make conversion to integers round

[StefanKarpinski]

This is easy, per se, but making sure everything that depends on the behavior of conversion to integers doesn't break is harder.

[JeffBezanson]

I have nightmares that we will not be able to get this inlined.

[StefanKarpinski]

Do you want to not do it instead?

[JeffBezanson]

It might be possible. If I can get llvm to inline it we can do it.

[JeffBezanson]

OK, I can make an intrinsic that efficiently computes itrunc(x+copysign(0.5, x)). Is that an acceptable way to compute this?

[StefanKarpinski]

Looks good to me. Seems to behave identically to x->int(round(x)). The following test code passes:

ir1(x) = int(round(x))
ir2(x) = int(x + copysign(0.5, x))

for x = {-0.5,
         -0.49999999999999999,
         -0.4999999999999999,
          0.0,
          0.4999999999999999,
          0.49999999999999999,
          0.5,
         -Inf,
          Inf,
          NaN,
         -NaN}
    @assert isequal(ir1(x),ir2(x))
end