patchsubmodule: fix fetch_in_submodule logic

This is a fix on top of ab/retire-parse-remote, which is now in next. I think submodule fetching is pretty broken, so we should do this or something like it soon.

-- >8 -- Commit 1c1518071c (submodule: use "fetch" logic instead of custom remote discovery, 2020-11-14) rewrote the logic in fetch_in_submodule to do:

  elif test "$2" -ne ""

But this is nonsense in shell: -ne is for numeric comparisons. This should be "=" or more idiomatically:

  elif test -n "$2"

But once we fix that, many tests start failing. Because that commit introduced another problem. The caller that passes 3 arguments looks like this:

    fetch_in_submodule "$sm_path" $depth "$sha1"

Note the unquoted $depth parameter. When it isn't set, the function will see only 2 arguments, and the function has no idea if what it sees in $2 is an option to go on the command line, or a refspec to pass on stdin. In the old code before that commit:

   fetch_in_submodule () (
        sanitize_submodule_env &&
        cd "$1" &&
  -     case "$2" in
  -     '')
  -             git fetch ;;
  -     *)
  -             shift
  -             git fetch $(get_default_remote) "$@" ;;
  -     esac

we treated those the same, so it didn't matter. But in the new logic (with my fix above):

  +     if test $# -eq 3
  +     then
  +             echo "$3" | git fetch --stdin "$2"
  +     elif test -n "$n"
  +     then
  +             git fetch "$2"
  +     else
  +             git fetch
  +     fi

we use the number of parameters to distinguish the two. Let's insist that the caller pass an empty string for positional parameter two if they want to have a third parameter after it.

But that still leaves one problem. In the --stdin block, we unconditionally pass "$2" to git-fetch, even if it's the empty string. Rather than add another conditional, we can use :+ parameter expansion to include it only if it's non-empty. In fact, we can do the same for the elif, too, simplifying it further. Technically this is overkill, since we know the --depth parameter will not have whitespace (and indeed, most callers do not bother quoting it), but it doesn't hurt for the function to be careful.

It's somewhat amazing that no tests were failing. I think what happened is that:

  - the 3-arg form rarely triggered; any call with a non-empty $depth
    and a $sha1 would work, but one with an empty $depth would only have
    2 arguments

  - because of the wrong arguments to "test", the shell would complain
    and exit non-zero. So we never ran the middle conditional at all

  - that left every call running "git fetch" with no arguments. A
    well-written test could have detected the distinction here, but in
    practice omitting --depth just means fetching more commits, and
    fetching everything (rather than a single sha1) works as long as the
    commit in question is reachable

Signed-off-by: Jeff King <peff@peff.net>
---
A few side notes:

  - I imagine the mixup between "ne" and "=" is because they have the
    opposite meaning in perl. :)

  - it probably wouldn't hurt to beef up the tests, especially around
    fetching an unreachable sha1, but after getting lost for an hour in
    the spaghetti of the submodule code and its tests, I gave up. I do
    at least feel this code is being exercised (because once the initial
    problem is fixed, tons of things fail).

 git-submodule.sh | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

Jeff King <peff@peff.net> writes:

Thanks for catching. I thought we stared at this part long enough to be updated between the rounds; it is embarrassing that we've missed it.

True.

Makes sense.  Thanks.