Hi Torsten,

On Tue, 9 Aug 2016, Torsten Bögershausen wrote:

> >   git -c core.autocrlf=$crlf add $fname >"${pfx}_$f.err" 2>&1
> > 
> > would make more sense. We _know_ that we have to do convert_to_git() in
> > that step because the content is changed. And then you can ignore the
> > warnings from "git commit" (which are racy), or you can simply commit as
> > a whole later, as some other loops do.
> > 
> > But like Dscho, I do not actually understand what this test is checking.
> > The function is called commit_chk_wrnNNO(), so perhaps you really are
> > interested in what "commit" has to say. But IMHO that is not an
> > interesting test. We know that if it has to read the content from disk,
> > it will call convert_to_git(), which is the exact same code path used by
> > "git add". So I do not understand what it is accomplishing to make a
> > commit at all here.
> 
> It seems as if the test has been written without understanding the raciness.
> It should commit files with different line endings on top of
> a file with mixed line endings.
> The warning should be checked (and here "git add" can be used,
> or the file can be commited directly).
> I'm not sure why the test ended up in doing both.
> 
> However, doing it the right way triggers a bug in convert.c,
> (some warnings are missing, so I need some days to come up
> with a proper patch)

FWIW I would strongly prefer to use the warning of `git add` and not even
bother with `git commit`. What we are interested in is the warning
message, generated by convert_to_git(). Not using the first one and
triggering a second one merely adds unnecessary churn that increases the
CO2 budget of running the test.

On that matter, I wonder whether there would be a chance to revamp t0027
in a major way, with the following goals:

- to make it very obvious to the casual reader what is being tested

- to combine Git invocations when possible, e.g. running one big `git add`
  on a couple of files and then verify the relevant parts of the output

- dramatically decreasing the time required to run the test, without
  sacrificing correctness (I would wager a bet that not only a few of
  those 1388 test cases essentially exercise identical code paths)

Ciao,
Dscho