LLVM based Windows build?

[kasper93]

Requesting base-llvmwin64 and base-llvmwin32 for parity with winarm64. LLVM build works for winarm64 so why not provide full build of it?

[BtbN]

Why though?
Building llvm takes MUCH longer than a gcc based toolchain, and the resulting image is quite a bit bigger.
And for all I'm aware, there is no discernable advantage for building with clang over gcc.
So it just seems like a waste of resources to me. The projects keeps scraping close to GitHubs size limits as it is right now.
And I'm actually consistently looking into replacing the winarm64 toolchain with gcc as well, now that gcc-16 finally supports it.

[BtbN]

If you just want a base image, I believe the winarm64 one can build for other windows targets as well?
LLVM/clang is multi-target, and I think the llvm mingw repo builds a toolchain for all windows targets, not specifically aarch64.

[kasper93]

Building llvm takes MUCH longer than a gcc based toolchain, and the resulting image is quite a bit bigger.

You can probably lower this time by reducing number of supported targets in build. LLVM by default supports everything.

And for all I'm aware, there is no discernable advantage for building with clang over gcc.

I don't have trust in GCC on Windows. Especially with autovectorization, where it still align stack incorrectly, forcing everyone to use unaligned loads always.

So it just seems like a waste of resources to me. The projects keeps scraping close to GitHubs size limits as it is right now.

Fair, looks like I will have to replicate whole infra. I wanted to avoid that.

And I'm actually consistently looking into replacing the winarm64 toolchain with gcc as well, now that gcc-16 finally supports it.

I would advice against that. LLVM is both better for aarch64 and Windows. Aiming into using inferior, less mature, toolchian is bizarre goal.

If you just want a base image, I believe the winarm64 one can build for other windows targets as well?
LLVM/clang is multi-target, and I think the llvm mingw repo builds a toolchain for all windows targets, not specifically aarch64.

I aim to use full image. However, as you noticed yourself. The LLVM build is resuable, so the base image would have to be built once for 3 variants. Which by itself compensate for the longer build time.

[BtbN]

Not quite, since building gcc 3 times in parallel is still multiple hours faster than building llvm once.
I'd also necessitate a whole other stage in the build process, as an in-between base.

I never looked into reducing the number of targets. Though I feel like the majority of code is shared by all.
I'll see how much faster it ends up being.
There's definitely no resources left on this repo to add another dozen or more images with all dependencies with llvm.
You can fork it and change the targets if you like.

[Andarwinux]

FYI, the upcoming LLVM 23.x introduced PCH support (clang-only), so build times are expected to be reduced by 30%.

Using a prebuilt PGOed clang (such as llvm-mingw's) as bootstrap compiler can further minimize build time.

If one can control LLVM build options, unnecessary components can be disabled to save time, see https://github.com/shinchiro/mpv-winbuild-cmake/blob/master/toolchain/llvm/llvm.cmake

https://llvm-compile-time-tracker.com/compare.php?from=1546138e6f18f2e7d0ca90eebe0ffc18c2364bf6&to=fb423ba07c3ff51fda3caa41d8bbb78d50f61906&stat=instructions%3Au