That's because it's banker's rounding, as previously established in this thread (correct?). When the fractional part is 0.5, banker's rounding rounds up if the integer part is odd and rounds down if the integer part is even. Your x1 has even integer part and x2 has odd integer part, hence the different rounding.
Okay but the engineers in 1995, when the 80587 has bugged, the guys did not want to fix such this joke ?
And they have build all the next SIMD architecture based with this failure ?
And the IA has been based with this failure ?
And now, the engineers are surprised by the new hardware circuits made by the IA ? These engineers do not understand the new components created by the IA...