hashcat/OpenCL/amp_a1.cl

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

#include "inc_hash_constants.h"
#include "inc_vendor.cl"
#include "inc_types.cl"

inline void switch_buffer_by_offset_1x64_le_S (u32 w[64], const u32 offset)
{
  #if defined IS_AMD || defined IS_GENERIC
  const int offset_mod_4 = offset & 3;

  const int offset_minus_4 = 4 - offset;

  switch (offset / 4)
  {
    case  0:
      w[63] = amd_bytealign_S (w[63], w[62], offset_minus_4);
      w[62] = amd_bytealign_S (w[62], w[61], offset_minus_4);
      w[61] = amd_bytealign_S (w[61], w[60], offset_minus_4);
      w[60] = amd_bytealign_S (w[60], w[59], offset_minus_4);
      w[59] = amd_bytealign_S (w[59], w[58], offset_minus_4);
      w[58] = amd_bytealign_S (w[58], w[57], offset_minus_4);
      w[57] = amd_bytealign_S (w[57], w[56], offset_minus_4);
      w[56] = amd_bytealign_S (w[56], w[55], offset_minus_4);
      w[55] = amd_bytealign_S (w[55], w[54], offset_minus_4);
      w[54] = amd_bytealign_S (w[54], w[53], offset_minus_4);
      w[53] = amd_bytealign_S (w[53], w[52], offset_minus_4);
      w[52] = amd_bytealign_S (w[52], w[51], offset_minus_4);
      w[51] = amd_bytealign_S (w[51], w[50], offset_minus_4);
      w[50] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[49] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[48] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[47] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[46] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[45] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[44] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[43] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[42] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[41] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[40] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[39] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[38] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[37] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[36] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[35] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[34] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[33] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[32] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[31] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[30] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[29] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[28] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[27] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[26] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[25] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[24] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[23] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[22] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[21] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[20] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[19] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[18] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[17] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[16] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[15] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[14] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[13] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[12] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[11] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[10] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[ 9] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[ 8] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[ 7] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[ 6] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[ 5] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[ 4] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[ 3] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[ 2] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[ 1] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[ 0] = amd_bytealign_S (w[ 0],     0, offset_minus_4);

      if (offset_mod_4 == 0)
      {
        w[ 0] = w[ 1];
        w[ 1] = w[ 2];
        w[ 2] = w[ 3];
        w[ 3] = w[ 4];
        w[ 4] = w[ 5];
        w[ 5] = w[ 6];
        w[ 6] = w[ 7];
        w[ 7] = w[ 8];
        w[ 8] = w[ 9];
        w[ 9] = w[10];
        w[10] = w[11];
        w[11] = w[12];
        w[12] = w[13];
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case  1:
      w[63] = amd_bytealign_S (w[62], w[61], offset_minus_4);
      w[62] = amd_bytealign_S (w[61], w[60], offset_minus_4);
      w[61] = amd_bytealign_S (w[60], w[59], offset_minus_4);
      w[60] = amd_bytealign_S (w[59], w[58], offset_minus_4);
      w[59] = amd_bytealign_S (w[58], w[57], offset_minus_4);
      w[58] = amd_bytealign_S (w[57], w[56], offset_minus_4);
      w[57] = amd_bytealign_S (w[56], w[55], offset_minus_4);
      w[56] = amd_bytealign_S (w[55], w[54], offset_minus_4);
      w[55] = amd_bytealign_S (w[54], w[53], offset_minus_4);
      w[54] = amd_bytealign_S (w[53], w[52], offset_minus_4);
      w[53] = amd_bytealign_S (w[52], w[51], offset_minus_4);
      w[52] = amd_bytealign_S (w[51], w[50], offset_minus_4);
      w[51] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[50] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[49] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[48] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[47] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[46] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[45] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[44] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[43] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[42] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[41] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[40] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[39] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[38] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[37] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[36] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[35] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[34] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[33] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[32] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[31] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[30] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[29] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[28] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[27] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[26] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[25] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[24] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[23] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[22] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[21] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[20] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[19] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[18] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[17] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[16] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[15] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[14] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[13] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[12] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[11] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[10] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[ 9] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[ 8] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[ 7] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[ 6] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[ 5] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[ 4] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[ 3] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[ 2] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[ 1] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[ 1] = w[ 2];
        w[ 2] = w[ 3];
        w[ 3] = w[ 4];
        w[ 4] = w[ 5];
        w[ 5] = w[ 6];
        w[ 6] = w[ 7];
        w[ 7] = w[ 8];
        w[ 8] = w[ 9];
        w[ 9] = w[10];
        w[10] = w[11];
        w[11] = w[12];
        w[12] = w[13];
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case  2:
      w[63] = amd_bytealign_S (w[61], w[60], offset_minus_4);
      w[62] = amd_bytealign_S (w[60], w[59], offset_minus_4);
      w[61] = amd_bytealign_S (w[59], w[58], offset_minus_4);
      w[60] = amd_bytealign_S (w[58], w[57], offset_minus_4);
      w[59] = amd_bytealign_S (w[57], w[56], offset_minus_4);
      w[58] = amd_bytealign_S (w[56], w[55], offset_minus_4);
      w[57] = amd_bytealign_S (w[55], w[54], offset_minus_4);
      w[56] = amd_bytealign_S (w[54], w[53], offset_minus_4);
      w[55] = amd_bytealign_S (w[53], w[52], offset_minus_4);
      w[54] = amd_bytealign_S (w[52], w[51], offset_minus_4);
      w[53] = amd_bytealign_S (w[51], w[50], offset_minus_4);
      w[52] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[51] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[50] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[49] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[48] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[47] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[46] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[45] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[44] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[43] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[42] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[41] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[40] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[39] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[38] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[37] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[36] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[35] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[34] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[33] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[32] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[31] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[30] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[29] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[28] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[27] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[26] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[25] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[24] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[23] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[22] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[21] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[20] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[19] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[18] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[17] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[16] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[15] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[14] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[13] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[12] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[11] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[10] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[ 9] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[ 8] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[ 7] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[ 6] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[ 5] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[ 4] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[ 3] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[ 2] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[ 2] = w[ 3];
        w[ 3] = w[ 4];
        w[ 4] = w[ 5];
        w[ 5] = w[ 6];
        w[ 6] = w[ 7];
        w[ 7] = w[ 8];
        w[ 8] = w[ 9];
        w[ 9] = w[10];
        w[10] = w[11];
        w[11] = w[12];
        w[12] = w[13];
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case  3:
      w[63] = amd_bytealign_S (w[60], w[59], offset_minus_4);
      w[62] = amd_bytealign_S (w[59], w[58], offset_minus_4);
      w[61] = amd_bytealign_S (w[58], w[57], offset_minus_4);
      w[60] = amd_bytealign_S (w[57], w[56], offset_minus_4);
      w[59] = amd_bytealign_S (w[56], w[55], offset_minus_4);
      w[58] = amd_bytealign_S (w[55], w[54], offset_minus_4);
      w[57] = amd_bytealign_S (w[54], w[53], offset_minus_4);
      w[56] = amd_bytealign_S (w[53], w[52], offset_minus_4);
      w[55] = amd_bytealign_S (w[52], w[51], offset_minus_4);
      w[54] = amd_bytealign_S (w[51], w[50], offset_minus_4);
      w[53] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[52] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[51] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[50] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[49] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[48] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[47] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[46] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[45] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[44] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[43] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[42] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[41] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[40] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[39] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[38] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[37] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[36] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[35] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[34] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[33] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[32] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[31] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[30] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[29] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[28] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[27] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[26] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[25] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[24] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[23] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[22] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[21] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[20] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[19] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[18] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[17] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[16] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[15] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[14] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[13] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[12] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[11] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[10] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[ 9] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[ 8] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[ 7] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[ 6] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[ 5] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[ 4] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[ 3] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[ 3] = w[ 4];
        w[ 4] = w[ 5];
        w[ 5] = w[ 6];
        w[ 6] = w[ 7];
        w[ 7] = w[ 8];
        w[ 8] = w[ 9];
        w[ 9] = w[10];
        w[10] = w[11];
        w[11] = w[12];
        w[12] = w[13];
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case  4:
      w[63] = amd_bytealign_S (w[59], w[58], offset_minus_4);
      w[62] = amd_bytealign_S (w[58], w[57], offset_minus_4);
      w[61] = amd_bytealign_S (w[57], w[56], offset_minus_4);
      w[60] = amd_bytealign_S (w[56], w[55], offset_minus_4);
      w[59] = amd_bytealign_S (w[55], w[54], offset_minus_4);
      w[58] = amd_bytealign_S (w[54], w[53], offset_minus_4);
      w[57] = amd_bytealign_S (w[53], w[52], offset_minus_4);
      w[56] = amd_bytealign_S (w[52], w[51], offset_minus_4);
      w[55] = amd_bytealign_S (w[51], w[50], offset_minus_4);
      w[54] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[53] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[52] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[51] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[50] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[49] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[48] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[47] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[46] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[45] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[44] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[43] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[42] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[41] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[40] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[39] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[38] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[37] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[36] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[35] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[34] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[33] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[32] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[31] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[30] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[29] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[28] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[27] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[26] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[25] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[24] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[23] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[22] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[21] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[20] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[19] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[18] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[17] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[16] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[15] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[14] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[13] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[12] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[11] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[10] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[ 9] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[ 8] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[ 7] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[ 6] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[ 5] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[ 4] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[ 4] = w[ 5];
        w[ 5] = w[ 6];
        w[ 6] = w[ 7];
        w[ 7] = w[ 8];
        w[ 8] = w[ 9];
        w[ 9] = w[10];
        w[10] = w[11];
        w[11] = w[12];
        w[12] = w[13];
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case  5:
      w[63] = amd_bytealign_S (w[58], w[57], offset_minus_4);
      w[62] = amd_bytealign_S (w[57], w[56], offset_minus_4);
      w[61] = amd_bytealign_S (w[56], w[55], offset_minus_4);
      w[60] = amd_bytealign_S (w[55], w[54], offset_minus_4);
      w[59] = amd_bytealign_S (w[54], w[53], offset_minus_4);
      w[58] = amd_bytealign_S (w[53], w[52], offset_minus_4);
      w[57] = amd_bytealign_S (w[52], w[51], offset_minus_4);
      w[56] = amd_bytealign_S (w[51], w[50], offset_minus_4);
      w[55] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[54] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[53] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[52] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[51] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[50] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[49] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[48] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[47] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[46] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[45] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[44] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[43] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[42] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[41] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[40] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[39] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[38] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[37] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[36] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[35] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[34] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[33] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[32] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[31] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[30] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[29] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[28] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[27] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[26] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[25] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[24] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[23] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[22] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[21] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[20] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[19] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[18] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[17] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[16] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[15] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[14] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[13] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[12] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[11] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[10] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[ 9] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[ 8] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[ 7] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[ 6] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[ 5] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[ 5] = w[ 6];
        w[ 6] = w[ 7];
        w[ 7] = w[ 8];
        w[ 8] = w[ 9];
        w[ 9] = w[10];
        w[10] = w[11];
        w[11] = w[12];
        w[12] = w[13];
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case  6:
      w[63] = amd_bytealign_S (w[57], w[56], offset_minus_4);
      w[62] = amd_bytealign_S (w[56], w[55], offset_minus_4);
      w[61] = amd_bytealign_S (w[55], w[54], offset_minus_4);
      w[60] = amd_bytealign_S (w[54], w[53], offset_minus_4);
      w[59] = amd_bytealign_S (w[53], w[52], offset_minus_4);
      w[58] = amd_bytealign_S (w[52], w[51], offset_minus_4);
      w[57] = amd_bytealign_S (w[51], w[50], offset_minus_4);
      w[56] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[55] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[54] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[53] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[52] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[51] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[50] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[49] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[48] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[47] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[46] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[45] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[44] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[43] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[42] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[41] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[40] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[39] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[38] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[37] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[36] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[35] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[34] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[33] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[32] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[31] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[30] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[29] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[28] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[27] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[26] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[25] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[24] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[23] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[22] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[21] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[20] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[19] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[18] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[17] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[16] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[15] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[14] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[13] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[12] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[11] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[10] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[ 9] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[ 8] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[ 7] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[ 6] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[ 6] = w[ 7];
        w[ 7] = w[ 8];
        w[ 8] = w[ 9];
        w[ 9] = w[10];
        w[10] = w[11];
        w[11] = w[12];
        w[12] = w[13];
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case  7:
      w[63] = amd_bytealign_S (w[56], w[55], offset_minus_4);
      w[62] = amd_bytealign_S (w[55], w[54], offset_minus_4);
      w[61] = amd_bytealign_S (w[54], w[53], offset_minus_4);
      w[60] = amd_bytealign_S (w[53], w[52], offset_minus_4);
      w[59] = amd_bytealign_S (w[52], w[51], offset_minus_4);
      w[58] = amd_bytealign_S (w[51], w[50], offset_minus_4);
      w[57] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[56] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[55] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[54] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[53] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[52] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[51] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[50] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[49] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[48] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[47] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[46] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[45] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[44] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[43] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[42] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[41] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[40] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[39] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[38] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[37] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[36] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[35] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[34] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[33] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[32] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[31] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[30] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[29] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[28] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[27] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[26] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[25] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[24] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[23] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[22] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[21] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[20] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[19] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[18] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[17] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[16] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[15] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[14] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[13] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[12] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[11] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[10] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[ 9] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[ 8] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[ 7] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[ 7] = w[ 8];
        w[ 8] = w[ 9];
        w[ 9] = w[10];
        w[10] = w[11];
        w[11] = w[12];
        w[12] = w[13];
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case  8:
      w[63] = amd_bytealign_S (w[55], w[54], offset_minus_4);
      w[62] = amd_bytealign_S (w[54], w[53], offset_minus_4);
      w[61] = amd_bytealign_S (w[53], w[52], offset_minus_4);
      w[60] = amd_bytealign_S (w[52], w[51], offset_minus_4);
      w[59] = amd_bytealign_S (w[51], w[50], offset_minus_4);
      w[58] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[57] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[56] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[55] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[54] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[53] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[52] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[51] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[50] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[49] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[48] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[47] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[46] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[45] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[44] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[43] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[42] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[41] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[40] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[39] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[38] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[37] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[36] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[35] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[34] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[33] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[32] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[31] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[30] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[29] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[28] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[27] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[26] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[25] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[24] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[23] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[22] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[21] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[20] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[19] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[18] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[17] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[16] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[15] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[14] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[13] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[12] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[11] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[10] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[ 9] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[ 8] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[ 8] = w[ 9];
        w[ 9] = w[10];
        w[10] = w[11];
        w[11] = w[12];
        w[12] = w[13];
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case  9:
      w[63] = amd_bytealign_S (w[54], w[53], offset_minus_4);
      w[62] = amd_bytealign_S (w[53], w[52], offset_minus_4);
      w[61] = amd_bytealign_S (w[52], w[51], offset_minus_4);
      w[60] = amd_bytealign_S (w[51], w[50], offset_minus_4);
      w[59] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[58] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[57] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[56] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[55] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[54] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[53] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[52] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[51] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[50] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[49] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[48] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[47] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[46] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[45] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[44] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[43] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[42] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[41] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[40] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[39] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[38] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[37] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[36] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[35] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[34] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[33] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[32] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[31] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[30] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[29] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[28] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[27] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[26] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[25] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[24] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[23] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[22] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[21] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[20] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[19] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[18] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[17] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[16] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[15] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[14] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[13] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[12] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[11] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[10] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[ 9] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[ 9] = w[10];
        w[10] = w[11];
        w[11] = w[12];
        w[12] = w[13];
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 10:
      w[63] = amd_bytealign_S (w[53], w[52], offset_minus_4);
      w[62] = amd_bytealign_S (w[52], w[51], offset_minus_4);
      w[61] = amd_bytealign_S (w[51], w[50], offset_minus_4);
      w[60] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[59] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[58] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[57] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[56] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[55] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[54] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[53] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[52] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[51] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[50] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[49] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[48] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[47] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[46] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[45] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[44] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[43] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[42] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[41] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[40] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[39] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[38] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[37] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[36] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[35] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[34] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[33] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[32] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[31] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[30] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[29] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[28] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[27] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[26] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[25] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[24] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[23] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[22] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[21] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[20] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[19] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[18] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[17] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[16] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[15] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[14] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[13] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[12] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[11] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[10] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[10] = w[11];
        w[11] = w[12];
        w[12] = w[13];
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 11:
      w[63] = amd_bytealign_S (w[52], w[51], offset_minus_4);
      w[62] = amd_bytealign_S (w[51], w[50], offset_minus_4);
      w[61] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[60] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[59] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[58] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[57] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[56] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[55] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[54] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[53] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[52] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[51] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[50] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[49] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[48] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[47] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[46] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[45] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[44] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[43] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[42] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[41] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[40] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[39] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[38] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[37] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[36] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[35] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[34] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[33] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[32] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[31] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[30] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[29] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[28] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[27] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[26] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[25] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[24] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[23] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[22] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[21] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[20] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[19] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[18] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[17] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[16] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[15] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[14] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[13] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[12] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[11] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[11] = w[12];
        w[12] = w[13];
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 12:
      w[63] = amd_bytealign_S (w[51], w[50], offset_minus_4);
      w[62] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[61] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[60] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[59] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[58] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[57] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[56] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[55] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[54] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[53] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[52] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[51] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[50] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[49] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[48] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[47] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[46] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[45] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[44] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[43] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[42] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[41] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[40] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[39] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[38] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[37] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[36] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[35] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[34] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[33] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[32] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[31] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[30] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[29] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[28] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[27] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[26] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[25] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[24] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[23] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[22] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[21] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[20] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[19] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[18] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[17] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[16] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[15] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[14] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[13] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[12] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[12] = w[13];
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 13:
      w[63] = amd_bytealign_S (w[50], w[49], offset_minus_4);
      w[62] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[61] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[60] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[59] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[58] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[57] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[56] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[55] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[54] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[53] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[52] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[51] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[50] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[49] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[48] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[47] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[46] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[45] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[44] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[43] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[42] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[41] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[40] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[39] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[38] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[37] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[36] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[35] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[34] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[33] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[32] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[31] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[30] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[29] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[28] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[27] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[26] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[25] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[24] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[23] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[22] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[21] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[20] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[19] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[18] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[17] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[16] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[15] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[14] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[13] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[13] = w[14];
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 14:
      w[63] = amd_bytealign_S (w[49], w[48], offset_minus_4);
      w[62] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[61] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[60] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[59] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[58] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[57] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[56] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[55] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[54] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[53] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[52] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[51] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[50] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[49] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[48] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[47] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[46] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[45] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[44] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[43] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[42] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[41] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[40] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[39] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[38] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[37] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[36] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[35] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[34] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[33] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[32] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[31] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[30] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[29] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[28] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[27] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[26] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[25] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[24] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[23] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[22] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[21] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[20] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[19] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[18] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[17] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[16] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[15] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[14] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[14] = w[15];
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 15:
      w[63] = amd_bytealign_S (w[48], w[47], offset_minus_4);
      w[62] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[61] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[60] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[59] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[58] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[57] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[56] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[55] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[54] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[53] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[52] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[51] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[50] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[49] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[48] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[47] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[46] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[45] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[44] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[43] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[42] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[41] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[40] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[39] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[38] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[37] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[36] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[35] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[34] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[33] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[32] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[31] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[30] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[29] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[28] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[27] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[26] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[25] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[24] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[23] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[22] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[21] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[20] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[19] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[18] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[17] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[16] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[15] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[15] = w[16];
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 16:
      w[63] = amd_bytealign_S (w[47], w[46], offset_minus_4);
      w[62] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[61] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[60] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[59] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[58] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[57] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[56] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[55] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[54] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[53] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[52] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[51] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[50] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[49] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[48] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[47] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[46] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[45] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[44] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[43] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[42] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[41] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[40] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[39] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[38] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[37] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[36] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[35] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[34] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[33] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[32] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[31] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[30] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[29] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[28] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[27] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[26] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[25] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[24] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[23] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[22] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[21] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[20] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[19] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[18] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[17] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[16] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[16] = w[17];
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 17:
      w[63] = amd_bytealign_S (w[46], w[45], offset_minus_4);
      w[62] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[61] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[60] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[59] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[58] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[57] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[56] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[55] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[54] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[53] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[52] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[51] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[50] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[49] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[48] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[47] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[46] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[45] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[44] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[43] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[42] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[41] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[40] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[39] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[38] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[37] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[36] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[35] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[34] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[33] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[32] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[31] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[30] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[29] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[28] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[27] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[26] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[25] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[24] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[23] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[22] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[21] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[20] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[19] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[18] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[17] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[17] = w[18];
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 18:
      w[63] = amd_bytealign_S (w[45], w[44], offset_minus_4);
      w[62] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[61] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[60] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[59] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[58] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[57] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[56] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[55] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[54] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[53] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[52] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[51] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[50] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[49] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[48] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[47] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[46] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[45] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[44] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[43] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[42] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[41] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[40] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[39] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[38] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[37] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[36] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[35] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[34] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[33] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[32] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[31] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[30] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[29] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[28] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[27] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[26] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[25] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[24] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[23] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[22] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[21] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[20] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[19] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[18] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[18] = w[19];
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 19:
      w[63] = amd_bytealign_S (w[44], w[43], offset_minus_4);
      w[62] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[61] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[60] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[59] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[58] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[57] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[56] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[55] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[54] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[53] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[52] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[51] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[50] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[49] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[48] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[47] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[46] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[45] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[44] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[43] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[42] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[41] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[40] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[39] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[38] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[37] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[36] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[35] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[34] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[33] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[32] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[31] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[30] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[29] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[28] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[27] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[26] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[25] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[24] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[23] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[22] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[21] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[20] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[19] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[19] = w[20];
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 20:
      w[63] = amd_bytealign_S (w[43], w[42], offset_minus_4);
      w[62] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[61] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[60] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[59] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[58] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[57] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[56] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[55] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[54] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[53] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[52] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[51] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[50] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[49] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[48] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[47] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[46] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[45] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[44] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[43] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[42] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[41] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[40] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[39] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[38] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[37] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[36] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[35] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[34] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[33] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[32] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[31] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[30] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[29] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[28] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[27] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[26] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[25] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[24] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[23] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[22] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[21] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[20] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[20] = w[21];
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 21:
      w[63] = amd_bytealign_S (w[42], w[41], offset_minus_4);
      w[62] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[61] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[60] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[59] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[58] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[57] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[56] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[55] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[54] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[53] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[52] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[51] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[50] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[49] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[48] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[47] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[46] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[45] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[44] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[43] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[42] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[41] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[40] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[39] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[38] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[37] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[36] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[35] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[34] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[33] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[32] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[31] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[30] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[29] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[28] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[27] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[26] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[25] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[24] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[23] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[22] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[21] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[21] = w[22];
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 22:
      w[63] = amd_bytealign_S (w[41], w[40], offset_minus_4);
      w[62] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[61] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[60] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[59] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[58] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[57] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[56] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[55] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[54] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[53] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[52] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[51] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[50] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[49] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[48] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[47] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[46] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[45] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[44] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[43] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[42] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[41] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[40] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[39] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[38] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[37] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[36] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[35] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[34] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[33] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[32] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[31] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[30] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[29] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[28] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[27] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[26] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[25] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[24] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[23] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[22] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[22] = w[23];
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 23:
      w[63] = amd_bytealign_S (w[40], w[39], offset_minus_4);
      w[62] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[61] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[60] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[59] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[58] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[57] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[56] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[55] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[54] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[53] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[52] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[51] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[50] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[49] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[48] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[47] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[46] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[45] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[44] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[43] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[42] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[41] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[40] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[39] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[38] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[37] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[36] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[35] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[34] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[33] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[32] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[31] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[30] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[29] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[28] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[27] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[26] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[25] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[24] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[23] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[23] = w[24];
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 24:
      w[63] = amd_bytealign_S (w[39], w[38], offset_minus_4);
      w[62] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[61] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[60] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[59] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[58] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[57] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[56] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[55] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[54] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[53] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[52] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[51] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[50] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[49] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[48] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[47] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[46] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[45] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[44] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[43] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[42] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[41] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[40] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[39] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[38] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[37] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[36] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[35] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[34] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[33] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[32] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[31] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[30] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[29] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[28] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[27] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[26] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[25] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[24] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[24] = w[25];
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 25:
      w[63] = amd_bytealign_S (w[38], w[37], offset_minus_4);
      w[62] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[61] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[60] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[59] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[58] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[57] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[56] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[55] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[54] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[53] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[52] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[51] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[50] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[49] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[48] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[47] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[46] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[45] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[44] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[43] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[42] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[41] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[40] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[39] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[38] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[37] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[36] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[35] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[34] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[33] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[32] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[31] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[30] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[29] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[28] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[27] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[26] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[25] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[25] = w[26];
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 26:
      w[63] = amd_bytealign_S (w[37], w[36], offset_minus_4);
      w[62] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[61] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[60] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[59] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[58] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[57] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[56] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[55] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[54] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[53] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[52] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[51] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[50] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[49] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[48] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[47] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[46] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[45] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[44] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[43] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[42] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[41] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[40] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[39] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[38] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[37] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[36] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[35] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[34] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[33] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[32] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[31] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[30] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[29] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[28] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[27] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[26] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[26] = w[27];
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 27:
      w[63] = amd_bytealign_S (w[36], w[35], offset_minus_4);
      w[62] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[61] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[60] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[59] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[58] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[57] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[56] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[55] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[54] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[53] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[52] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[51] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[50] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[49] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[48] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[47] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[46] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[45] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[44] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[43] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[42] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[41] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[40] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[39] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[38] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[37] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[36] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[35] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[34] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[33] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[32] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[31] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[30] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[29] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[28] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[27] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[27] = w[28];
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 28:
      w[63] = amd_bytealign_S (w[35], w[34], offset_minus_4);
      w[62] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[61] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[60] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[59] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[58] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[57] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[56] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[55] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[54] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[53] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[52] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[51] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[50] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[49] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[48] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[47] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[46] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[45] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[44] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[43] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[42] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[41] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[40] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[39] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[38] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[37] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[36] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[35] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[34] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[33] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[32] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[31] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[30] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[29] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[28] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[28] = w[29];
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 29:
      w[63] = amd_bytealign_S (w[34], w[33], offset_minus_4);
      w[62] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[61] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[60] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[59] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[58] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[57] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[56] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[55] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[54] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[53] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[52] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[51] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[50] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[49] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[48] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[47] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[46] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[45] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[44] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[43] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[42] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[41] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[40] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[39] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[38] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[37] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[36] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[35] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[34] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[33] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[32] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[31] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[30] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[29] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[29] = w[30];
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 30:
      w[63] = amd_bytealign_S (w[33], w[32], offset_minus_4);
      w[62] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[61] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[60] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[59] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[58] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[57] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[56] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[55] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[54] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[53] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[52] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[51] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[50] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[49] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[48] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[47] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[46] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[45] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[44] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[43] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[42] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[41] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[40] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[39] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[38] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[37] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[36] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[35] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[34] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[33] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[32] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[31] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[30] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[30] = w[31];
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 31:
      w[63] = amd_bytealign_S (w[32], w[31], offset_minus_4);
      w[62] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[61] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[60] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[59] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[58] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[57] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[56] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[55] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[54] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[53] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[52] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[51] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[50] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[49] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[48] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[47] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[46] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[45] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[44] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[43] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[42] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[41] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[40] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[39] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[38] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[37] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[36] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[35] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[34] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[33] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[32] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[31] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[31] = w[32];
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 32:
      w[63] = amd_bytealign_S (w[31], w[30], offset_minus_4);
      w[62] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[61] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[60] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[59] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[58] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[57] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[56] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[55] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[54] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[53] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[52] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[51] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[50] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[49] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[48] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[47] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[46] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[45] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[44] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[43] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[42] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[41] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[40] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[39] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[38] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[37] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[36] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[35] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[34] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[33] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[32] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[32] = w[33];
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 33:
      w[63] = amd_bytealign_S (w[30], w[29], offset_minus_4);
      w[62] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[61] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[60] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[59] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[58] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[57] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[56] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[55] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[54] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[53] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[52] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[51] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[50] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[49] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[48] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[47] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[46] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[45] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[44] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[43] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[42] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[41] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[40] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[39] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[38] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[37] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[36] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[35] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[34] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[33] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[33] = w[34];
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 34:
      w[63] = amd_bytealign_S (w[29], w[28], offset_minus_4);
      w[62] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[61] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[60] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[59] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[58] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[57] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[56] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[55] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[54] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[53] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[52] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[51] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[50] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[49] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[48] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[47] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[46] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[45] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[44] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[43] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[42] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[41] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[40] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[39] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[38] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[37] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[36] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[35] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[34] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[34] = w[35];
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 35:
      w[63] = amd_bytealign_S (w[28], w[27], offset_minus_4);
      w[62] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[61] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[60] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[59] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[58] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[57] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[56] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[55] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[54] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[53] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[52] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[51] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[50] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[49] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[48] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[47] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[46] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[45] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[44] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[43] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[42] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[41] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[40] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[39] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[38] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[37] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[36] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[35] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[35] = w[36];
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 36:
      w[63] = amd_bytealign_S (w[27], w[26], offset_minus_4);
      w[62] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[61] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[60] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[59] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[58] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[57] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[56] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[55] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[54] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[53] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[52] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[51] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[50] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[49] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[48] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[47] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[46] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[45] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[44] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[43] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[42] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[41] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[40] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[39] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[38] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[37] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[36] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[36] = w[37];
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 37:
      w[63] = amd_bytealign_S (w[26], w[25], offset_minus_4);
      w[62] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[61] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[60] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[59] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[58] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[57] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[56] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[55] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[54] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[53] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[52] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[51] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[50] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[49] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[48] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[47] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[46] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[45] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[44] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[43] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[42] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[41] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[40] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[39] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[38] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[37] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[37] = w[38];
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 38:
      w[63] = amd_bytealign_S (w[25], w[24], offset_minus_4);
      w[62] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[61] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[60] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[59] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[58] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[57] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[56] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[55] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[54] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[53] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[52] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[51] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[50] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[49] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[48] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[47] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[46] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[45] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[44] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[43] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[42] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[41] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[40] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[39] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[38] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[38] = w[39];
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 39:
      w[63] = amd_bytealign_S (w[24], w[23], offset_minus_4);
      w[62] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[61] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[60] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[59] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[58] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[57] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[56] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[55] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[54] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[53] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[52] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[51] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[50] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[49] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[48] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[47] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[46] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[45] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[44] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[43] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[42] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[41] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[40] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[39] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[39] = w[40];
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 40:
      w[63] = amd_bytealign_S (w[23], w[22], offset_minus_4);
      w[62] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[61] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[60] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[59] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[58] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[57] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[56] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[55] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[54] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[53] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[52] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[51] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[50] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[49] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[48] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[47] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[46] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[45] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[44] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[43] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[42] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[41] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[40] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[40] = w[41];
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 41:
      w[63] = amd_bytealign_S (w[22], w[21], offset_minus_4);
      w[62] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[61] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[60] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[59] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[58] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[57] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[56] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[55] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[54] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[53] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[52] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[51] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[50] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[49] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[48] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[47] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[46] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[45] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[44] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[43] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[42] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[41] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[41] = w[42];
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 42:
      w[63] = amd_bytealign_S (w[21], w[20], offset_minus_4);
      w[62] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[61] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[60] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[59] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[58] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[57] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[56] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[55] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[54] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[53] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[52] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[51] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[50] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[49] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[48] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[47] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[46] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[45] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[44] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[43] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[42] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[42] = w[43];
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 43:
      w[63] = amd_bytealign_S (w[20], w[19], offset_minus_4);
      w[62] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[61] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[60] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[59] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[58] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[57] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[56] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[55] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[54] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[53] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[52] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[51] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[50] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[49] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[48] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[47] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[46] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[45] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[44] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[43] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[43] = w[44];
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 44:
      w[63] = amd_bytealign_S (w[19], w[18], offset_minus_4);
      w[62] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[61] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[60] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[59] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[58] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[57] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[56] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[55] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[54] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[53] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[52] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[51] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[50] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[49] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[48] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[47] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[46] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[45] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[44] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[44] = w[45];
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 45:
      w[63] = amd_bytealign_S (w[18], w[17], offset_minus_4);
      w[62] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[61] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[60] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[59] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[58] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[57] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[56] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[55] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[54] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[53] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[52] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[51] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[50] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[49] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[48] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[47] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[46] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[45] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[45] = w[46];
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 46:
      w[63] = amd_bytealign_S (w[17], w[16], offset_minus_4);
      w[62] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[61] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[60] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[59] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[58] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[57] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[56] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[55] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[54] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[53] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[52] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[51] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[50] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[49] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[48] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[47] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[46] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[46] = w[47];
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 47:
      w[63] = amd_bytealign_S (w[16], w[15], offset_minus_4);
      w[62] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[61] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[60] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[59] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[58] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[57] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[56] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[55] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[54] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[53] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[52] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[51] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[50] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[49] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[48] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[47] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[47] = w[48];
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 48:
      w[63] = amd_bytealign_S (w[15], w[14], offset_minus_4);
      w[62] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[61] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[60] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[59] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[58] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[57] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[56] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[55] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[54] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[53] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[52] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[51] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[50] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[49] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[48] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[48] = w[49];
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 49:
      w[63] = amd_bytealign_S (w[14], w[13], offset_minus_4);
      w[62] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[61] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[60] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[59] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[58] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[57] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[56] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[55] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[54] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[53] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[52] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[51] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[50] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[49] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[49] = w[50];
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 50:
      w[63] = amd_bytealign_S (w[13], w[12], offset_minus_4);
      w[62] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[61] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[60] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[59] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[58] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[57] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[56] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[55] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[54] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[53] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[52] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[51] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[50] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[50] = w[51];
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 51:
      w[63] = amd_bytealign_S (w[12], w[11], offset_minus_4);
      w[62] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[61] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[60] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[59] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[58] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[57] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[56] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[55] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[54] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[53] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[52] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[51] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[51] = w[52];
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 52:
      w[63] = amd_bytealign_S (w[11], w[10], offset_minus_4);
      w[62] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[61] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[60] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[59] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[58] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[57] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[56] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[55] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[54] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[53] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[52] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[52] = w[53];
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 53:
      w[63] = amd_bytealign_S (w[10], w[ 9], offset_minus_4);
      w[62] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[61] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[60] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[59] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[58] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[57] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[56] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[55] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[54] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[53] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[53] = w[54];
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 54:
      w[63] = amd_bytealign_S (w[ 9], w[ 8], offset_minus_4);
      w[62] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[61] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[60] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[59] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[58] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[57] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[56] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[55] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[54] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[54] = w[55];
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 55:
      w[63] = amd_bytealign_S (w[ 8], w[ 7], offset_minus_4);
      w[62] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[61] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[60] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[59] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[58] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[57] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[56] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[55] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[55] = w[56];
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 56:
      w[63] = amd_bytealign_S (w[ 7], w[ 6], offset_minus_4);
      w[62] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[61] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[60] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[59] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[58] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[57] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[56] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[56] = w[57];
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 57:
      w[63] = amd_bytealign_S (w[ 6], w[ 5], offset_minus_4);
      w[62] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[61] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[60] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[59] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[58] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[57] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[57] = w[58];
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 58:
      w[63] = amd_bytealign_S (w[ 5], w[ 4], offset_minus_4);
      w[62] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[61] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[60] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[59] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[58] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[57] = 0;
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[58] = w[59];
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 59:
      w[63] = amd_bytealign_S (w[ 4], w[ 3], offset_minus_4);
      w[62] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[61] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[60] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[59] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[58] = 0;
      w[57] = 0;
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[59] = w[60];
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 60:
      w[63] = amd_bytealign_S (w[ 3], w[ 2], offset_minus_4);
      w[62] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[61] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[60] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[59] = 0;
      w[58] = 0;
      w[57] = 0;
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[60] = w[61];
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 61:
      w[63] = amd_bytealign_S (w[ 2], w[ 1], offset_minus_4);
      w[62] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[61] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[60] = 0;
      w[59] = 0;
      w[58] = 0;
      w[57] = 0;
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[61] = w[62];
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 62:
      w[63] = amd_bytealign_S (w[ 1], w[ 0], offset_minus_4);
      w[62] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[61] = 0;
      w[60] = 0;
      w[59] = 0;
      w[58] = 0;
      w[57] = 0;
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[62] = w[63];
        w[63] = 0;
      }

      break;

    case 63:
      w[63] = amd_bytealign_S (w[ 0],     0, offset_minus_4);
      w[62] = 0;
      w[61] = 0;
      w[60] = 0;
      w[59] = 0;
      w[58] = 0;
      w[57] = 0;
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      if (offset_mod_4 == 0)
      {
        w[63] = 0;
      }

      break;
  }
  #endif

  #ifdef IS_NV
  const int offset_minus_4 = 4 - (offset % 4);

  const int selector = (0x76543210 >> (offset_minus_4 * 4)) & 0xffff;

  switch (offset / 4)
  {
    case  0:
      w[63] = __byte_perm_S (w[62], w[63], selector);
      w[62] = __byte_perm_S (w[61], w[62], selector);
      w[61] = __byte_perm_S (w[60], w[61], selector);
      w[60] = __byte_perm_S (w[59], w[60], selector);
      w[59] = __byte_perm_S (w[58], w[59], selector);
      w[58] = __byte_perm_S (w[57], w[58], selector);
      w[57] = __byte_perm_S (w[56], w[57], selector);
      w[56] = __byte_perm_S (w[55], w[56], selector);
      w[55] = __byte_perm_S (w[54], w[55], selector);
      w[54] = __byte_perm_S (w[53], w[54], selector);
      w[53] = __byte_perm_S (w[52], w[53], selector);
      w[52] = __byte_perm_S (w[51], w[52], selector);
      w[51] = __byte_perm_S (w[50], w[51], selector);
      w[50] = __byte_perm_S (w[49], w[50], selector);
      w[49] = __byte_perm_S (w[48], w[49], selector);
      w[48] = __byte_perm_S (w[47], w[48], selector);
      w[47] = __byte_perm_S (w[46], w[47], selector);
      w[46] = __byte_perm_S (w[45], w[46], selector);
      w[45] = __byte_perm_S (w[44], w[45], selector);
      w[44] = __byte_perm_S (w[43], w[44], selector);
      w[43] = __byte_perm_S (w[42], w[43], selector);
      w[42] = __byte_perm_S (w[41], w[42], selector);
      w[41] = __byte_perm_S (w[40], w[41], selector);
      w[40] = __byte_perm_S (w[39], w[40], selector);
      w[39] = __byte_perm_S (w[38], w[39], selector);
      w[38] = __byte_perm_S (w[37], w[38], selector);
      w[37] = __byte_perm_S (w[36], w[37], selector);
      w[36] = __byte_perm_S (w[35], w[36], selector);
      w[35] = __byte_perm_S (w[34], w[35], selector);
      w[34] = __byte_perm_S (w[33], w[34], selector);
      w[33] = __byte_perm_S (w[32], w[33], selector);
      w[32] = __byte_perm_S (w[31], w[32], selector);
      w[31] = __byte_perm_S (w[30], w[31], selector);
      w[30] = __byte_perm_S (w[29], w[30], selector);
      w[29] = __byte_perm_S (w[28], w[29], selector);
      w[28] = __byte_perm_S (w[27], w[28], selector);
      w[27] = __byte_perm_S (w[26], w[27], selector);
      w[26] = __byte_perm_S (w[25], w[26], selector);
      w[25] = __byte_perm_S (w[24], w[25], selector);
      w[24] = __byte_perm_S (w[23], w[24], selector);
      w[23] = __byte_perm_S (w[22], w[23], selector);
      w[22] = __byte_perm_S (w[21], w[22], selector);
      w[21] = __byte_perm_S (w[20], w[21], selector);
      w[20] = __byte_perm_S (w[19], w[20], selector);
      w[19] = __byte_perm_S (w[18], w[19], selector);
      w[18] = __byte_perm_S (w[17], w[18], selector);
      w[17] = __byte_perm_S (w[16], w[17], selector);
      w[16] = __byte_perm_S (w[15], w[16], selector);
      w[15] = __byte_perm_S (w[14], w[15], selector);
      w[14] = __byte_perm_S (w[13], w[14], selector);
      w[13] = __byte_perm_S (w[12], w[13], selector);
      w[12] = __byte_perm_S (w[11], w[12], selector);
      w[11] = __byte_perm_S (w[10], w[11], selector);
      w[10] = __byte_perm_S (w[ 9], w[10], selector);
      w[ 9] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[ 8] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[ 7] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[ 6] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[ 5] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[ 4] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[ 3] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[ 2] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[ 1] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[ 0] = __byte_perm_S (    0, w[ 0], selector);

      break;

    case  1:
      w[63] = __byte_perm_S (w[61], w[62], selector);
      w[62] = __byte_perm_S (w[60], w[61], selector);
      w[61] = __byte_perm_S (w[59], w[60], selector);
      w[60] = __byte_perm_S (w[58], w[59], selector);
      w[59] = __byte_perm_S (w[57], w[58], selector);
      w[58] = __byte_perm_S (w[56], w[57], selector);
      w[57] = __byte_perm_S (w[55], w[56], selector);
      w[56] = __byte_perm_S (w[54], w[55], selector);
      w[55] = __byte_perm_S (w[53], w[54], selector);
      w[54] = __byte_perm_S (w[52], w[53], selector);
      w[53] = __byte_perm_S (w[51], w[52], selector);
      w[52] = __byte_perm_S (w[50], w[51], selector);
      w[51] = __byte_perm_S (w[49], w[50], selector);
      w[50] = __byte_perm_S (w[48], w[49], selector);
      w[49] = __byte_perm_S (w[47], w[48], selector);
      w[48] = __byte_perm_S (w[46], w[47], selector);
      w[47] = __byte_perm_S (w[45], w[46], selector);
      w[46] = __byte_perm_S (w[44], w[45], selector);
      w[45] = __byte_perm_S (w[43], w[44], selector);
      w[44] = __byte_perm_S (w[42], w[43], selector);
      w[43] = __byte_perm_S (w[41], w[42], selector);
      w[42] = __byte_perm_S (w[40], w[41], selector);
      w[41] = __byte_perm_S (w[39], w[40], selector);
      w[40] = __byte_perm_S (w[38], w[39], selector);
      w[39] = __byte_perm_S (w[37], w[38], selector);
      w[38] = __byte_perm_S (w[36], w[37], selector);
      w[37] = __byte_perm_S (w[35], w[36], selector);
      w[36] = __byte_perm_S (w[34], w[35], selector);
      w[35] = __byte_perm_S (w[33], w[34], selector);
      w[34] = __byte_perm_S (w[32], w[33], selector);
      w[33] = __byte_perm_S (w[31], w[32], selector);
      w[32] = __byte_perm_S (w[30], w[31], selector);
      w[31] = __byte_perm_S (w[29], w[30], selector);
      w[30] = __byte_perm_S (w[28], w[29], selector);
      w[29] = __byte_perm_S (w[27], w[28], selector);
      w[28] = __byte_perm_S (w[26], w[27], selector);
      w[27] = __byte_perm_S (w[25], w[26], selector);
      w[26] = __byte_perm_S (w[24], w[25], selector);
      w[25] = __byte_perm_S (w[23], w[24], selector);
      w[24] = __byte_perm_S (w[22], w[23], selector);
      w[23] = __byte_perm_S (w[21], w[22], selector);
      w[22] = __byte_perm_S (w[20], w[21], selector);
      w[21] = __byte_perm_S (w[19], w[20], selector);
      w[20] = __byte_perm_S (w[18], w[19], selector);
      w[19] = __byte_perm_S (w[17], w[18], selector);
      w[18] = __byte_perm_S (w[16], w[17], selector);
      w[17] = __byte_perm_S (w[15], w[16], selector);
      w[16] = __byte_perm_S (w[14], w[15], selector);
      w[15] = __byte_perm_S (w[13], w[14], selector);
      w[14] = __byte_perm_S (w[12], w[13], selector);
      w[13] = __byte_perm_S (w[11], w[12], selector);
      w[12] = __byte_perm_S (w[10], w[11], selector);
      w[11] = __byte_perm_S (w[ 9], w[10], selector);
      w[10] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[ 9] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[ 8] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[ 7] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[ 6] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[ 5] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[ 4] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[ 3] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[ 2] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[ 1] = __byte_perm_S (    0, w[ 0], selector);
      w[ 0] = 0;

      break;

    case  2:
      w[63] = __byte_perm_S (w[60], w[61], selector);
      w[62] = __byte_perm_S (w[59], w[60], selector);
      w[61] = __byte_perm_S (w[58], w[59], selector);
      w[60] = __byte_perm_S (w[57], w[58], selector);
      w[59] = __byte_perm_S (w[56], w[57], selector);
      w[58] = __byte_perm_S (w[55], w[56], selector);
      w[57] = __byte_perm_S (w[54], w[55], selector);
      w[56] = __byte_perm_S (w[53], w[54], selector);
      w[55] = __byte_perm_S (w[52], w[53], selector);
      w[54] = __byte_perm_S (w[51], w[52], selector);
      w[53] = __byte_perm_S (w[50], w[51], selector);
      w[52] = __byte_perm_S (w[49], w[50], selector);
      w[51] = __byte_perm_S (w[48], w[49], selector);
      w[50] = __byte_perm_S (w[47], w[48], selector);
      w[49] = __byte_perm_S (w[46], w[47], selector);
      w[48] = __byte_perm_S (w[45], w[46], selector);
      w[47] = __byte_perm_S (w[44], w[45], selector);
      w[46] = __byte_perm_S (w[43], w[44], selector);
      w[45] = __byte_perm_S (w[42], w[43], selector);
      w[44] = __byte_perm_S (w[41], w[42], selector);
      w[43] = __byte_perm_S (w[40], w[41], selector);
      w[42] = __byte_perm_S (w[39], w[40], selector);
      w[41] = __byte_perm_S (w[38], w[39], selector);
      w[40] = __byte_perm_S (w[37], w[38], selector);
      w[39] = __byte_perm_S (w[36], w[37], selector);
      w[38] = __byte_perm_S (w[35], w[36], selector);
      w[37] = __byte_perm_S (w[34], w[35], selector);
      w[36] = __byte_perm_S (w[33], w[34], selector);
      w[35] = __byte_perm_S (w[32], w[33], selector);
      w[34] = __byte_perm_S (w[31], w[32], selector);
      w[33] = __byte_perm_S (w[30], w[31], selector);
      w[32] = __byte_perm_S (w[29], w[30], selector);
      w[31] = __byte_perm_S (w[28], w[29], selector);
      w[30] = __byte_perm_S (w[27], w[28], selector);
      w[29] = __byte_perm_S (w[26], w[27], selector);
      w[28] = __byte_perm_S (w[25], w[26], selector);
      w[27] = __byte_perm_S (w[24], w[25], selector);
      w[26] = __byte_perm_S (w[23], w[24], selector);
      w[25] = __byte_perm_S (w[22], w[23], selector);
      w[24] = __byte_perm_S (w[21], w[22], selector);
      w[23] = __byte_perm_S (w[20], w[21], selector);
      w[22] = __byte_perm_S (w[19], w[20], selector);
      w[21] = __byte_perm_S (w[18], w[19], selector);
      w[20] = __byte_perm_S (w[17], w[18], selector);
      w[19] = __byte_perm_S (w[16], w[17], selector);
      w[18] = __byte_perm_S (w[15], w[16], selector);
      w[17] = __byte_perm_S (w[14], w[15], selector);
      w[16] = __byte_perm_S (w[13], w[14], selector);
      w[15] = __byte_perm_S (w[12], w[13], selector);
      w[14] = __byte_perm_S (w[11], w[12], selector);
      w[13] = __byte_perm_S (w[10], w[11], selector);
      w[12] = __byte_perm_S (w[ 9], w[10], selector);
      w[11] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[10] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[ 9] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[ 8] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[ 7] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[ 6] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[ 5] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[ 4] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[ 3] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[ 2] = __byte_perm_S (    0, w[ 0], selector);
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case  3:
      w[63] = __byte_perm_S (w[59], w[60], selector);
      w[62] = __byte_perm_S (w[58], w[59], selector);
      w[61] = __byte_perm_S (w[57], w[58], selector);
      w[60] = __byte_perm_S (w[56], w[57], selector);
      w[59] = __byte_perm_S (w[55], w[56], selector);
      w[58] = __byte_perm_S (w[54], w[55], selector);
      w[57] = __byte_perm_S (w[53], w[54], selector);
      w[56] = __byte_perm_S (w[52], w[53], selector);
      w[55] = __byte_perm_S (w[51], w[52], selector);
      w[54] = __byte_perm_S (w[50], w[51], selector);
      w[53] = __byte_perm_S (w[49], w[50], selector);
      w[52] = __byte_perm_S (w[48], w[49], selector);
      w[51] = __byte_perm_S (w[47], w[48], selector);
      w[50] = __byte_perm_S (w[46], w[47], selector);
      w[49] = __byte_perm_S (w[45], w[46], selector);
      w[48] = __byte_perm_S (w[44], w[45], selector);
      w[47] = __byte_perm_S (w[43], w[44], selector);
      w[46] = __byte_perm_S (w[42], w[43], selector);
      w[45] = __byte_perm_S (w[41], w[42], selector);
      w[44] = __byte_perm_S (w[40], w[41], selector);
      w[43] = __byte_perm_S (w[39], w[40], selector);
      w[42] = __byte_perm_S (w[38], w[39], selector);
      w[41] = __byte_perm_S (w[37], w[38], selector);
      w[40] = __byte_perm_S (w[36], w[37], selector);
      w[39] = __byte_perm_S (w[35], w[36], selector);
      w[38] = __byte_perm_S (w[34], w[35], selector);
      w[37] = __byte_perm_S (w[33], w[34], selector);
      w[36] = __byte_perm_S (w[32], w[33], selector);
      w[35] = __byte_perm_S (w[31], w[32], selector);
      w[34] = __byte_perm_S (w[30], w[31], selector);
      w[33] = __byte_perm_S (w[29], w[30], selector);
      w[32] = __byte_perm_S (w[28], w[29], selector);
      w[31] = __byte_perm_S (w[27], w[28], selector);
      w[30] = __byte_perm_S (w[26], w[27], selector);
      w[29] = __byte_perm_S (w[25], w[26], selector);
      w[28] = __byte_perm_S (w[24], w[25], selector);
      w[27] = __byte_perm_S (w[23], w[24], selector);
      w[26] = __byte_perm_S (w[22], w[23], selector);
      w[25] = __byte_perm_S (w[21], w[22], selector);
      w[24] = __byte_perm_S (w[20], w[21], selector);
      w[23] = __byte_perm_S (w[19], w[20], selector);
      w[22] = __byte_perm_S (w[18], w[19], selector);
      w[21] = __byte_perm_S (w[17], w[18], selector);
      w[20] = __byte_perm_S (w[16], w[17], selector);
      w[19] = __byte_perm_S (w[15], w[16], selector);
      w[18] = __byte_perm_S (w[14], w[15], selector);
      w[17] = __byte_perm_S (w[13], w[14], selector);
      w[16] = __byte_perm_S (w[12], w[13], selector);
      w[15] = __byte_perm_S (w[11], w[12], selector);
      w[14] = __byte_perm_S (w[10], w[11], selector);
      w[13] = __byte_perm_S (w[ 9], w[10], selector);
      w[12] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[11] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[10] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[ 9] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[ 8] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[ 7] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[ 6] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[ 5] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[ 4] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[ 3] = __byte_perm_S (    0, w[ 0], selector);
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case  4:
      w[63] = __byte_perm_S (w[58], w[59], selector);
      w[62] = __byte_perm_S (w[57], w[58], selector);
      w[61] = __byte_perm_S (w[56], w[57], selector);
      w[60] = __byte_perm_S (w[55], w[56], selector);
      w[59] = __byte_perm_S (w[54], w[55], selector);
      w[58] = __byte_perm_S (w[53], w[54], selector);
      w[57] = __byte_perm_S (w[52], w[53], selector);
      w[56] = __byte_perm_S (w[51], w[52], selector);
      w[55] = __byte_perm_S (w[50], w[51], selector);
      w[54] = __byte_perm_S (w[49], w[50], selector);
      w[53] = __byte_perm_S (w[48], w[49], selector);
      w[52] = __byte_perm_S (w[47], w[48], selector);
      w[51] = __byte_perm_S (w[46], w[47], selector);
      w[50] = __byte_perm_S (w[45], w[46], selector);
      w[49] = __byte_perm_S (w[44], w[45], selector);
      w[48] = __byte_perm_S (w[43], w[44], selector);
      w[47] = __byte_perm_S (w[42], w[43], selector);
      w[46] = __byte_perm_S (w[41], w[42], selector);
      w[45] = __byte_perm_S (w[40], w[41], selector);
      w[44] = __byte_perm_S (w[39], w[40], selector);
      w[43] = __byte_perm_S (w[38], w[39], selector);
      w[42] = __byte_perm_S (w[37], w[38], selector);
      w[41] = __byte_perm_S (w[36], w[37], selector);
      w[40] = __byte_perm_S (w[35], w[36], selector);
      w[39] = __byte_perm_S (w[34], w[35], selector);
      w[38] = __byte_perm_S (w[33], w[34], selector);
      w[37] = __byte_perm_S (w[32], w[33], selector);
      w[36] = __byte_perm_S (w[31], w[32], selector);
      w[35] = __byte_perm_S (w[30], w[31], selector);
      w[34] = __byte_perm_S (w[29], w[30], selector);
      w[33] = __byte_perm_S (w[28], w[29], selector);
      w[32] = __byte_perm_S (w[27], w[28], selector);
      w[31] = __byte_perm_S (w[26], w[27], selector);
      w[30] = __byte_perm_S (w[25], w[26], selector);
      w[29] = __byte_perm_S (w[24], w[25], selector);
      w[28] = __byte_perm_S (w[23], w[24], selector);
      w[27] = __byte_perm_S (w[22], w[23], selector);
      w[26] = __byte_perm_S (w[21], w[22], selector);
      w[25] = __byte_perm_S (w[20], w[21], selector);
      w[24] = __byte_perm_S (w[19], w[20], selector);
      w[23] = __byte_perm_S (w[18], w[19], selector);
      w[22] = __byte_perm_S (w[17], w[18], selector);
      w[21] = __byte_perm_S (w[16], w[17], selector);
      w[20] = __byte_perm_S (w[15], w[16], selector);
      w[19] = __byte_perm_S (w[14], w[15], selector);
      w[18] = __byte_perm_S (w[13], w[14], selector);
      w[17] = __byte_perm_S (w[12], w[13], selector);
      w[16] = __byte_perm_S (w[11], w[12], selector);
      w[15] = __byte_perm_S (w[10], w[11], selector);
      w[14] = __byte_perm_S (w[ 9], w[10], selector);
      w[13] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[12] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[11] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[10] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[ 9] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[ 8] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[ 7] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[ 6] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[ 5] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[ 4] = __byte_perm_S (    0, w[ 0], selector);
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case  5:
      w[63] = __byte_perm_S (w[57], w[58], selector);
      w[62] = __byte_perm_S (w[56], w[57], selector);
      w[61] = __byte_perm_S (w[55], w[56], selector);
      w[60] = __byte_perm_S (w[54], w[55], selector);
      w[59] = __byte_perm_S (w[53], w[54], selector);
      w[58] = __byte_perm_S (w[52], w[53], selector);
      w[57] = __byte_perm_S (w[51], w[52], selector);
      w[56] = __byte_perm_S (w[50], w[51], selector);
      w[55] = __byte_perm_S (w[49], w[50], selector);
      w[54] = __byte_perm_S (w[48], w[49], selector);
      w[53] = __byte_perm_S (w[47], w[48], selector);
      w[52] = __byte_perm_S (w[46], w[47], selector);
      w[51] = __byte_perm_S (w[45], w[46], selector);
      w[50] = __byte_perm_S (w[44], w[45], selector);
      w[49] = __byte_perm_S (w[43], w[44], selector);
      w[48] = __byte_perm_S (w[42], w[43], selector);
      w[47] = __byte_perm_S (w[41], w[42], selector);
      w[46] = __byte_perm_S (w[40], w[41], selector);
      w[45] = __byte_perm_S (w[39], w[40], selector);
      w[44] = __byte_perm_S (w[38], w[39], selector);
      w[43] = __byte_perm_S (w[37], w[38], selector);
      w[42] = __byte_perm_S (w[36], w[37], selector);
      w[41] = __byte_perm_S (w[35], w[36], selector);
      w[40] = __byte_perm_S (w[34], w[35], selector);
      w[39] = __byte_perm_S (w[33], w[34], selector);
      w[38] = __byte_perm_S (w[32], w[33], selector);
      w[37] = __byte_perm_S (w[31], w[32], selector);
      w[36] = __byte_perm_S (w[30], w[31], selector);
      w[35] = __byte_perm_S (w[29], w[30], selector);
      w[34] = __byte_perm_S (w[28], w[29], selector);
      w[33] = __byte_perm_S (w[27], w[28], selector);
      w[32] = __byte_perm_S (w[26], w[27], selector);
      w[31] = __byte_perm_S (w[25], w[26], selector);
      w[30] = __byte_perm_S (w[24], w[25], selector);
      w[29] = __byte_perm_S (w[23], w[24], selector);
      w[28] = __byte_perm_S (w[22], w[23], selector);
      w[27] = __byte_perm_S (w[21], w[22], selector);
      w[26] = __byte_perm_S (w[20], w[21], selector);
      w[25] = __byte_perm_S (w[19], w[20], selector);
      w[24] = __byte_perm_S (w[18], w[19], selector);
      w[23] = __byte_perm_S (w[17], w[18], selector);
      w[22] = __byte_perm_S (w[16], w[17], selector);
      w[21] = __byte_perm_S (w[15], w[16], selector);
      w[20] = __byte_perm_S (w[14], w[15], selector);
      w[19] = __byte_perm_S (w[13], w[14], selector);
      w[18] = __byte_perm_S (w[12], w[13], selector);
      w[17] = __byte_perm_S (w[11], w[12], selector);
      w[16] = __byte_perm_S (w[10], w[11], selector);
      w[15] = __byte_perm_S (w[ 9], w[10], selector);
      w[14] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[13] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[12] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[11] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[10] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[ 9] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[ 8] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[ 7] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[ 6] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[ 5] = __byte_perm_S (    0, w[ 0], selector);
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case  6:
      w[63] = __byte_perm_S (w[56], w[57], selector);
      w[62] = __byte_perm_S (w[55], w[56], selector);
      w[61] = __byte_perm_S (w[54], w[55], selector);
      w[60] = __byte_perm_S (w[53], w[54], selector);
      w[59] = __byte_perm_S (w[52], w[53], selector);
      w[58] = __byte_perm_S (w[51], w[52], selector);
      w[57] = __byte_perm_S (w[50], w[51], selector);
      w[56] = __byte_perm_S (w[49], w[50], selector);
      w[55] = __byte_perm_S (w[48], w[49], selector);
      w[54] = __byte_perm_S (w[47], w[48], selector);
      w[53] = __byte_perm_S (w[46], w[47], selector);
      w[52] = __byte_perm_S (w[45], w[46], selector);
      w[51] = __byte_perm_S (w[44], w[45], selector);
      w[50] = __byte_perm_S (w[43], w[44], selector);
      w[49] = __byte_perm_S (w[42], w[43], selector);
      w[48] = __byte_perm_S (w[41], w[42], selector);
      w[47] = __byte_perm_S (w[40], w[41], selector);
      w[46] = __byte_perm_S (w[39], w[40], selector);
      w[45] = __byte_perm_S (w[38], w[39], selector);
      w[44] = __byte_perm_S (w[37], w[38], selector);
      w[43] = __byte_perm_S (w[36], w[37], selector);
      w[42] = __byte_perm_S (w[35], w[36], selector);
      w[41] = __byte_perm_S (w[34], w[35], selector);
      w[40] = __byte_perm_S (w[33], w[34], selector);
      w[39] = __byte_perm_S (w[32], w[33], selector);
      w[38] = __byte_perm_S (w[31], w[32], selector);
      w[37] = __byte_perm_S (w[30], w[31], selector);
      w[36] = __byte_perm_S (w[29], w[30], selector);
      w[35] = __byte_perm_S (w[28], w[29], selector);
      w[34] = __byte_perm_S (w[27], w[28], selector);
      w[33] = __byte_perm_S (w[26], w[27], selector);
      w[32] = __byte_perm_S (w[25], w[26], selector);
      w[31] = __byte_perm_S (w[24], w[25], selector);
      w[30] = __byte_perm_S (w[23], w[24], selector);
      w[29] = __byte_perm_S (w[22], w[23], selector);
      w[28] = __byte_perm_S (w[21], w[22], selector);
      w[27] = __byte_perm_S (w[20], w[21], selector);
      w[26] = __byte_perm_S (w[19], w[20], selector);
      w[25] = __byte_perm_S (w[18], w[19], selector);
      w[24] = __byte_perm_S (w[17], w[18], selector);
      w[23] = __byte_perm_S (w[16], w[17], selector);
      w[22] = __byte_perm_S (w[15], w[16], selector);
      w[21] = __byte_perm_S (w[14], w[15], selector);
      w[20] = __byte_perm_S (w[13], w[14], selector);
      w[19] = __byte_perm_S (w[12], w[13], selector);
      w[18] = __byte_perm_S (w[11], w[12], selector);
      w[17] = __byte_perm_S (w[10], w[11], selector);
      w[16] = __byte_perm_S (w[ 9], w[10], selector);
      w[15] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[14] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[13] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[12] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[11] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[10] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[ 9] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[ 8] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[ 7] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[ 6] = __byte_perm_S (    0, w[ 0], selector);
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case  7:
      w[63] = __byte_perm_S (w[55], w[56], selector);
      w[62] = __byte_perm_S (w[54], w[55], selector);
      w[61] = __byte_perm_S (w[53], w[54], selector);
      w[60] = __byte_perm_S (w[52], w[53], selector);
      w[59] = __byte_perm_S (w[51], w[52], selector);
      w[58] = __byte_perm_S (w[50], w[51], selector);
      w[57] = __byte_perm_S (w[49], w[50], selector);
      w[56] = __byte_perm_S (w[48], w[49], selector);
      w[55] = __byte_perm_S (w[47], w[48], selector);
      w[54] = __byte_perm_S (w[46], w[47], selector);
      w[53] = __byte_perm_S (w[45], w[46], selector);
      w[52] = __byte_perm_S (w[44], w[45], selector);
      w[51] = __byte_perm_S (w[43], w[44], selector);
      w[50] = __byte_perm_S (w[42], w[43], selector);
      w[49] = __byte_perm_S (w[41], w[42], selector);
      w[48] = __byte_perm_S (w[40], w[41], selector);
      w[47] = __byte_perm_S (w[39], w[40], selector);
      w[46] = __byte_perm_S (w[38], w[39], selector);
      w[45] = __byte_perm_S (w[37], w[38], selector);
      w[44] = __byte_perm_S (w[36], w[37], selector);
      w[43] = __byte_perm_S (w[35], w[36], selector);
      w[42] = __byte_perm_S (w[34], w[35], selector);
      w[41] = __byte_perm_S (w[33], w[34], selector);
      w[40] = __byte_perm_S (w[32], w[33], selector);
      w[39] = __byte_perm_S (w[31], w[32], selector);
      w[38] = __byte_perm_S (w[30], w[31], selector);
      w[37] = __byte_perm_S (w[29], w[30], selector);
      w[36] = __byte_perm_S (w[28], w[29], selector);
      w[35] = __byte_perm_S (w[27], w[28], selector);
      w[34] = __byte_perm_S (w[26], w[27], selector);
      w[33] = __byte_perm_S (w[25], w[26], selector);
      w[32] = __byte_perm_S (w[24], w[25], selector);
      w[31] = __byte_perm_S (w[23], w[24], selector);
      w[30] = __byte_perm_S (w[22], w[23], selector);
      w[29] = __byte_perm_S (w[21], w[22], selector);
      w[28] = __byte_perm_S (w[20], w[21], selector);
      w[27] = __byte_perm_S (w[19], w[20], selector);
      w[26] = __byte_perm_S (w[18], w[19], selector);
      w[25] = __byte_perm_S (w[17], w[18], selector);
      w[24] = __byte_perm_S (w[16], w[17], selector);
      w[23] = __byte_perm_S (w[15], w[16], selector);
      w[22] = __byte_perm_S (w[14], w[15], selector);
      w[21] = __byte_perm_S (w[13], w[14], selector);
      w[20] = __byte_perm_S (w[12], w[13], selector);
      w[19] = __byte_perm_S (w[11], w[12], selector);
      w[18] = __byte_perm_S (w[10], w[11], selector);
      w[17] = __byte_perm_S (w[ 9], w[10], selector);
      w[16] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[15] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[14] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[13] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[12] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[11] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[10] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[ 9] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[ 8] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[ 7] = __byte_perm_S (    0, w[ 0], selector);
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case  8:
      w[63] = __byte_perm_S (w[54], w[55], selector);
      w[62] = __byte_perm_S (w[53], w[54], selector);
      w[61] = __byte_perm_S (w[52], w[53], selector);
      w[60] = __byte_perm_S (w[51], w[52], selector);
      w[59] = __byte_perm_S (w[50], w[51], selector);
      w[58] = __byte_perm_S (w[49], w[50], selector);
      w[57] = __byte_perm_S (w[48], w[49], selector);
      w[56] = __byte_perm_S (w[47], w[48], selector);
      w[55] = __byte_perm_S (w[46], w[47], selector);
      w[54] = __byte_perm_S (w[45], w[46], selector);
      w[53] = __byte_perm_S (w[44], w[45], selector);
      w[52] = __byte_perm_S (w[43], w[44], selector);
      w[51] = __byte_perm_S (w[42], w[43], selector);
      w[50] = __byte_perm_S (w[41], w[42], selector);
      w[49] = __byte_perm_S (w[40], w[41], selector);
      w[48] = __byte_perm_S (w[39], w[40], selector);
      w[47] = __byte_perm_S (w[38], w[39], selector);
      w[46] = __byte_perm_S (w[37], w[38], selector);
      w[45] = __byte_perm_S (w[36], w[37], selector);
      w[44] = __byte_perm_S (w[35], w[36], selector);
      w[43] = __byte_perm_S (w[34], w[35], selector);
      w[42] = __byte_perm_S (w[33], w[34], selector);
      w[41] = __byte_perm_S (w[32], w[33], selector);
      w[40] = __byte_perm_S (w[31], w[32], selector);
      w[39] = __byte_perm_S (w[30], w[31], selector);
      w[38] = __byte_perm_S (w[29], w[30], selector);
      w[37] = __byte_perm_S (w[28], w[29], selector);
      w[36] = __byte_perm_S (w[27], w[28], selector);
      w[35] = __byte_perm_S (w[26], w[27], selector);
      w[34] = __byte_perm_S (w[25], w[26], selector);
      w[33] = __byte_perm_S (w[24], w[25], selector);
      w[32] = __byte_perm_S (w[23], w[24], selector);
      w[31] = __byte_perm_S (w[22], w[23], selector);
      w[30] = __byte_perm_S (w[21], w[22], selector);
      w[29] = __byte_perm_S (w[20], w[21], selector);
      w[28] = __byte_perm_S (w[19], w[20], selector);
      w[27] = __byte_perm_S (w[18], w[19], selector);
      w[26] = __byte_perm_S (w[17], w[18], selector);
      w[25] = __byte_perm_S (w[16], w[17], selector);
      w[24] = __byte_perm_S (w[15], w[16], selector);
      w[23] = __byte_perm_S (w[14], w[15], selector);
      w[22] = __byte_perm_S (w[13], w[14], selector);
      w[21] = __byte_perm_S (w[12], w[13], selector);
      w[20] = __byte_perm_S (w[11], w[12], selector);
      w[19] = __byte_perm_S (w[10], w[11], selector);
      w[18] = __byte_perm_S (w[ 9], w[10], selector);
      w[17] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[16] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[15] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[14] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[13] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[12] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[11] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[10] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[ 9] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[ 8] = __byte_perm_S (    0, w[ 0], selector);
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case  9:
      w[63] = __byte_perm_S (w[53], w[54], selector);
      w[62] = __byte_perm_S (w[52], w[53], selector);
      w[61] = __byte_perm_S (w[51], w[52], selector);
      w[60] = __byte_perm_S (w[50], w[51], selector);
      w[59] = __byte_perm_S (w[49], w[50], selector);
      w[58] = __byte_perm_S (w[48], w[49], selector);
      w[57] = __byte_perm_S (w[47], w[48], selector);
      w[56] = __byte_perm_S (w[46], w[47], selector);
      w[55] = __byte_perm_S (w[45], w[46], selector);
      w[54] = __byte_perm_S (w[44], w[45], selector);
      w[53] = __byte_perm_S (w[43], w[44], selector);
      w[52] = __byte_perm_S (w[42], w[43], selector);
      w[51] = __byte_perm_S (w[41], w[42], selector);
      w[50] = __byte_perm_S (w[40], w[41], selector);
      w[49] = __byte_perm_S (w[39], w[40], selector);
      w[48] = __byte_perm_S (w[38], w[39], selector);
      w[47] = __byte_perm_S (w[37], w[38], selector);
      w[46] = __byte_perm_S (w[36], w[37], selector);
      w[45] = __byte_perm_S (w[35], w[36], selector);
      w[44] = __byte_perm_S (w[34], w[35], selector);
      w[43] = __byte_perm_S (w[33], w[34], selector);
      w[42] = __byte_perm_S (w[32], w[33], selector);
      w[41] = __byte_perm_S (w[31], w[32], selector);
      w[40] = __byte_perm_S (w[30], w[31], selector);
      w[39] = __byte_perm_S (w[29], w[30], selector);
      w[38] = __byte_perm_S (w[28], w[29], selector);
      w[37] = __byte_perm_S (w[27], w[28], selector);
      w[36] = __byte_perm_S (w[26], w[27], selector);
      w[35] = __byte_perm_S (w[25], w[26], selector);
      w[34] = __byte_perm_S (w[24], w[25], selector);
      w[33] = __byte_perm_S (w[23], w[24], selector);
      w[32] = __byte_perm_S (w[22], w[23], selector);
      w[31] = __byte_perm_S (w[21], w[22], selector);
      w[30] = __byte_perm_S (w[20], w[21], selector);
      w[29] = __byte_perm_S (w[19], w[20], selector);
      w[28] = __byte_perm_S (w[18], w[19], selector);
      w[27] = __byte_perm_S (w[17], w[18], selector);
      w[26] = __byte_perm_S (w[16], w[17], selector);
      w[25] = __byte_perm_S (w[15], w[16], selector);
      w[24] = __byte_perm_S (w[14], w[15], selector);
      w[23] = __byte_perm_S (w[13], w[14], selector);
      w[22] = __byte_perm_S (w[12], w[13], selector);
      w[21] = __byte_perm_S (w[11], w[12], selector);
      w[20] = __byte_perm_S (w[10], w[11], selector);
      w[19] = __byte_perm_S (w[ 9], w[10], selector);
      w[18] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[17] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[16] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[15] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[14] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[13] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[12] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[11] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[10] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[ 9] = __byte_perm_S (    0, w[ 0], selector);
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 10:
      w[63] = __byte_perm_S (w[52], w[53], selector);
      w[62] = __byte_perm_S (w[51], w[52], selector);
      w[61] = __byte_perm_S (w[50], w[51], selector);
      w[60] = __byte_perm_S (w[49], w[50], selector);
      w[59] = __byte_perm_S (w[48], w[49], selector);
      w[58] = __byte_perm_S (w[47], w[48], selector);
      w[57] = __byte_perm_S (w[46], w[47], selector);
      w[56] = __byte_perm_S (w[45], w[46], selector);
      w[55] = __byte_perm_S (w[44], w[45], selector);
      w[54] = __byte_perm_S (w[43], w[44], selector);
      w[53] = __byte_perm_S (w[42], w[43], selector);
      w[52] = __byte_perm_S (w[41], w[42], selector);
      w[51] = __byte_perm_S (w[40], w[41], selector);
      w[50] = __byte_perm_S (w[39], w[40], selector);
      w[49] = __byte_perm_S (w[38], w[39], selector);
      w[48] = __byte_perm_S (w[37], w[38], selector);
      w[47] = __byte_perm_S (w[36], w[37], selector);
      w[46] = __byte_perm_S (w[35], w[36], selector);
      w[45] = __byte_perm_S (w[34], w[35], selector);
      w[44] = __byte_perm_S (w[33], w[34], selector);
      w[43] = __byte_perm_S (w[32], w[33], selector);
      w[42] = __byte_perm_S (w[31], w[32], selector);
      w[41] = __byte_perm_S (w[30], w[31], selector);
      w[40] = __byte_perm_S (w[29], w[30], selector);
      w[39] = __byte_perm_S (w[28], w[29], selector);
      w[38] = __byte_perm_S (w[27], w[28], selector);
      w[37] = __byte_perm_S (w[26], w[27], selector);
      w[36] = __byte_perm_S (w[25], w[26], selector);
      w[35] = __byte_perm_S (w[24], w[25], selector);
      w[34] = __byte_perm_S (w[23], w[24], selector);
      w[33] = __byte_perm_S (w[22], w[23], selector);
      w[32] = __byte_perm_S (w[21], w[22], selector);
      w[31] = __byte_perm_S (w[20], w[21], selector);
      w[30] = __byte_perm_S (w[19], w[20], selector);
      w[29] = __byte_perm_S (w[18], w[19], selector);
      w[28] = __byte_perm_S (w[17], w[18], selector);
      w[27] = __byte_perm_S (w[16], w[17], selector);
      w[26] = __byte_perm_S (w[15], w[16], selector);
      w[25] = __byte_perm_S (w[14], w[15], selector);
      w[24] = __byte_perm_S (w[13], w[14], selector);
      w[23] = __byte_perm_S (w[12], w[13], selector);
      w[22] = __byte_perm_S (w[11], w[12], selector);
      w[21] = __byte_perm_S (w[10], w[11], selector);
      w[20] = __byte_perm_S (w[ 9], w[10], selector);
      w[19] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[18] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[17] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[16] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[15] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[14] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[13] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[12] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[11] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[10] = __byte_perm_S (    0, w[ 0], selector);
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 11:
      w[63] = __byte_perm_S (w[51], w[52], selector);
      w[62] = __byte_perm_S (w[50], w[51], selector);
      w[61] = __byte_perm_S (w[49], w[50], selector);
      w[60] = __byte_perm_S (w[48], w[49], selector);
      w[59] = __byte_perm_S (w[47], w[48], selector);
      w[58] = __byte_perm_S (w[46], w[47], selector);
      w[57] = __byte_perm_S (w[45], w[46], selector);
      w[56] = __byte_perm_S (w[44], w[45], selector);
      w[55] = __byte_perm_S (w[43], w[44], selector);
      w[54] = __byte_perm_S (w[42], w[43], selector);
      w[53] = __byte_perm_S (w[41], w[42], selector);
      w[52] = __byte_perm_S (w[40], w[41], selector);
      w[51] = __byte_perm_S (w[39], w[40], selector);
      w[50] = __byte_perm_S (w[38], w[39], selector);
      w[49] = __byte_perm_S (w[37], w[38], selector);
      w[48] = __byte_perm_S (w[36], w[37], selector);
      w[47] = __byte_perm_S (w[35], w[36], selector);
      w[46] = __byte_perm_S (w[34], w[35], selector);
      w[45] = __byte_perm_S (w[33], w[34], selector);
      w[44] = __byte_perm_S (w[32], w[33], selector);
      w[43] = __byte_perm_S (w[31], w[32], selector);
      w[42] = __byte_perm_S (w[30], w[31], selector);
      w[41] = __byte_perm_S (w[29], w[30], selector);
      w[40] = __byte_perm_S (w[28], w[29], selector);
      w[39] = __byte_perm_S (w[27], w[28], selector);
      w[38] = __byte_perm_S (w[26], w[27], selector);
      w[37] = __byte_perm_S (w[25], w[26], selector);
      w[36] = __byte_perm_S (w[24], w[25], selector);
      w[35] = __byte_perm_S (w[23], w[24], selector);
      w[34] = __byte_perm_S (w[22], w[23], selector);
      w[33] = __byte_perm_S (w[21], w[22], selector);
      w[32] = __byte_perm_S (w[20], w[21], selector);
      w[31] = __byte_perm_S (w[19], w[20], selector);
      w[30] = __byte_perm_S (w[18], w[19], selector);
      w[29] = __byte_perm_S (w[17], w[18], selector);
      w[28] = __byte_perm_S (w[16], w[17], selector);
      w[27] = __byte_perm_S (w[15], w[16], selector);
      w[26] = __byte_perm_S (w[14], w[15], selector);
      w[25] = __byte_perm_S (w[13], w[14], selector);
      w[24] = __byte_perm_S (w[12], w[13], selector);
      w[23] = __byte_perm_S (w[11], w[12], selector);
      w[22] = __byte_perm_S (w[10], w[11], selector);
      w[21] = __byte_perm_S (w[ 9], w[10], selector);
      w[20] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[19] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[18] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[17] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[16] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[15] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[14] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[13] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[12] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[11] = __byte_perm_S (    0, w[ 0], selector);
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 12:
      w[63] = __byte_perm_S (w[50], w[51], selector);
      w[62] = __byte_perm_S (w[49], w[50], selector);
      w[61] = __byte_perm_S (w[48], w[49], selector);
      w[60] = __byte_perm_S (w[47], w[48], selector);
      w[59] = __byte_perm_S (w[46], w[47], selector);
      w[58] = __byte_perm_S (w[45], w[46], selector);
      w[57] = __byte_perm_S (w[44], w[45], selector);
      w[56] = __byte_perm_S (w[43], w[44], selector);
      w[55] = __byte_perm_S (w[42], w[43], selector);
      w[54] = __byte_perm_S (w[41], w[42], selector);
      w[53] = __byte_perm_S (w[40], w[41], selector);
      w[52] = __byte_perm_S (w[39], w[40], selector);
      w[51] = __byte_perm_S (w[38], w[39], selector);
      w[50] = __byte_perm_S (w[37], w[38], selector);
      w[49] = __byte_perm_S (w[36], w[37], selector);
      w[48] = __byte_perm_S (w[35], w[36], selector);
      w[47] = __byte_perm_S (w[34], w[35], selector);
      w[46] = __byte_perm_S (w[33], w[34], selector);
      w[45] = __byte_perm_S (w[32], w[33], selector);
      w[44] = __byte_perm_S (w[31], w[32], selector);
      w[43] = __byte_perm_S (w[30], w[31], selector);
      w[42] = __byte_perm_S (w[29], w[30], selector);
      w[41] = __byte_perm_S (w[28], w[29], selector);
      w[40] = __byte_perm_S (w[27], w[28], selector);
      w[39] = __byte_perm_S (w[26], w[27], selector);
      w[38] = __byte_perm_S (w[25], w[26], selector);
      w[37] = __byte_perm_S (w[24], w[25], selector);
      w[36] = __byte_perm_S (w[23], w[24], selector);
      w[35] = __byte_perm_S (w[22], w[23], selector);
      w[34] = __byte_perm_S (w[21], w[22], selector);
      w[33] = __byte_perm_S (w[20], w[21], selector);
      w[32] = __byte_perm_S (w[19], w[20], selector);
      w[31] = __byte_perm_S (w[18], w[19], selector);
      w[30] = __byte_perm_S (w[17], w[18], selector);
      w[29] = __byte_perm_S (w[16], w[17], selector);
      w[28] = __byte_perm_S (w[15], w[16], selector);
      w[27] = __byte_perm_S (w[14], w[15], selector);
      w[26] = __byte_perm_S (w[13], w[14], selector);
      w[25] = __byte_perm_S (w[12], w[13], selector);
      w[24] = __byte_perm_S (w[11], w[12], selector);
      w[23] = __byte_perm_S (w[10], w[11], selector);
      w[22] = __byte_perm_S (w[ 9], w[10], selector);
      w[21] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[20] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[19] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[18] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[17] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[16] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[15] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[14] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[13] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[12] = __byte_perm_S (    0, w[ 0], selector);
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 13:
      w[63] = __byte_perm_S (w[49], w[50], selector);
      w[62] = __byte_perm_S (w[48], w[49], selector);
      w[61] = __byte_perm_S (w[47], w[48], selector);
      w[60] = __byte_perm_S (w[46], w[47], selector);
      w[59] = __byte_perm_S (w[45], w[46], selector);
      w[58] = __byte_perm_S (w[44], w[45], selector);
      w[57] = __byte_perm_S (w[43], w[44], selector);
      w[56] = __byte_perm_S (w[42], w[43], selector);
      w[55] = __byte_perm_S (w[41], w[42], selector);
      w[54] = __byte_perm_S (w[40], w[41], selector);
      w[53] = __byte_perm_S (w[39], w[40], selector);
      w[52] = __byte_perm_S (w[38], w[39], selector);
      w[51] = __byte_perm_S (w[37], w[38], selector);
      w[50] = __byte_perm_S (w[36], w[37], selector);
      w[49] = __byte_perm_S (w[35], w[36], selector);
      w[48] = __byte_perm_S (w[34], w[35], selector);
      w[47] = __byte_perm_S (w[33], w[34], selector);
      w[46] = __byte_perm_S (w[32], w[33], selector);
      w[45] = __byte_perm_S (w[31], w[32], selector);
      w[44] = __byte_perm_S (w[30], w[31], selector);
      w[43] = __byte_perm_S (w[29], w[30], selector);
      w[42] = __byte_perm_S (w[28], w[29], selector);
      w[41] = __byte_perm_S (w[27], w[28], selector);
      w[40] = __byte_perm_S (w[26], w[27], selector);
      w[39] = __byte_perm_S (w[25], w[26], selector);
      w[38] = __byte_perm_S (w[24], w[25], selector);
      w[37] = __byte_perm_S (w[23], w[24], selector);
      w[36] = __byte_perm_S (w[22], w[23], selector);
      w[35] = __byte_perm_S (w[21], w[22], selector);
      w[34] = __byte_perm_S (w[20], w[21], selector);
      w[33] = __byte_perm_S (w[19], w[20], selector);
      w[32] = __byte_perm_S (w[18], w[19], selector);
      w[31] = __byte_perm_S (w[17], w[18], selector);
      w[30] = __byte_perm_S (w[16], w[17], selector);
      w[29] = __byte_perm_S (w[15], w[16], selector);
      w[28] = __byte_perm_S (w[14], w[15], selector);
      w[27] = __byte_perm_S (w[13], w[14], selector);
      w[26] = __byte_perm_S (w[12], w[13], selector);
      w[25] = __byte_perm_S (w[11], w[12], selector);
      w[24] = __byte_perm_S (w[10], w[11], selector);
      w[23] = __byte_perm_S (w[ 9], w[10], selector);
      w[22] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[21] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[20] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[19] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[18] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[17] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[16] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[15] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[14] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[13] = __byte_perm_S (    0, w[ 0], selector);
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 14:
      w[63] = __byte_perm_S (w[48], w[49], selector);
      w[62] = __byte_perm_S (w[47], w[48], selector);
      w[61] = __byte_perm_S (w[46], w[47], selector);
      w[60] = __byte_perm_S (w[45], w[46], selector);
      w[59] = __byte_perm_S (w[44], w[45], selector);
      w[58] = __byte_perm_S (w[43], w[44], selector);
      w[57] = __byte_perm_S (w[42], w[43], selector);
      w[56] = __byte_perm_S (w[41], w[42], selector);
      w[55] = __byte_perm_S (w[40], w[41], selector);
      w[54] = __byte_perm_S (w[39], w[40], selector);
      w[53] = __byte_perm_S (w[38], w[39], selector);
      w[52] = __byte_perm_S (w[37], w[38], selector);
      w[51] = __byte_perm_S (w[36], w[37], selector);
      w[50] = __byte_perm_S (w[35], w[36], selector);
      w[49] = __byte_perm_S (w[34], w[35], selector);
      w[48] = __byte_perm_S (w[33], w[34], selector);
      w[47] = __byte_perm_S (w[32], w[33], selector);
      w[46] = __byte_perm_S (w[31], w[32], selector);
      w[45] = __byte_perm_S (w[30], w[31], selector);
      w[44] = __byte_perm_S (w[29], w[30], selector);
      w[43] = __byte_perm_S (w[28], w[29], selector);
      w[42] = __byte_perm_S (w[27], w[28], selector);
      w[41] = __byte_perm_S (w[26], w[27], selector);
      w[40] = __byte_perm_S (w[25], w[26], selector);
      w[39] = __byte_perm_S (w[24], w[25], selector);
      w[38] = __byte_perm_S (w[23], w[24], selector);
      w[37] = __byte_perm_S (w[22], w[23], selector);
      w[36] = __byte_perm_S (w[21], w[22], selector);
      w[35] = __byte_perm_S (w[20], w[21], selector);
      w[34] = __byte_perm_S (w[19], w[20], selector);
      w[33] = __byte_perm_S (w[18], w[19], selector);
      w[32] = __byte_perm_S (w[17], w[18], selector);
      w[31] = __byte_perm_S (w[16], w[17], selector);
      w[30] = __byte_perm_S (w[15], w[16], selector);
      w[29] = __byte_perm_S (w[14], w[15], selector);
      w[28] = __byte_perm_S (w[13], w[14], selector);
      w[27] = __byte_perm_S (w[12], w[13], selector);
      w[26] = __byte_perm_S (w[11], w[12], selector);
      w[25] = __byte_perm_S (w[10], w[11], selector);
      w[24] = __byte_perm_S (w[ 9], w[10], selector);
      w[23] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[22] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[21] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[20] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[19] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[18] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[17] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[16] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[15] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[14] = __byte_perm_S (    0, w[ 0], selector);
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 15:
      w[63] = __byte_perm_S (w[47], w[48], selector);
      w[62] = __byte_perm_S (w[46], w[47], selector);
      w[61] = __byte_perm_S (w[45], w[46], selector);
      w[60] = __byte_perm_S (w[44], w[45], selector);
      w[59] = __byte_perm_S (w[43], w[44], selector);
      w[58] = __byte_perm_S (w[42], w[43], selector);
      w[57] = __byte_perm_S (w[41], w[42], selector);
      w[56] = __byte_perm_S (w[40], w[41], selector);
      w[55] = __byte_perm_S (w[39], w[40], selector);
      w[54] = __byte_perm_S (w[38], w[39], selector);
      w[53] = __byte_perm_S (w[37], w[38], selector);
      w[52] = __byte_perm_S (w[36], w[37], selector);
      w[51] = __byte_perm_S (w[35], w[36], selector);
      w[50] = __byte_perm_S (w[34], w[35], selector);
      w[49] = __byte_perm_S (w[33], w[34], selector);
      w[48] = __byte_perm_S (w[32], w[33], selector);
      w[47] = __byte_perm_S (w[31], w[32], selector);
      w[46] = __byte_perm_S (w[30], w[31], selector);
      w[45] = __byte_perm_S (w[29], w[30], selector);
      w[44] = __byte_perm_S (w[28], w[29], selector);
      w[43] = __byte_perm_S (w[27], w[28], selector);
      w[42] = __byte_perm_S (w[26], w[27], selector);
      w[41] = __byte_perm_S (w[25], w[26], selector);
      w[40] = __byte_perm_S (w[24], w[25], selector);
      w[39] = __byte_perm_S (w[23], w[24], selector);
      w[38] = __byte_perm_S (w[22], w[23], selector);
      w[37] = __byte_perm_S (w[21], w[22], selector);
      w[36] = __byte_perm_S (w[20], w[21], selector);
      w[35] = __byte_perm_S (w[19], w[20], selector);
      w[34] = __byte_perm_S (w[18], w[19], selector);
      w[33] = __byte_perm_S (w[17], w[18], selector);
      w[32] = __byte_perm_S (w[16], w[17], selector);
      w[31] = __byte_perm_S (w[15], w[16], selector);
      w[30] = __byte_perm_S (w[14], w[15], selector);
      w[29] = __byte_perm_S (w[13], w[14], selector);
      w[28] = __byte_perm_S (w[12], w[13], selector);
      w[27] = __byte_perm_S (w[11], w[12], selector);
      w[26] = __byte_perm_S (w[10], w[11], selector);
      w[25] = __byte_perm_S (w[ 9], w[10], selector);
      w[24] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[23] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[22] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[21] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[20] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[19] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[18] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[17] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[16] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[15] = __byte_perm_S (    0, w[ 0], selector);
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 16:
      w[63] = __byte_perm_S (w[46], w[47], selector);
      w[62] = __byte_perm_S (w[45], w[46], selector);
      w[61] = __byte_perm_S (w[44], w[45], selector);
      w[60] = __byte_perm_S (w[43], w[44], selector);
      w[59] = __byte_perm_S (w[42], w[43], selector);
      w[58] = __byte_perm_S (w[41], w[42], selector);
      w[57] = __byte_perm_S (w[40], w[41], selector);
      w[56] = __byte_perm_S (w[39], w[40], selector);
      w[55] = __byte_perm_S (w[38], w[39], selector);
      w[54] = __byte_perm_S (w[37], w[38], selector);
      w[53] = __byte_perm_S (w[36], w[37], selector);
      w[52] = __byte_perm_S (w[35], w[36], selector);
      w[51] = __byte_perm_S (w[34], w[35], selector);
      w[50] = __byte_perm_S (w[33], w[34], selector);
      w[49] = __byte_perm_S (w[32], w[33], selector);
      w[48] = __byte_perm_S (w[31], w[32], selector);
      w[47] = __byte_perm_S (w[30], w[31], selector);
      w[46] = __byte_perm_S (w[29], w[30], selector);
      w[45] = __byte_perm_S (w[28], w[29], selector);
      w[44] = __byte_perm_S (w[27], w[28], selector);
      w[43] = __byte_perm_S (w[26], w[27], selector);
      w[42] = __byte_perm_S (w[25], w[26], selector);
      w[41] = __byte_perm_S (w[24], w[25], selector);
      w[40] = __byte_perm_S (w[23], w[24], selector);
      w[39] = __byte_perm_S (w[22], w[23], selector);
      w[38] = __byte_perm_S (w[21], w[22], selector);
      w[37] = __byte_perm_S (w[20], w[21], selector);
      w[36] = __byte_perm_S (w[19], w[20], selector);
      w[35] = __byte_perm_S (w[18], w[19], selector);
      w[34] = __byte_perm_S (w[17], w[18], selector);
      w[33] = __byte_perm_S (w[16], w[17], selector);
      w[32] = __byte_perm_S (w[15], w[16], selector);
      w[31] = __byte_perm_S (w[14], w[15], selector);
      w[30] = __byte_perm_S (w[13], w[14], selector);
      w[29] = __byte_perm_S (w[12], w[13], selector);
      w[28] = __byte_perm_S (w[11], w[12], selector);
      w[27] = __byte_perm_S (w[10], w[11], selector);
      w[26] = __byte_perm_S (w[ 9], w[10], selector);
      w[25] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[24] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[23] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[22] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[21] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[20] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[19] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[18] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[17] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[16] = __byte_perm_S (    0, w[ 0], selector);
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 17:
      w[63] = __byte_perm_S (w[45], w[46], selector);
      w[62] = __byte_perm_S (w[44], w[45], selector);
      w[61] = __byte_perm_S (w[43], w[44], selector);
      w[60] = __byte_perm_S (w[42], w[43], selector);
      w[59] = __byte_perm_S (w[41], w[42], selector);
      w[58] = __byte_perm_S (w[40], w[41], selector);
      w[57] = __byte_perm_S (w[39], w[40], selector);
      w[56] = __byte_perm_S (w[38], w[39], selector);
      w[55] = __byte_perm_S (w[37], w[38], selector);
      w[54] = __byte_perm_S (w[36], w[37], selector);
      w[53] = __byte_perm_S (w[35], w[36], selector);
      w[52] = __byte_perm_S (w[34], w[35], selector);
      w[51] = __byte_perm_S (w[33], w[34], selector);
      w[50] = __byte_perm_S (w[32], w[33], selector);
      w[49] = __byte_perm_S (w[31], w[32], selector);
      w[48] = __byte_perm_S (w[30], w[31], selector);
      w[47] = __byte_perm_S (w[29], w[30], selector);
      w[46] = __byte_perm_S (w[28], w[29], selector);
      w[45] = __byte_perm_S (w[27], w[28], selector);
      w[44] = __byte_perm_S (w[26], w[27], selector);
      w[43] = __byte_perm_S (w[25], w[26], selector);
      w[42] = __byte_perm_S (w[24], w[25], selector);
      w[41] = __byte_perm_S (w[23], w[24], selector);
      w[40] = __byte_perm_S (w[22], w[23], selector);
      w[39] = __byte_perm_S (w[21], w[22], selector);
      w[38] = __byte_perm_S (w[20], w[21], selector);
      w[37] = __byte_perm_S (w[19], w[20], selector);
      w[36] = __byte_perm_S (w[18], w[19], selector);
      w[35] = __byte_perm_S (w[17], w[18], selector);
      w[34] = __byte_perm_S (w[16], w[17], selector);
      w[33] = __byte_perm_S (w[15], w[16], selector);
      w[32] = __byte_perm_S (w[14], w[15], selector);
      w[31] = __byte_perm_S (w[13], w[14], selector);
      w[30] = __byte_perm_S (w[12], w[13], selector);
      w[29] = __byte_perm_S (w[11], w[12], selector);
      w[28] = __byte_perm_S (w[10], w[11], selector);
      w[27] = __byte_perm_S (w[ 9], w[10], selector);
      w[26] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[25] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[24] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[23] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[22] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[21] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[20] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[19] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[18] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[17] = __byte_perm_S (    0, w[ 0], selector);
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 18:
      w[63] = __byte_perm_S (w[44], w[45], selector);
      w[62] = __byte_perm_S (w[43], w[44], selector);
      w[61] = __byte_perm_S (w[42], w[43], selector);
      w[60] = __byte_perm_S (w[41], w[42], selector);
      w[59] = __byte_perm_S (w[40], w[41], selector);
      w[58] = __byte_perm_S (w[39], w[40], selector);
      w[57] = __byte_perm_S (w[38], w[39], selector);
      w[56] = __byte_perm_S (w[37], w[38], selector);
      w[55] = __byte_perm_S (w[36], w[37], selector);
      w[54] = __byte_perm_S (w[35], w[36], selector);
      w[53] = __byte_perm_S (w[34], w[35], selector);
      w[52] = __byte_perm_S (w[33], w[34], selector);
      w[51] = __byte_perm_S (w[32], w[33], selector);
      w[50] = __byte_perm_S (w[31], w[32], selector);
      w[49] = __byte_perm_S (w[30], w[31], selector);
      w[48] = __byte_perm_S (w[29], w[30], selector);
      w[47] = __byte_perm_S (w[28], w[29], selector);
      w[46] = __byte_perm_S (w[27], w[28], selector);
      w[45] = __byte_perm_S (w[26], w[27], selector);
      w[44] = __byte_perm_S (w[25], w[26], selector);
      w[43] = __byte_perm_S (w[24], w[25], selector);
      w[42] = __byte_perm_S (w[23], w[24], selector);
      w[41] = __byte_perm_S (w[22], w[23], selector);
      w[40] = __byte_perm_S (w[21], w[22], selector);
      w[39] = __byte_perm_S (w[20], w[21], selector);
      w[38] = __byte_perm_S (w[19], w[20], selector);
      w[37] = __byte_perm_S (w[18], w[19], selector);
      w[36] = __byte_perm_S (w[17], w[18], selector);
      w[35] = __byte_perm_S (w[16], w[17], selector);
      w[34] = __byte_perm_S (w[15], w[16], selector);
      w[33] = __byte_perm_S (w[14], w[15], selector);
      w[32] = __byte_perm_S (w[13], w[14], selector);
      w[31] = __byte_perm_S (w[12], w[13], selector);
      w[30] = __byte_perm_S (w[11], w[12], selector);
      w[29] = __byte_perm_S (w[10], w[11], selector);
      w[28] = __byte_perm_S (w[ 9], w[10], selector);
      w[27] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[26] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[25] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[24] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[23] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[22] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[21] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[20] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[19] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[18] = __byte_perm_S (    0, w[ 0], selector);
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 19:
      w[63] = __byte_perm_S (w[43], w[44], selector);
      w[62] = __byte_perm_S (w[42], w[43], selector);
      w[61] = __byte_perm_S (w[41], w[42], selector);
      w[60] = __byte_perm_S (w[40], w[41], selector);
      w[59] = __byte_perm_S (w[39], w[40], selector);
      w[58] = __byte_perm_S (w[38], w[39], selector);
      w[57] = __byte_perm_S (w[37], w[38], selector);
      w[56] = __byte_perm_S (w[36], w[37], selector);
      w[55] = __byte_perm_S (w[35], w[36], selector);
      w[54] = __byte_perm_S (w[34], w[35], selector);
      w[53] = __byte_perm_S (w[33], w[34], selector);
      w[52] = __byte_perm_S (w[32], w[33], selector);
      w[51] = __byte_perm_S (w[31], w[32], selector);
      w[50] = __byte_perm_S (w[30], w[31], selector);
      w[49] = __byte_perm_S (w[29], w[30], selector);
      w[48] = __byte_perm_S (w[28], w[29], selector);
      w[47] = __byte_perm_S (w[27], w[28], selector);
      w[46] = __byte_perm_S (w[26], w[27], selector);
      w[45] = __byte_perm_S (w[25], w[26], selector);
      w[44] = __byte_perm_S (w[24], w[25], selector);
      w[43] = __byte_perm_S (w[23], w[24], selector);
      w[42] = __byte_perm_S (w[22], w[23], selector);
      w[41] = __byte_perm_S (w[21], w[22], selector);
      w[40] = __byte_perm_S (w[20], w[21], selector);
      w[39] = __byte_perm_S (w[19], w[20], selector);
      w[38] = __byte_perm_S (w[18], w[19], selector);
      w[37] = __byte_perm_S (w[17], w[18], selector);
      w[36] = __byte_perm_S (w[16], w[17], selector);
      w[35] = __byte_perm_S (w[15], w[16], selector);
      w[34] = __byte_perm_S (w[14], w[15], selector);
      w[33] = __byte_perm_S (w[13], w[14], selector);
      w[32] = __byte_perm_S (w[12], w[13], selector);
      w[31] = __byte_perm_S (w[11], w[12], selector);
      w[30] = __byte_perm_S (w[10], w[11], selector);
      w[29] = __byte_perm_S (w[ 9], w[10], selector);
      w[28] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[27] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[26] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[25] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[24] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[23] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[22] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[21] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[20] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[19] = __byte_perm_S (    0, w[ 0], selector);
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 20:
      w[63] = __byte_perm_S (w[42], w[43], selector);
      w[62] = __byte_perm_S (w[41], w[42], selector);
      w[61] = __byte_perm_S (w[40], w[41], selector);
      w[60] = __byte_perm_S (w[39], w[40], selector);
      w[59] = __byte_perm_S (w[38], w[39], selector);
      w[58] = __byte_perm_S (w[37], w[38], selector);
      w[57] = __byte_perm_S (w[36], w[37], selector);
      w[56] = __byte_perm_S (w[35], w[36], selector);
      w[55] = __byte_perm_S (w[34], w[35], selector);
      w[54] = __byte_perm_S (w[33], w[34], selector);
      w[53] = __byte_perm_S (w[32], w[33], selector);
      w[52] = __byte_perm_S (w[31], w[32], selector);
      w[51] = __byte_perm_S (w[30], w[31], selector);
      w[50] = __byte_perm_S (w[29], w[30], selector);
      w[49] = __byte_perm_S (w[28], w[29], selector);
      w[48] = __byte_perm_S (w[27], w[28], selector);
      w[47] = __byte_perm_S (w[26], w[27], selector);
      w[46] = __byte_perm_S (w[25], w[26], selector);
      w[45] = __byte_perm_S (w[24], w[25], selector);
      w[44] = __byte_perm_S (w[23], w[24], selector);
      w[43] = __byte_perm_S (w[22], w[23], selector);
      w[42] = __byte_perm_S (w[21], w[22], selector);
      w[41] = __byte_perm_S (w[20], w[21], selector);
      w[40] = __byte_perm_S (w[19], w[20], selector);
      w[39] = __byte_perm_S (w[18], w[19], selector);
      w[38] = __byte_perm_S (w[17], w[18], selector);
      w[37] = __byte_perm_S (w[16], w[17], selector);
      w[36] = __byte_perm_S (w[15], w[16], selector);
      w[35] = __byte_perm_S (w[14], w[15], selector);
      w[34] = __byte_perm_S (w[13], w[14], selector);
      w[33] = __byte_perm_S (w[12], w[13], selector);
      w[32] = __byte_perm_S (w[11], w[12], selector);
      w[31] = __byte_perm_S (w[10], w[11], selector);
      w[30] = __byte_perm_S (w[ 9], w[10], selector);
      w[29] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[28] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[27] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[26] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[25] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[24] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[23] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[22] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[21] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[20] = __byte_perm_S (    0, w[ 0], selector);
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 21:
      w[63] = __byte_perm_S (w[41], w[42], selector);
      w[62] = __byte_perm_S (w[40], w[41], selector);
      w[61] = __byte_perm_S (w[39], w[40], selector);
      w[60] = __byte_perm_S (w[38], w[39], selector);
      w[59] = __byte_perm_S (w[37], w[38], selector);
      w[58] = __byte_perm_S (w[36], w[37], selector);
      w[57] = __byte_perm_S (w[35], w[36], selector);
      w[56] = __byte_perm_S (w[34], w[35], selector);
      w[55] = __byte_perm_S (w[33], w[34], selector);
      w[54] = __byte_perm_S (w[32], w[33], selector);
      w[53] = __byte_perm_S (w[31], w[32], selector);
      w[52] = __byte_perm_S (w[30], w[31], selector);
      w[51] = __byte_perm_S (w[29], w[30], selector);
      w[50] = __byte_perm_S (w[28], w[29], selector);
      w[49] = __byte_perm_S (w[27], w[28], selector);
      w[48] = __byte_perm_S (w[26], w[27], selector);
      w[47] = __byte_perm_S (w[25], w[26], selector);
      w[46] = __byte_perm_S (w[24], w[25], selector);
      w[45] = __byte_perm_S (w[23], w[24], selector);
      w[44] = __byte_perm_S (w[22], w[23], selector);
      w[43] = __byte_perm_S (w[21], w[22], selector);
      w[42] = __byte_perm_S (w[20], w[21], selector);
      w[41] = __byte_perm_S (w[19], w[20], selector);
      w[40] = __byte_perm_S (w[18], w[19], selector);
      w[39] = __byte_perm_S (w[17], w[18], selector);
      w[38] = __byte_perm_S (w[16], w[17], selector);
      w[37] = __byte_perm_S (w[15], w[16], selector);
      w[36] = __byte_perm_S (w[14], w[15], selector);
      w[35] = __byte_perm_S (w[13], w[14], selector);
      w[34] = __byte_perm_S (w[12], w[13], selector);
      w[33] = __byte_perm_S (w[11], w[12], selector);
      w[32] = __byte_perm_S (w[10], w[11], selector);
      w[31] = __byte_perm_S (w[ 9], w[10], selector);
      w[30] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[29] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[28] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[27] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[26] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[25] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[24] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[23] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[22] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[21] = __byte_perm_S (    0, w[ 0], selector);
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 22:
      w[63] = __byte_perm_S (w[40], w[41], selector);
      w[62] = __byte_perm_S (w[39], w[40], selector);
      w[61] = __byte_perm_S (w[38], w[39], selector);
      w[60] = __byte_perm_S (w[37], w[38], selector);
      w[59] = __byte_perm_S (w[36], w[37], selector);
      w[58] = __byte_perm_S (w[35], w[36], selector);
      w[57] = __byte_perm_S (w[34], w[35], selector);
      w[56] = __byte_perm_S (w[33], w[34], selector);
      w[55] = __byte_perm_S (w[32], w[33], selector);
      w[54] = __byte_perm_S (w[31], w[32], selector);
      w[53] = __byte_perm_S (w[30], w[31], selector);
      w[52] = __byte_perm_S (w[29], w[30], selector);
      w[51] = __byte_perm_S (w[28], w[29], selector);
      w[50] = __byte_perm_S (w[27], w[28], selector);
      w[49] = __byte_perm_S (w[26], w[27], selector);
      w[48] = __byte_perm_S (w[25], w[26], selector);
      w[47] = __byte_perm_S (w[24], w[25], selector);
      w[46] = __byte_perm_S (w[23], w[24], selector);
      w[45] = __byte_perm_S (w[22], w[23], selector);
      w[44] = __byte_perm_S (w[21], w[22], selector);
      w[43] = __byte_perm_S (w[20], w[21], selector);
      w[42] = __byte_perm_S (w[19], w[20], selector);
      w[41] = __byte_perm_S (w[18], w[19], selector);
      w[40] = __byte_perm_S (w[17], w[18], selector);
      w[39] = __byte_perm_S (w[16], w[17], selector);
      w[38] = __byte_perm_S (w[15], w[16], selector);
      w[37] = __byte_perm_S (w[14], w[15], selector);
      w[36] = __byte_perm_S (w[13], w[14], selector);
      w[35] = __byte_perm_S (w[12], w[13], selector);
      w[34] = __byte_perm_S (w[11], w[12], selector);
      w[33] = __byte_perm_S (w[10], w[11], selector);
      w[32] = __byte_perm_S (w[ 9], w[10], selector);
      w[31] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[30] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[29] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[28] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[27] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[26] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[25] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[24] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[23] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[22] = __byte_perm_S (    0, w[ 0], selector);
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 23:
      w[63] = __byte_perm_S (w[39], w[40], selector);
      w[62] = __byte_perm_S (w[38], w[39], selector);
      w[61] = __byte_perm_S (w[37], w[38], selector);
      w[60] = __byte_perm_S (w[36], w[37], selector);
      w[59] = __byte_perm_S (w[35], w[36], selector);
      w[58] = __byte_perm_S (w[34], w[35], selector);
      w[57] = __byte_perm_S (w[33], w[34], selector);
      w[56] = __byte_perm_S (w[32], w[33], selector);
      w[55] = __byte_perm_S (w[31], w[32], selector);
      w[54] = __byte_perm_S (w[30], w[31], selector);
      w[53] = __byte_perm_S (w[29], w[30], selector);
      w[52] = __byte_perm_S (w[28], w[29], selector);
      w[51] = __byte_perm_S (w[27], w[28], selector);
      w[50] = __byte_perm_S (w[26], w[27], selector);
      w[49] = __byte_perm_S (w[25], w[26], selector);
      w[48] = __byte_perm_S (w[24], w[25], selector);
      w[47] = __byte_perm_S (w[23], w[24], selector);
      w[46] = __byte_perm_S (w[22], w[23], selector);
      w[45] = __byte_perm_S (w[21], w[22], selector);
      w[44] = __byte_perm_S (w[20], w[21], selector);
      w[43] = __byte_perm_S (w[19], w[20], selector);
      w[42] = __byte_perm_S (w[18], w[19], selector);
      w[41] = __byte_perm_S (w[17], w[18], selector);
      w[40] = __byte_perm_S (w[16], w[17], selector);
      w[39] = __byte_perm_S (w[15], w[16], selector);
      w[38] = __byte_perm_S (w[14], w[15], selector);
      w[37] = __byte_perm_S (w[13], w[14], selector);
      w[36] = __byte_perm_S (w[12], w[13], selector);
      w[35] = __byte_perm_S (w[11], w[12], selector);
      w[34] = __byte_perm_S (w[10], w[11], selector);
      w[33] = __byte_perm_S (w[ 9], w[10], selector);
      w[32] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[31] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[30] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[29] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[28] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[27] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[26] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[25] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[24] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[23] = __byte_perm_S (    0, w[ 0], selector);
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 24:
      w[63] = __byte_perm_S (w[38], w[39], selector);
      w[62] = __byte_perm_S (w[37], w[38], selector);
      w[61] = __byte_perm_S (w[36], w[37], selector);
      w[60] = __byte_perm_S (w[35], w[36], selector);
      w[59] = __byte_perm_S (w[34], w[35], selector);
      w[58] = __byte_perm_S (w[33], w[34], selector);
      w[57] = __byte_perm_S (w[32], w[33], selector);
      w[56] = __byte_perm_S (w[31], w[32], selector);
      w[55] = __byte_perm_S (w[30], w[31], selector);
      w[54] = __byte_perm_S (w[29], w[30], selector);
      w[53] = __byte_perm_S (w[28], w[29], selector);
      w[52] = __byte_perm_S (w[27], w[28], selector);
      w[51] = __byte_perm_S (w[26], w[27], selector);
      w[50] = __byte_perm_S (w[25], w[26], selector);
      w[49] = __byte_perm_S (w[24], w[25], selector);
      w[48] = __byte_perm_S (w[23], w[24], selector);
      w[47] = __byte_perm_S (w[22], w[23], selector);
      w[46] = __byte_perm_S (w[21], w[22], selector);
      w[45] = __byte_perm_S (w[20], w[21], selector);
      w[44] = __byte_perm_S (w[19], w[20], selector);
      w[43] = __byte_perm_S (w[18], w[19], selector);
      w[42] = __byte_perm_S (w[17], w[18], selector);
      w[41] = __byte_perm_S (w[16], w[17], selector);
      w[40] = __byte_perm_S (w[15], w[16], selector);
      w[39] = __byte_perm_S (w[14], w[15], selector);
      w[38] = __byte_perm_S (w[13], w[14], selector);
      w[37] = __byte_perm_S (w[12], w[13], selector);
      w[36] = __byte_perm_S (w[11], w[12], selector);
      w[35] = __byte_perm_S (w[10], w[11], selector);
      w[34] = __byte_perm_S (w[ 9], w[10], selector);
      w[33] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[32] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[31] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[30] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[29] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[28] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[27] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[26] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[25] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[24] = __byte_perm_S (    0, w[ 0], selector);
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 25:
      w[63] = __byte_perm_S (w[37], w[38], selector);
      w[62] = __byte_perm_S (w[36], w[37], selector);
      w[61] = __byte_perm_S (w[35], w[36], selector);
      w[60] = __byte_perm_S (w[34], w[35], selector);
      w[59] = __byte_perm_S (w[33], w[34], selector);
      w[58] = __byte_perm_S (w[32], w[33], selector);
      w[57] = __byte_perm_S (w[31], w[32], selector);
      w[56] = __byte_perm_S (w[30], w[31], selector);
      w[55] = __byte_perm_S (w[29], w[30], selector);
      w[54] = __byte_perm_S (w[28], w[29], selector);
      w[53] = __byte_perm_S (w[27], w[28], selector);
      w[52] = __byte_perm_S (w[26], w[27], selector);
      w[51] = __byte_perm_S (w[25], w[26], selector);
      w[50] = __byte_perm_S (w[24], w[25], selector);
      w[49] = __byte_perm_S (w[23], w[24], selector);
      w[48] = __byte_perm_S (w[22], w[23], selector);
      w[47] = __byte_perm_S (w[21], w[22], selector);
      w[46] = __byte_perm_S (w[20], w[21], selector);
      w[45] = __byte_perm_S (w[19], w[20], selector);
      w[44] = __byte_perm_S (w[18], w[19], selector);
      w[43] = __byte_perm_S (w[17], w[18], selector);
      w[42] = __byte_perm_S (w[16], w[17], selector);
      w[41] = __byte_perm_S (w[15], w[16], selector);
      w[40] = __byte_perm_S (w[14], w[15], selector);
      w[39] = __byte_perm_S (w[13], w[14], selector);
      w[38] = __byte_perm_S (w[12], w[13], selector);
      w[37] = __byte_perm_S (w[11], w[12], selector);
      w[36] = __byte_perm_S (w[10], w[11], selector);
      w[35] = __byte_perm_S (w[ 9], w[10], selector);
      w[34] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[33] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[32] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[31] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[30] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[29] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[28] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[27] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[26] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[25] = __byte_perm_S (    0, w[ 0], selector);
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 26:
      w[63] = __byte_perm_S (w[36], w[37], selector);
      w[62] = __byte_perm_S (w[35], w[36], selector);
      w[61] = __byte_perm_S (w[34], w[35], selector);
      w[60] = __byte_perm_S (w[33], w[34], selector);
      w[59] = __byte_perm_S (w[32], w[33], selector);
      w[58] = __byte_perm_S (w[31], w[32], selector);
      w[57] = __byte_perm_S (w[30], w[31], selector);
      w[56] = __byte_perm_S (w[29], w[30], selector);
      w[55] = __byte_perm_S (w[28], w[29], selector);
      w[54] = __byte_perm_S (w[27], w[28], selector);
      w[53] = __byte_perm_S (w[26], w[27], selector);
      w[52] = __byte_perm_S (w[25], w[26], selector);
      w[51] = __byte_perm_S (w[24], w[25], selector);
      w[50] = __byte_perm_S (w[23], w[24], selector);
      w[49] = __byte_perm_S (w[22], w[23], selector);
      w[48] = __byte_perm_S (w[21], w[22], selector);
      w[47] = __byte_perm_S (w[20], w[21], selector);
      w[46] = __byte_perm_S (w[19], w[20], selector);
      w[45] = __byte_perm_S (w[18], w[19], selector);
      w[44] = __byte_perm_S (w[17], w[18], selector);
      w[43] = __byte_perm_S (w[16], w[17], selector);
      w[42] = __byte_perm_S (w[15], w[16], selector);
      w[41] = __byte_perm_S (w[14], w[15], selector);
      w[40] = __byte_perm_S (w[13], w[14], selector);
      w[39] = __byte_perm_S (w[12], w[13], selector);
      w[38] = __byte_perm_S (w[11], w[12], selector);
      w[37] = __byte_perm_S (w[10], w[11], selector);
      w[36] = __byte_perm_S (w[ 9], w[10], selector);
      w[35] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[34] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[33] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[32] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[31] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[30] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[29] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[28] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[27] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[26] = __byte_perm_S (    0, w[ 0], selector);
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 27:
      w[63] = __byte_perm_S (w[35], w[36], selector);
      w[62] = __byte_perm_S (w[34], w[35], selector);
      w[61] = __byte_perm_S (w[33], w[34], selector);
      w[60] = __byte_perm_S (w[32], w[33], selector);
      w[59] = __byte_perm_S (w[31], w[32], selector);
      w[58] = __byte_perm_S (w[30], w[31], selector);
      w[57] = __byte_perm_S (w[29], w[30], selector);
      w[56] = __byte_perm_S (w[28], w[29], selector);
      w[55] = __byte_perm_S (w[27], w[28], selector);
      w[54] = __byte_perm_S (w[26], w[27], selector);
      w[53] = __byte_perm_S (w[25], w[26], selector);
      w[52] = __byte_perm_S (w[24], w[25], selector);
      w[51] = __byte_perm_S (w[23], w[24], selector);
      w[50] = __byte_perm_S (w[22], w[23], selector);
      w[49] = __byte_perm_S (w[21], w[22], selector);
      w[48] = __byte_perm_S (w[20], w[21], selector);
      w[47] = __byte_perm_S (w[19], w[20], selector);
      w[46] = __byte_perm_S (w[18], w[19], selector);
      w[45] = __byte_perm_S (w[17], w[18], selector);
      w[44] = __byte_perm_S (w[16], w[17], selector);
      w[43] = __byte_perm_S (w[15], w[16], selector);
      w[42] = __byte_perm_S (w[14], w[15], selector);
      w[41] = __byte_perm_S (w[13], w[14], selector);
      w[40] = __byte_perm_S (w[12], w[13], selector);
      w[39] = __byte_perm_S (w[11], w[12], selector);
      w[38] = __byte_perm_S (w[10], w[11], selector);
      w[37] = __byte_perm_S (w[ 9], w[10], selector);
      w[36] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[35] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[34] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[33] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[32] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[31] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[30] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[29] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[28] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[27] = __byte_perm_S (    0, w[ 0], selector);
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 28:
      w[63] = __byte_perm_S (w[34], w[35], selector);
      w[62] = __byte_perm_S (w[33], w[34], selector);
      w[61] = __byte_perm_S (w[32], w[33], selector);
      w[60] = __byte_perm_S (w[31], w[32], selector);
      w[59] = __byte_perm_S (w[30], w[31], selector);
      w[58] = __byte_perm_S (w[29], w[30], selector);
      w[57] = __byte_perm_S (w[28], w[29], selector);
      w[56] = __byte_perm_S (w[27], w[28], selector);
      w[55] = __byte_perm_S (w[26], w[27], selector);
      w[54] = __byte_perm_S (w[25], w[26], selector);
      w[53] = __byte_perm_S (w[24], w[25], selector);
      w[52] = __byte_perm_S (w[23], w[24], selector);
      w[51] = __byte_perm_S (w[22], w[23], selector);
      w[50] = __byte_perm_S (w[21], w[22], selector);
      w[49] = __byte_perm_S (w[20], w[21], selector);
      w[48] = __byte_perm_S (w[19], w[20], selector);
      w[47] = __byte_perm_S (w[18], w[19], selector);
      w[46] = __byte_perm_S (w[17], w[18], selector);
      w[45] = __byte_perm_S (w[16], w[17], selector);
      w[44] = __byte_perm_S (w[15], w[16], selector);
      w[43] = __byte_perm_S (w[14], w[15], selector);
      w[42] = __byte_perm_S (w[13], w[14], selector);
      w[41] = __byte_perm_S (w[12], w[13], selector);
      w[40] = __byte_perm_S (w[11], w[12], selector);
      w[39] = __byte_perm_S (w[10], w[11], selector);
      w[38] = __byte_perm_S (w[ 9], w[10], selector);
      w[37] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[36] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[35] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[34] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[33] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[32] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[31] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[30] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[29] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[28] = __byte_perm_S (    0, w[ 0], selector);
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 29:
      w[63] = __byte_perm_S (w[33], w[34], selector);
      w[62] = __byte_perm_S (w[32], w[33], selector);
      w[61] = __byte_perm_S (w[31], w[32], selector);
      w[60] = __byte_perm_S (w[30], w[31], selector);
      w[59] = __byte_perm_S (w[29], w[30], selector);
      w[58] = __byte_perm_S (w[28], w[29], selector);
      w[57] = __byte_perm_S (w[27], w[28], selector);
      w[56] = __byte_perm_S (w[26], w[27], selector);
      w[55] = __byte_perm_S (w[25], w[26], selector);
      w[54] = __byte_perm_S (w[24], w[25], selector);
      w[53] = __byte_perm_S (w[23], w[24], selector);
      w[52] = __byte_perm_S (w[22], w[23], selector);
      w[51] = __byte_perm_S (w[21], w[22], selector);
      w[50] = __byte_perm_S (w[20], w[21], selector);
      w[49] = __byte_perm_S (w[19], w[20], selector);
      w[48] = __byte_perm_S (w[18], w[19], selector);
      w[47] = __byte_perm_S (w[17], w[18], selector);
      w[46] = __byte_perm_S (w[16], w[17], selector);
      w[45] = __byte_perm_S (w[15], w[16], selector);
      w[44] = __byte_perm_S (w[14], w[15], selector);
      w[43] = __byte_perm_S (w[13], w[14], selector);
      w[42] = __byte_perm_S (w[12], w[13], selector);
      w[41] = __byte_perm_S (w[11], w[12], selector);
      w[40] = __byte_perm_S (w[10], w[11], selector);
      w[39] = __byte_perm_S (w[ 9], w[10], selector);
      w[38] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[37] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[36] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[35] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[34] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[33] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[32] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[31] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[30] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[29] = __byte_perm_S (    0, w[ 0], selector);
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 30:
      w[63] = __byte_perm_S (w[32], w[33], selector);
      w[62] = __byte_perm_S (w[31], w[32], selector);
      w[61] = __byte_perm_S (w[30], w[31], selector);
      w[60] = __byte_perm_S (w[29], w[30], selector);
      w[59] = __byte_perm_S (w[28], w[29], selector);
      w[58] = __byte_perm_S (w[27], w[28], selector);
      w[57] = __byte_perm_S (w[26], w[27], selector);
      w[56] = __byte_perm_S (w[25], w[26], selector);
      w[55] = __byte_perm_S (w[24], w[25], selector);
      w[54] = __byte_perm_S (w[23], w[24], selector);
      w[53] = __byte_perm_S (w[22], w[23], selector);
      w[52] = __byte_perm_S (w[21], w[22], selector);
      w[51] = __byte_perm_S (w[20], w[21], selector);
      w[50] = __byte_perm_S (w[19], w[20], selector);
      w[49] = __byte_perm_S (w[18], w[19], selector);
      w[48] = __byte_perm_S (w[17], w[18], selector);
      w[47] = __byte_perm_S (w[16], w[17], selector);
      w[46] = __byte_perm_S (w[15], w[16], selector);
      w[45] = __byte_perm_S (w[14], w[15], selector);
      w[44] = __byte_perm_S (w[13], w[14], selector);
      w[43] = __byte_perm_S (w[12], w[13], selector);
      w[42] = __byte_perm_S (w[11], w[12], selector);
      w[41] = __byte_perm_S (w[10], w[11], selector);
      w[40] = __byte_perm_S (w[ 9], w[10], selector);
      w[39] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[38] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[37] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[36] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[35] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[34] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[33] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[32] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[31] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[30] = __byte_perm_S (    0, w[ 0], selector);
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 31:
      w[63] = __byte_perm_S (w[31], w[32], selector);
      w[62] = __byte_perm_S (w[30], w[31], selector);
      w[61] = __byte_perm_S (w[29], w[30], selector);
      w[60] = __byte_perm_S (w[28], w[29], selector);
      w[59] = __byte_perm_S (w[27], w[28], selector);
      w[58] = __byte_perm_S (w[26], w[27], selector);
      w[57] = __byte_perm_S (w[25], w[26], selector);
      w[56] = __byte_perm_S (w[24], w[25], selector);
      w[55] = __byte_perm_S (w[23], w[24], selector);
      w[54] = __byte_perm_S (w[22], w[23], selector);
      w[53] = __byte_perm_S (w[21], w[22], selector);
      w[52] = __byte_perm_S (w[20], w[21], selector);
      w[51] = __byte_perm_S (w[19], w[20], selector);
      w[50] = __byte_perm_S (w[18], w[19], selector);
      w[49] = __byte_perm_S (w[17], w[18], selector);
      w[48] = __byte_perm_S (w[16], w[17], selector);
      w[47] = __byte_perm_S (w[15], w[16], selector);
      w[46] = __byte_perm_S (w[14], w[15], selector);
      w[45] = __byte_perm_S (w[13], w[14], selector);
      w[44] = __byte_perm_S (w[12], w[13], selector);
      w[43] = __byte_perm_S (w[11], w[12], selector);
      w[42] = __byte_perm_S (w[10], w[11], selector);
      w[41] = __byte_perm_S (w[ 9], w[10], selector);
      w[40] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[39] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[38] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[37] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[36] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[35] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[34] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[33] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[32] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[31] = __byte_perm_S (    0, w[ 0], selector);
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 32:
      w[63] = __byte_perm_S (w[30], w[31], selector);
      w[62] = __byte_perm_S (w[29], w[30], selector);
      w[61] = __byte_perm_S (w[28], w[29], selector);
      w[60] = __byte_perm_S (w[27], w[28], selector);
      w[59] = __byte_perm_S (w[26], w[27], selector);
      w[58] = __byte_perm_S (w[25], w[26], selector);
      w[57] = __byte_perm_S (w[24], w[25], selector);
      w[56] = __byte_perm_S (w[23], w[24], selector);
      w[55] = __byte_perm_S (w[22], w[23], selector);
      w[54] = __byte_perm_S (w[21], w[22], selector);
      w[53] = __byte_perm_S (w[20], w[21], selector);
      w[52] = __byte_perm_S (w[19], w[20], selector);
      w[51] = __byte_perm_S (w[18], w[19], selector);
      w[50] = __byte_perm_S (w[17], w[18], selector);
      w[49] = __byte_perm_S (w[16], w[17], selector);
      w[48] = __byte_perm_S (w[15], w[16], selector);
      w[47] = __byte_perm_S (w[14], w[15], selector);
      w[46] = __byte_perm_S (w[13], w[14], selector);
      w[45] = __byte_perm_S (w[12], w[13], selector);
      w[44] = __byte_perm_S (w[11], w[12], selector);
      w[43] = __byte_perm_S (w[10], w[11], selector);
      w[42] = __byte_perm_S (w[ 9], w[10], selector);
      w[41] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[40] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[39] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[38] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[37] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[36] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[35] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[34] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[33] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[32] = __byte_perm_S (    0, w[ 0], selector);
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 33:
      w[63] = __byte_perm_S (w[29], w[30], selector);
      w[62] = __byte_perm_S (w[28], w[29], selector);
      w[61] = __byte_perm_S (w[27], w[28], selector);
      w[60] = __byte_perm_S (w[26], w[27], selector);
      w[59] = __byte_perm_S (w[25], w[26], selector);
      w[58] = __byte_perm_S (w[24], w[25], selector);
      w[57] = __byte_perm_S (w[23], w[24], selector);
      w[56] = __byte_perm_S (w[22], w[23], selector);
      w[55] = __byte_perm_S (w[21], w[22], selector);
      w[54] = __byte_perm_S (w[20], w[21], selector);
      w[53] = __byte_perm_S (w[19], w[20], selector);
      w[52] = __byte_perm_S (w[18], w[19], selector);
      w[51] = __byte_perm_S (w[17], w[18], selector);
      w[50] = __byte_perm_S (w[16], w[17], selector);
      w[49] = __byte_perm_S (w[15], w[16], selector);
      w[48] = __byte_perm_S (w[14], w[15], selector);
      w[47] = __byte_perm_S (w[13], w[14], selector);
      w[46] = __byte_perm_S (w[12], w[13], selector);
      w[45] = __byte_perm_S (w[11], w[12], selector);
      w[44] = __byte_perm_S (w[10], w[11], selector);
      w[43] = __byte_perm_S (w[ 9], w[10], selector);
      w[42] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[41] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[40] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[39] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[38] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[37] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[36] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[35] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[34] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[33] = __byte_perm_S (    0, w[ 0], selector);
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 34:
      w[63] = __byte_perm_S (w[28], w[29], selector);
      w[62] = __byte_perm_S (w[27], w[28], selector);
      w[61] = __byte_perm_S (w[26], w[27], selector);
      w[60] = __byte_perm_S (w[25], w[26], selector);
      w[59] = __byte_perm_S (w[24], w[25], selector);
      w[58] = __byte_perm_S (w[23], w[24], selector);
      w[57] = __byte_perm_S (w[22], w[23], selector);
      w[56] = __byte_perm_S (w[21], w[22], selector);
      w[55] = __byte_perm_S (w[20], w[21], selector);
      w[54] = __byte_perm_S (w[19], w[20], selector);
      w[53] = __byte_perm_S (w[18], w[19], selector);
      w[52] = __byte_perm_S (w[17], w[18], selector);
      w[51] = __byte_perm_S (w[16], w[17], selector);
      w[50] = __byte_perm_S (w[15], w[16], selector);
      w[49] = __byte_perm_S (w[14], w[15], selector);
      w[48] = __byte_perm_S (w[13], w[14], selector);
      w[47] = __byte_perm_S (w[12], w[13], selector);
      w[46] = __byte_perm_S (w[11], w[12], selector);
      w[45] = __byte_perm_S (w[10], w[11], selector);
      w[44] = __byte_perm_S (w[ 9], w[10], selector);
      w[43] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[42] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[41] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[40] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[39] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[38] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[37] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[36] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[35] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[34] = __byte_perm_S (    0, w[ 0], selector);
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 35:
      w[63] = __byte_perm_S (w[27], w[28], selector);
      w[62] = __byte_perm_S (w[26], w[27], selector);
      w[61] = __byte_perm_S (w[25], w[26], selector);
      w[60] = __byte_perm_S (w[24], w[25], selector);
      w[59] = __byte_perm_S (w[23], w[24], selector);
      w[58] = __byte_perm_S (w[22], w[23], selector);
      w[57] = __byte_perm_S (w[21], w[22], selector);
      w[56] = __byte_perm_S (w[20], w[21], selector);
      w[55] = __byte_perm_S (w[19], w[20], selector);
      w[54] = __byte_perm_S (w[18], w[19], selector);
      w[53] = __byte_perm_S (w[17], w[18], selector);
      w[52] = __byte_perm_S (w[16], w[17], selector);
      w[51] = __byte_perm_S (w[15], w[16], selector);
      w[50] = __byte_perm_S (w[14], w[15], selector);
      w[49] = __byte_perm_S (w[13], w[14], selector);
      w[48] = __byte_perm_S (w[12], w[13], selector);
      w[47] = __byte_perm_S (w[11], w[12], selector);
      w[46] = __byte_perm_S (w[10], w[11], selector);
      w[45] = __byte_perm_S (w[ 9], w[10], selector);
      w[44] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[43] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[42] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[41] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[40] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[39] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[38] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[37] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[36] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[35] = __byte_perm_S (    0, w[ 0], selector);
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 36:
      w[63] = __byte_perm_S (w[26], w[27], selector);
      w[62] = __byte_perm_S (w[25], w[26], selector);
      w[61] = __byte_perm_S (w[24], w[25], selector);
      w[60] = __byte_perm_S (w[23], w[24], selector);
      w[59] = __byte_perm_S (w[22], w[23], selector);
      w[58] = __byte_perm_S (w[21], w[22], selector);
      w[57] = __byte_perm_S (w[20], w[21], selector);
      w[56] = __byte_perm_S (w[19], w[20], selector);
      w[55] = __byte_perm_S (w[18], w[19], selector);
      w[54] = __byte_perm_S (w[17], w[18], selector);
      w[53] = __byte_perm_S (w[16], w[17], selector);
      w[52] = __byte_perm_S (w[15], w[16], selector);
      w[51] = __byte_perm_S (w[14], w[15], selector);
      w[50] = __byte_perm_S (w[13], w[14], selector);
      w[49] = __byte_perm_S (w[12], w[13], selector);
      w[48] = __byte_perm_S (w[11], w[12], selector);
      w[47] = __byte_perm_S (w[10], w[11], selector);
      w[46] = __byte_perm_S (w[ 9], w[10], selector);
      w[45] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[44] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[43] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[42] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[41] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[40] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[39] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[38] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[37] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[36] = __byte_perm_S (    0, w[ 0], selector);
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 37:
      w[63] = __byte_perm_S (w[25], w[26], selector);
      w[62] = __byte_perm_S (w[24], w[25], selector);
      w[61] = __byte_perm_S (w[23], w[24], selector);
      w[60] = __byte_perm_S (w[22], w[23], selector);
      w[59] = __byte_perm_S (w[21], w[22], selector);
      w[58] = __byte_perm_S (w[20], w[21], selector);
      w[57] = __byte_perm_S (w[19], w[20], selector);
      w[56] = __byte_perm_S (w[18], w[19], selector);
      w[55] = __byte_perm_S (w[17], w[18], selector);
      w[54] = __byte_perm_S (w[16], w[17], selector);
      w[53] = __byte_perm_S (w[15], w[16], selector);
      w[52] = __byte_perm_S (w[14], w[15], selector);
      w[51] = __byte_perm_S (w[13], w[14], selector);
      w[50] = __byte_perm_S (w[12], w[13], selector);
      w[49] = __byte_perm_S (w[11], w[12], selector);
      w[48] = __byte_perm_S (w[10], w[11], selector);
      w[47] = __byte_perm_S (w[ 9], w[10], selector);
      w[46] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[45] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[44] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[43] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[42] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[41] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[40] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[39] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[38] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[37] = __byte_perm_S (    0, w[ 0], selector);
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 38:
      w[63] = __byte_perm_S (w[24], w[25], selector);
      w[62] = __byte_perm_S (w[23], w[24], selector);
      w[61] = __byte_perm_S (w[22], w[23], selector);
      w[60] = __byte_perm_S (w[21], w[22], selector);
      w[59] = __byte_perm_S (w[20], w[21], selector);
      w[58] = __byte_perm_S (w[19], w[20], selector);
      w[57] = __byte_perm_S (w[18], w[19], selector);
      w[56] = __byte_perm_S (w[17], w[18], selector);
      w[55] = __byte_perm_S (w[16], w[17], selector);
      w[54] = __byte_perm_S (w[15], w[16], selector);
      w[53] = __byte_perm_S (w[14], w[15], selector);
      w[52] = __byte_perm_S (w[13], w[14], selector);
      w[51] = __byte_perm_S (w[12], w[13], selector);
      w[50] = __byte_perm_S (w[11], w[12], selector);
      w[49] = __byte_perm_S (w[10], w[11], selector);
      w[48] = __byte_perm_S (w[ 9], w[10], selector);
      w[47] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[46] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[45] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[44] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[43] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[42] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[41] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[40] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[39] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[38] = __byte_perm_S (    0, w[ 0], selector);
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 39:
      w[63] = __byte_perm_S (w[23], w[24], selector);
      w[62] = __byte_perm_S (w[22], w[23], selector);
      w[61] = __byte_perm_S (w[21], w[22], selector);
      w[60] = __byte_perm_S (w[20], w[21], selector);
      w[59] = __byte_perm_S (w[19], w[20], selector);
      w[58] = __byte_perm_S (w[18], w[19], selector);
      w[57] = __byte_perm_S (w[17], w[18], selector);
      w[56] = __byte_perm_S (w[16], w[17], selector);
      w[55] = __byte_perm_S (w[15], w[16], selector);
      w[54] = __byte_perm_S (w[14], w[15], selector);
      w[53] = __byte_perm_S (w[13], w[14], selector);
      w[52] = __byte_perm_S (w[12], w[13], selector);
      w[51] = __byte_perm_S (w[11], w[12], selector);
      w[50] = __byte_perm_S (w[10], w[11], selector);
      w[49] = __byte_perm_S (w[ 9], w[10], selector);
      w[48] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[47] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[46] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[45] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[44] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[43] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[42] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[41] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[40] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[39] = __byte_perm_S (    0, w[ 0], selector);
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 40:
      w[63] = __byte_perm_S (w[22], w[23], selector);
      w[62] = __byte_perm_S (w[21], w[22], selector);
      w[61] = __byte_perm_S (w[20], w[21], selector);
      w[60] = __byte_perm_S (w[19], w[20], selector);
      w[59] = __byte_perm_S (w[18], w[19], selector);
      w[58] = __byte_perm_S (w[17], w[18], selector);
      w[57] = __byte_perm_S (w[16], w[17], selector);
      w[56] = __byte_perm_S (w[15], w[16], selector);
      w[55] = __byte_perm_S (w[14], w[15], selector);
      w[54] = __byte_perm_S (w[13], w[14], selector);
      w[53] = __byte_perm_S (w[12], w[13], selector);
      w[52] = __byte_perm_S (w[11], w[12], selector);
      w[51] = __byte_perm_S (w[10], w[11], selector);
      w[50] = __byte_perm_S (w[ 9], w[10], selector);
      w[49] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[48] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[47] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[46] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[45] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[44] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[43] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[42] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[41] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[40] = __byte_perm_S (    0, w[ 0], selector);
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 41:
      w[63] = __byte_perm_S (w[21], w[22], selector);
      w[62] = __byte_perm_S (w[20], w[21], selector);
      w[61] = __byte_perm_S (w[19], w[20], selector);
      w[60] = __byte_perm_S (w[18], w[19], selector);
      w[59] = __byte_perm_S (w[17], w[18], selector);
      w[58] = __byte_perm_S (w[16], w[17], selector);
      w[57] = __byte_perm_S (w[15], w[16], selector);
      w[56] = __byte_perm_S (w[14], w[15], selector);
      w[55] = __byte_perm_S (w[13], w[14], selector);
      w[54] = __byte_perm_S (w[12], w[13], selector);
      w[53] = __byte_perm_S (w[11], w[12], selector);
      w[52] = __byte_perm_S (w[10], w[11], selector);
      w[51] = __byte_perm_S (w[ 9], w[10], selector);
      w[50] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[49] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[48] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[47] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[46] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[45] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[44] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[43] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[42] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[41] = __byte_perm_S (    0, w[ 0], selector);
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 42:
      w[63] = __byte_perm_S (w[20], w[21], selector);
      w[62] = __byte_perm_S (w[19], w[20], selector);
      w[61] = __byte_perm_S (w[18], w[19], selector);
      w[60] = __byte_perm_S (w[17], w[18], selector);
      w[59] = __byte_perm_S (w[16], w[17], selector);
      w[58] = __byte_perm_S (w[15], w[16], selector);
      w[57] = __byte_perm_S (w[14], w[15], selector);
      w[56] = __byte_perm_S (w[13], w[14], selector);
      w[55] = __byte_perm_S (w[12], w[13], selector);
      w[54] = __byte_perm_S (w[11], w[12], selector);
      w[53] = __byte_perm_S (w[10], w[11], selector);
      w[52] = __byte_perm_S (w[ 9], w[10], selector);
      w[51] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[50] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[49] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[48] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[47] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[46] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[45] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[44] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[43] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[42] = __byte_perm_S (    0, w[ 0], selector);
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 43:
      w[63] = __byte_perm_S (w[19], w[20], selector);
      w[62] = __byte_perm_S (w[18], w[19], selector);
      w[61] = __byte_perm_S (w[17], w[18], selector);
      w[60] = __byte_perm_S (w[16], w[17], selector);
      w[59] = __byte_perm_S (w[15], w[16], selector);
      w[58] = __byte_perm_S (w[14], w[15], selector);
      w[57] = __byte_perm_S (w[13], w[14], selector);
      w[56] = __byte_perm_S (w[12], w[13], selector);
      w[55] = __byte_perm_S (w[11], w[12], selector);
      w[54] = __byte_perm_S (w[10], w[11], selector);
      w[53] = __byte_perm_S (w[ 9], w[10], selector);
      w[52] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[51] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[50] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[49] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[48] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[47] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[46] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[45] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[44] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[43] = __byte_perm_S (    0, w[ 0], selector);
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 44:
      w[63] = __byte_perm_S (w[18], w[19], selector);
      w[62] = __byte_perm_S (w[17], w[18], selector);
      w[61] = __byte_perm_S (w[16], w[17], selector);
      w[60] = __byte_perm_S (w[15], w[16], selector);
      w[59] = __byte_perm_S (w[14], w[15], selector);
      w[58] = __byte_perm_S (w[13], w[14], selector);
      w[57] = __byte_perm_S (w[12], w[13], selector);
      w[56] = __byte_perm_S (w[11], w[12], selector);
      w[55] = __byte_perm_S (w[10], w[11], selector);
      w[54] = __byte_perm_S (w[ 9], w[10], selector);
      w[53] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[52] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[51] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[50] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[49] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[48] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[47] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[46] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[45] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[44] = __byte_perm_S (    0, w[ 0], selector);
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 45:
      w[63] = __byte_perm_S (w[17], w[18], selector);
      w[62] = __byte_perm_S (w[16], w[17], selector);
      w[61] = __byte_perm_S (w[15], w[16], selector);
      w[60] = __byte_perm_S (w[14], w[15], selector);
      w[59] = __byte_perm_S (w[13], w[14], selector);
      w[58] = __byte_perm_S (w[12], w[13], selector);
      w[57] = __byte_perm_S (w[11], w[12], selector);
      w[56] = __byte_perm_S (w[10], w[11], selector);
      w[55] = __byte_perm_S (w[ 9], w[10], selector);
      w[54] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[53] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[52] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[51] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[50] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[49] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[48] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[47] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[46] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[45] = __byte_perm_S (    0, w[ 0], selector);
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 46:
      w[63] = __byte_perm_S (w[16], w[17], selector);
      w[62] = __byte_perm_S (w[15], w[16], selector);
      w[61] = __byte_perm_S (w[14], w[15], selector);
      w[60] = __byte_perm_S (w[13], w[14], selector);
      w[59] = __byte_perm_S (w[12], w[13], selector);
      w[58] = __byte_perm_S (w[11], w[12], selector);
      w[57] = __byte_perm_S (w[10], w[11], selector);
      w[56] = __byte_perm_S (w[ 9], w[10], selector);
      w[55] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[54] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[53] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[52] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[51] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[50] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[49] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[48] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[47] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[46] = __byte_perm_S (    0, w[ 0], selector);
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 47:
      w[63] = __byte_perm_S (w[15], w[16], selector);
      w[62] = __byte_perm_S (w[14], w[15], selector);
      w[61] = __byte_perm_S (w[13], w[14], selector);
      w[60] = __byte_perm_S (w[12], w[13], selector);
      w[59] = __byte_perm_S (w[11], w[12], selector);
      w[58] = __byte_perm_S (w[10], w[11], selector);
      w[57] = __byte_perm_S (w[ 9], w[10], selector);
      w[56] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[55] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[54] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[53] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[52] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[51] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[50] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[49] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[48] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[47] = __byte_perm_S (    0, w[ 0], selector);
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 48:
      w[63] = __byte_perm_S (w[14], w[15], selector);
      w[62] = __byte_perm_S (w[13], w[14], selector);
      w[61] = __byte_perm_S (w[12], w[13], selector);
      w[60] = __byte_perm_S (w[11], w[12], selector);
      w[59] = __byte_perm_S (w[10], w[11], selector);
      w[58] = __byte_perm_S (w[ 9], w[10], selector);
      w[57] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[56] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[55] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[54] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[53] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[52] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[51] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[50] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[49] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[48] = __byte_perm_S (    0, w[ 0], selector);
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 49:
      w[63] = __byte_perm_S (w[13], w[14], selector);
      w[62] = __byte_perm_S (w[12], w[13], selector);
      w[61] = __byte_perm_S (w[11], w[12], selector);
      w[60] = __byte_perm_S (w[10], w[11], selector);
      w[59] = __byte_perm_S (w[ 9], w[10], selector);
      w[58] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[57] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[56] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[55] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[54] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[53] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[52] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[51] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[50] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[49] = __byte_perm_S (    0, w[ 0], selector);
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 50:
      w[63] = __byte_perm_S (w[12], w[13], selector);
      w[62] = __byte_perm_S (w[11], w[12], selector);
      w[61] = __byte_perm_S (w[10], w[11], selector);
      w[60] = __byte_perm_S (w[ 9], w[10], selector);
      w[59] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[58] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[57] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[56] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[55] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[54] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[53] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[52] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[51] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[50] = __byte_perm_S (    0, w[ 0], selector);
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 51:
      w[63] = __byte_perm_S (w[11], w[12], selector);
      w[62] = __byte_perm_S (w[10], w[11], selector);
      w[61] = __byte_perm_S (w[ 9], w[10], selector);
      w[60] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[59] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[58] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[57] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[56] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[55] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[54] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[53] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[52] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[51] = __byte_perm_S (    0, w[ 0], selector);
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 52:
      w[63] = __byte_perm_S (w[10], w[11], selector);
      w[62] = __byte_perm_S (w[ 9], w[10], selector);
      w[61] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[60] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[59] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[58] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[57] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[56] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[55] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[54] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[53] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[52] = __byte_perm_S (    0, w[ 0], selector);
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 53:
      w[63] = __byte_perm_S (w[ 9], w[10], selector);
      w[62] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[61] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[60] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[59] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[58] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[57] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[56] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[55] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[54] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[53] = __byte_perm_S (    0, w[ 0], selector);
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 54:
      w[63] = __byte_perm_S (w[ 8], w[ 9], selector);
      w[62] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[61] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[60] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[59] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[58] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[57] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[56] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[55] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[54] = __byte_perm_S (    0, w[ 0], selector);
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 55:
      w[63] = __byte_perm_S (w[ 7], w[ 8], selector);
      w[62] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[61] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[60] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[59] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[58] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[57] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[56] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[55] = __byte_perm_S (    0, w[ 0], selector);
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 56:
      w[63] = __byte_perm_S (w[ 6], w[ 7], selector);
      w[62] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[61] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[60] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[59] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[58] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[57] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[56] = __byte_perm_S (    0, w[ 0], selector);
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 57:
      w[63] = __byte_perm_S (w[ 5], w[ 6], selector);
      w[62] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[61] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[60] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[59] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[58] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[57] = __byte_perm_S (    0, w[ 0], selector);
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 58:
      w[63] = __byte_perm_S (w[ 4], w[ 5], selector);
      w[62] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[61] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[60] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[59] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[58] = __byte_perm_S (    0, w[ 0], selector);
      w[57] = 0;
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 59:
      w[63] = __byte_perm_S (w[ 3], w[ 4], selector);
      w[62] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[61] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[60] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[59] = __byte_perm_S (    0, w[ 0], selector);
      w[58] = 0;
      w[57] = 0;
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 60:
      w[63] = __byte_perm_S (w[ 2], w[ 3], selector);
      w[62] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[61] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[60] = __byte_perm_S (    0, w[ 0], selector);
      w[59] = 0;
      w[58] = 0;
      w[57] = 0;
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 61:
      w[63] = __byte_perm_S (w[ 1], w[ 2], selector);
      w[62] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[61] = __byte_perm_S (    0, w[ 0], selector);
      w[60] = 0;
      w[59] = 0;
      w[58] = 0;
      w[57] = 0;
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 62:
      w[63] = __byte_perm_S (w[ 0], w[ 1], selector);
      w[62] = __byte_perm_S (    0, w[ 0], selector);
      w[61] = 0;
      w[60] = 0;
      w[59] = 0;
      w[58] = 0;
      w[57] = 0;
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;

    case 63:
      w[63] = __byte_perm_S (    0, w[ 0], selector);
      w[62] = 0;
      w[61] = 0;
      w[60] = 0;
      w[59] = 0;
      w[58] = 0;
      w[57] = 0;
      w[56] = 0;
      w[55] = 0;
      w[54] = 0;
      w[53] = 0;
      w[52] = 0;
      w[51] = 0;
      w[50] = 0;
      w[49] = 0;
      w[48] = 0;
      w[47] = 0;
      w[46] = 0;
      w[45] = 0;
      w[44] = 0;
      w[43] = 0;
      w[42] = 0;
      w[41] = 0;
      w[40] = 0;
      w[39] = 0;
      w[38] = 0;
      w[37] = 0;
      w[36] = 0;
      w[35] = 0;
      w[34] = 0;
      w[33] = 0;
      w[32] = 0;
      w[31] = 0;
      w[30] = 0;
      w[29] = 0;
      w[28] = 0;
      w[27] = 0;
      w[26] = 0;
      w[25] = 0;
      w[24] = 0;
      w[23] = 0;
      w[22] = 0;
      w[21] = 0;
      w[20] = 0;
      w[19] = 0;
      w[18] = 0;
      w[17] = 0;
      w[16] = 0;
      w[15] = 0;
      w[14] = 0;
      w[13] = 0;
      w[12] = 0;
      w[11] = 0;
      w[10] = 0;
      w[ 9] = 0;
      w[ 8] = 0;
      w[ 7] = 0;
      w[ 6] = 0;
      w[ 5] = 0;
      w[ 4] = 0;
      w[ 3] = 0;
      w[ 2] = 0;
      w[ 1] = 0;
      w[ 0] = 0;

      break;
  }
  #endif
}

__kernel void amp (__global pw_t *pws, __global pw_t *pws_amp, __global const kernel_rule_t *rules_buf, __global const pw_t *combs_buf, __global const bf_t *bfs_buf, const u32 combs_mode, const u32 gid_max)
{
  const u32 gid = get_global_id (0);

  if (gid >= gid_max) return;

  pw_t pw = pws[gid];

  pw_t comb = combs_buf[0];

  const u32 pw_len = pw.pw_len;

  const u32 comb_len = comb.pw_len;

  if (combs_mode == COMBINATOR_MODE_BASE_LEFT)
  {
    switch_buffer_by_offset_1x64_le_S (comb.i, pw_len);
  }

  if (combs_mode == COMBINATOR_MODE_BASE_RIGHT)
  {
    switch_buffer_by_offset_1x64_le_S (pw.i, comb_len);
  }

  #pragma unroll
  for (int i = 0; i < 64; i++)
  {
    pw.i[i] |= comb.i[i];
  }

  pw.pw_len = pw_len + comb_len;

  pws_amp[gid] = pw;
}