- Update mpfr-4.2.0-cummulative.patch, adds patches fixing the following bugs
* the mpfr_reldiff function, which computes |b-c|/b, is buggy on special
values, e.g. on the following (b,c) values: (+Inf,+Inf) gives ±0
instead of NaN (like NaN/Inf); (+0,+0) gives 1 instead of NaN (like 0/0);
(+0,1) gives 1 instead of Inf (like 1/0). Moreover, the sign of 0 for
(+Inf,+Inf) or (-Inf,-Inf) is not set, i.e. it is just the sign of the
destination before the call; as a consequence, results are not even
consistent.
* the reuse tests are incomplete: the sign of a result zero is not checked,
so that it can miss bugs (one of the mpfr_reldiff bugs mentioned above,
in particular).
* the general code for the power function (mpfr_pow_general internal
function) has two bugs in particular cases: the first one is an incorrect
computation of the error bound when there has been an intermediate
underflow or overflow (in such a case, the computation is performed
again with a rescaling, thus with an additional error term, but there is
a bug in the computation of this term), so that the result may be
rounded incorrectly (in particular, a spurious overflow is possible);
the second one occurs in a corner case (destination precision 1,
rounding to nearest, and where the rounded result assuming an unbounded
exponent range would be 2emin-2 and the exact result is larger than this
value), with the only consequence being a missing underflow exception
(the underflow flag is not set).
* the mpfr_compound_si function can take a huge amount of memory and time
in some cases (when the argument x is a large even integer and xn is
represented exactly in the target precision) and does not correctly
detect overflows and underflows
* MPFR can crash when a formatted output function is called with
%.2147483648Rg in the format string.