#include <emmintrin.h>
#include <smmintrin.h>

#define H_PRIME   1000003
#define HASH_LEN  11

// my_m128_t mimics Intel __m128i type
typedef struct my_m128_t my_m128_t;
struct my_m128_t { short m128i_i16[8]; };

int inner(
   int startval, int endval,
   int m2, int d2, int c2, int l2, int x2, int v2, int i2,
   my_m128_t* ps)
{
  __m128i s2 = _mm_set_epi32(x2, l2, c2, d2);
  __m128i hp = _mm_set1_epi32(H_PRIME);
  __m128i s3, s4, s5, s6, s7;

  int m3, m4, m5, m6, m7, m9;
  int d8, d9;
  int c9;
  int l9;
  int x9;
  int q3, q4, q5, q6, q7, q8, q9;
  int dist;
  int iret = 0;

  for (q3 = startval; q3 < endval; ++q3) {
    if (q3 == 10 || q3 == 13) continue;
    m3 = (m2 ^ q3) * H_PRIME;
    s3 = _mm_mullo_epi32(_mm_xor_si128(s2, _mm_set1_epi32(q3)), hp);
    for (q4 = 1; q4 < 128; ++q4) {
      if (q4 == 10 || q4 == 13) continue;
      m4 = (m3 ^ q4) * H_PRIME;
      s4 = _mm_mullo_epi32(_mm_xor_si128(s3, _mm_set1_epi32(q4)), hp);
      for (q5 = 1; q5 < 128; ++q5) {
        if (q5 == 10 || q5 == 13) continue;
        m5 = (m4 ^ q5) * H_PRIME;
        s5 = _mm_mullo_epi32(_mm_xor_si128(s4, _mm_set1_epi32(q5)), hp);
        for (q6 = 1; q6 < 128; ++q6) {
          if (q6 == 10 || q6 == 13) continue;
          m6 = (m5 ^ q6) * H_PRIME;
          s6 = _mm_mullo_epi32(_mm_xor_si128(s5, _mm_set1_epi32(q6)), hp);
          for (q7 = 1; q7 < 128; ++q7) {
            if (q7 == 10 || q7 == 13) continue;
            m7 = (m6 ^ q7) * H_PRIME;
            s7 = _mm_mullo_epi32(_mm_xor_si128(s6, _mm_set1_epi32(q7)), hp);
            for (q8 = 1; q8 < 128; ++q8) {
              if (q8 == 10 || q8 == 13) continue;

              d8 = (s7.m128i_i32[0] ^ q8) * H_PRIME ^ HASH_LEN;
              d9 = d8 & ~127;
              d9 %= 1001; if (d9 < 0) d9 += 1001;
              if (d9 < 500 - 127) continue;
              if (d9 > 500 + 127) continue;
              dist = d9 < 500 ? 500 - d9 : d9 - 500;
              q9 = (d8 & 127) ^ dist;
              if (q9 == 0 || q9 == 10 || q9 == 13) continue;

              c9 = (s7.m128i_i32[1] ^ q8) * H_PRIME ^ HASH_LEN ^ q9;
              c9 %= 1001; if (c9 < 0) c9 += 1001;
              if (c9 != 100) continue;

              l9 = (s7.m128i_i32[2] ^ q8) * H_PRIME ^ HASH_LEN ^ q9;
              l9 %= 1001; if (l9 < 0) l9 += 1001;
              if (l9 != 50) continue;

              x9 = (s7.m128i_i32[3] ^ q8) * H_PRIME ^ HASH_LEN ^ q9;
              x9 %= 1001; if (x9 < 0) x9 += 1001;
              if (x9 != 10) continue;

              m9 = (m7 ^ q8) * H_PRIME ^ HASH_LEN ^ q9;
              if (m9 == -1) m9 = -2;
              m9 %= 1001; if (m9 < 0) m9 += 1001;
              if (m9 != 1000) continue;

              ps[iret].m128i_i16[0] = q3;
              ps[iret].m128i_i16[1] = q4;
              ps[iret].m128i_i16[2] = q5;
              ps[iret].m128i_i16[3] = q6;
              ps[iret].m128i_i16[4] = q7;
              ps[iret].m128i_i16[5] = q8;
              ps[iret].m128i_i16[6] = q9;
              ++iret;
            }
          }
        }
      }
    }
  }
  return iret;
}