RtlCopyMemoryNonTemporal

VOID __stdcall RtlCopyMemoryNonTemporal(VOID *Destination, const VOID *Source, SIZE_T Length){
  signed __int64 v3; 
  char *v4; 
  char *v5; 
  __int64 v6; 
  signed __int64 v7; 
  __m128i v8; 
  __m128i v9; 
  bool v10; 
  SIZE_T v11; 
  bool v12; 
  v3 = (_BYTE *)Source - (_BYTE *)Destination;
  if( Length >= 0x4F )
  {
    if( ((unsigned __int8)Destination & 0xF) != 0 )
    {
      v4 = (char *)Destination + Length;
      _mm_stream_si32((int *)Destination, *(_QWORD *)((char *)Destination + v3));
      if( ((unsigned __int8)Destination & 8) == 0 )
        _mm_stream_si32((int *)Destination + 2, *(_QWORD *)((char *)Destination + v3 + 8));
      Destination = (VOID *)(((unsigned __int64)Destination + 16) & 0xFFFFFFFFFFFFFFF0ui64);
      Length = v4 - (_BYTE *)Destination;
    }
    do
    {
      v5 = (char *)Destination;
      v6 = 16i64;
      v10 = Length < 0x400;
      Length -= 1024i64;
      if( v10 )
      {
        v11 = Length + 1024;
        v6 = v11 >> 6;
        Length = v11 & 0x3F;
      }
      do
      {
        _mm_prefetchnta(&v5[v3]);
        v5 += 64;
        --v6;
      }
      while( v6 );
      v7 = v5 - (_BYTE *)Destination;
      do
      {
        v8 = _mm_loadu_si128((const __m128i *)((char *)Destination + v3 + 16));
        _mm_stream_si128((__m128i *)Destination, _mm_loadu_si128((const __m128i *)((char *)Destination + v3)));
        _mm_stream_si128((__m128i *)Destination + 1, v8);
        Destination = (char *)Destination + 64;
        v9 = _mm_loadu_si128((const __m128i *)((char *)Destination + v3 - 16));
        _mm_stream_si128((__m128i *)Destination - 2, _mm_loadu_si128((const __m128i *)((char *)Destination + v3 - 32)));
        _mm_stream_si128((__m128i *)Destination - 1, v9);
        v7 -= 64i64;
      }
      while( v7 );
    }
    while( Length >= 0x40 );
  }
  else if( Length < 8 )
  {
    for( ; Length; --Length )
    {
      *(_BYTE *)Destination = *((_BYTE *)Destination + v3);
      Destination = (char *)Destination + 1;
    }
    return;
  }
  while( 1 )
  {
    v12 = Length <= 8;
    Length -= 8i64;
    if( v12 )
      break;
    _mm_stream_si32((int *)Destination, *(_QWORD *)((char *)Destination + v3));
    Destination = (char *)Destination + 8;
  }
  _mm_stream_si32((int *)((char *)Destination + Length), *(_QWORD *)((char *)Destination + Length + v3));
  _mm_sfence();
}

Referenced by: