void complex_fir(float _Complex * restrict in, float _Complex * restrict out, unsigned int n, float * restrict coeff, unsigned int coeff_len)
{
    unsigned int i;
    float _Complex s;
    float _Complex * restrict in_p;
    float * restrict coeff_p;

#ifdef LOL
    in  = __builtin_assume_aligned(in, 8);
    out = __builtin_assume_aligned(out, 8);
    in_p = __builtin_assume_aligned(in_p, 8);
#endif

    do {
        s = 0;
        in_p = in++;
        coeff_p = coeff;

        i = coeff_len;
        do {
            s += *in_p++ * *coeff_p++;
        } while (--i);

        *out++ = s;
    } while (--n);
}