WIP: Improve performance of windows-aarch64 builds

[zanieb]

Extending #10306 — trying to reach an acceptable perf

[oberrich]

The "Install LLVM" step already installs LLVM/Clang 19 and puts it in path.
Visual Studio comes with Clang 18 which has a bindgen bug affecting aarch64, so I would advise not installing Microsoft.VisualStudio.ComponentGroup.NativeDesktop.Llvm.Clang unless you need Clang 18 specifically.

This will save some time on the installer as well, only getting the build tools and SDK takes around 5m 23s; although that shouldn't matter once you get the caching working 👍

[zanieb]

Interesting, thanks for the tips!

[zanieb]

Right now the problem is that the cache is huge, 7GB is going to cause cache churn since we have a 10GB limit in the repository.

[oberrich]

Right now the problem is that the cache is huge, 7GB is going to cause cache churn since we have a 10GB limit in the repository.

Oh wow even with caching it takes 3m 21s because of the download I assume.

I think it might be worth it to only use the native ARM64 runner in a release CI, Rust can use --target aarch64... just fine on x86_64 hosts and this whole ">10 min CI just for getting a the bare minimum setup to compile rust" is kind of just burning money for no upside.

Probably best to wait for Microsoft/GitHub to ship a usable VM image.

[zanieb]

I think it might be worth it to only use the native ARM64 runner in a release CI, Rust can use --target aarch64... just fine on x86_64 hosts and this whole ">10 min CI just for getting a the bare minimum setup to compile rust" is kind of just burning money for no upside.

Yeah, but then it can regress and then I have to burn a bunch of cycles fixing a release :( trade-offs

Probably best to wait for Microsoft/GitHub to ship a usable VM image.

Sort of agree.