A transitive dependency of spin (llama.cpp via ggml) uses the vld1q_s8_x4 and vld1q_u8_x4 compiler built-in SIMD intrinsics.
These intrinsics are missing for aarch64 in gcc < 10.3, while cross-rs ships with gcc 9 for aarch64-unknown-linux-musl as of writing.
The code in this folder does a feature check and patches the arm_neon.h header with polyfills if the functions are missing.
See #1786 for the upstream issue.