[llvm] [workflows] Port buildkite Windows config to GitHub actions (PR #82093)

[Mehdi Amini via llvm-commits]

https://github.com/joker-eph commented:

It's worth noting that there are a few variables here:

- Which path in the repo this triggers on
- The mapping from a path to a given project
- The **dependent** projects that are kicked in based on the affected projects from the path mapping.

That is: minimally speaking it is very possible to enable this for every single project, while turning of the **dependent** projects calculation. For example a PR touching `lld` should only build and test `lld`: this should be cheap overall.

Your problem (as far as I understand) is rather: "when project X is touched: how many other dependent project can we bundle here" ; when X is `llvm` it becomes very large but when X is MLIR (for example) then it is just adding Flang on top of MLIR.

If the above is correct, then a good starting point could be to start with a very limited set of dependent projects, based on the coupling ("how much are tests of project Y likely to provide extra coverage to project X?"). LLVM can break potentially all the other projects in theory, but probably not with the same likeliness.

If the goal is to re-establish some amount of Windows coverage ASAP, then the most straightforward way to me seems like: "test only the project that is touched by a PR". From there we can observe breakages and increate the number of dependent projects based on actually impact.

https://github.com/llvm/llvm-project/pull/82093