Handling of floating point numbers changed from GMP 4.1.4 to GMP 4.2.1

Roberto Bagnara bagnara at cs.unipr.it
Wed Sep 12 10:42:48 CEST 2007

When we installed GMP 4.2.1 on some machines, the regression testing procedure
we use for the Parma Polyhedra Library started signaling a problem.  It boils
down to the following test program:

#include <gmpxx.h>
#include <iostream>
#include <cstdio>
#include <fenv.h>

int main() {
   // Rounding upward.

   mpz_class num("137171200400403985");
   mpz_class den("1125899906842624");
   mpq_class rational(num, den);
   std::cout << "real_coeff = " << rational << "\n";

   double floating = rational.get_d();
   printf("float = %.20g\n", floating);

   return 0;

The results are different depending on whether it is linked to GMP 4.1.4
or with GMP 4.2.1.  For example, on my Fedora 7, x86_64 machine where I
have GMP 4.1.4 in /usr/lib64 and GMP 4.2.1 in /usr/local/lib, I use the

g++ -I/usr/local/include -o bug.o -c bug.cc
g++ -static -o bug-gmp-4.2.1 bug.o -L/usr/local/lib -lm -lgmpxx -lgmp
g++ -static -o bug-gmp-4.1.4 bug.o -L/usr/lib64 -lm -lgmpxx -lgmp

This prints

real_coeff = 137171200400403985/1125899906842624
float = 121.83249999999999602
real_coeff = 137171200400403985/1125899906842624
float = 121.83250000000001023

The mismatch does not happen if one does not touch the rounding mode
So, on this example, while GMP 4.1.4 results do not depend on the
rounding mode, with GMP 4.2.1 the rounding mode matters.
Perhaps the answer is simply that GMP provides no guarantee
when invoked with the rounding mode set to anything different from
FE_TONEAREST, but I have not found that in the documentation.
Instead, the documentation says, e.g., that

   double mpq_get_d (mpq_t op)
   Convert op to a double, truncating if necessary (ie.: rounding towards zero).

We are uncertain about what to do: for our computations we need to set
the rounding mode;  for efficiency we would like to avoid setting it back
and forth;  and, on the other hand, we now have code that behaves
differently depending on the version of GMP it is linked to.
Any advice?
All the best,


Prof. Roberto Bagnara
Computer Science Group
Department of Mathematics, University of Parma, Italy
mailto:bagnara at cs.unipr.it

More information about the gmp-discuss mailing list