Rounding error in (**)/2

[ridgeworks]

@jan

Sounds good. Please note that I did merge the bigint-to-float branch. I think that is still ok as that includes the mpq_to_double() patches we want anyway. Your code should add a fast-path implementation for mpz_to_double() that now uses mpq_to_double().

Almost. There are a couple of additional patches to mpq_to_double; one to avoid over-rounding in the FE_UPWARD and FE_DOWNWARD rounding modes, and one to ensure that infinity is returned the result will overflow a 64 bit float.

The new function is named mpz_float which is called from mpz_to_double rather than mpz_fdiv with a divisor argument of MPZ_ONE.

Changes are confined to gmp.c as described and Tests/rational/test_ieee_754.pl which will have some additional rounding tests cases for floating various “types” of integers.