Re: [PATCH 1/2] add: warn when adding an embedded repository

[Jeff King <peff@peff.net>]

On Wed, Jun 14, 2017 at 10:53:12AM -0700, Stefan Beller wrote:

> >> In my ideal dream world of submodules we would have the following:
> >>
> >>   $ cat .gitmodules
> >>   [submodule "sub42"]
> >>     path = foo
> >>   # path only in tree!
> >
> > TBH, I am not sure why we need "path"; couldn't we just use the
> > subsection name as an implicit path?
> 
> That is what was done back in the time. But then people wanted to rename
> submodules (i.e. move them around in the worktree), so the path is not
> constant, so either we'd have to move around the git dir whenever the
> submodule is renamed (bad idea IMO), or instead introduce a mapping
> between (constant name <-> variable path). So that was done.

Ah, right. That makes sense. I forgot that in addition to the in-tree
path, we have to store the submodule repository itself as some name. The
extra level of indirection there isn't strictly necessary, but it lets
the "name" act as a unique id.

> Historically (IIUC) we had submodule.path.url which then was changed
> to submodule.name.url + name->path resolution. And as a hack(?) or
> easy way out of a problem then, the name is often the same as the path
> hence confusing people, when they see:
> 
>     [submodule "foo"]
>         path = foo
>         url = dadada/foo
> 
> What foo means what now? ;)

Right, I am such a person that has been confused. ;)

Thanks for explaining.

> Talking about another tangent:
> 
>   For files there is a rename detection available. For submodules
>   It is hard to imagine that there ever will be such a rename detection
>   as files have because of the explciit name<->path mapping.
> 
>   We *know* when a submodule was moved. So why even try
>   to do rename detection? As we record only sha1s for a submodule
>   you could swap two submodule object names by accident.
>   Consider a superproject that contains different kernels, such as
>   a kernel for your phone/embedded device and then a kernel for
>   your workstation or other device. And these two kernels are different
>   for technical reasons but share the same history.

Do you mean during the rename detection phase of a diff, check to see if
.gitmodules registered a change in path for a particular module (by
finding its entry in the diff and looking at both sides), and if so then
mark that as a rename for the submodule paths?

From a cursory glance, that sounds like an interesting approach.

-Peff