I am developing a library that requires avx and fma instructions to be fast, If simd isn't enabled it performs bad and I don't want to implement other versions of my functions without simd. Should I use compiler_error macro to fail compilation if avx and fma isn't enabled or is telling users to enable target-cpu=native is enough?
Personally I usecfg'd compile_errors. Your library should not allow execution of SIMD intrinsics without compile- or run-time checks of required target feature.