My (kinda works) fix to cross-compile android-llvm for Windows on Linux using MSVCConfig

[Naville Zhang]

Hi:

As title suggested, 

I attempted to fix up enough of llvm_android, thought it would be a good idea to show the initial work, get some code-review comments and attempt to submit my patches to AOSP.

My patch / comments / status quo is here https://gist.github.com/Naville/10e36bfcd38890ac266f5c9290ed97dc

[Naville Zhang]

As to why MingW is not enough, please reference to https://github.com/msys2/MINGW-packages/issues/9771

Basically MingW's exception handling is very broken, and in our downstream build we build LLVM with Z3, a project with exception semantics everywhere and thus breaks

[Naville Zhang]

I've successfully built everything on Linux using MSVC SDK using the attached patch.

Main changes include:

# win_sdk

- Fix broken path handling, now we require win_sdk's path to point to the root of the sdk package packed by aosp depot_tools

- Fix the algorithm used to expand environment variables in SetEnv.json

- Remove the half-ass baked broken Windows SDK Case Fixing component, and instead rely on the SDK being stored in a case-insensitive filesystem

# MSVCConfig

- Fix typo in code that results in  llvm_triple returning a triple instead of a string

- Use Host Stage2 toolchain to avoid building LLVM Host Tools in WindowsBuilder, which is broken without MinGW and too much work to fix.

- Add llvm-rc.exe and llvm-ml.exe

# LLVM

- Disable LLVM's InlineAsm BLAKE3 Implementation, which last time I check is broken on ClangCL, resulting in runtime crash.

It seems to me that I can't upload this patch to AOSP for a review directly.

Zhang

[Naville Zhang]

What does't work yet:

- I didn't check if MinGW build still works with this patch.

- LLVM on Windows is not able to detect the prebuilt Python, probably some code is still broken.

- Packaging the final zip for Windows is broken, because the packaging component code still expects static library to have .a extension instead of .lib extension. 

As I was saying, I'd love to get some feedback and maybe get the initial patch merged into NDK upstream before attempting to fix the remaining problems