GH-47909: [C++] Fix MSVC ARM64 build

[jgiannuzzi]

Rationale for this change

#47573 broke the Windows ARM64 MSVC build by trying to compile xsimd even when ARROW_SIMD_LEVEL is set to NONE.

What changes are included in this PR?

Make sure that xsimd is only included when SIMD is requested.

Are these changes tested?

Yes, in the Windows ARM64 MSVC build pipeline that I was developing on #47811 when I noticed the change.

Are there any user-facing changes?

No.

• GitHub Issue: [C++] Windows ARM64 MSVC build is broken #47909

[github-actions]

⚠️ GitHub issue #47909 has been automatically assigned in GitHub to PR creator.

[pitrou]

@AntoinePrv Could you take a look at this?

[AntoinePrv]

Thanks for looking into this!

For reference, I did no create a _neon file because to reproduce the pattern used for byte_stream_split.

cpp/src/arrow/util/byte_stream_split_internal.cc
cpp/src/arrow/util/byte_stream_split_internal.h
cpp/src/arrow/util/byte_stream_split_internal_avx2.cc
cpp/src/arrow/util/byte_stream_split_test.cc

The fix works, but I am also working on this files, and may add an SSE compilation that will share the same xsimd code as the Neon one (hence the undescriptive _default).

I believe that, equivalently to the byte_stream_split, we can only gate the #includes in bpacking_simd_default.cc behind #if defined(ARROW_HAVE_NEON) to resolve this.
(I'm open to other ways of doing this, but I believe we should keep the code base consistent).

[AntoinePrv]

Hope it solves the issue, fix is looking good to me.

[jgiannuzzi]

Thanks for the insight, @AntoinePrv! I have tested it locally and have updated #47811 so we can see it working in the CI.

[pitrou]

+1, let's wait for CI

[jgiannuzzi]

Thanks for the merge @pitrou! I have just rebased #47811 and it's ready for review.

[conbench-apache-arrow]

After merging your PR, Conbench analyzed the 4 benchmarking runs that have been run so far on merge-commit a625340.

There were no benchmark performance regressions. 🎉

The full Conbench report has more details. It also includes information about 15 possible false positives for unstable benchmarks that are known to sometimes produce them.