The issue is visible on Apple Silicon when building Cycles.
Cycles includes sse2neon.h via two code paths: own CPU headers and
Blender's headers. The Blender headers will request higher precision
defines and Cycles does not need it (not for the kernel anyway, as
it has measurable performance penalty).
The solution is to wrap defines in the BLI_simd.h with check, so
that the flags are not re-defined.
The Blender's integration in Cycles does not really care if those
operations are precise or not, as the actual computations are done
elsewhere.
Pull Request: https://projects.blender.org/blender/blender/pulls/110953
6afbb66a53