meson build: SSE2/AVX2 Blosc shuffle is never enabled with MSVC (Windows branch is dead code) #866

🤖 AI text below 🤖

Problem description

In 0.17.0's src/numcodecs/meson.build, the x86_64 SIMD detection uses GCC-style flags:

have_sse2 = cc.has_argument('-msse2')
...
have_avx2 = cc.has_argument('-mavx2')

MSVC doesn't accept -msse2 or -mavx2, so with cl.exe both checks always fail. This line is from the conda-forge win-64 build log:

C compiler for the host machine: cl.exe (msvc 19.44.35229 ...)
...
Compiler for C supports arguments -msse2: NO

As a result, the Windows-specific blocks further down can never run:

if have_sse2
  ...
  blosc_c_args += ['-DSHUFFLE_SSE2_ENABLED', '-msse2']
  if host_machine.system() == 'windows'
    blosc_c_args += ['-D__SSE2__']
  endif
endif

The same applies to the AVX2 block and -D__AVX2__. The Windows build always uses the generic shuffle/bitshuffle code.

This is not a regression from 0.16.x. The old setup.py gated SIMD on cpuinfo.platform.machine() == 'x86_64', which is 'AMD64' on Windows, so SSE2/AVX2 were never enabled there either. It does look like the meson build meant to support SIMD on Windows, since it adds the __SSE2__/__AVX2__ defines for that case.

Expected behaviour

On x86_64 with MSVC:

• SSE2 is part of the x64 baseline, so have_sse2 can be true without any compiler flag.
• MSVC allows AVX2 intrinsics without /arch:AVX2, and c-blosc picks the implementation at runtime (__cpuid under _MSC_VER in shuffle.c). So the AVX2 sources can be compiled without a flag as well. /arch:AVX2 should not be used, because it would let the compiler emit AVX2 in the rest of the translation unit.

A possible shape for the fix (untested):

is_msvc = cc.get_argument_syntax() == 'msvc'

if get_option('sse2').disabled() or disable_sse2_env
  have_sse2 = false
else
  have_sse2 = is_msvc or cc.has_argument('-msse2')
endif
# same for avx2

The -msse2/-mavx2 additions to blosc_c_args would then need to be skipped for MSVC, keeping only the -D__SSE2__/-D__AVX2__ defines.

Related: zstd assembly source breaks MSVC builds

Also on x86_64 + MSVC, meson setup currently fails outright:

src\numcodecs\meson.build:96:13: ERROR: No host machine compiler for
'src/numcodecs/../../c-blosc/internal-complibs/zstd-1.5.6/decompress/huf_decompress_amd64.S'

zstd already turns off its assembly paths under MSVC (ZSTD_ASM_SUPPORTED in common/portability_macros.h), so the .S file can be skipped there. conda-forge is currently carrying this patch:

if host_machine.cpu_family() == 'x86_64' and cc.get_argument_syntax() != 'msvc'

See conda-forge/numcodecs-feedstock#126.

Version and installation information

• numcodecs.__version__: 0.17.0 (sdist from PyPI)
• Python: 3.12–3.14
• OS: Windows (win-64, MSVC 19.44 / VS 2022)
• Installation: conda-forge build (pip install . --no-build-isolation with meson-python)