d8 = (d7 ^ 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;

##</code><code>##

c9 = (c7 ^ q8) * H_PRIME ^ HASH_LEN ^ q9;
c9 %= 1001; if (c9 < 0) c9 += 1001;
if (c9 != 100) continue;

##</code><code>##

#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(
   const unsigned char* bytevecM,
   int startval, int endval,
   int m2, int d2, int c2, int l2, int x2,
   my_m128_t* ps)
{
  int m3, m4, m5, m6, m7;
  int d3, d4, d5, d6, d7, d9;
  int c3, c4, c5, c6, c7, c9;
  int l3, l4, l5, l6, l7, l9;
  int x3, x4, x5, x6, x7, x9;
  int q3, q4, q5, q6, q7, q8, q9;
  int iret = 0;
  for (q3 = startval; q3 < endval; ++q3) {
    if (q3 == 10 || q3 == 13) continue;
    m3 = (m2 ^ q3) * H_PRIME;
    d3 = (d2 ^ q3) * H_PRIME;
    c3 = (c2 ^ q3) * H_PRIME;
    l3 = (l2 ^ q3) * H_PRIME;
    x3 = (x2 ^ q3) * H_PRIME;
    for (q4 = 1; q4 < 128; ++q4) {
      if (q4 == 10 || q4 == 13) continue;
      m4 = (m3 ^ q4) * H_PRIME;
      d4 = (d3 ^ q4) * H_PRIME;
      c4 = (c3 ^ q4) * H_PRIME;
      l4 = (l3 ^ q4) * H_PRIME;
      x4 = (x3 ^ q4) * H_PRIME;
      for (q5 = 1; q5 < 128; ++q5) {
        if (q5 == 10 || q5 == 13) continue;
        m5 = (m4 ^ q5) * H_PRIME;
        d5 = (d4 ^ q5) * H_PRIME;
        c5 = (c4 ^ q5) * H_PRIME;
        l5 = (l4 ^ q5) * H_PRIME;
        x5 = (x4 ^ q5) * H_PRIME;
        for (q6 = 1; q6 < 128; ++q6) {
          if (q6 == 10 || q6 == 13) continue;
          m6 = (m5 ^ q6) * H_PRIME;
          d6 = (d5 ^ q6) * H_PRIME;
          c6 = (c5 ^ q6) * H_PRIME;
          l6 = (l5 ^ q6) * H_PRIME;
          x6 = (x5 ^ q6) * H_PRIME;
          for (q7 = 1; q7 < 128; ++q7) {
            if (q7 == 10 || q7 == 13) continue;
            m7 = (m6 ^ q7) * H_PRIME;
            d7 = (d6 ^ q7) * H_PRIME;
            c7 = (c6 ^ q7) * H_PRIME;
            l7 = (l6 ^ q7) * H_PRIME;
            x7 = (x6 ^ q7) * H_PRIME;
            for (q8 = 1; q8 < 128; ++q8) {
              if (q8 == 10 || q8 == 13) continue;
              q9 = bytevecM[(unsigned int)(m7 ^ q8)];
              if (q9 == 0) continue;
              d9 = (d7 ^ q8) * H_PRIME ^ HASH_LEN ^ q9;
              d9 %= 1001; if (d9 < 0) d9 += 1001;
              if (d9 != 500) continue;
              c9 = (c7 ^ q8) * H_PRIME ^ HASH_LEN ^ q9;
              c9 %= 1001; if (c9 < 0) c9 += 1001;
              if (c9 != 100) continue;
              l9 = (l7 ^ q8) * H_PRIME ^ HASH_LEN ^ q9;
              l9 %= 1001; if (l9 < 0) l9 += 1001;
              if (l9 != 50) continue;
              x9 = (x7 ^ q8) * H_PRIME ^ HASH_LEN ^ q9;
              x9 %= 1001; if (x9 < 0) x9 += 1001;
              if (x9 != 10) 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;
}

##</code><code>##

_mm_prefetch(&bytevecM[(unsigned int)m7 & 0xffffff80], _MM_HINT_T0);
_mm_prefetch(&bytevecM[64+((unsigned int)m7 & 0xffffff80)], _MM_HINT_T0);

##</code><code>##

BOOL MySetLockPagesPrivilege(HANDLE hProcess, BOOL bEnable)
{
   struct {
      DWORD Count;
      LUID_AND_ATTRIBUTES Privilege[1];
   } Info;
   HANDLE Token;
   if ( !OpenProcessToken(hProcess, TOKEN_ADJUST_PRIVILEGES, &Token) ) {
      printf("Cannot open process token.\n");
      return FALSE;
   }
   Info.Count = 1;
   if (bEnable) {
      Info.Privilege[0].Attributes = SE_PRIVILEGE_ENABLED;
   }
   else {
      Info.Privilege[0].Attributes = 0;
   }
   if ( !LookupPrivilegeValue(NULL, SE_LOCK_MEMORY_NAME, &(Info.Privilege[0].Luid)) ) {
      printf("Cannot get privilege for %s.\n", SE_LOCK_MEMORY_NAME);
      return FALSE;
   }
   if ( !AdjustTokenPrivileges(Token, FALSE, (PTOKEN_PRIVILEGES)&Info, 0, NULL, NULL) )  {
      printf ("Cannot adjust token privileges (%u)\n", GetLastError());
      return FALSE;
   }
   // ERROR_NOT_ALL_ASSIGNED
   // 1300 (0x514)
   // Not all privileges or groups referenced are assigned to the caller.
   DWORD dwErr = GetLastError();
   printf("AdjustTokenPrivileges dwErr=%u\n", dwErr);
   if (dwErr != ERROR_SUCCESS) {
      printf("Cannot enable the SE_LOCK_MEMORY_NAME privilege");
      return FALSE;
   }
   CloseHandle(Token);
   return TRUE;
}

void LargeMemInit()
{
   if ( !MySetLockPagesPrivilege(GetCurrentProcess(), TRUE) ) {
      fprintf(stderr, "MySetLockPagesPrivilege failed.\n");
      exit(1);
   }
}

// VirtualAlloc may fail with: ERROR_NO_SYSTEM_RESOURCES = 1450. To fix, reboot.
static void* my_big_malloc(size_t s)
{
   DWORD dwErr;
   void* pMem;
   pMem = VirtualAlloc(
      NULL,                           // Let system decide where to allocate it
      s,                              // size of memory block
      MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES,
      PAGE_READWRITE
   );
   if (pMem == NULL) {
      // Error 1314 = ERROR_PRIVILEGE_NOT_HELD
      dwErr = GetLastError();
      fprintf(stderr, "VirtualAlloc failed: dwError=%d\n", (int)dwErr);
      exit(1);
   }
   printf("VirtualAlloc ok p=%p\n", pMem);
   return pMem;
}

##</code><code>##

#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(
   const unsigned char* bytevecM,
   int startval, int endval,
   int m2, int d2, int c2, int l2, int x2,
   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;
  int d9;
  int c9;
  int l9;
  int x9;
  int q3, q4, q5, q6, q7, q8, q9;
  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);
            _mm_prefetch(&bytevecM[(unsigned int)m7 & 0xffffff80], _MM_HINT_T0);
            _mm_prefetch(&bytevecM[64+((unsigned int)m7 & 0xffffff80)], _MM_HINT_T0);
            for (q8 = 1; q8 < 128; ++q8) {
              if (q8 == 10 || q8 == 13) continue;
              q9 = bytevecM[(unsigned int)(m7 ^ q8)];
              if (q9 == 0) continue;
              d9 = (s7.m128i_i32[0] ^ q8) * H_PRIME ^ HASH_LEN ^ q9;
              d9 %= 1001; if (d9 < 0) d9 += 1001;
              if (d9 != 500) 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;
              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;
}

##</code><code>##

#define UNROLL(qx) \
  q9 = bytevecM[(unsigned int)(m7 ^ qx)];                         \
  if (q9 != 0) {                                                  \
    d9 = (s7.m128i_i32[0] ^ qx) * H_PRIME ^ HASH_LEN ^ q9;        \
    d9 %= 1001; if (d9 < 0) d9 += 1001;                           \
    if (d9 == 500) {                                              \
      c9 = (s7.m128i_i32[1] ^ qx) * H_PRIME ^ HASH_LEN ^ q9;      \
      c9 %= 1001; if (c9 < 0) c9 += 1001;                         \
      if (c9 == 100) {                                            \
        l9 = (s7.m128i_i32[2] ^ qx) * H_PRIME ^ HASH_LEN ^ q9;    \
        l9 %= 1001; if (l9 < 0) l9 += 1001;                       \
        if (l9 == 50) {                                           \
          x9 = (s7.m128i_i32[3] ^ qx) * H_PRIME ^ HASH_LEN ^ q9;  \
          x9 %= 1001; if (x9 < 0) x9 += 1001;                     \
          if (x9 == 10) {                                         \
            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] = qx;                           \
            ps[iret].m128i_i16[6] = q9;                           \
            ++iret;                                               \
          }                                                       \
        }                                                         \
      }                                                           \
    }                                                             \
  }

##</code><code>##

for (q8 = 1; q8 < 128; ++q8) {
  if (q8 == 10 || q8 == 13) continue;
  q9 = bytevecM[(unsigned int)(m7 ^ q8)];
  if (q9 == 0) continue;
  d9 = (s7.m128i_i32[0] ^ q8) * H_PRIME ^ HASH_LEN ^ q9;
  d9 %= 1001; if (d9 < 0) d9 += 1001;
  if (d9 != 500) 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;
  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;
}

##</code><code>##

UNROLL(1)
UNROLL(2)
UNROLL(3)
UNROLL(4)
UNROLL(5)
UNROLL(6)
UNROLL(7)
UNROLL(8)
UNROLL(9)
UNROLL(11)
UNROLL(12)
for (q8 = 14; q8 < 128; ++q8) {
  UNROLL(q8)
}

##</code><code>##

UNROLL(1)
UNROLL(2)
UNROLL(3)
...
UNROLL(127)