TODO:
----

1. Check for denormal and pseudodenormal operands in ALL instructions
   I hope all instructions return the same values as real CPU.

2. Unmasked underflow/overflow should correct the result by magic number
   for all operations, including float32 and float64.

3. Elliminate floa128 use, Intel uses only 67-bit precision calculations
   when float128 has at least 112-bit. Replacement of float128 with for
   example 96-bit precision number could significantly speed up 
   calculations.
