Downstream syncs are broken

[lnicola]

Josh created a non-roundtrip push! Do NOT merge this into rustc!
Expected 21f8445ea523e83cd4f11b0a67a3a5ced2b1f56f, got 79aa5a3dea3879af10ca2e58925177febdacce8e.

Caused by the library subtree shenanigans.

[lnicola]

@Veykril did you use git-subtree?

Not really sure how to investigate this, but the first difference I can see is:

commit a4e71b13646cd9c61a1b1bba085cc8b6f1a8add5
Author:     Lukas Wirth <me@lukaswirth.dev>
AuthorDate: Fri Nov 7 20:01:13 2025 +0100
Commit:     Lukas Wirth <me@lukaswirth.dev>
CommitDate: Fri Nov 7 20:01:13 2025 +0100

    Subtree smol_str

commit f46ac05406fda4c4a86747c0188aa5aeff84b6a1
Merge: 4bf516ee5a 02336c59be
Author:     Lukas Wirth <me@lukaswirth.dev>
AuthorDate: Fri Nov 7 20:01:13 2025 +0100
Commit:     Lukas Wirth <me@lukaswirth.dev>
CommitDate: Fri Nov 7 20:01:13 2025 +0100

    Subtree smol_str

So the expected commit has no history for smol_str, while the actual one does. I'm guessing we need to tell josh about the new subtree, but I'm not sure how. CC @RalfJung do you have any tips?

[Veykril]

Oh I did, I figured it wouldn't matter here given this was going to be a one time thing 😅

[RalfJung]

Sorry, which subtree? I thought you're using josh?

Cc @Kobzol

[lnicola]

We've transplanted three more repositories into this one, under /lib, so we might have to update the filter. I tried :rev(55d9a533b309119c8acd13061581b43ae8840823:prefix=src/tools/rust-analyzer):/src/tools/rust-analyzer:rev(4bf516ee5a960c1e2eee9fedd9b1c9e976a19c86:prefix=lib/smol_str):/lib/smol_str but that's probably not the right SHA.

[RalfJung]

Ah, you've added other subtrees into this repo with git subtree? That's a brave experiment.^^ That puts the subtree commits into this repo unchanged, as ancestors of the subtree merge commit, so the history is really odd-looking as the files seem to all suddenly move from one place to another with that merge commit.

I have no idea how to make such commits round-trip through josh, sorry. A non-roundtrip josh sync is generally a josh bug so you may want to report this upstream.

[RalfJung]

I tried :rev(55d9a53:prefix=src/tools/rust-analyzer):/src/tools/rust-analyzer:rev(4bf516e:prefix=lib/smol_str):/lib/smol_str but that's probably not the right SHA.

That doesn't make sense, this would extract only the src/tools/rust-analyzer/lib/smol_str history, you want the entire RA history. The filter describes how the rustc repo history should be transformed to compute this repo's history.

I don't see how a different filter could help. The weird smol_str commits (02336c5 and its parents) should probably end up in rustc under the src/tools/rust-analyzer folder (otherwise there's no way they will end up in this repo again). But whether josh is able to construct such a history, I don't know -- you repo has multiple commits without parents now, so josh has to synthesize a new commit without parent on a push, and I don't know if it is capable of that or what the algorithm would even be for it to figure that out.

[lnicola]

It's a one-off, so I think what we want should be achievable in some way:

• rewrite the history of the other repo to put everything into a subdirectory (git filter-branch and maybe git subtree can probably do this already?)
• import that into the rust-analyzer repo, anchored at whatever commit

So we don't have to use josh and we can tolerate a force-push if needed and there's another way to do this.

[RalfJung]

I don't know if the root cause is the new no-parent commit, or the strange shape of that merge commit. So, hard to say without trying it. Might be best to create a test branch from before those syncs that adds a single new no-parent commit and merges it with a normal-looking merge, just to see what Josh does with that.

[Veykril]

Ugh so sorry for that ....

[Kobzol]

I'd suggest pausing any rustc sync pulls and pushes of RA before this is resolved, because in the worst case it might require a force push to RA 😅

Regarding how to fix this, I'd try what Inicola suggested. If it's a one-off, we can rewrite the history of the subtrees and then just merge them into this repo as a normal merge commit, rather than messing with subtrees.

[RalfJung]

Fwiw it might be worth trying to merge the subtrees in the rustc repo - josh pull is a much easier operation than josh push. But that would definitely involve an RA repo force push then. And obviously it should be tested before landing in the rusct repo as force pushes there are not really a thing.

[RalfJung]

What would likely happen then is that the smol_str history only ends up in the rustc repo since it doesn't look like it has anything to do with RA.

[lnicola]

Might be best to create a test branch from before those syncs that adds a single new no-parent commit and merges it with a normal-looking merge, just to see what Josh does with that.

I tried this:

$ git branch test-merge 4bf516ee5a960c1e2eee9fedd9b1c9e976a19c86 # last commit before smol_str
$ git checkout --orphan empty
$ git rm -rf .
$ touch xyzzy
$ git add -A xyzzy
$ git ci -am "First"
$ git switch test-merge
$ git merge --allow-unrelated-histories empty
$ rustc-josh-sync push [...]
[...]
    Josh created a non-roundtrip push! Do NOT merge this into rustc!
    Expected 6d7e78d196a4b8482d8537283d2b04dddc1e4fbc, got 7ff61f7cdd2b10c2bcf3379911180cb8a9e3ca9a.