INSTALL - master - githubmirror/mpfr

INSTALL
768 строк33 KB
Vincent Lefevre
Corrected documentation about the failing check_null() tests (bug in GMP)
14 мар 2025, 20:47
14 мар 2025, 20:47888d7be
Copyright 1999-2025 Free Software Foundation, Inc.
Contributed by the Pascaline and Caramba projects, INRIA.

This file is part of the GNU MPFR Library.

The GNU MPFR Library is free software; you can redistribute it and/or modify
it under the terms of the GNU Lesser General Public License as published by
the Free Software Foundation; either version 3 of the License, or (at your
option) any later version.

The GNU MPFR Library is distributed in the hope that it will be useful, but
WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
License for more details.

You should have received a copy of the GNU Lesser General Public License
along with the GNU MPFR Library; see the file COPYING.LESSER.
If not, see <https://www.gnu.org/licenses/>.



                        Installing GNU MPFR
                        ===================

Note: In case of problem, please read this INSTALL file carefully before
reporting a bug, in particular Section "In case of problem" below. Some
problems are due to bad configuration on the user side (not specific to
MPFR).

0. You first need to install GMP. See <https://gmplib.org/>.
   MPFR requires GMP version 5.0.0 or later.
   You need a C compiler, preferably GCC, but any reasonable compiler should
   work (C++ compilers should work too, under the condition that they do not
   break type punning via union).
   And you need the standard Unix "make" command, plus some other standard
   Unix utility commands.

1. Extract the files from the archive.

2. It is strongly advised to apply the latest patches if this has
   not been done yet and if patches are available. You can check
   on the release page:

     https://www.mpfr.org/mpfr-4.3.0/

   which may have additional information. The patches can be applied
   with commands like:

     wget --no-config https://www.mpfr.org/mpfr-4.3.0/allpatches
     patch -N -Z -p1 < allpatches

   or

     curl https://www.mpfr.org/mpfr-4.3.0/allpatches | patch -N -Z -p1

   (Those instructions are for the GNU patch command, for example
   /usr/bin/gpatch on Solaris.)

3. In the MPFR directory, to detect your system, type:
     ./configure
   possibly with options (see below, in particular if this step or
   one of the following fails). You should also check whether WARNING
   lines have been output (such a problem may cause a failure in one
   of the following steps).
   Note: paths provided in configure options must always be absolute
   (relative paths are not supported).

4. To build the library, type:
     make

   [optional] if you want to tune MPFR for your specific architecture, see
   the section "Tuning MPFR" below. Note that for most common architectures,
   MPFR includes some default tuning parameters which should be near from
   optimal.

5. To check the built library (runs the test files), type:
     make check

   Note: If any test fails, information about this failure can be found in
   the tests/test-suite.log file; you should provide this file in your bug
   reports (in addition to other useful information, as mentioned later).
   If you want the contents of this file to be automatically output in case
   of failure, you can set the VERBOSE environment variable to 1 before
   running "make check", for instance by typing:
     VERBOSE=1 make check

6. [Optional / experimental] Binary distributions may also want to run:
     make check-gmp-symbols

   This will check that MPFR does not use GMP internal symbols, which
   could yield failures in case of GMP upgrade without a MPFR rebuild.
   But note that this is a heuristic and might give false positives or
   false negatives. Please report any problem to the MPFR developers.
   End users may also be interested in this check, unless they have
   allowed GMP internals with configure options (see below).

7. To install it (default "/usr/local" | see "--prefix" option), type:
     make install

If you installed MPFR (header and library) in directories that are
not searched by default by the compiler and/or linking tools, then,
like with other libraries, you may need to set up some environment
variables such as C_INCLUDE_PATH (to find the header mpfr.h),
LIBRARY_PATH (to find the library), and if the shared library has
been installed, LD_LIBRARY_PATH (before execution) or LD_RUN_PATH
(before linking); this list is not exhaustive and some environment
variables may be specific to your system. "make install" gives some
instructions; please read them. You can also find more information
in the manuals of your compiler and linker. The MPFR FAQ may also
give some information.

Remember that if you have several MPFR (or GMP) versions installed
(e.g., one with the system, and one, newer, by you), you will not
necessarily get a compilation/linking error if a wrong library is
used (e.g., because LD_LIBRARY_PATH has not been set correctly).
But unexpected results may occur.

Under Mac OS X, if the shared library was not installed and you use
Apple's linker (this is the default), you will also need to provide
the -search_paths_first linker flag ("-Wl,-search_paths_first" when
you link via gcc) to make sure that the right library is selected,
as by default, Apple's linker selects a shared library preferably,
even when it is farther in the library paths. We recall that if a
wrong library is selected due to this behavior, unexpected results
may occur.


Building the documentation
==========================

To build the documentation in various formats, you may first need to
install recent versions of some utilities such as texinfo.

* Type "make info" to produce the documentation in the info format.

* Type "make pdf" to produce the documentation in the PDF format.

* Type "make dvi" to produce the documentation in the DVI format.

* Type "make ps" to produce the documentation in the PostScript format.

* Type "make html" to produce the documentation in the HTML format
  (in several pages); if you want only one output HTML file, then
  type "makeinfo --html --no-split mpfr.texi" from the doc directory
  instead.


Building MPFR with internal GMP header files
============================================

MPFR built with internal GMP header files is a bit faster, so you may want
to build it with them. Just do this in step 1:

  ./configure --with-gmp-build=GMPBUILD

where GMPBUILD is the GMP build directory. The needed header files are:
gmp-impl.h, longlong.h and all the necessary headers to use them, which
may be located either in the GMP source directory or in the GMP build
directory, in case they are different (MPFR takes care of that, as of
MPFR 3.1.0).

Warning: the library obtained in this way may use some internal GMP
symbols, and thus dynamically linking your software with a different
version of GMP might fail, even though it is declared as compatible
by Libtool's versioning system.


Tuning MPFR
===========

[For the current GMP version (6.1.0), a Unix-like OS is required.]

For this, you need to build MPFR with a GMP build directory (see above).
In the GMP build directory, you also need to go into the "tune" subdirectory
and type "make speed". This will build the GMP speed library, which is used
by the MPFR tuning mechanism.

Then go back to the MPFR build directory, go into the "tune" subdirectory and
type "make tune". This will build an optimized file "mparam.h" for your
specific architecture.


./configure options
===================

--prefix=DIR            installs MPFR headers and library in DIR/include and
                        DIR/lib respectively (the default is "/usr/local").

--with-gmp-include=DIR  assumes that DIR contains gmp.h
--with-gmp-lib=DIR      assumes that DIR contains the GMP library
--with-gmp=DIR          assumes that DIR is where you have installed GMP.
                        same as --with-gmp-lib=DIR/lib
                        and     --with-gmp-include=DIR/include
                        (use either --with-gmp alone or one or both of
                        --with-gmp-lib/--with-gmp-include)
                        Warning! Do not use these options if you have
                        CPPFLAGS and/or LDFLAGS containing a -I or -L
                        option with a directory that contains a GMP
                        header or library file, as these options just
                        add -I and -L options to CPPFLAGS and LDFLAGS
                        *after* the ones that are currently declared,
                        so that DIR will have a lower precedence. Also,
                        this may not work if DIR is a system directory
                        (typically /usr or /usr/local); see below.

--with-gmp-build=DIR    assumes that DIR contains the GMP build directory,
                        and enables the use of GMP internals (see above).
                        Warning! This option and the group of options
                        --with-gmp are mutually exclusive.

--enable-assert         build MPFR with assertions.

--enable-thread-safe    build MPFR as thread safe, using compiler-level
                        Thread-Local Storage (TLS). Note: TLS support is
                        roughly tested by configure. If configure detects
                        that TLS does not work (because of the compiler,
                        linker or system libraries), it will output an
                        error message, telling you to build MPFR without
                        thread safe. For instance, though Mac OS X uses
                        GCC, it may not currently support GCC's __thread
                        storage class.

--disable-thread-safe   build MPFR without TLS. By default, TLS support
                        is detected automatically, and MPFR is built as
                        thread safe if supported. However, this detection
                        is only a heuristic: TLS can be detected as
                        supported while its support is incomplete or
                        buggy (MPFR tests may fail). In such a case,
                        this option is useful.

--enable-decimal-float  build conversion functions from/to decimal floats.
                        Note that detection by the configure script is
                        limited in case of cross-compilation.
                        Accepted arguments:
                   yes  Decimal support is requested and the configure
                        script fails if it detects that decimals do not
                        work.
                        The encoding (BID or DPD) will automatically be
                        detected at configure time or at compile time if
                        possible (if not, generic code will be used).
                    no  Decimal support is explicitly disabled.
                  auto  Decimal support is enabled if the configure script
                        detects that it works. This is the default when
                        --{enable,disable}-decimal-float is not given.
                   bid  Decimal support is requested and the encoding is
                        assumed to be BID (some check may be done).
                   dpd  Decimal support is requested and the encoding is
                        assumed to be DPD (some check may be done).
               generic  Decimal support is requested and the generic code
                        is used (mainly for developers).

--enable-gmp-internals  allows the MPFR build to use GMP's undocumented
                        functions (not from the public API). Note that
                        library versioning is not guaranteed to work if
                        this option is used. Thus it must not be used in
                        binary distributions.

--with-sysroot=DIR      Search for dependent libraries within DIR (which
                        may be useful in cross-compilation). If you use
                        this option, you need to have Libtool 2.4+ on
                        the target system. See Libtool 2.4+'s NEWS file.

Note: By default, the configure script tries to set CC / CFLAGS to GMP's
ones from gmp.h (__GMP_CC / __GMP_CFLAGS) in order to ensure that MPFR is
built with the same ABI as GMP. The reason is that when GMP is built, it
may set CC / CFLAGS to select an ABI that is not the default one in order
to have a better performance. The -pedantic option in GMP's CFLAGS, when
present (which is the case by default), is removed, because the MPFR
build system uses some C extensions (when this script detects that they
are supported) and -pedantic yields too many useless warnings. However,
this setting from GMP is not guaranteed to work as the configure script
does some compiler tests earlier, and a conflict may arise. Also, the
values obtained from GMP may be incorrect for the MPFR build if GMP has
been built on a different machine; in such a case, the user may need to
specify CC / CFLAGS, as explained below.

Moreover, even without --with-gmp-build and --enable-gmp-internals,
MPFR might use some GMP internals by mistake. This would be a bug,
which should be reported to the MPFR developers.

Run "./configure --help" to see the other options (default options
from Autoconf and Automake).


If 'gmp.h' and 'libgmp' do not match
====================================

Under some circumstances, the configure script may output a message
saying:

  'gmp.h' and 'libgmp' seem to have different versions or
  we cannot run a program linked with GMP (if you cannot
  see the version numbers above). [...]

Even though the configure script does not fail in such a case, this
message most often indicates a real error, which needs to be avoided.
Possible causes are:

* The wanted GMP library does not have the same ABI as the one chosen
  to build MPFR. The consequences may be:

    - A different GMP library (with the correct ABI) has been found,
      but does not have the same version as 'gmp.h'.

    - No other GMP libraries have been found (in this case, no version
      numbers should be printed above the warning message).

  This is incorrect and one of the following steps (make, make check)
  will probably fail. GMP (actually gmp.h) now provides CC and CFLAGS
  information to select the correct ABI, so that this problem should
  no longer occur; but if GMP was provided by a binary package, such
  information may not be valid. See the

    checking for CC and CFLAGS in gmp.h...

  line in the configure output (about the 11th line) and the following
  few ones for more information. You may need to reinstall GMP or to
  provide your own CC and/or CFLAGS. See also the remaining of this
  INSTALL file.

* A configure option like --with-gmp or --with-gmp-include was used
  with a system include directory, e.g. one of the following:

    --with-gmp=/usr
    --with-gmp=/usr/local
    --with-gmp-include=/usr/include
    --with-gmp-include=/usr/local/include

  GCC (and possibly other compilers) will ignore such a directive for
  include directories (but this rule is not applied for the library
  itself!). This means that the library search paths will be reordered
  as declared, but the specified include directory will still be near
  the end of the include search paths (thus with a low precedence).
  This is not a problem if only one GMP version is installed, but
  otherwise, a wrong gmp.h may be chosen, so that the versions of
  gmp.h and libgmp may not match. The suggestions are the following:

    - If you want to use the GMP version under /usr, then you should
      uninstall all the other GMP versions (header and library files)
      that may be seen in the search paths, in particular those under
      /usr/local.

    - If you want to use the GMP version under /usr/local, then you
      should uninstall all the other GMP versions (header and library
      files) that may be seen in the search paths, but *NOT* the one
      under /usr (the version under /usr is provided by the OS vendor,
      and changing/removing anything related to it may break your
      system, and /usr should have a lower precedence than /usr/local
      anyway).

  To find where GMP versions have been installed:

    $ locate libgmp  (if you have a locate database)

  and if the compiler is GCC:

    $ gcc -print-file-name=libgmp.so     (under most systems)
    $ gcc -print-file-name=libgmp.dylib  (under Mac OS X)

  and if this does not work, you may try:

    $ gcc -print-search-dirs

* An official GCC version was used under Debian GNU/Linux. Problems
  may come from the fact that Debian chose a different convention
  for library search paths concerning 32-bit and 64-bit libraries.
  A possible problem can be, for instance:

  [Debian's GCC, correct library path]
  $ gcc -print-file-name=libgmp.so
  /home/vlefevre/gmp/athlon64/lib/../lib/libgmp.so

  [Official GCC, incorrect library path]
  $ gcc-4.3.1 -print-file-name=libgmp.so
  /usr/lib/../lib64/libgmp.so

  The solution: use a GCC provided by Debian or add symbolic links
  such as lib64 -> lib (on 64-bit machines) for your library paths.

* The problem may also be temporary and only due to the fact that
  libtool could not be used at this time. This is unlikely, though.


In case of problem
==================

First, look for any warning message in the configure output.

Several documents may help you to solve the problem:
  * this INSTALL file, in particular information given below;
  * the FAQ (either the FAQ.html file distributed with MPFR, or the
    on-line version <https://www.mpfr.org/faq.html>, which may be more
    up-to-date);
  * the MPFR web page for this version <https://www.mpfr.org/mpfr-4.3.0/>,
    which lists bugs found in this version and provides some patches.

If the "configure" fails, please check that the C compiler and its
options are the same as those for the GMP build (specially the ABI).
You can see the latter with the following command:

  grep "^CC\|^CFLAGS" GMPBUILD/Makefile

if the GMP build directory is available. Then type:

  ./configure <configure options> CC=<C compiler> CFLAGS="<compiler options>"

(quotes are needed when there are spaces or other special characters
in the CC/CFLAGS value) and continue the install. On some platforms,
you should provide further options to match those used by GMP, or set
some environment variables. For instance, see the "Notes on AIX/PowerPC"
section below.

Warning! Do NOT use optimization options that can change the semantics
of math operations, such as GCC's -ffast-math or Sun CC's -fast.
Otherwise conversions from/to double's may be incorrect on infinities,
NaN's and signed zeros. Since native FP arithmetic is used in a few
places only, such options would not make MPFR faster anyway.

If some "make check" tests fail, you can try the --disable-thread-safe
configure option (see the configure options above): it has been reported
that some platforms have buggy TLS support. Before trying this option,
you may want to check in the configure output whether MPFR was built
with TLS support; if yes, you will have a line:
  checking for TLS support... yes
Alternatively "grep MPFR_USE_THREAD_SAFE config.log" will show that
MPFR_USE_THREAD_SAFE is defined to 1. If it is "no" (or the variable
is not defined), the --disable-thread-safe option would be useless.

Some tests failure may be due to other compiler bugs, in particular
in optimization code. You can try to build MPFR without compiler
optimizations by giving -O0 (letter O, digit 0) in CFLAGS. If the
MPFR tests no longer fail, this was probably due to a compiler bug,
though we cannot exclude a bug in MPFR. You may want to contact us
(see below), possibly after looking at:
  https://members.loria.fr/PZimmermann/software/compilerbugs.html

On some platforms, try with "gmake" (GNU make) instead of "make".
Problems have been reported with the Tru64 make.

If the configure script reports that gmp.h version and libgmp version
are different, or if the build was OK, but the tests failed to link
with GMP or gave an error like

  undefined reference to `__gmp_get_memory_functions'

meaning that the GMP library was not found or a wrong GMP library was
selected by the linker, then your library search paths are probably
not correctly set (some paths are missing or they are specified in an
incorrect order).

Such problems commonly occur under some GNU/Linux machines, where the
default header and library search paths may be inconsistent: GCC is
configured to search /usr/local/include and /usr/local/lib by default,
while the dynamic linker ignores /usr/local/lib. If you have a GMP
version installed in /usr (provided by the OS vendor) and a new one
installed in /usr/local, then the header of the new GMP version and
the library of the old GMP version will be used! The best solution
is to make sure that the dynamic linker configuration is consistent
with GCC's behavior, for instance by having /usr/local/lib in
/etc/ld.so.conf or in some file from /etc/ld.so.conf.d (as Debian
did: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=395177). See
also https://gcc.gnu.org/legacy-ml/gcc-help/2010-01/msg00171.html for
more information. Alternatively you can use:
  * environment variables. This may sometimes be necessary. If DIR
    is the installation directory of GMP, add DIR/include to your
    CPATH or C_INCLUDE_PATH (for compilers other than GCC, please
    check the manual of your compiler), and add DIR/lib to your
    LIBRARY_PATH and LD_LIBRARY_PATH (and/or LD_RUN_PATH);
  * --with-gmp* configure options (described above), e.g.
    --with-gmp=/opt/local (to use /opt/local/include for headers and
    /opt/local/lib for libraries), but other software that uses GMP
    and/or MPFR will need correct paths too, and environment variables
    allow one to set such search paths in a global way.
    Note about "--with-gmp=/usr/local". This option may appear to
    solve the above inconsistency problem, but does not work as you
    expect. Indeed it affects the library search path, in particular,
    the one used by the dynamic linker (thus adding the missing
    /usr/local/lib directory as wanted), but since /usr/local/include
    is a "standard system include directory" for GCC, the include
    search patch is not changed; this is often not a problem in this
    particular case because usually, /usr/local/include is already
    last in the include search patch, but this may fail under some
    occasions and may trigger obscure errors.

For instance, under Unix, where paths are separated by a colon:

  * With POSIX sh-compatible shells (e.g. sh, ksh, bash, zsh):
    export C_INCLUDE_PATH="/usr/local/include:/other/path/include"
    export LIBRARY_PATH="/usr/local/lib:/other/path/lib"
    export LD_LIBRARY_PATH="$LIBRARY_PATH"

  * With csh or tcsh:
    setenv C_INCLUDE_PATH "/usr/local/include:/other/path/include"
    setenv LIBRARY_PATH "/usr/local/lib:/other/path/lib"
    setenv LD_LIBRARY_PATH "$LIBRARY_PATH"

If almost all the tests fail and the messages in the test-suite.log file
(or in the output, when running individual tests from the command line)
start with a line of the form:

  Incorrect MPFR version! (xxx header vs yyy library)

then this means that an installed MPFR version is tested instead of the
one that has just been built. This is probably not a bug in MPFR, but a
problem caused by the user or system configuration (particular options,
environment variables, etc.) or a bug in the toolchain. In particular,
if LD_LIBRARY_PATH overrides the run path (set up by libtool) and an
installed ABI-compatible version of MPFR is in a directory listed in
the LD_LIBRARY_PATH search path, then this will break. An example with
GNU ld:

  https://sourceware.org/bugzilla/show_bug.cgi?id=21476

If you can't solve your problem, you should contact us via the MPFR
mailing-list <mpfr@inria.fr>, indicating the machine and operating system
used (uname -a), the compiler and version used (gcc -v if you use gcc),
the configure options used if any (including variables such as CC and
CFLAGS), the version of GMP and MPFR used, and a description of the
problem encountered, in particular the tests/test-suite.log file if
"make check" failed. Please send us also the log of the "configure"
(config.log).

tfprintf and tsprintf failures in check_null
--------------------------------------------

If you get the failures

$ ./tfprintf
Error in test #40: mpfr_vfprintf printed 1 characters instead of 4

$ ./tsprintf
Error in check_null, i = 0
expected r = 1, s = { 0, 0, 1 }
got      r = 0, s = { 0, 1, 1 }

the cause is probably that GMP uses a buggy vsnprintf replacement.
The reason GMP uses its own vsnprintf replacement is that during
configure, it detected that the vsnprintf function from the C library
is buggy. But the replacement from GMP 6.3.0 and below also has its own
bugs. This means that using the formatted output functions with %c and
the value 0 will yield an erratic behavior.

For the above issue, this is fixed in the GMP development branch:

  https://gmplib.org/repo/gmp/rev/4ac76064639e

You can have details at
  https://gmplib.org/list-archives/gmp-bugs/2023-December/005420.html
  https://gitlab.inria.fr/mpfr/mpfr/-/commit/390e51ef85

These tests can be skipped by defining the MPFR_TESTS_SKIP_CHECK_NULL macro,
but you should rather get a fixed vsnprintf replacement.


Notes about ABI
===============

On 64-bit computers, it may happen that the default ABI (Application Binary
Interface) chosen by MPFR does not correspond to the default one chosen by
the compiler.

In particular, this kind of message may indicate the problem:
/usr/bin/ld: skipping incompatible mpfr/src/.libs/libmpfr.a when searching for -lmpfr

In fact, since MPFR relies on GMP, it uses the same ABI as GMP.
To be sure that your program compiles and links correctly, use the same
compiler flags as MPFR does (look for CFLAGS in config.log).

You might also recompile GMP with a different ABI, with for example
./configure ABI=32.


Notes on Mac OS X
=================

If you get an error of the form

  ld: pointer in read-only segment not allowed in slidable image...

this can mean that the link is done against a static (GMP) library.
In such a case, you should configure MPFR with --disable-shared to
disable the build of the shared library.


Notes on FreeBSD 4.3
====================

FreeBSD 4.3 is provided with an incorrect <float.h> header file, and
MPFR tests related to long double's may fail. If you cannot upgrade
the system, you can still use MPFR with FreeBSD 4.3, but you should
not use conversions with the long double type.


Notes on AIX/PowerPC
====================

The following has been tested on AIX 7.1.3 (gcc111.fsffrance.org)
with gcc 4.8.1 and GMP 6.1.0.

If GMP was built with the 64-bit ABI, before building and testing MPFR,
you should set the OBJECT_MODE environment variable to 64, e.g., with:

  export OBJECT_MODE=64

(in a sh-compatible shell). Alternatively add the following to the configure
line: AR="ar -X64" NM="nm -B -X64".


Notes on Solaris
================

Do not put a -R<dir> option in the LD_OPTIONS environment variable, at least
if the directory <dir> can contain an MPFR library. Otherwise this MPFR
library may be chosen for the tests (make check) instead of the one that has
just been built, in which case, either you will get errors due to unmatched
versions or this problem may remain undetected. The reason is that this
option will appear before the -R options added by libtool, such as the one
to the src/.libs directory containing the MPFR library that has just been
built, and will have the precedence.


Notes for the CompCert compiler
===============================

[Tested with CompCert 3.10 and MPFR master-11992-f75b0c388 on x86_64-linux]

CompCert (https://compcert.org/) is a formally verified compiler.
To compile MPFR with CompCert:

$ ./configure --disable-shared CC=ccomp CFLAGS="-flongdouble -fstruct-passing"

You also need to unset LD_LIBRARY_PATH, and/or you might need to change
wl="" into wl="-Wl," in the libtool file (after running configure).

All tests (make check) should pass (tget_set_d64, tget_set_d128 and
tset_float128 are skipped, since CompCert does not support _Decimal64,
_Decimal128 nor _Float128).


Notes for the Intel OneApi compiler
===================================

When using the Intel OneApi compiler (icx), one should add -fp-model=strict
to CFLAGS so that the conversion routines from/to native floating-point
types (float, double, ...) work properly. Otherwise some tests will fail.


MPFR for use with Windows Applications
======================================

There are several ways of building MPFR on Windows, the most appropriate
approach depending on how you intend to use the resulting libraries.

a. Using MinGW
==============

1 - We advise to use Mingw-w64 (https://www.mingw-w64.org/), which is
    simpler and less demanding than Cygwin. Contrary to Cygwin, it also
    provides native Windows code.

    If you also use MSYS, you should use "make" for MSYS instead of
    the "make" utility from Mingw-w64 or from GCC, which causes the
    following error:

      libtool: warning: libobj name 'extract.Tpo -c -o extract.lo extract.lo'
      may not contain shell special characters.
      rm: unknown option -- c

    References about this issue and solution:
      https://sourceforge.net/p/msys2/tickets/223/
      https://sympa.inria.fr/sympa/arc/mpfr/2016-07/msg00009.html

2 - If you just want to make a binary with gcc, there is nothing to do:
    GMP, MPFR and the program compile exactly as under Linux. (It is
    recommended to pass --build=xxx-yyy-mingw64 to the GMP configure command,
    or --build=xxx with xxx containing mingw.)

3 - To avoid using the Microsoft runtime (which might not be conform to ISO C),
    you can use the MinGW runtime package (which is an integral part of MinGW).

    With MinGW version v8.0.0 and later, the formatted output functions
    (printf, etc.) are ISO/POSIX-conforming by default; however, this is
    no longer true if -std=c89 is used at build time. Conversely, with
    earlier MinGW versions, it is possible to get conforming functions
    with either '-ansi', '-posix' or '-D__USE_MINGW_ANSI_STDIO'. Note
    that if there is a conformity mismatch between the options used for
    the GMP build (from which the MPFR build gets the output) and those
    used for the MPFR tests, the tsprintf test may fail with one of the
    errors below. Be careful that a non-conforming output may yield a
    buffer overflow.

    The following applies to old MinGW versions, and may be discouraged
    with recent MinGW versions.

    In order to have the MPFR formatted output functions based on an
    ISO-conforming printf(), you need to compile GMP (not MPFR) with
    CC="gcc -D__USE_MINGW_ANSI_STDIO" (since the standard printf modifiers
    %e, %Ld and %td are passed to GMP).

    To avoid failures of some of the printf-related tests, MPFR needs to
    be compiled with the same __USE_MINGW_ANSI_STDIO as with GMP, i.e.
    this macro should be defined for both or undefined for both; this
    should be the case by default, unless CC or CFLAGS has been redefined.
    For instance, if __USE_MINGW_ANSI_STDIO is defined in GMP but not in
    MPFR, the following error may occur:

      Error in mpfr_vsprintf (s, "%e", ...);
      expected: "-1.250000e+000"
      got:      "-1.250000e+00"
      FAIL tsprintf.exe (exit status: 1)

    and in the opposite case:

      Error in mpfr_vsprintf (s, "%e", ...);
      expected: "-1.250000e+00"
      got:      "-1.250000e+000"
      FAIL tsprintf.exe (exit status: 1)

    Note with old GMP versions: Other issues could arise due to the fact
    that the C functions vsnprintf and vsprintf both used internally in
    old GMP versions do not produce the same output:

      https://sympa.inria.fr/sympa/arc/mpfr/2016-03/msg00045.html
      https://sympa.inria.fr/sympa/arc/mpfr/2016-03/msg00051.html
      https://sympa.inria.fr/sympa/arc/mpfr/2016-03/msg00053.html

    If __USE_MINGW_ANSI_STDIO has not been defined when building GMP,
    then the length modifiers j, L and t are not supported with the GMP
    formatted output functions, and as a consequence, also with MPFR.
    This is automatically detected by the configure script, except when
    cross-compiling (e.g. under Linux), in which case some macros need
    to be defined explicitly, e.g. with

      "CPPFLAGS=-DNPRINTF_J -DNPRINTF_L -DNPRINTF_T"

    in order to avoid potential issues with the MPFR library and failures
    in the test suite (the corresponding tests are disabled explicitly).

b. Using Cygwin
===============

This build should be similar to that for MinGW except that the resulting
library depends on the Cygwin DLL and cannot therefore be used with
Visual Studio as with MinGW. Indeed, the binaries compiled with Cygwin
require a dynamic library (cygwin.dll) to work; there is a Cygwin option
-mno-cygwin to build native code, but it may require some non-portable tricks.

In case of failure, you may need to pass LDFLAGS='-shared-libgcc' at the
end of the configure line due to a bug in GCC. Otherwise, if threading
support is not needed, you can configure MPFR with --disable-thread-safe.

c. Using Microsoft Visual C++ and Intel C++ Compilers
=====================================================

Static and dynamic MPFR libraries for the 32- and 64-bit versions of
Windows can be built with Microsoft Visual Studio 2015 using the
Microsoft Visual C++ compiler, see:

  https://visualstudio.microsoft.com/

The Intel C++ compiler provided as a part of Intel Parallel Studio XE
can also be used:

  https://software.intel.com/en-us/parallel-studio-xe

The relevant build projects are available here:

  https://github.com/BrianGladman