Issue with printing?

[bmsherman]

# let x = 4294967296 in x^2;;
- : real = 1.8446744074e19 +- 0.0
# let x = 4294967296 in x^2 - 18446744073709551616.;;
- : real = 0.0 +- 0.0
# 18446744073709551616. - 1.8446744074e19;;
- : real = -290448384. +- 0.0

[bmsherman]

Changing the digits argument in the Mpfr.get_str call to a larger number, like 100, makes it return the correct answer, indicating that this is indeed an issue with printing. It's surprising, because the contract for setting digits to 0 says:

If n is zero, the number of digits of the significand is chosen large enough so that re-reading the printed value with the same precision, assuming both output and input use rounding to nearest, will recover the original value of op. More precisely, in most cases, the chosen precision of str is the minimal precision m depending only on p = PREC(op) and b that satisfies the above property, i.e., m = 1 + ceil(p*log(2)/log(b)), with p replaced by p-1 if b is a power of 2, but in some very rare cases, it might be m+1 (the smallest case for bases up to 62 is when p equals 186564318007 for bases 7 and 49).

Either I'm misunderstanding that interface, or there is a bug with MPFR.