantics", when the behaviour of command line and
> >> "--stdin" differ, it is a bug in the latter.
> >
> > Agreed. It also helps in this case that the command-line behavior is
> > sensible and the --stdin one is not. :)
> >
> > I think the solution is probably something like:
> 
> You beat me to it while I was wondering what to do between the local
> got_rev_arg variable and the revs->rev_input_given field.

> > You beat me to it while I was wondering what to do between the local
> > got_rev_arg variable and the revs->rev_input_given field.
> 
> That makes me wonder why we need got_rev_arg at all if we have
> revs->rev_input_given. But I suspect an answer can be found by digging
> into git-blame. I probably won't do that immediately, so if you want to,
> you can do so without worrying that we're duplicating work. :)

> Subject: [PATCH] revision: set rev_input_given in handle_revision_arg()
>
> Commit 7ba826290a (revision: add rev_input_given flag, 2017-08-02) added
> a flag to rev_info to tell whether we got any revision arguments. As
> explained there, this is necessary because some revision arguments may
> not produce any pending traversal objects, but should still inhibit
> default behaviors (e.g., a glob that matches nothing).

> However, it only set the flag in the globbing code, but not for
> revisions we get on the command-line or via stdin. This leads to two
> problems:
>
>   - the command-line code keeps its own separate got_rev_arg flag; this
>     isn't wrong, but it's confusing and an extra maintenance burden
>
>   - even specifically-named rev arguments might end up not adding any
>     pending objects: if --ignore-missing is set, then specifying a
>     missing object is a noop rather than an error.