You can do things that way, but unless you're manipulating things properly, you won't get the right answer. A procedeure for doing a N=32 point example is:
1) generate 32 point time domain data,
2) forward transform it,
3) divide results by 32,
4) multiply positive frequencies by 2 (frequencies f= 1 to 15),
5) zero negative frequencies (frequencies f=17 to 31),
6) inverse transform it,
7) ignore the imaginary outputs.
The real results of the above procedure will match the original input waveform.
Note that in the above, you don't multiply the f=0 or f=N/2 outputs of the FFT by 2. For this case (N=32 - an even number), they are unique - they don't have negative frequency counterparts. Now if you're using an appropriate FFT routine that allows for odd N, then you have to account for the fact that you'll have an f=0 point, but no f=N/2 point.
Unless you are doing very large FFT's, you're probably not going to save much in the way of processing time by using just the positive frequencies.