Re: [PATCH] merge-recursive: use fspathcmp() in path_hashmap_cmp()

[Johannes Schindelin <Johannes.Schindelin@gmx.de>]

Hi Peff,

On Mon, 13 Sep 2021, Jeff King wrote:

> On Mon, Sep 13, 2021 at 01:37:48PM +0200, Johannes Schindelin wrote:
>
> > > Good point.  So fspathcmp() and friends would need a repo parameter. :-|
> >
> > Yes, we will eventually have to pass `struct repository *r` into a _lot_
> > of call chains. It'll be a disruptive change, yet if the submodule folks
> > truly want to aim for in-process recursive treatment of submodules, there
> > is no alternative.
> >
> > FWIW on Windows there are other potentially repository-specific settings
> > that are relevant in similar situations. For example, there is
> > `core.symlinks`.
>
> Another approach is to stuff the appropriate globals into the repository
> struct, and then "push" onto the global the_repository pointer, treating
> it like a stack. And then low-level code is free to use that global
> context, even if it wasn't passed in.
>
> That helps the primary use case of "now I need to do something in a
> sub-module, but I'd like to do it in-process". But it's not without
> challenges:
>
>   - code which acts at the boundary of a submodule and a superproject
>     may be more awkward (since only one of them can be "the current
>     repository" at a time).
>
>   - it's a challenge with threading (an obvious problem would be a
>     multi-threaded grep which wanted to descend into a submodule). Using
>     a thread-local global for the_repository might solve that.
>
> It's possible that this is a terrible direction to go, so I'm not
> necessarily endorsing it, but just offering it as a possibility to think
> about. The trickiest thing is that any devil would likely be in the
> details, and we wouldn't know until proceeding for a while along that
> path. Whereas passing around a context struct, while verbose and
> annoying, is a well-understood construct.

I would not so far as to call it a terrible direction. It is definitely
worth a thought or two.

At the end of the day, I fear that it is too tricky in practice, though.

Seeing as there seems to be some appetite for refactoring Git's code on
this list, I am thinking that the `struct repository *r` direction might
be the one to go for. And I mean like "move the globals into that struct"
as opposed to introducing that stack you talked about. It would even be a
refactoring where I would understand the motivation, and agree with it,
too.

Ciao,
Dscho