Enable MSVC (cl.exe) build on Windows ARM64 by pdeep854 · Pull Request #1285 · xtensor-stack/xsimd

pdeep854 · 2026-04-01T09:26:48Z

PR: Enable MSVC (cl.exe) build on Windows ARM64

Summary

This PR restores full build compatibility for xsimd with the MSVC compiler (cl.exe) on
Windows ARM64 (_M_ARM64). Prior to this change, the project failed to compile with MSVC
on ARM64 due to fundamental differences in how MSVC exposes ARM NEON intrinsics compared
to GCC and Clang.

All changes are strictly additive and backward-compatible: every existing GCC/Clang code
path is preserved unchanged inside #else branches. No behaviour is altered for any
non-MSVC-ARM64 target.

Testing

The build was verified locally on a Windows ARM64 machine using Visual Studio 2022:

cmake .. -DBUILD_TESTS=ON -DDOWNLOAD_DOCTEST=ON -DBUILD_BENCHMARK=ON ^
         -DBUILD_EXAMPLES=ON -DCMAKE_BUILD_TYPE=Release ^
         -G "Visual Studio 17 2022"
cmake --build . --config Release

Result: All three targets build successfully with zero errors:

benchmark_xsimd.vcxproj -> ...\Release\benchmark_xsimd.exe
mandelbrot.vcxproj      -> ...\Release\mandelbrot.exe
test_xsimd.vcxproj      -> ...\Release\test_xsimd.exe

serge-sans-paille

First set of review, more to come later but at least it will trigger the discussion

serge-sans-paille · 2026-04-01T10:28:57Z

include/xsimd/arch/common/xsimd_common_memory.hpp

 #include <algorithm>
 #include <array>
 #include <complex>
+#include <cstdio>


unused header, probably a leftover from a debug session.

thanks will remove

serge-sans-paille · 2026-04-01T10:31:29Z

include/xsimd/arch/common/xsimd_common_memory.hpp

                for (size_t i = 0; i < sizeof...(Is); ++i)
                    if ((bitmask >> i) & 1u)
                        std::swap(mask_buffer[inserted++], mask_buffer[i]);
+                // Fill remaining positions with the last valid index to avoid undefined behavior


why not always setting zero?

after swapping some unused entries at the end of mask_buffer still contain old index values. This mask is later used by swizzle() to gather vector elements. on MSVC arm64, swizzle() returns 0 for invalid indices, so these leftover values cause incorrect results.
to fix this the unused entries are filled with the last valid index, which acts as a safe placeholder. can also set it to 0

serge-sans-paille · 2026-04-01T10:33:44Z

include/xsimd/arch/common/xsimd_common_memory.hpp

+            compress_mask.store_aligned(&mask_out[0]);
+            alignas(A::alignment()) T z_out[size];
+            z.store_aligned(&z_out[0]);
+            auto res = swizzle(z, compress_mask);


what if the overload we were using previously actually exists?

thanks will remove it as it was added for debugging

serge-sans-paille · 2026-04-01T10:34:08Z

include/xsimd/arch/xsimd_neon.hpp

+            alignas(16) T data[] = { static_cast<T>(args)... };
+            if constexpr (sizeof(T) == 1)
+            {
+                if constexpr (std::is_unsigned<T>::value)


xsimd requires only C++14 as of this PR :-/

will make changes as per c++14

serge-sans-paille · 2026-04-01T10:37:09Z

include/xsimd/arch/xsimd_neon64.hpp

        {
+#if defined(_MSC_VER) && defined(_M_ARM64)
+            alignas(16) double data[] = { d0, d1 };
+            return vld1q_f64(data);


we could make this the default if it's more prtable: https://godbolt.org/z/fP3Tvjv5f

sure will make it default

serge-sans-paille · 2026-04-01T10:37:24Z

include/xsimd/arch/xsimd_neon64.hpp

+                static_cast<unsigned_type>(b0 ? -1LL : 0LL),
+                static_cast<unsigned_type>(b1 ? -1LL : 0LL)
+            };
+            return vld1q_u64(data);


will make default

pdeep854 · 2026-04-02T11:48:05Z

@serge-sans-paille i have addressed all review comments. request you to review

serge-sans-paille

Another set of review.

serge-sans-paille · 2026-04-03T09:19:02Z

include/xsimd/arch/common/xsimd_common_memory.hpp

                    if ((bitmask >> i) & 1u)
                        std::swap(mask_buffer[inserted++], mask_buffer[i]);
+                // Fill remaining (don't-care) tail positions with index 0.
+                for (size_t i = inserted; i < sizeof...(Is); ++i)


use std::fill or memset instead.

serge-sans-paille · 2026-04-03T09:20:27Z

include/xsimd/arch/xsimd_neon.hpp

-            {
-                using type = uint64x2_t;
-            };
-


I like this change

added this change

serge-sans-paille · 2026-04-03T09:22:44Z

include/xsimd/arch/xsimd_neon.hpp

        template <class A, class T, class... Args, detail::enable_integral_t<T> = 0>
        XSIMD_INLINE batch<T, A> set(batch<T, A> const&, requires_arch<neon>, Args... args) noexcept
        {
+#if defined(_MSC_VER) && defined(_M_ARM64) && !defined(__clang__)


we could probably syndicate this long guard in a single macro variable somewhere.

serge-sans-paille · 2026-04-03T09:23:30Z

include/xsimd/arch/xsimd_neon.hpp

+            alignas(16) T data[] = { static_cast<T>(args)... };
+            return detail::msvc_arm64_load<T>(data);
+#else
            return xsimd::types::detail::neon_vector_type<T> { args... };


I have the feeling that we could use detail::msvc_arm64_load<T> in all context there, avoiding the macro guard.

added as per suggestion

serge-sans-paille · 2026-04-03T09:23:48Z

include/xsimd/arch/xsimd_neon.hpp

+            using unsigned_type = as_unsigned_integer_t<T>;
+            alignas(16) unsigned_type data[] = { static_cast<unsigned_type>(args ? -1LL : 0LL)... };
+            return detail::msvc_arm64_load_u<T>(data);
+#else


pdeep854 marked this pull request as ready for review April 1, 2026 09:28

This was referenced Apr 1, 2026

Enable compiling arm/neon with MSVC for windows on arm64 #612

Open

MSVC compile error for NEON (win/arm64) #611

Open

MugundanMCW mentioned this pull request Apr 1, 2026

GH-47195: [Python][CI] Add support for building PyArrow library on Windows ARM64 apache/arrow#48539

Open

serge-sans-paille requested changes Apr 1, 2026

View reviewed changes

pdeep854 force-pushed the msvc-arm64-enablement branch from 9862d57 to 920fafc Compare April 2, 2026 08:08

pdeep854 closed this Apr 2, 2026

pdeep854 force-pushed the msvc-arm64-enablement branch from 920fafc to e2f0536 Compare April 2, 2026 08:39

pdeep854 added 2 commits April 2, 2026 14:11

enable MSVC cl.exe build changes as per c++ 14

cb24d65

enable MSVC cl.exe build changes

6437cdd

pdeep854 reopened this Apr 2, 2026

pdeep854 requested a review from serge-sans-paille April 2, 2026 08:52

pdeep854 added 2 commits April 2, 2026 14:24

Merge branch 'xtensor-stack:master' into msvc-arm64-enablement

4af7256

fix for clang build failure

7e3fe1a

serge-sans-paille reviewed Apr 3, 2026

View reviewed changes

removed unneccesary macro

7337ff8

pdeep854 requested a review from serge-sans-paille April 3, 2026 12:04

serge-sans-paille mentioned this pull request Apr 3, 2026

Fix windows CI arm runner #1288

Draft

Conversation

pdeep854 commented Apr 1, 2026

PR: Enable MSVC (cl.exe) build on Windows ARM64

Summary

Testing

Uh oh!

serge-sans-paille left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pdeep854 Apr 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pdeep854 commented Apr 2, 2026

Uh oh!

serge-sans-paille left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

pdeep854 Apr 1, 2026 •

edited

Loading