RFC patchAdd test_repo_expect_success for running tests in a new repository

This could be convenient when tests are independent from the rest in the same file. Normally we would do this

test_expect_success '...' '
	git init foo &&
	(
		cd foo &&
		<script>
	)
'

Now we can write a shorter version

test_repo_expect_success '...' '
	<script>
'

The other function, test_subdir_expect_success, expands the script to "( cd <repo> && <script> )", which can be useful for grouping a series of tests that operate on the same repository in a subdir, e.g.

test_expect_success 'create repo abc' 'test_create_repo abc' test_subdir_expect_success abc '...' <script> test_subdir_expect_success abc '...' <another-script>

Signed-off-by: Nguyễn Thái Ngọc Duy <pclouds@gmail.com>
---
 Lately I start to add more and more tests in this style. So this
 looks like a good change to me.

 t/README                | 15 +++++++++++++++
 t/test-lib-functions.sh | 20 ++++++++++++++++++++
 2 files changed, 35 insertions(+)

Nguyễn Thái Ngọc Duy  <pclouds@gmail.com> writes:

Implementations of these helper functions are not all that interesting for reviewers (except for finding unacceptable bits, that is), I would think.

More interesting are how much cleaner the existing code would become. I know we have tons of them that do "create a new, do this and that in the repository".

Need a good doc here for a function that takes variable number of parameters in a funny way.

This bash-ism will never be applied to my tree.

Mental note: Two-arg form is for title and script.

By the way, unquoted subdir makes it too sloppy to live.

Mental note: Three-arg form is for title, prereq and script.

What the heck is this one? That is why I said "implementations are not interesting, the way they help existing ones more readable is". It is totally unclear where the $4 comes from and has to be given in a funny order to the helper (and I am sure it will make perfect sense when it is explained).

Too deep an indent level here.

I do not like a few things in here (besides the same kind of unacceptable stuff as the other one).

Often we need a new repository for testing a handful steps, and we would want to be able to do this sequence:

    - create a new repo
    - do one thing in that repo
    - do another thing in that repo
    - do yet another thing in that repo

That would force tests to say "test_repo_expect_success" to do the first thing (creating and running the first command) and then subsequently do "test_subdir_expect_success $there" to run the remainder.

I think we would only want to have test_expect_success_there (I am avoiding "subdir" because the word has connotations that you do not want in your use case. When we say "subdir" we typically do not mean a separate repository or submodule) without the "auto creation of a repository with unknown name" bit.

	test_expect_success 'set up a new repository for testing' '
		test_create_repo myrepo
	'

        test_expect_success_there mytest 'do one thing there' '
		>empty &&
                git add empty
                git commit -m "add empty"
	'

        test_expect_success_there mytest 'do another thing there' '
		git ls-files >actual &&
                echo empty >expect &&
                test_cmp expect actual
	'

On Sun, Sep 20, 2015 at 08:25:02AM +0700, Nguyễn Thái Ngọc Duy wrote:

I dunno. Like Junio, I would prefer to see some proposed conversions to be able to evaluate it.

But I do not find the first one all that bad. Sure, it is more lines, but it is also a lot more _flexible_. You can "init --bare". You can create a repo, do some setup in it, then do some more stuff outside, and then do some more setup inside it.

I also find it weird to happen at the "test_expect_success" layer, rather than inside, as it seems somewhat orthogonal. E.g., how do I expect failure?

It seems like it would be more flexible as a helper inside the test snippet. But then I cannot think of a way to write it that is much shorter than "(cd x && ...)".

For single git commands, I do find myself increasingly writing the first one as:

  test_expect_success '...' '
	git init foo &&
	git -C foo something
  '

which is pretty short and clear, I think. It also keeps ancillary files at the top-level. So if you do:

  git clone . foo &&
  git rev-parse refs/heads/master >expect &&
  git -C foo rev-parse refs/remotes/origin/master >actual &&
  test_cmp expect actual

all of the "expect" and "actual" appear at the same top-level, and you do not have to worry about mentioning "../expect", etc.

My biggest problem with that technique is that you cannot run test_* functions this way, so:

  git -C foo test_commit bar

does not work, and you have to write:

  (cd foo && test_commit bar)

though IMHO that is not so bad.

Like I said, though, I'd reserve judgement until I saw some proposed conversions.

If we have a whole series of tests that are all supposed to be inside "foo", I wonder if it would be that distasteful to simply "cd foo", run the tests, and then cd back. Of course we'd want to notice if the "cd" failed for whatever reason. Perhaps something like:

  test_pushdir foo
  test_expect_success ...
  test_popdir

where "pushdir" would set an "error" flag if it fails, and any tests inside it would auto-fail without running (which is much safer than trying to run them in the wrong directory!).

This could also be extended to any other type of state mutation besides "cd" if we wanted, though I do not have anything in mind (I guess we could set env variables, but we usually just do that inside a test_expect and assume it worked in subsequent tests).

-Peff