/**
 * Author......: See docs/credits.txt
 * License.....: MIT
 */

//too much register pressure
//#define NEW_SIMD_CODE

#ifdef KERNEL_STATIC
#include "inc_vendor.h"
#include "inc_types.h"
#include "inc_common.cl"
#include "inc_rp_optimized.h"
#include "inc_rp_optimized.cl"
#include "inc_simd.cl"
#include "inc_hash_sha1.cl"
#endif

typedef struct pstoken
{
  u32 salt_buf[128];
  u32 salt_len;

  u32 pc_digest[5];
  u32 pc_offset;

} pstoken_t;

KERNEL_FQ void m13500_m04 (KERN_ATTR_RULES_ESALT (pstoken_t))
{
  /**
   * modifier
   */

  const u64 lid = get_local_id (0);

  /**
   * base
   */

  const u64 gid = get_global_id (0);

  if (gid >= gid_max) return;

  u32 pw_buf0[4];
  u32 pw_buf1[4];

  pw_buf0[0] = pws[gid].i[0];
  pw_buf0[1] = pws[gid].i[1];
  pw_buf0[2] = pws[gid].i[2];
  pw_buf0[3] = pws[gid].i[3];
  pw_buf1[0] = pws[gid].i[4];
  pw_buf1[1] = pws[gid].i[5];
  pw_buf1[2] = pws[gid].i[6];
  pw_buf1[3] = pws[gid].i[7];

  const u32 pw_len = pws[gid].pw_len & 63;

  /**
   * salt
   */

  const u32 pc_offset = esalt_bufs[digests_offset].pc_offset;

  const u32 pc_offset4 = pc_offset * 4;

  u32 pc_digest[5];

  pc_digest[0] = esalt_bufs[digests_offset].pc_digest[0];
  pc_digest[1] = esalt_bufs[digests_offset].pc_digest[1];
  pc_digest[2] = esalt_bufs[digests_offset].pc_digest[2];
  pc_digest[3] = esalt_bufs[digests_offset].pc_digest[3];
  pc_digest[4] = esalt_bufs[digests_offset].pc_digest[4];

  u32 salt_buf0[4];
  u32 salt_buf1[4];
  u32 salt_buf2[4];
  u32 salt_buf3[4];

  salt_buf0[0] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  0]);
  salt_buf0[1] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  1]);
  salt_buf0[2] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  2]);
  salt_buf0[3] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  3]);
  salt_buf1[0] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  4]);
  salt_buf1[1] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  5]);
  salt_buf1[2] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  6]);
  salt_buf1[3] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  7]);
  salt_buf2[0] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  8]);
  salt_buf2[1] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  9]);
  salt_buf2[2] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset + 10]);
  salt_buf2[3] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset + 11]);
  salt_buf3[0] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset + 12]);
  salt_buf3[1] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset + 13]);
  salt_buf3[2] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset + 14]);
  salt_buf3[3] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset + 15]);

  const u32 salt_len = esalt_bufs[digests_offset].salt_len;

  /**
   * loop
   */

  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
  {
    u32x w0[4] = { 0 };
    u32x w1[4] = { 0 };
    u32x w2[4] = { 0 };
    u32x w3[4] = { 0 };

    const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);

    append_0x80_2x4_VV (w0, w1, out_len);

    make_utf16le (w1, w2, w3);
    make_utf16le (w0, w0, w1);

    const u32x out_len2 = out_len * 2;

    const u32x out_salt_len = out_len2 + salt_len;

    /**
     * prepend salt -- can't stay outside the loop this time
     */

    w0[0] = swap32 (w0[0]);
    w0[1] = swap32 (w0[1]);
    w0[2] = swap32 (w0[2]);
    w0[3] = swap32 (w0[3]);
    w1[0] = swap32 (w1[0]);
    w1[1] = swap32 (w1[1]);
    w1[2] = swap32 (w1[2]);
    w1[3] = swap32 (w1[3]);
    w2[0] = swap32 (w2[0]);
    w2[1] = swap32 (w2[1]);
    w2[2] = swap32 (w2[2]);
    w2[3] = swap32 (w2[3]);
    w3[0] = swap32 (w3[0]);
    w3[1] = swap32 (w3[1]);
    w3[2] = swap32 (w3[2]);
    w3[3] = swap32 (w3[3]);

    u32x c0[4] = { 0 };
    u32x c1[4] = { 0 };
    u32x c2[4] = { 0 };
    u32x c3[4] = { 0 };

    switch_buffer_by_offset_carry_be (w0, w1, w2, w3, c0, c1, c2, c3, salt_len - pc_offset4);

    w0[0] |= salt_buf0[0];
    w0[1] |= salt_buf0[1];
    w0[2] |= salt_buf0[2];
    w0[3] |= salt_buf0[3];
    w1[0] |= salt_buf1[0];
    w1[1] |= salt_buf1[1];
    w1[2] |= salt_buf1[2];
    w1[3] |= salt_buf1[3];
    w2[0] |= salt_buf2[0];
    w2[1] |= salt_buf2[1];
    w2[2] |= salt_buf2[2];
    w2[3] |= salt_buf2[3];
    w3[0] |= salt_buf3[0];
    w3[1] |= salt_buf3[1];
    w3[2] |= salt_buf3[2];
    w3[3] |= salt_buf3[3];

    u32x a = pc_digest[0];
    u32x b = pc_digest[1];
    u32x c = pc_digest[2];
    u32x d = pc_digest[3];
    u32x e = pc_digest[4];

    const u32 r = (salt_len - pc_offset4) + out_len2;

    if (r >= 56)
    {
      u32x w0_t = w0[0];
      u32x w1_t = w0[1];
      u32x w2_t = w0[2];
      u32x w3_t = w0[3];
      u32x w4_t = w1[0];
      u32x w5_t = w1[1];
      u32x w6_t = w1[2];
      u32x w7_t = w1[3];
      u32x w8_t = w2[0];
      u32x w9_t = w2[1];
      u32x wa_t = w2[2];
      u32x wb_t = w2[3];
      u32x wc_t = w3[0];
      u32x wd_t = w3[1];
      u32x we_t = w3[2];
      u32x wf_t = w3[3];

      #undef K
      #define K SHA1C00

      SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0_t);
      SHA1_STEP (SHA1_F0o, e, a, b, c, d, w1_t);
      SHA1_STEP (SHA1_F0o, d, e, a, b, c, w2_t);
      SHA1_STEP (SHA1_F0o, c, d, e, a, b, w3_t);
      SHA1_STEP (SHA1_F0o, b, c, d, e, a, w4_t);
      SHA1_STEP (SHA1_F0o, a, b, c, d, e, w5_t);
      SHA1_STEP (SHA1_F0o, e, a, b, c, d, w6_t);
      SHA1_STEP (SHA1_F0o, d, e, a, b, c, w7_t);
      SHA1_STEP (SHA1_F0o, c, d, e, a, b, w8_t);
      SHA1_STEP (SHA1_F0o, b, c, d, e, a, w9_t);
      SHA1_STEP (SHA1_F0o, a, b, c, d, e, wa_t);
      SHA1_STEP (SHA1_F0o, e, a, b, c, d, wb_t);
      SHA1_STEP (SHA1_F0o, d, e, a, b, c, wc_t);
      SHA1_STEP (SHA1_F0o, c, d, e, a, b, wd_t);
      SHA1_STEP (SHA1_F0o, b, c, d, e, a, we_t);
      SHA1_STEP (SHA1_F0o, a, b, c, d, e, wf_t);
      w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t);
      w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, d, e, a, b, c, w1_t);
      w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, c, d, e, a, b, w2_t);
      w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, b, c, d, e, a, w3_t);

      #undef K
      #define K SHA1C01

      w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w4_t);
      w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w5_t);
      w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w6_t);
      w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w7_t);
      w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w8_t);
      w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w9_t);
      wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wa_t);
      wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wb_t);
      wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wc_t);
      wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wd_t);
      we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, we_t);
      wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wf_t);
      w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t);
      w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t);
      w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t);
      w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t);
      w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w4_t);
      w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w5_t);
      w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w6_t);
      w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w7_t);

      #undef K
      #define K SHA1C02

      w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w8_t);
      w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w9_t);
      wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wa_t);
      wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wb_t);
      wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wc_t);
      wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, wd_t);
      we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, we_t);
      wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wf_t);
      w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w0_t);
      w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w1_t);
      w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w2_t);
      w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w3_t);
      w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w4_t);
      w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w5_t);
      w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w6_t);
      w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w7_t);
      w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w8_t);
      w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w9_t);
      wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wa_t);
      wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wb_t);

      #undef K
      #define K SHA1C03

      wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wc_t);
      wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wd_t);
      we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, we_t);
      wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wf_t);
      w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t);
      w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t);
      w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w2_t);
      w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t);
      w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w4_t);
      w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w5_t);
      w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w6_t);
      w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w7_t);
      w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w8_t);
      w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w9_t);
      wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wa_t);
      wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wb_t);
      wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wc_t);
      wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wd_t);
      we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, we_t);
      wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wf_t);

      a += pc_digest[0];
      b += pc_digest[1];
      c += pc_digest[2];
      d += pc_digest[3];
      e += pc_digest[4];

      w0[0] = c0[0];
      w0[1] = c0[1];
      w0[2] = c0[2];
      w0[3] = c0[3];
      w1[0] = c1[0];
      w1[1] = c1[1];
      w1[2] = c1[2];
      w1[3] = c1[3];
      w2[0] = c2[0];
      w2[1] = c2[1];
      w2[2] = c2[2];
      w2[3] = c2[3];
      w3[0] = c3[0];
      w3[1] = c3[1];
      w3[2] = c3[2];
      w3[3] = c3[3];
    }

    /**
     * final sha1
     */

    u32x w0_t = w0[0];
    u32x w1_t = w0[1];
    u32x w2_t = w0[2];
    u32x w3_t = w0[3];
    u32x w4_t = w1[0];
    u32x w5_t = w1[1];
    u32x w6_t = w1[2];
    u32x w7_t = w1[3];
    u32x w8_t = w2[0];
    u32x w9_t = w2[1];
    u32x wa_t = w2[2];
    u32x wb_t = w2[3];
    u32x wc_t = w3[0];
    u32x wd_t = w3[1];
    u32x we_t = 0;
    u32x wf_t = out_salt_len * 8;

    u32x r_a = a;
    u32x r_b = b;
    u32x r_c = c;
    u32x r_d = d;
    u32x r_e = e;

    #undef K
    #define K SHA1C00

    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0_t);
    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w1_t);
    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w2_t);
    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w3_t);
    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w4_t);
    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w5_t);
    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w6_t);
    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w7_t);
    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w8_t);
    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w9_t);
    SHA1_STEP (SHA1_F0o, a, b, c, d, e, wa_t);
    SHA1_STEP (SHA1_F0o, e, a, b, c, d, wb_t);
    SHA1_STEP (SHA1_F0o, d, e, a, b, c, wc_t);
    SHA1_STEP (SHA1_F0o, c, d, e, a, b, wd_t);
    SHA1_STEP (SHA1_F0o, b, c, d, e, a, we_t);
    SHA1_STEP (SHA1_F0o, a, b, c, d, e, wf_t);
    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t);
    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, d, e, a, b, c, w1_t);
    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, c, d, e, a, b, w2_t);
    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, b, c, d, e, a, w3_t);

    #undef K
    #define K SHA1C01

    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w4_t);
    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w5_t);
    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w6_t);
    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w7_t);
    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w8_t);
    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w9_t);
    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wa_t);
    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wb_t);
    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wc_t);
    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wd_t);
    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, we_t);
    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wf_t);
    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t);
    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t);
    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t);
    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t);
    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w4_t);
    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w5_t);
    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w6_t);
    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w7_t);

    #undef K
    #define K SHA1C02

    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w8_t);
    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w9_t);
    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wa_t);
    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wb_t);
    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wc_t);
    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, wd_t);
    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, we_t);
    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wf_t);
    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w0_t);
    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w1_t);
    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w2_t);
    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w3_t);
    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w4_t);
    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w5_t);
    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w6_t);
    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w7_t);
    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w8_t);
    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w9_t);
    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wa_t);
    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wb_t);

    #undef K
    #define K SHA1C03

    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wc_t);
    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wd_t);
    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, we_t);
    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wf_t);
    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t);
    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t);
    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w2_t);
    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t);
    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w4_t);
    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w5_t);
    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w6_t);
    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w7_t);
    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w8_t);
    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w9_t);
    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wa_t);
    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wb_t);
    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wc_t);
    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wd_t);
    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, we_t);
    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wf_t);

    a += r_a;
    b += r_b;
    c += r_c;
    d += r_d;
    e += r_e;

    COMPARE_M_SIMD (d, e, c, b);
  }
}

KERNEL_FQ void m13500_m08 (KERN_ATTR_RULES_ESALT (pstoken_t))
{
}

KERNEL_FQ void m13500_m16 (KERN_ATTR_RULES_ESALT (pstoken_t))
{
}

KERNEL_FQ void m13500_s04 (KERN_ATTR_RULES_ESALT (pstoken_t))
{
  /**
   * modifier
   */

  const u64 lid = get_local_id (0);

  /**
   * base
   */

  const u64 gid = get_global_id (0);

  if (gid >= gid_max) return;

  u32 pw_buf0[4];
  u32 pw_buf1[4];

  pw_buf0[0] = pws[gid].i[0];
  pw_buf0[1] = pws[gid].i[1];
  pw_buf0[2] = pws[gid].i[2];
  pw_buf0[3] = pws[gid].i[3];
  pw_buf1[0] = pws[gid].i[4];
  pw_buf1[1] = pws[gid].i[5];
  pw_buf1[2] = pws[gid].i[6];
  pw_buf1[3] = pws[gid].i[7];

  const u32 pw_len = pws[gid].pw_len & 63;

  /**
   * salt
   */

  const u32 pc_offset = esalt_bufs[digests_offset].pc_offset;

  const u32 pc_offset4 = pc_offset * 4;

  u32 pc_digest[5];

  pc_digest[0] = esalt_bufs[digests_offset].pc_digest[0];
  pc_digest[1] = esalt_bufs[digests_offset].pc_digest[1];
  pc_digest[2] = esalt_bufs[digests_offset].pc_digest[2];
  pc_digest[3] = esalt_bufs[digests_offset].pc_digest[3];
  pc_digest[4] = esalt_bufs[digests_offset].pc_digest[4];

  u32 salt_buf0[4];
  u32 salt_buf1[4];
  u32 salt_buf2[4];
  u32 salt_buf3[4];

  salt_buf0[0] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  0]);
  salt_buf0[1] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  1]);
  salt_buf0[2] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  2]);
  salt_buf0[3] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  3]);
  salt_buf1[0] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  4]);
  salt_buf1[1] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  5]);
  salt_buf1[2] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  6]);
  salt_buf1[3] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  7]);
  salt_buf2[0] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  8]);
  salt_buf2[1] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset +  9]);
  salt_buf2[2] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset + 10]);
  salt_buf2[3] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset + 11]);
  salt_buf3[0] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset + 12]);
  salt_buf3[1] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset + 13]);
  salt_buf3[2] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset + 14]);
  salt_buf3[3] = swap32_S (esalt_bufs[digests_offset].salt_buf[pc_offset + 15]);

  const u32 salt_len = esalt_bufs[digests_offset].salt_len;

  /**
   * digest
   */

  const u32 search[4] =
  {
    digests_buf[digests_offset].digest_buf[DGST_R0],
    digests_buf[digests_offset].digest_buf[DGST_R1],
    digests_buf[digests_offset].digest_buf[DGST_R2],
    digests_buf[digests_offset].digest_buf[DGST_R3]
  };

  /**
   * loop
   */

  for (u32 il_pos = 0; il_pos < il_cnt; il_pos += VECT_SIZE)
  {
    u32x w0[4] = { 0 };
    u32x w1[4] = { 0 };
    u32x w2[4] = { 0 };
    u32x w3[4] = { 0 };

    const u32x out_len = apply_rules_vect (pw_buf0, pw_buf1, pw_len, rules_buf, il_pos, w0, w1);

    append_0x80_2x4_VV (w0, w1, out_len);

    make_utf16le (w1, w2, w3);
    make_utf16le (w0, w0, w1);

    const u32x out_len2 = out_len * 2;

    const u32x out_salt_len = out_len2 + salt_len;

    /**
     * prepend salt -- can't stay outside the loop this time
     */

    w0[0] = swap32 (w0[0]);
    w0[1] = swap32 (w0[1]);
    w0[2] = swap32 (w0[2]);
    w0[3] = swap32 (w0[3]);
    w1[0] = swap32 (w1[0]);
    w1[1] = swap32 (w1[1]);
    w1[2] = swap32 (w1[2]);
    w1[3] = swap32 (w1[3]);
    w2[0] = swap32 (w2[0]);
    w2[1] = swap32 (w2[1]);
    w2[2] = swap32 (w2[2]);
    w2[3] = swap32 (w2[3]);
    w3[0] = swap32 (w3[0]);
    w3[1] = swap32 (w3[1]);
    w3[2] = swap32 (w3[2]);
    w3[3] = swap32 (w3[3]);

    u32x c0[4] = { 0 };
    u32x c1[4] = { 0 };
    u32x c2[4] = { 0 };
    u32x c3[4] = { 0 };

    switch_buffer_by_offset_carry_be (w0, w1, w2, w3, c0, c1, c2, c3, salt_len - pc_offset4);

    w0[0] |= salt_buf0[0];
    w0[1] |= salt_buf0[1];
    w0[2] |= salt_buf0[2];
    w0[3] |= salt_buf0[3];
    w1[0] |= salt_buf1[0];
    w1[1] |= salt_buf1[1];
    w1[2] |= salt_buf1[2];
    w1[3] |= salt_buf1[3];
    w2[0] |= salt_buf2[0];
    w2[1] |= salt_buf2[1];
    w2[2] |= salt_buf2[2];
    w2[3] |= salt_buf2[3];
    w3[0] |= salt_buf3[0];
    w3[1] |= salt_buf3[1];
    w3[2] |= salt_buf3[2];
    w3[3] |= salt_buf3[3];

    u32x a = pc_digest[0];
    u32x b = pc_digest[1];
    u32x c = pc_digest[2];
    u32x d = pc_digest[3];
    u32x e = pc_digest[4];

    const u32 r = (salt_len - pc_offset4) + out_len2;

    if (r >= 56)
    {
      u32x w0_t = w0[0];
      u32x w1_t = w0[1];
      u32x w2_t = w0[2];
      u32x w3_t = w0[3];
      u32x w4_t = w1[0];
      u32x w5_t = w1[1];
      u32x w6_t = w1[2];
      u32x w7_t = w1[3];
      u32x w8_t = w2[0];
      u32x w9_t = w2[1];
      u32x wa_t = w2[2];
      u32x wb_t = w2[3];
      u32x wc_t = w3[0];
      u32x wd_t = w3[1];
      u32x we_t = w3[2];
      u32x wf_t = w3[3];

      #undef K
      #define K SHA1C00

      SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0_t);
      SHA1_STEP (SHA1_F0o, e, a, b, c, d, w1_t);
      SHA1_STEP (SHA1_F0o, d, e, a, b, c, w2_t);
      SHA1_STEP (SHA1_F0o, c, d, e, a, b, w3_t);
      SHA1_STEP (SHA1_F0o, b, c, d, e, a, w4_t);
      SHA1_STEP (SHA1_F0o, a, b, c, d, e, w5_t);
      SHA1_STEP (SHA1_F0o, e, a, b, c, d, w6_t);
      SHA1_STEP (SHA1_F0o, d, e, a, b, c, w7_t);
      SHA1_STEP (SHA1_F0o, c, d, e, a, b, w8_t);
      SHA1_STEP (SHA1_F0o, b, c, d, e, a, w9_t);
      SHA1_STEP (SHA1_F0o, a, b, c, d, e, wa_t);
      SHA1_STEP (SHA1_F0o, e, a, b, c, d, wb_t);
      SHA1_STEP (SHA1_F0o, d, e, a, b, c, wc_t);
      SHA1_STEP (SHA1_F0o, c, d, e, a, b, wd_t);
      SHA1_STEP (SHA1_F0o, b, c, d, e, a, we_t);
      SHA1_STEP (SHA1_F0o, a, b, c, d, e, wf_t);
      w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t);
      w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, d, e, a, b, c, w1_t);
      w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, c, d, e, a, b, w2_t);
      w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, b, c, d, e, a, w3_t);

      #undef K
      #define K SHA1C01

      w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w4_t);
      w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w5_t);
      w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w6_t);
      w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w7_t);
      w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w8_t);
      w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w9_t);
      wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wa_t);
      wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wb_t);
      wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wc_t);
      wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wd_t);
      we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, we_t);
      wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wf_t);
      w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t);
      w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t);
      w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t);
      w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t);
      w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w4_t);
      w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w5_t);
      w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w6_t);
      w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w7_t);

      #undef K
      #define K SHA1C02

      w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w8_t);
      w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w9_t);
      wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wa_t);
      wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wb_t);
      wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wc_t);
      wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, wd_t);
      we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, we_t);
      wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wf_t);
      w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w0_t);
      w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w1_t);
      w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w2_t);
      w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w3_t);
      w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w4_t);
      w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w5_t);
      w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w6_t);
      w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w7_t);
      w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w8_t);
      w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w9_t);
      wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wa_t);
      wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wb_t);

      #undef K
      #define K SHA1C03

      wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wc_t);
      wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wd_t);
      we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, we_t);
      wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wf_t);
      w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t);
      w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t);
      w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w2_t);
      w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t);
      w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w4_t);
      w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w5_t);
      w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w6_t);
      w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w7_t);
      w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w8_t);
      w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w9_t);
      wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wa_t);
      wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wb_t);
      wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wc_t);
      wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wd_t);
      we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, we_t);
      wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wf_t);

      a += pc_digest[0];
      b += pc_digest[1];
      c += pc_digest[2];
      d += pc_digest[3];
      e += pc_digest[4];

      w0[0] = c0[0];
      w0[1] = c0[1];
      w0[2] = c0[2];
      w0[3] = c0[3];
      w1[0] = c1[0];
      w1[1] = c1[1];
      w1[2] = c1[2];
      w1[3] = c1[3];
      w2[0] = c2[0];
      w2[1] = c2[1];
      w2[2] = c2[2];
      w2[3] = c2[3];
      w3[0] = c3[0];
      w3[1] = c3[1];
      w3[2] = c3[2];
      w3[3] = c3[3];
    }

    /**
     * final sha1
     */

    u32x w0_t = w0[0];
    u32x w1_t = w0[1];
    u32x w2_t = w0[2];
    u32x w3_t = w0[3];
    u32x w4_t = w1[0];
    u32x w5_t = w1[1];
    u32x w6_t = w1[2];
    u32x w7_t = w1[3];
    u32x w8_t = w2[0];
    u32x w9_t = w2[1];
    u32x wa_t = w2[2];
    u32x wb_t = w2[3];
    u32x wc_t = w3[0];
    u32x wd_t = w3[1];
    u32x we_t = 0;
    u32x wf_t = out_salt_len * 8;

    u32x r_a = a;
    u32x r_b = b;
    u32x r_c = c;
    u32x r_d = d;
    u32x r_e = e;

    #undef K
    #define K SHA1C00

    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w0_t);
    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w1_t);
    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w2_t);
    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w3_t);
    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w4_t);
    SHA1_STEP (SHA1_F0o, a, b, c, d, e, w5_t);
    SHA1_STEP (SHA1_F0o, e, a, b, c, d, w6_t);
    SHA1_STEP (SHA1_F0o, d, e, a, b, c, w7_t);
    SHA1_STEP (SHA1_F0o, c, d, e, a, b, w8_t);
    SHA1_STEP (SHA1_F0o, b, c, d, e, a, w9_t);
    SHA1_STEP (SHA1_F0o, a, b, c, d, e, wa_t);
    SHA1_STEP (SHA1_F0o, e, a, b, c, d, wb_t);
    SHA1_STEP (SHA1_F0o, d, e, a, b, c, wc_t);
    SHA1_STEP (SHA1_F0o, c, d, e, a, b, wd_t);
    SHA1_STEP (SHA1_F0o, b, c, d, e, a, we_t);
    SHA1_STEP (SHA1_F0o, a, b, c, d, e, wf_t);
    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F0o, e, a, b, c, d, w0_t);
    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F0o, d, e, a, b, c, w1_t);
    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F0o, c, d, e, a, b, w2_t);
    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F0o, b, c, d, e, a, w3_t);

    #undef K
    #define K SHA1C01

    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w4_t);
    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w5_t);
    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w6_t);
    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w7_t);
    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w8_t);
    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w9_t);
    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wa_t);
    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wb_t);
    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wc_t);
    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wd_t);
    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, we_t);
    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wf_t);
    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w0_t);
    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w1_t);
    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w2_t);
    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w3_t);
    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w4_t);
    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w5_t);
    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w6_t);
    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w7_t);

    #undef K
    #define K SHA1C02

    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w8_t);
    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w9_t);
    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wa_t);
    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wb_t);
    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wc_t);
    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, wd_t);
    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, we_t);
    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, wf_t);
    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w0_t);
    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w1_t);
    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w2_t);
    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w3_t);
    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w4_t);
    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, w5_t);
    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, w6_t);
    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F2o, a, b, c, d, e, w7_t);
    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F2o, e, a, b, c, d, w8_t);
    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F2o, d, e, a, b, c, w9_t);
    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F2o, c, d, e, a, b, wa_t);
    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F2o, b, c, d, e, a, wb_t);

    #undef K
    #define K SHA1C03

    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wc_t);
    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wd_t);
    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, we_t);
    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, wf_t);
    w0_t = rotl32 ((wd_t ^ w8_t ^ w2_t ^ w0_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w0_t);
    w1_t = rotl32 ((we_t ^ w9_t ^ w3_t ^ w1_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w1_t);
    w2_t = rotl32 ((wf_t ^ wa_t ^ w4_t ^ w2_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w2_t);
    w3_t = rotl32 ((w0_t ^ wb_t ^ w5_t ^ w3_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w3_t);
    w4_t = rotl32 ((w1_t ^ wc_t ^ w6_t ^ w4_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w4_t);
    w5_t = rotl32 ((w2_t ^ wd_t ^ w7_t ^ w5_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, w5_t);
    w6_t = rotl32 ((w3_t ^ we_t ^ w8_t ^ w6_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, w6_t);
    w7_t = rotl32 ((w4_t ^ wf_t ^ w9_t ^ w7_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, w7_t);
    w8_t = rotl32 ((w5_t ^ w0_t ^ wa_t ^ w8_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, w8_t);
    w9_t = rotl32 ((w6_t ^ w1_t ^ wb_t ^ w9_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, w9_t);
    wa_t = rotl32 ((w7_t ^ w2_t ^ wc_t ^ wa_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wa_t);
    wb_t = rotl32 ((w8_t ^ w3_t ^ wd_t ^ wb_t), 1u); SHA1_STEP (SHA1_F1, a, b, c, d, e, wb_t);
    wc_t = rotl32 ((w9_t ^ w4_t ^ we_t ^ wc_t), 1u); SHA1_STEP (SHA1_F1, e, a, b, c, d, wc_t);
    wd_t = rotl32 ((wa_t ^ w5_t ^ wf_t ^ wd_t), 1u); SHA1_STEP (SHA1_F1, d, e, a, b, c, wd_t);
    we_t = rotl32 ((wb_t ^ w6_t ^ w0_t ^ we_t), 1u); SHA1_STEP (SHA1_F1, c, d, e, a, b, we_t);
    wf_t = rotl32 ((wc_t ^ w7_t ^ w1_t ^ wf_t), 1u); SHA1_STEP (SHA1_F1, b, c, d, e, a, wf_t);

    a += r_a;
    b += r_b;
    c += r_c;
    d += r_d;
    e += r_e;

    COMPARE_S_SIMD (d, e, c, b);
  }
}

KERNEL_FQ void m13500_s08 (KERN_ATTR_RULES_ESALT (pstoken_t))
{
}

KERNEL_FQ void m13500_s16 (KERN_ATTR_RULES_ESALT (pstoken_t))
{
}