Re: [PATCH 0/8] CMake build system for git

Hi Junio,

On Fri, 24 Apr 2020, Junio C Hamano wrote:

Would your position change if Sibi was interested in maintaining this, and helping with keeping CMake support in shape?

If you have a look at https://github.com/git-for-windows/git/pull/2580 and at https://github.com/gitgitgadget/git/pull/614, you will probably get the same impression as I: in contrast to earlier contributors who just talked about adding CMake support, Sibi went ahead and did it, and then polished the heck out of the patches, was very responsive and active.

I have a pretty good feeling about this.

To be honest, I don't think that that happens all that often, and quite honestly, I think anybody who has had a look at the Makefile and at CMakeLists.txt will have a pretty good idea where new files should go.

When it comes to new Makefile knobs, I do agree that it would place an unacceptable burden on contributors if we expected them to add the same knob to CMakeLists.txt. But we already don't do that for our autoconf support, so why would we expect it for CMake?

When it comes to adding new, and/or removing, files, I fail to see the problem. It is dead easy to keep the Makefile and CMakeLists.txt in sync when it comes to lists of files.

Another thing that strikes me as much more favorable this time as compared to earlier attempts at adding CMake support is that we now have pretty good automated builds that will definitely help with keeping CMake support up to date.

So now that I have covered the "it's not all that bad to maintain this" angle, let's cover the much more important part: "why would we even bother with this?". And that's pretty easy to answer for me: first-class Visual Studio support.

It is totally unacceptable to expect a contributor to download three quarters of a gigabyte and then let the Git for Windows SDK occupy around two full gigabytes on disk, just to get going. This is an upfront cost we currently expect from any developer on Windows, whether someone familiar with Visual Studio or more familiar with the command-line and GCC or not.

And guess what? Whenever I try to encourage a developer who uses Git and who is pretty good with C to "just check out the Git for Windows SDK, run `make vcxproj`, then load the generated Visual Studio solution and then start developing", the odds are really, really, really high that that's the last of it.

This is frustrating. And not only for me: Why should a contributor fight so hard just to get started, when all they want to do is to tinker with the code in order to see whether there is a chance that they can easily make it do what they want?

For some time now, I have an automated build which auto-generates the `vs/master` branch from Git for Windows' `master` branch, which has all of this preconfigured. But it is too hard to find, and it is definitely too hard to go from there to contributing the patches back to the Git project, so I think that maybe one contributor used that venue in the last years. One. Just one. That's very little return for the effort spent on getting Visual Studio support back to functional ever after it was broken with Visual Studio 2010's release. And it demonstrates a level of being uninviting that I am simply uncomfortable with.

There's more.

If you are a developer who feels at home in IDEs, running `make` in the command-line, and finding out about options in the Makefile and how to specify them on the command-line, is just very awkward. CMake recognizes this, and offers GUIs to configure the builds. So that's nice.

But there's more.

It would appear that in recent years, ARM64-based Windows laptops have made great progress. They are affordable, they are fast, and they have long battery life. If I could use and develop Git for Windows on one of these beauties, I would.

But I can't, because the obstacles to building Git on ARM64 Windows are just too high. The mingw-w64 version of GCC "does not really support" that. We have literally no ARM64 support whatsoever from MSYS2's side, i.e. no Bash, no Perl, no OpenSSH.

And then Sibi comes along and offers CMake support for Git.

CMake can target Visual Studio and clang. Both have pretty good support for ARM64 Windows.

So while this is not a complete solution for the ARM64 Windows problem, it is at least the most exciting opportunity in that regard that I have seen in _years_.

Given all these benefits, I would like to believe that the cost is rather minimal in comparison.

And that's coming from me, who is not a fan of CMake at all.

Of course, I do hope that I have convinced you above that it might not look all that bleak.

But even if I haven't, I would like to propose to run with Sibi's patches and merge them first to `pu`, and then to `next`, and let that cook for a while (I will merge them into Git for Windows early so that there is also some support stream from that side).

I know you think that the maintenance cost is too high, but I think it might be more than just manageable. And I don't think that the risk is too high to give the patches a try, at least inside `next`, for a couple of weeks or even months.

If it turns out that they _do_ add too much of a maintenance burden, big deal: we just drop the patches, and that's that. No hard feelings, we gave it a try, a scientific test, if you want, and we now have evidence to back up your initial suspicion.

If it turns out that they _do_ add value _and_ are easy to maintain, then that's good, right? And then, at your leisurely leisure, you can merge them down to `master` and eventually into an official release.

What do you think? Doesn't that sound like a good plan?

Ciao, Dscho