Windows ARM64 (MSVC) build broken by NEON SIMD templatization (D95392150)

[kyamagu]

Summary

The NEON SIMD templatization landed via D95392150 (commit 8d8268c8a, GitHub PR #4866) breaks the build on Windows ARM64 with MSVC. The previous code (simdlib_neon.h before templatization) compiled successfully.

Environment

• Compiler: MSVC (Visual Studio 2022), targeting arm64
• OS: Windows 11 ARM64
• faiss version: main branch at commit 8d8268c8a and later

Errors

Three categories of errors occur:

1. NEON intrinsics cannot be used as non-type template parameters

MSVC implements NEON intrinsics (vsubq_u16, vandq_u16, vorrq_u16, etc.) as compiler builtins, not as real functions. Their address cannot be taken, so the binary_func_impl::call<&vsubq_u16>() pattern in simdlib_neon.h fails:

faiss/impl/simdlib/simdlib_neon.h(540): error C2065: 'vsubq_u16': undeclared identifier
faiss/impl/simdlib/simdlib_neon.h(540): error C2975: 'F': invalid template argument, expected compile-time constant expression

This affects all lines using .call<&intrinsic>() (lines 540, 552, 564, etc.).

2. Duplicate function definitions due to MSVC ARM64 type aliasing

On MSVC ARM64, uint8x16x2_t and uint16x8x2_t (and other NEON struct types) resolve to the same underlying __n128x2 type. This causes overload ambiguity:

faiss/impl/simdlib/simdlib_neon.h(34): error C2084: function 'uint8x16_t faiss::detail::reinterpret_u8(const uint8x16x2_t &)' already has a body

3. Missing members/operators on NEON-specialized template types

Several members and operators are not defined for the NEON specializations:

faiss/impl/fast_scan/simd_result_handlers.h(184): error C2039: 'get_scalar_0': is not a member of 'faiss::simd16uint16_tpl<faiss::SIMDLevel::ARM_NEON>'
faiss/IndexIVFPQFastScan.cpp(206): error C2039: 'store': is not a member of 'faiss::simd8float32_tpl<faiss::SIMDLevel::ARM_NEON>'

Operators >> and & are also missing for NEON types in kernels_simd256.h.

Suggested Fixes

1. Wrap NEON builtins in real functions so their addresses can be taken (or replace the function-pointer template pattern with a different dispatch mechanism for MSVC).
2. Guard overloaded reinterpret_* functions with #ifndef _MSC_VER or use SFINAE / if constexpr to avoid ambiguity from collapsed NEON types on MSVC.
3. Ensure all required members (get_scalar_0, store, operators >>, &) are defined for the ARM_NEON specializations.

Alternatively, falling back to SIMDLevel::NONE (emulated) on MSVC ARM64 would be a quick workaround until proper MSVC NEON support is added.

Related

• MSVC link failure: __restrict mismatch in fvec_madd AVX2/AVX512 specializations #4857 — MSVC __restrict mismatch from the same templatization effort (x86, different issue)
• C++20 features, while only C++17 may be implied #4890 — C++20 features in templatized code while only C++17 is required

Reproducer

Build faiss with CMake on Windows 11 ARM64 using MSVC:

cmake -B build -DFAISS_ENABLE_GPU=OFF -DFAISS_ENABLE_PYTHON=OFF -DBUILD_TESTING=OFF -A ARM64
cmake --build build

CI logs showing the failure: https://github.com/faiss-wheels/faiss-wheels/actions/runs/23570080174/job/68630487844?pr=158

[mnorris11]

cc @algoriddle