fix: missing compilation flag on recent msvc version

[mfuntowicz]

In recent MSVC compiler versions it needs to specify which preprocessor to use.

This PR addresses this adding the required flag for MSVC.

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[Titus-von-Koeller]

Thanks for this — the /fp:fast de-duplication is a nice simplification and behavior-identical, so that part is good to go. Two things on the new flag, plus one ask.

1. It also hits compilers that don't want it. CMake's MSVC variable is true for any MSVC-emulating compiler, which includes icx (used by the XPU backend) and clang-cl. Clang's preprocessor is already conformant, so the flag is rejected. This PR's own build-xpu (windows-2022, 2026) job logs it on every translation unit:

icx: warning: argument unused during compilation: '-Zc:preprocessor' [-Wunused-command-line-argument]

Suggest gating on the compiler ID instead:

if(MSVC)
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /fp:fast")

    # Only real cl.exe needs (and accepts) the conformant preprocessor switch;
    # clang-cl / icx are already conformant and warn on it.
    if(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /Zc:preprocessor")
    endif()

    # /arch:AVX2 is only valid for x86/x64 targets, not ARM64
    string(TOLOWER "${CMAKE_SYSTEM_PROCESSOR}" _msvc_arch)
    if(_msvc_arch MATCHES "x86|x64|amd64")
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
    endif()
endif()

2. It may not fix the failure it's aimed at. The flag lands in CMAKE_CXX_FLAGS, which never reaches the .cu translation units — CMake compiles those with CMAKE_CUDA_FLAGS. Given the branch name (fix/cuda_121_win), if the error was in a CUDA file this patch is a no-op for it. The CUDA-side equivalent already exists, but is gated to CUDA 13.2+:

https://github.com/bitsandbytes-foundation/bitsandbytes/blob/a2b90e6/CMakeLists.txt#L197-L200

If the failure is nvcc-side, loosening that gate (e.g. keying on MSVC_VERSION rather than the CUDA version) is the actual fix. Either way, the two halves of the build currently disagree on preprocessor conformance for CUDA < 13.2, which is worth aligning.

3. Could you add the error to the description? All Windows jobs are green on main — including build-cpu (windows-2025) and build-cuda (windows-2025, 12.1.1) — so nothing in CI reproduces this, and it's hard to tell which of the two fixes above is correct. The failing file, the full error text, and your cl.exe + CUDA versions would settle it.

If it turns out to be BNB_OMP_PARALLEL_FOR in csrc/cpu_ops.cpp (it uses _Pragma("omp parallel for"), historically MSVC-conformant-preprocessor-only), then __pragma(omp parallel for) guarded by defined(_MSC_VER) && !defined(__clang__) would be a more surgical fix than a global conformance switch — though I'm fine with the switch too.