I have done some simple speed and accuracy tests on floating point maths to compare X87 code with XMM code. The comments in the attached file should be reasonably self explanatory.
I have read comments that XMM maths is faster than X87. My tests confirm that XMM can at least match X87 for a fairly simple case. In 64 bit mode, the parameter passing would be more efficient than here so XMM should then be a bit faster.