Detailed description 
* Possible incorrect results due to internal underflow, which can lead to a huge loss of accuracy while the error analysis doesn't take that into account. If the underflow occurs at the last function call (just before the MPFR_CAN_ROUND), the result should be correct (or MPFR gets into an infinite loop). TODO: check the code and the error analysis.
* Possible integer overflows on some machines.
* Possible bugs with huge precisions (> 2^30).
* Possible bugs if the chosen exponent range does not allow to represent the range [1/16, 16].
* Possible infinite loop in some functions for particular cases: when the exact result is an exactly representable number or the middle of consecutive two such numbers. However for nonalgebraic functions, it is believed that no such case exists, except the wellknown cases like cos(0)=1, exp(0)=1, and so on, and the x^y function when y is an integer or y=1/2^k.
* The mpfr_set_ld function may be quite slow if the long double type has an exponent of more than 15 bits.
* mpfr_set_d may give wrong results on some nonIEEE architectures.
* Error analysis for some functions may be incorrect (outofdate due to modifications in the code?).
* Possible use of nonportable feature (preC99) of the integer division with negative result.

