Notes for the MPFR developers and Subversion users
==================================================

To compile source code obtained from the Subversion repository, you
need some GNU development utilities: aclocal, autoheader, automake,
autoconf 2.50 (at least) and libtoolize. As some files like "configure"
are not part of the Subversion repository, you first need to run
"autoreconf -i"; if you have both autoconf 2.13 and 2.50 installed and
use a wrapper, you may need to run "autoreconf" a second time (as with
old Debian packages, due to a bug in the wrapper). Then you can run
"configure" in the usual way (see the INSTALL file, but note that
there are no patches to apply, and the URLs are not valid since the
corresponding version has not been released yet).

If for some reason, this doesn't work, there's still the old way: run
the "prepare" script to generate these files; as the "prepare" script
also runs "configure", you must give the options that will be passed
to "configure", for instance: ./prepare --with-gmp=/path/to/gmp
Give "--help" if you don't know (./prepare --help). Then, you can
use the configure script as usual. Read the INSTALL file for more
information. If you use GNU tools, you can give "-dev" as the first
option to the "prepare" script to get better dependency tracking.

To generate mpfr.info, you need texinfo version 4.2 (or higher).

===========================================================================

The VERSION file contains the number of the next release version, i.e.
the version currently being developed. A suffix can be attached for the
development versions (in general, "-dev") or pre-release versions (e.g.
"-rc1"). It must be updated with the update-version script.

If nightly snapshots are built, the date in the yyyymmdd format and/or
the Subversion revision number (giving more accurate information) must
be added to the version as a suffix, for instance: "2.3.0-20070621" or
"2.3.0-dev-r4553".

Patches can be tracked by adding a chunk of the form

--- PATCHES~    Tue Nov  6 19:59:33 2001
+++ PATCHES     Tue Nov  6 19:59:42 2001
@@ -1,0 +1 @@
+<your-id-here>

to the patch file[*]. After such patches have been applied, the file
get_patches.c providing the mpfr_get_patches() function will be rebuilt
by "make". MPFR distributors can still modify the version suffix from
the applied patches according to their version naming scheme; for
instance, for their own patches, MPFR developers do:

  ./update-version 2 4 0 p1 -

[*] This idea comes from Thomas Roessler, who implemented it in Mutt.

For patches from MPFR developers, e.g. for MPFR 2.4.0:
1. Unarchive the tarball: a directory mpfr-2.4.0 is created.
2. Go into this directory (cd mpfr-2.4.0).
3. Apply the current patches with "patch -N -Z -p1 < /path/to/patches".
4. Reset the PATCHES file with "true >| PATCHES".
5. Rename mpfr-2.4.0 as mpfr-2.4.0-a and duplicate it as mpfr-2.4.0-b
   without changing the timestamps (e.g. with cp -a).
6. In mpfr-2.4.0-b, apply the patch obtained with "svn diff", e.g.
     patch --no-backup-if-mismatch -p0 < /path/to/new_patch
7. Update the version information:
     /path/to/mpfr-working-copy/update-version 2 4 0 p<n> -
   where <n> is the patch number.
8. Update PATCHES file: echo >> PATCHES <patch_name>
9. Make the patch: TZ=UTC diff -Naurd mpfr-2.4.0-a mpfr-2.4.0-b

===========================================================================

When submitting patches, unified diffs (option -u) are recommended,
as they are more readable. You can also use the option -d to generate
a smaller set of changes. See diff(1) for more information.

===========================================================================

Copyright Notices: For easier maintainability, make sure that the
copyright notices match the regexp "Copyright.* yyyy Free Software"
where yyyy is the year of the latest modification in the branch
(and nothing else should match it).

The latest rules for GNU software can be found here:

  http://www.gnu.org/prep/maintain/maintain.html#Copyright-Notices

===========================================================================

To make a release (for the MPFR team):

  *** Please read this section entirely before making any release. ***

  0) Make sure that the mpfr-longlong.h file (from GMP's longlong.h) and
     the libtool-related files (config.guess, etc.) are up-to-date. An
     "autoreconf -f -i" may be necessary.

  1) Generate the tuning parameters on different architectures
     and put them in mparam_h.in. For each architecture:

     a) download the latest release of GMP on gmplib.org
     b) build GMP with --disable-shared in say /tmp/gmp-x.y.z
        (there is no need in tuning GMP, since most users will build MPFR
	 with a vanilla GMP installation, i.e., with the default GMP tuning)
     c) configure MPFR with --disable-shared --with-gmp-build=/tmp/gmp-x.y.z
     d) run "make" and "make tune"
     e) put the resulting mparam.h file into mparam_h.in (please include
        the version of GMP and the compiler used)

     You can produce time graphs to check the thresholds are correct (and
     compare to the corresponding mpf functions) with mbench. For example:

     $ cd mpfr/mbench
     $ make mpfr-gfx GMP=... MPFR=...
     $  ./mpfr-gfx -b16 -e320 -s16 -f2 -x3 # compares mpfr_mul and mpf_mul
                                           # from 16 to 320 bits with increment
                                           # of 16 bits
     $ gnuplot -persist plot.gnuplot

     Another example, comparing mpfr_mul and mpf_mul from 2 to 1000000 bits,
     with ratio 1.1 between two sizes, 10 random values, and 10 smoothness
     checks:

     $ ./mpfr-gfx -b2 -e1000000 -r1.1 -f10 -x3 -m10
     $ gnuplot -persist plot.gnuplot

     Check the coverage of each source file by the test suite is at least 90%
     (or clearly justify any value under this threshold), and publish (for
     example in NEWS) the global coverage of this release. The individual
     coverage of each source file might also be published on the release web
     page.

  2) Check the version and change the suffix to "rc1", "rc2", etc. with
     update-version for the release candidates; remove the suffix for
     the final release.
     Update the libtool version too (see Makefile.am).
     Update the date in mpfr.texi.

  3) Update the NEWS file.
     Update the FAQ.html file with update-faq (and check it).

  4) Update the ChangeLog file with "TZ=UTC svn log -rHEAD:0 -v" in
     UTF-8 locales, e.g. "LC_ALL=en_US.UTF8 TZ=UTC svn log -rHEAD:0 -v".

  5) Make sure that any change has been committed (with "svn st").
     Generate the release version with "make dist", but see below.

  6) Test the release version on different machines, with and without
     --enable-assert (--enable-assert, though doing more checks, may
     hide bugs due to the fact that <stdio.h> is always included) with
     and without GMP's --disable-alloca configure option (or compile GMP
     with --enable-alloca=debug and MPFR with --with-gmp-build to be
     able to get the memory leak errors), with and without -DXDEBUG in
     $CFLAGS, with and without gmp internal files, with and without
     GMP built as a shared library, with and without srcdir equal to
     objdir (../mpfr/configure --srcdir=/users/spaces/pelissip/mpfr/).

     Try different gcc versions with different options: with and without
     "-std=c99 -O3 -D_XOPEN_SOURCE=500", with and without "-ansi" (which
     allows to turn off features that are incompatible with ISO C90),
     with and without [-ansi] -pedantic-errors (which has the effect to
     disable extensions, such as long long when used together with -ansi)
     with and without --enable-thread-safe, in various FPU precisions
     (double, double extended and single) if the platform supports that
     (e.g. under Linux/x86), and in various locales (LC_ALL=tr_TR in
     particular, if installed). On x86, test with -m96bit-long-double
     and -m128bit-long-double. Try also with gcc's -fno-common option.
     Check also with "-Wformat=2", but without logging support (in order
     to avoid too many spurious warnings).

     Check that make and make check pass with a C++ compiler, for example:
     ./configure CC=g++ (MPFR 2.3.2 did not).

     Try different compilers, e.g., icc, opencc (x86_64 machines),
     tcc <http://bellard.org/tcc/>.

     On 64-bit PowerPC, test against GMP built with the different ABI's:
     32, mode32 and mode64 (in particular mode32, where long's have
     32 bits and limbs have 64 bits [long long]).

     Test with both "make check" and the worst cases.

     Check that there are no abnormal regressions in the timings (both
     for 100, 1000, 10000 digits, http://mpfr.org/mpfr-current/timings.html,
     and for small precision, using the mbench program, see mpfr/mbench).

     Test the library interface compatibility by running the test suite
     compiled against an old library version and dynamically linked with
     the new library version: for instance, build the shared library of
     old and new MPFR versions with the same configure options, and from
     the build directory of the old version, do something like:
     (cd .libs && ln -nsf ../../mpfr-new/.libs/libmpfr.so.1.* libmpfr.so.1)
     then "make check".

  7) For the release itself (not the release candidates), if no problems
     have been found, create a tag with:
       svn cp .../mpfr/branches/x.y .../mpfr/tags/x.y.z

  8) Update the version with the update-version script to indicate the
     next version (use the "dev" suffix).

For major or minor releases (but not patchlevels), a branch may be
created first to allow new features to be committed to the trunk.

A libtool bug (with icc) to be fixed before the next MPFR releases:
  http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=485421

Before a release, add AM_MAINTAINER_MODE(enable) to configure.in,
autoreconf -i with patched automake (2008-09-01), then comment out
AM_MAINTAINER_MODE(enable) while keeping the timestamp.

===========================================================================

To check the coverage of the Test Suite, you can use GCOV.
   ./configure CFLAGS="-fprofile-arcs -ftest-coverage"
   make clean
   make check
   find . -name '*.c' -exec gcov '{}' ';' | grep "lines executed" | sort

For each source file, there is a .c.gcov file which contains much more
information.
Another solution is to run the script 'coverage' from MPFR source directory.

===========================================================================

List of the used macros for building MPFR:

+ HAVE_STRCASECMP:      Define if the system supports 'strcasecmp' function.
+ HAVE_STRNCASECMP:     Define if the system supports 'strncasecmp' function.

+ HAVE_CONFIG_H:        Define if we have to include 'config.h' first.
+ MPFR_HAVE_GMP_IMPL:   Define if we have the gmp internal files.
                         ('gmp-impl.h', 'gmp-maparam.h', ...).

+ HAVE_ALLOCA_H:        Define if the function 'alloca' is in alloca.h.
+ HAVE_LONG_LONG:       Define if the system supports 'long long'

+ HAVE_STDARG:          Define if the system supports 'stdarg.h'.
                         Otherwise it is assumed it is 'vararg.h'.
+ HAVE_STDINT_H:        Define if 'intmax_t' is supported (ISO C99).

 Define format of long double.
+ HAVE_LDOUBLE_IEEE_EXT_LITTLE:         IEEE extended, little endian.
+ HAVE_LDOUBLE_IEEE_QUAD_BIG:           IEEE quad, big endian.

+ XDEBUG:               Use generic 'double' code instead of IEEE specific one.
                        The IEEE code for double needs GMP internal files.
+ WANT_ASSERT:          Define if we want to turn on the assertions.
+ MPFR_EXP_CHECK:       Define if we want to check the exp field.

+ MPFR_PREC_FORMAT:     Define the internal format of prec field
                         (For experimented users).

+ IEEE_DBL_MANT_DIG:    Number of bits in the mantissa of a double.
                        (Default: 53).
+ LDBL_MANT_DIG:        Number of bits in the mantissa of a long double

List of the used macros for checking MPFR:

+ MPFR_HAVE_FESETROUND: Define if the fesetround function is defined
                         (and in header fenv.h).
+ HAVE_DENORMS:         Define if denormalized floats work.
+ HAVE_SYS_TIME_H:      Define if the header sys/time.h is usable.
+ HAVE_GETTIMEOFDAY:    Define if the function gettimeofday is present.
+ MPFR_HAVE_TESTS_x86:  Define if we are on x86.

===========================================================================

The GNU Coding standards can be read at:
  http://www.gnu.org/prep/standards_toc.html
ISO C Names and corresponding headers:
  http://www.schweikhardt.net/identifiers.html
Language C:
  http://www.vmunix.com/~gabor/c/draft.html

To allow MPFR to be built on some buggy compiler, try to follow
theses rules:

   =====================================================================

Don't write:
  mp_limb_t l;
  [...]
  if (l) do_action ();
But:
  mp_limb_t l;
  [...]
  if (l != 0) do_action ();

 since mp_limb_t may be "unsigned long long", and some buggy compiler
produce illegal codes with the first form.

   =====================================================================

Don't use "near" and "far" as variable names since they are "Keywords"
for some C compiler (Old DOS compiler). Also don't use "pm", which is used
by the C compiler 'sharp' to design variables that should be stored in the
flash memory. Don't use "new", which is reserved in C++.

   =====================================================================

Avoid variable names "l", "I" and "O", which look like "1" and "0" with
some fonts.

   =====================================================================

Quoted from <http://www.gnu.org/software/gcc/codingconventions.html>:

  Avoid the use of identifiers or idioms that would prevent code
  compiling with a C++ compiler. Identifiers such as new or class,
  that are reserved words in C++, should not be used as variables
  or field names. Explicit casts should be used to convert between
  void* and other pointer types.

   =====================================================================

Try to avoid "LONG_MIN/1" since it produces a SIGNAL on (old) FreeBsd.
Don't forget that LONG_MIN/-1 is not representable (specially
with code like MPFR_EXP_MIN/n).

   =====================================================================

Though the ISO C standard requires that <stdlib.h> defines NULL,
do not use NULL with #include <stdlib.h> only, because this will
not work with the native SunOS 4 C compiler, whose headers are
not conform to the standard (even with C90); other problems may
occur on this architecture. You can either include <stdio.h> or
use 0 (possibly casted to the target pointer type).

   ====================================================================

Use locale-dependent functions when the result needs to depend on the
locales, e.g. the fractional point in mpfr_out_str.

Conversely, do not use locale-dependent functions when the result must
not depend on the locales. In particular, the alphanumeric characters
used in number strings (as created by mpfr_get_str) must be those of
the required characters from the basic character set (see ISO C99
standard Section 5.2.1 "Character sets"). And tolower(letter) does
not necessarily return the corresponding lowercase letter from these
required characters. For instance, tolower('I') returns a dotless 'i'
in Turkish locales.

   ====================================================================

In the tests, do not use `mpfr_set_d` (except when testing it), as the
result will depend on the floating-point arithmetic of the system;
this has shown many problems in the past and problems may still occur
with new systems. Use `mpfr_set_si` or `mpfr_set_str` instead.

Also, make sure that the tests run against previous MPFR versions,
possibly by disabling some tests with code like

  #if MPFR_VERSION >= MPFR_VERSION_NUM(2,3,0)

Indeed one can now easily run the trunk tests in a branch by executing

  svn switch .../svn/mpfr/trunk/tests tests

from the working copy. One can know when the tests directory has been
switched, thanks to

  $ svn status
      S  tests

In case of failure, freeing the memory explicitly is not necessary.
We do this in case of success just to be able to detect memory leaks
in MPFR.

   ====================================================================

If you have to mix TMP_DECL and MPFR_SAVE_EXPO_DECL in the declaring
section of your function, please declare MPFR_SAVE_EXPO_DECL before
TMP_DECL, since TMP_DECL may be replace by nothing:

Instead of:                                 Usually preprocessed as:
  unsigned long t                             unsigned long t;
  TMP_DECL (maker);                           ;
  MPFR_SAVE_EXPO_DECL (expo);                 mpfr_save_expo_t expo;
use:
  unsigned long t                             unsigned long t;
  MPFR_SAVE_EXPO_DECL (expo);                 mpfr_save_expo_t expo;
  TMP_DECL (maker);                           ;

   ====================================================================

Do not use TMP_DECL / TMP_ALLOC, ... but MPFR_TMP_DECL, MPFR_TMP_ALLOC, ...

   ====================================================================

Do not use C99-only features, such as empty macro arguments or C++-style
comments.

   ====================================================================

When testing a "boolean" macro M (i.e. which is normally either equal
to 1 or undefined), do not use #if M, but #ifdef M or #if defined(M).
With icc, the form #if M triggers a warning ("remark #193: zero used
for undefined preprocessing identifier").

===========================================================================

If wou want to use the logging of MPFR, you need to enable it:
 ./configure --enable-logging
 make clean
 make
Then link your program with this new build of MPFR.

Warning! The logging code for functions sometimes output an "inexact"
value, but in case of exception, this value may be meaningless. In
fact, the output value is the value of some variable; please check
the source code of the function to understand its real meaning.

You can control what is logged using the environment variables:

MPFR_LOG_FILE:  Name of the LOG file (default: mpfr.log).
MPFR_LOG_BASE:  Base of the outputs (default: 10).
MPFR_LOG_PREC:  # of digits of the outputs (default: 0, ie. maximum).
MPFR_LOG_LEVEL: Max recursive level (default: 7).

MPFR_LOG_INPUT:    Log the inputs
MPFR_LOG_OUTPUT:   Log the outputs
MPFR_LOG_TIME:     Log the time spent inside the function.
MPFR_LOG_INTERNAL: Log the intermediary variables if any.
MPFR_LOG_MSG:      Log the messages sent by MPFR if any.
MPFR_LOG_ZIV:      Log what the Ziv Loops do.
MPFR_LOG_STAT:     Log how many times Ziv failed.
MPFR_LOG_ALL:      Log everything

Define them. Run your program, and view `mpfr.log`.

   ====================================================================

This feature is available only for gcc >= 3.0 and glibc >= 2.0.
To achieve this, theses macros have been added:

+++ MPFR_LOG_VAR(y)
 Log a MPFR variable if requested (INTERNAL).
 Example:
   mpfr_t y;
   MPFR_LOG_VAR (y);

+++ MPFR_LOG_MSG(x)
 Log another message (a warning for example)
Example:
 MPFR_LOG_MSG (("WARNING: Unchecked code\n", 0));
The 0 is here a dummy value, because there must be at least an argument
after the format string.

+++ MPFR_LOG_BEGIN(x)
 Add this macro at the beginning of a function.
Example:
 int dodo (mpfr_t x, mpfr_t op, int cnt, mp_rnd_t rnd) {
  [decl]
  MPFR_LOG_BEGIN (("op[%#R]=%R rnd=%s", op, op, RND2STR(rnd)));

+++ MPFR_LOG_END(x)
 Add this macro at the end of a function.
Example:
 MPFR_LOG_END (("x[%#R]=%R i=%d", x, x, i));
 return i;
}

+++ MPFR_LOG_FUNC (begin,end)
  Add this macro at the beginning of a function. It does
the same job as MPFR_LOG_BEGIN and MPFR_LOG_END but it is smatter
since it intercepts the return itself to put the end statement.
Example
 MPFR_LOG_FUNC (("op[%#R]=%R rnd=%d", op, op"),
                ("x[%#R]=%R inexact=%d", x, x, i));


The double brackets "((" and "))" are needed since MPFR must still
compile with non GNU compiler, so Macros with variable # of args
are not allowed.

It registers a glibc printf extension %R to display a mpfr_t.
%#R is used to display the precision of a mpfr_t.
It uses some extended attributes of GCC (constructor, etc.) to achieve
its goals too.

===========================================================================
ZivLoop Controler

Ziv strategy is quite used in MPFR. In order to factorize the code, you
could use theses macros:

+++ MPFR_ZIV_DECL(_x)
 Declare a ZivLoop controller

+++ MPFR_ZIV_INIT(_x, _prec)
 Init a ZivLoop controller according to the initial value of _prec.

+++ MPFR_ZIV_NEXT(_x, _prec)
 Increase the precision _prec according to the ZivLoop controller.

+++ MPFR_ZIV_FREE(_x)
 Free the ZivLoop controller.

===========================================================================

If you plan to add a new function, you could follow this schema:

int
mpfr_toto (mpfr_ptr rop, mpfr_srcptr op, mp_rnd_t rnd)
{
  [Declare all used variables]
  int inexact;
  mp_prec_t prec;
  MPFR_ZIV_DECL (loop);
  MPFR_SAVE_EXPO_DECL (expo);

  /* Log it if requested */
  MPFR_LOG_BEGIN (("op[%#R]=%R rnd=%d", op, op, rnd),
                  ("rop[%#R]=%R inexact=%d", rop, rop, inexact));

  /* First deal with particular cases */
  if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (op)))
    {
      if (MPFR_IS_NAN (op))
        {
         MPFR_SET_NAN (rop);
         MPFR_RET_NAN;
        }
      else if (MPFR_IS_INF (op))
        {
         [Code to deal with Infinity]
        }
      else
        {
          MPFR_ASSERTD (MPFR_IS_ZERO (op));
          [Code to deal with Zero]
        }
    }
  [Other particular case: For example, op<0 or op == 1]

  [Compute the first estimation of the used precision `prec`]
  [Initialize the intermediate variables using mpfr_init2]
  MPFR_SAVE_EXPO_MARK (expo);            /* Maximal range for exponent */

  MPFR_ZIV_INIT (loop, prec);            /* Initialize the ZivLoop controler */
  for (;;)                               /* Infinite loop */
    {
      [Compute an estimation of the function and]
      [an estimate of the error.]
      if (MPFR_CAN_ROUND (...))          /* If we can round, quit the loop */
        break;
      MPFR_ZIV_NEXT (loop, prec);        /* Increase used precision */
      [Use `mpfr_set_prec` to resize all needed intermediate variables]
    }
  MPFR_ZIV_FREE (loop);                  /* Free the ZivLoop Controler */

  inexact = mpfr_set (rop, temp, rnd);   /* Set rop to the computed value */
  [Clear all intermediate variables]

  MPFR_SAVE_EXPO_FREE (expo);            /* Restore exponent range */
  return mpfr_check_range (rop, inexact, rnd);  /* Check range and quit */
}

Make sure that Ziv loops cannot increase the precision forever because of
internal exception. Otherwise one gets either a segmentation fault (with
limited stack size) or an assertion failure (with unlimited stack size,
e.g. with "make check").

Exception handling (overflow/underflow in particular):
  * Warning: To detect exceptions and/or possible error loss due to
    internal exceptions, testing whether some variable is singular with
    MPFR_IS_SINGULAR is generally not sufficient! Indeed, in case of
    overflow (resp. underflow), the value may be rounded (in absolute
    value) to the largest finite number (resp. to the smallest non-zero
    number, possible even in round-to-nearest mode).
  * The MPFR_BLOCK* macros can be useful, e.g.
      {
        MPFR_BLOCK_DECL (flags);
        /* ... */
        MPFR_BLOCK (flags, /* expression or statements */)
        /* ... */
        if (MPFR_OVERFLOW (flags))
          { /* case of overflow in expression or statements */ }
        /* ... */
      }
    See mpfr-impl.h (search for MPFR_BLOCK) for more information.

===========================================================================

If you plan to add a new threshold in MPFR which could be tuned,
you should add its default value in the file `mparam_h.in'. When the
script configure finishes, it creates the file `mparam.h' from `mparam_h.in'.

Then you needs to modify the program `tuneup.c' to allow it to compute
the new threshold. If it is a classical threshold (not complex), you could
use this method (example of mpfr_exp):

/* Define the threshold as a variable instead of a constant */
mp_prec_t mpfr_exp_threshold;
#undef  MPFR_EXP_THRESHOLD
#define MPFR_EXP_THRESHOLD mpfr_exp_threshold
/* Include the test function to threshold directly in the test
   program. It will overide the mpfr_exp coming from libmpfr.a */
#include "exp.c"
/* Define the speed function related to mpfr_exp */
static double speed_mpfr_exp (struct speed_params *s) {
  SPEED_MPFR_FUNC (mpfr_exp);
}

Then in the function `all', you will have to call the tune function,
and write the new THRESHOLD in the file `mparam.h':

  /* Tune mpfr_exp */
  if (verbose)
    printf ("Tuning mpfr_exp...\n");
  tune_simple_func (&mpfr_exp_threshold, speed_mpfr_exp);
  fprintf (f, "#define MPFR_EXP_THRESHOLD %lu\n",
           (unsigned long) mpfr_exp_threshold);

More complex tuning is possible but needs special attention.

===========================================================================

Bit Twiddling Hacks - Sean Eron Anderson maintain a list of tricks to get
efficient code on <http://graphics.stanford.edu/~seander/bithacks.html>.
WARNING: some of those tricks may not take into account possible overflows,
and may not be portable.

===========================================================================

MPFR manual (mpfr.texi):
  * Use "significand", not "mantissa".
  * Use "@minus{}" for the minus character, not "-".
  * Follow the English typography, not the French one!

===========================================================================

Running "make" outputs a lot of information, and warnings are not very
visible. The following tool "eet" allows a copy of warning messages to
be output to a different window (e.g. xterm or zenity):

  http://www.vinc17.org/unix/#eet

Direct link to the tarball: http://www.vinc17.org/unix/eet.tar.bz2

===========================================================================

Be careful when avoiding "'var' may be used uninitialized in this function"
warnings from gcc. Initializing such variables to a dummy value has several
drawbacks:
  * this may prevent other tools (that do static or dynamic analysis) from
    detecting bugs;
  * this makes code maintenance more difficult (e.g. when modifying the
    code, one may more easily forget a real initialization);
  * this makes the compiler add useless code (though this should not be
    significant).

The INITIALIZED macro can be used to avoid such warnings with gcc, e.g.

  int INITIALIZED(i);

It uses the "int i = i;" pseudo-initialization trick, disabled with other
compilers as this is undefined behavior. See:

    http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36296

If a dummy initialization must be added, use preferably an "invalid" value
(e.g. NULL for pointers, or a value that can be checked with MPFR_ASSERTN
before using it) that could make the program abort instead of returning an
incorrect value in case of a bug in MPFR.

===========================================================================

Avoid mixing signed and unsigned types, in particular mp_exp_t, which is
signed, and mpfr_prec_t (mp_prec_t), which is unsigned, as this can lead
signed types to be converted into unsigned types. If such a signed type
contains a negative value, the result will probably be incorrect. In
general, you will need to cast a mpfr_prec_t into a mp_exp_t.

Note that such bugs are difficult to detect because they may depend on
the platform (e.g., on LP64, 32-bit unsigned int + 64-bit long is OK,
but on ILP32, 32-bit int + 32-bit unsigned long is incorrect), but also
on the input values. So, do not rely on tests very much. However, if
a test works on 32 bits but fails on 64 bits in the extended exponent
range (or conversely), the cause may be related to the integer types
(e.g. a signness problem or an integer overflow due to different type
sizes).

===========================================================================

Because of a bug in the Mac OS X 10.5 linker, avoid tentative definitions
(C99, 6.9.2). Depending on the context, use either a simple declaration
(with the "extern" storage-class specifier) or an external definition.
This is also cleaner.