memset

INT64 __stdcall memset(INT64 result, INT64 a3){
  unsigned __int64 v2; 
  INT64 v3; 
  __int64 v4; 
  unsigned __int64 v5; 
  unsigned __int64 v6; 
  unsigned __int64 v7; 
  __m128 v8; 
  unsigned __int64 v9; 
  __m128 *v10; 
  unsigned __int64 v11; 
  unsigned __int64 v12; 
  unsigned __int64 i; 
  unsigned __int64 v14; 
  v3 = result;
  if( v2 < 8 )
  {
    for( ; v2; --v2 )
      *(_BYTE *)(result + v2 - 1) = a3;
  }
  else
  {
    v4 = 0x101010101010101i64 * (unsigned __int8)a3;
    if( v2 >= 0x4F )
    {
      v8 = _mm_movelh_ps((__m128)(unsigned __int64)v4, (__m128)(unsigned __int64)v4);
      *(__m128 *)result = v8;
      v9 = result + v2;
      v10 = (__m128 *)((result + 16) & 0xFFFFFFFFFFFFFFF0ui64);
      v11 = v9 - (_QWORD)v10;
      v12 = v11 >> 7;
      if( v11 >> 7 )
      {
        do
        {
          *v10 = v8;
          v10[1] = v8;
          v10 += 8;
          v10[-6] = v8;
          v10[-5] = v8;
          --v12;
          v10[-4] = v8;
          v10[-3] = v8;
          v10[-2] = v8;
          v10[-1] = v8;
        }
        while( v12 );
        v11 &= 0x7Fu;
      }
      for( i = v11 >> 4; i; --i )
        *v10++ = v8;
      v14 = v11 & 0xF;
      if( v14 )
        *(__m128 *)((char *)v10 + v14 - 16) = v8;
    }
    else
    {
      v5 = v2 & 0x78;
      v6 = (v2 & 0xFFFFFFFFFFFFFFF8ui64) + result;
      do
      {
        *(_QWORD *)(v3 + v5 - 8) = v4;
        v5 -= 8i64;
      }
      while( v5 );
      v7 = v2 & 7;
      if( v7 )
        *(_QWORD *)(v6 + v7 - 8) = v4;
    }
  }
  return v3;
}

Referenced by: