git: set update = none for all submodules so that cargo git dependencies don't clone them

[kkysen]

• c2rust-side fix for the git dependency version of Avoid /opt path dependency for c2rust-bitfields GaloisInc/Tractor-Crisp#46.

cargo always recursively clones git dependencies, even though c2rust git dependencies don't need any of the big submodules here used just for testing. Setting update = none is workaround (see rust-lang/cargo#10717). However, other users need to explicitly update submodules now.

With this change, it takes ~6 s to resolve a c2rust-bitfields git dependency, vs. many minutes before with all of the large submodules.

[ahomescu]

Do we really need this? I feel like it could create other problems, only to optimize a rare download.

[kkysen]

Do we really need this? I feel like it could create other problems, only to optimize a rare download.

I'm not sure if we really need this, but git dependencies aren't that rare. A bunch of other crates have done this workaround, too. So IMO it seems useful to me on its own, and didn't use to have this behavior before integrating c2rust-testsuite into this repo, so it's new breakage as well.

[randomPoison]

Unfortunately this is fallout from merging the test suite into the main repo. We did that to improve the development workflow, but it does have the consequence that when using the repo as a git dependency cargo is going to try pull down all the submodules. The cargo docs specifically call out update = none as the way to prevent this. Considering that using git dependencies isn't all that uncommon, especially for in-development projects where the latest code may not be on crates.io yet, it seems reasonable that we setup the repo to not download submodules automatically.

One suggestion to make things a bit nice to work with: Could we update test.py to automatically run git submodule update --init --checkout if the repo dir is empty? That'd save us a bit of repetition in the CI script and would make the workflow for running integration tests in a fresh checkout less error-prone.

[kkysen]

One suggestion to make things a bit nice to work with: Could we update test.py to automatically run git submodule update --init --checkout if the repo dir is empty? That'd save us a bit of repetition in the CI script and would make the workflow for running integration tests in a fresh checkout less error-prone.

Yeah, that could probably work.

Also, I realized that cargo install --git https://github.com/immunant/c2rust/, which I feel is actually pretty common, now takes forever.

[ahomescu]

Could we add a comment somewhere that we're doing this because cargo doesn't let us control whether it updates the submodules?

[ahomescu]

For reference: I think cargo did the wrong thing here, and should have provided a way to download a dependency without submodule.

[kkysen]

Could we add a comment somewhere that we're doing this because cargo doesn't let us control whether it updates the submodules?

Yeah, sure. You want me to add more here: https://github.com/immunant/c2rust/blob/a101fbe53ee2902531bfa6df756e79c792703ea7/.gitmodules?

For reference: I think cargo did the wrong thing here, and should have provided a way to download a dependency without submodule.

I agree. Harder to change that, though. Especially since we're stuck on an older nightly.

[ahomescu]

The cargo issue discussing submodules: rust-lang/cargo#4247 (still open).

[ahomescu]

LGTM (other than wishing cargo let us do it the other way). Handing it over to @randomPoison since there's some more feedback from her (I wish github had gerrit's +1 votes).