SNK Comrpession

xmmword_140867F90 = FC FF FF FF 01 00 00 00 02 00 00 00 03 00 00 00
xmmword_140867FA0 = 00 00 00 00 00 00 00 00 00 00 00 00 FF FF FF FF

__int64 __fastcall decompress(unsigned __int8 *scr_buff, unsigned __int8 *dst_buff, int decompressed_size)
{
  int v3; // ebp
  unsigned __int8 *v4; // rbx
  unsigned __int8 *v5; // rsi
  unsigned __int8 *v6; // rax
  __int64 result; // rax
  unsigned int v8; // ecx
  unsigned __int64 v9; // r9
  size_t v10; // rdi
  __int64 v11; // xmm0_8
  unsigned __int8 *v12; // rbx
  unsigned __int8 *v13; // rax
  unsigned __int64 v14; // r9
  unsigned __int64 v15; // r11
  unsigned __int8 *v16; // r8
  __int64 v17; // rcx
  unsigned __int8 *v18; // r8
  signed __int64 v19; // rdx
  unsigned __int8 *v20; // rbx
  __int64 v21; // rdx
  unsigned __int64 v22; // r9
  unsigned __int64 v23; // r10
  unsigned __int8 *v24; // r8
  unsigned __int8 *v25; // r8
  __int64 v26; // rcx
  unsigned __int8 *v27; // rax
  unsigned __int8 *v28; // r9
  unsigned __int8 *v29; // rdx
  signed __int64 v30; // rcx
  unsigned __int64 v31; // rdx
  unsigned __int64 v32; // r9
  _QWORD *v33; // rcx
  signed __int64 v34; // r8
  int v35[3]; // [rsp+20h] [rbp-78h]
  __int64 v36; // [rsp+2Ch] [rbp-6Ch]
  int v37; // [rsp+34h] [rbp-64h]
  int v38; // [rsp+38h] [rbp-60h]
  int v39; // [rsp+3Ch] [rbp-5Ch]
  __int128 v40[2]; // [rsp+40h] [rbp-58h]

  v3 = scr_buff;
  v4 = scr_buff;
  v5 = &dst_buff[decompressed_size];
  v35[0] = 0;
  v35[2] = 2;
  v35[1] = 1;
  v6 = dst_buff;
  v36 = 1i64;
  v37 = 4;
  v38 = 4;
  v39 = 4;
  v40[0] = _mm_load_si128(&xmmword_140867FA0);
  v40[1] = _mm_load_si128(&xmmword_140867F90);
  if ( !decompressed_size )
  {
    result = 0xFFFFFFFFi64;
    if ( !*scr_buff )
      result = 1i64;
    return result;
  }
  while ( 1 )
  {
    while ( 1 )
    {
      v8 = *v4++;
      v9 = v8;
      v10 = v8 & 0xF;
      if ( v10 > 8 || v6 > v5 - 26 )
        break;
      v11 = *v4;
      v12 = &v4[v10];
      *v6 = v11;
      v13 = &v6[v10];
      v14 = v8 >> 4;
      v15 = *v12;
      v4 = v12 + 2;
      v16 = &v13[-v15];
      if ( v14 == 15 || v15 < 8 )
        goto LABEL_16;
      *v13 = *v16;
      *(v13 + 1) = *(v16 + 1);
      *(v13 + 8) = *(v16 + 8);
      v6 = &v13[v14 + 4];
    }
    if ( v10 == 15 )
    {
      do
      {
        v17 = *v4++;
        v10 += v17;
      }
      while ( v17 == 255 );
    }
    v18 = &v6[v10];
    if ( &v6[v10] > v5 - 8 )
      break;
    v19 = v4 - v6;
    do
    {
      *v6 = *&v6[v19];
      v6 += 8;
    }
    while ( v6 < v18 );
    v20 = &v4[v10];
    v13 = v18;
    v15 = *v20;
    v4 = v20 + 2;
    v16 = &v18[-v15];
    v14 = v9 >> 4;
LABEL_16:
    *v13 = v15;
    if ( v14 == 15 )
    {
      do
      {
        v21 = *v4++;
        v14 += v21;
      }
      while ( v21 == 255 );
    }
    v22 = v14 + 4;
    v23 = &v13[v22];
    if ( v15 >= 8 )
    {
      v26 = *v16;
      v25 = v16 + 8;
      *v13 = v26;
    }
    else
    {
      *v13 = *v16;
      v13[1] = v16[1];
      v13[2] = v16[2];
      v13[3] = v16[3];
      v24 = &v16[v35[v15]];
      *(v13 + 1) = *v24;
      v25 = &v24[-*(v40 + v15)];
    }
    v27 = v13 + 8;
    if ( v23 <= (v5 - 12) )
    {
      *v27 = *v25;
      if ( v22 > 0x10 )
      {
        v33 = v27 + 8;
        v34 = v25 - (v27 + 8);
        do
        {
          *v33 = *(v33 + v34 + 8);
          ++v33;
        }
        while ( v33 < v23 );
      }
      goto LABEL_35;
    }
    v28 = v5 - 7;
    if ( v23 > (v5 - 5) )
      return (v3 - v4 - 1);
    if ( v27 < v28 )
    {
      v29 = v27;
      do
      {
        *v29 = *&v29[v25 - v27];
        v29 += 8;
      }
      while ( v29 < v28 );
      v30 = v28 - v27;
      v27 = v5 - 7;
      v25 += v30;
    }
    v31 = 0i64;
    v32 = v23 - v27;
    if ( v27 > v23 )
      v32 = 0i64;
    if ( v32 )
    {
      do
      {
        ++v31;
        *v27++ = *v25++;
      }
      while ( v31 < v32 );
      v6 = v23;
    }
    else
    {
LABEL_35:
      v6 = v23;
    }
  }
  if ( v18 == v5 )
  {
    memmove(v6, v4, v10);
    return (v4 + v10 - v3);
  }
  return (v3 - v4 - 1);
}