Important notice: Problems with a particular version of MPFR are discussed in the corresponding bugs page.
The latest version of this FAQ is available at http://www.mpfr.org/faq.html. Please look at this version if possible.
The main differences are:
The precision of a MPFR variable is the exact number of bits used for its mantissa, whereas in MPF, the precision requested by the user is a minimum value (MPF generally uses a higher precision). With the additional difference below, this implies that the MPFR results do not depend on the number of bits (16, 32, 64 or more) of the underlying architecture.
As a consequence, MPFR uses a base-2 exponent, whereas in MPF, this is a base-232 or base-264 exponent, depending on the limb size. For this reason (and other internal ones), the maximum exponent range in MPFR is different (and smaller, if the exponent is represented by the same type as in MPF).
MPFR provides an additional rounding mode argument to its functions; furthermore, it is guaranteed that the result of any operation is the nearest possible floating-point value from the exact result (considering the input variables as exact values), taking into account the precision of the destination variable and the rounding mode. MPFR also says whether the rounded result is above or below the exact result.
MPFR supports much more functions (in particular transcendental functions such as exponentials, logarithms, trigonometric functions and so on) and special values: signed zeros, infinities, not-a-number (NaN).
You need to add
to the function names, and to
specify the rounding mode (rGMP_RNDN for rounding to nearest,
GMP_RNDZ for rounding towards zero, GMP_RNDU
for rounding towards plus infinity, GMP_RNDD for rounding
towards minus infinity). You can also define macros as follows:
#define mpf_add(a, b, c) mpfr_add(a, b, c, GMP_RNDN)
The header file mpf2mpfr.h from the
MPFR distribution automatically
redefines all MPF functions in this
way, using the default MPFR rounding
mode. Thus you simply need to add the following line in all your files
using MPF functions:
#include <mpf2mpfr.h>
just after the gmp.h and mpfr.h
header files. If the program uses MPF
internals (such as direct access to __mpf_struct members),
additional changes will be needed.
__gmpXXXX.Link your program with GMP. Assuming that your program is foo.c, you should link it using: cc link.c -lmpfr -lgmp MPFR library reference (-lmpfr) should be before GMP's one (-lgmp). Another solution is, with GNU ld, to give all the libraries inside a group: gcc link.c -Wl,--start-group libgmp.a libmpfr.a -Wl,--end-group See INSTALL file and ld manual for more details.
Moreover, if several GMP versions are installed (e.g., one provided by the system and a new one installed by some user), you must make sure that the include and library search paths are consistent. Unfortunately, on various GNU/Linux machines, they aren't by default. Typical errors are: undefined reference to `__gmp_get_memory_functions' in make check when GMP 4.1.4 is installed in /usr/{include,lib} (provided by the system) and GMP 4.2.1 is installed in /usr/local/{include,lib} (installed by the user with configure, make, make install).
Your stack size limit may be too small; indeed, by default, GMP 4.1.4 and below allocates all temporary results on the stack, and in very high precisions, this limit may be reached. You can solve this problem in different ways:
You can upgrade to GMP 4.2 (or above), which now makes temporary allocations on the stack only when they are small.
You can increase the stack size limit with the limit, unlimit or ulimit command, depending on your shell. This may fail on some systems, where the maximum stack size cannot be increased above some value.
You can rebuild both GMP and MPFR to use another allocation method.
The reason may be the use of C floating-point numbers. If you want
to store a floating-point constant to a mpfr_t, you should use
mpfr_set_str (or one of the MPFR
constant functions, such as mpfr_const_pi for π) instead
of mpfr_set_d or mpfr_set_ld. Otherwise the
floating-point constant will be first converted into a reduced-precision
(e.g., 53-bit) binary number before
MPFR can work with it. This is the case
in particular for most exact decimal numbers, such as 0.17, which are
not exactly representable in binary.
Also remember that MPFR does not track
the accuracy of the results: copying a value x to y
with mpfr_set (y, x, GMP_RNDN) where the variable y
is more precise than the variable x will not make it more
accurate; the (binary) value will remain unchanged.
The MPFR team does not currently recommend any autoconf code, but a section will later be added to the MPFR manual. The MPFR team does not wish to support pkg-config yet.