RtlCopyMemoryNonTemporal

VOID *__stdcall RtlCopyMemoryNonTemporal(VOID *Destination, const VOID *Source, _SIZE_T Length){
  VOID *result; 
  signed __int64 v4; 
  char *v5; 
  char *v6; 
  __int64 v7; 
  signed __int64 v8; 
  __m128i v9; 
  __m128i v10; 
  bool v11; 
  unsigned __int64 v12; 
  bool v13; 
  int *v14; 

  v4 = (_BYTE *)Source - (_BYTE *)Destination;
  if( *(_QWORD *)&Length >= 0x4Fui64 )
  {
    if( ((unsigned __int8)Destination & 0xF) != 0 )
    {
      v5 = (char *)Destination + *(_QWORD *)&Length;
      _mm_stream_si32((int *)Destination, *(_QWORD *)((char *)Destination + v4));
      if( ((unsigned __int8)Destination & 8) == 0 )
        _mm_stream_si32((int *)Destination + 2, *(_QWORD *)((char *)Destination + v4 + 8));
      Destination = (VOID *)(((unsigned __int64)Destination + 16) & 0xFFFFFFFFFFFFFFF0ui64);
      *(_QWORD *)&Length = v5 - (_BYTE *)Destination;
    }
    do
    {
      v6 = (char *)Destination;
      v7 = 16i64;
      v11 = *(_QWORD *)&Length < 0x400ui64;
      *(_QWORD *)&Length -= 1024i64;
      if( v11 )
      {
        v12 = *(_QWORD *)&Length + 1024i64;
        v7 = v12 >> 6;
        *(_QWORD *)&Length = v12 & 0x3F;
      }
      do
      {
        _mm_prefetchnta(&v6[v4]);
        v6 += 64;
        --v7;
      }
      while( v7 );
      v8 = v6 - (_BYTE *)Destination;
      do
      {
        v9 = _mm_loadu_si128((const __m128i *)((char *)Destination + v4 + 16));
        _mm_stream_si128((__m128i *)Destination, _mm_loadu_si128((const __m128i *)((char *)Destination + v4)));
        _mm_stream_si128((__m128i *)Destination + 1, v9);
        Destination = (char *)Destination + 64;
        v10 = _mm_loadu_si128((const __m128i *)((char *)Destination + v4 - 16));
        _mm_stream_si128((__m128i *)Destination - 2, _mm_loadu_si128((const __m128i *)((char *)Destination + v4 - 32)));
        _mm_stream_si128((__m128i *)Destination - 1, v10);
        v8 -= 64i64;
      }
      while( v8 );
    }
    while( *(_QWORD *)&Length >= 0x40ui64 );
  }
  else if( *(_QWORD *)&Length < 8ui64 )
  {
    for( ; *(_QWORD *)&Length; --*(_QWORD *)&Length )
    {
      LOBYTE(result) = *((_BYTE *)Destination + v4);
      *(_BYTE *)Destination = (_BYTE)result;
      Destination = (char *)Destination + 1;
    }
    return result;
  }
  while( 1 )
  {
    v13 = *(_QWORD *)&Length <= 8ui64;
    *(_QWORD *)&Length -= 8i64;
    if( v13 )
      break;
    _mm_stream_si32((int *)Destination, *(_QWORD *)((char *)Destination + v4));
    Destination = (char *)Destination + 8;
  }
  v14 = (int *)((char *)Destination + *(_QWORD *)&Length);
  result = *(VOID **)((char *)v14 + v4);
  _mm_stream_si32(v14, (int)result);
  _mm_sfence();
  return result;
}

Referenced by: