mpir/doc/devel/configuration

/* doc/configuration (in Emacs -*-outline-*- format). */

Copyright 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
Copyright 2008 William Hart

This file is part of the MPIR Library.

The MPIR 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 2.1 of the License, or (at your
option) any later version.

The MPIR 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 MPIR Library; see the file COPYING.LIB.  If not, write to
the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
02110-1301, USA.

* Adding a new file

** Adding a top-level file

  i) Add it to libgmp_la_SOURCES in Makefile.am.

  ii) If libmp.la needs it (usually doesn't), then add it to
      libmp_la_SOURCES too.

** Adding a subdirectory file

For instance for mpz,

  i) Add file.c to libmpz_la_SOURCES in mpz/Makefile.am.

  ii) Add mpz/file$U.lo to MPZ_OBJECTS in the top-level Makefile.am

  iii) If for some reason libmp.la needs it (usually doesn't) then add
       mpz/file$U.lo to libmp_la_DEPENDENCIES in the top-level
       Makefile.am too.

The same applies to mpf, mpq, scanf and printf.

** Adding an mpn file

The way we build libmpn (in the `mpn' subdirectory) is quite special.

Currently only mpn/mp_bases.c is truely generic and included in every
configuration.  All other files are linked at build time into the mpn
build directory from one of the CPU specific sub-directories, or from
the mpn/generic directory.

There are five types of mpn source files.

  .asm	  Assembly code preprocessed with m4
  .S	  Assembly code preprocessed with cpp
  .s	  Assembly code not preprocessed at all
  .c	  C code
  .as   Yasm format assembly file - yasm macros only, no preprocessing

There are two types of .asm files.

  i) ``Normal'' files containing one function, though possibly with
     more than one entry point.

  ii) Multi-function files that generate one of a set of functions
      according to build options.

To add a new implementation of an existing function,

  i) Put it in the appropriate CPU-specific mpn subdirectory, it'll be
     detected and used (if it is for an architecture supported by yasm
     simply call it .as and yasm will automatically be used to assemble 
     it).

  ii) Any entrypoints tested by HAVE_NATIVE_func in other code must
      have PROLOGUE(func) for configure to grep.  This is normal for
      .asm or .S files, but for .c and .as files a dummy comment like 
      the following will be needed.

              /*
              PROLOGUE(func)
              */

              OR

              ;
              ; PROLOGUE(func)
              ;

To add a new implementation using a multi-function file, in addition
do the following,

  i) Use a MULFUNC_PROLOGUE(func1 func2 ...) in the .asm, declaring
     all the functions implemented, including carry-in variants.

     If there's a separate PROLOGUE(func) for each possible function
     (but this is usually not the case), then MULFUNC_PROLOGUE isn't
     necessary.

     Currently we don't use multifunction files with yasm

To add a new style of multi-function file, in addition do the
following,

  i) Add to the GMP_MULFUNC_CHOICES "case" statement in configure.in
     which lists each multi-function filename and what function files
     it can provide.

To add a completely new mpn function file, do the following,

  i) Ensure the filename is a valid C identifier, due to the
     -DOPERATION_$* used to support multi-function files.  This means
     "-" can't be used (but "_" can).

  ii) Add it to configure.in under one of the following

      a) `gmp_mpn_functions' if it exists for every target.  This
         means there must be a C version in mpn/generic.  (Eg. mul_1)

      b) `gmp_mpn_functions_optional' if it's a standard function, but
         doesn't need to exist for every target.  Code wanting to use
         this will test HAVE_NATIVE_func to see if it's available.
         (Eg. copyi)

      c) `extra_functions' for some targets, if it's a special
         function that only ever needs to exist for certain targets.
         Code wanting to use it can test either HAVE_NATIVE_func or
         HAVE_HOST_CPU_foo, as desired.

  iii) If HAVE_NATIVE_func is going to be used, then add a #undef to
       the AH_VERBATIM([HAVE_NATIVE] block in configure.in.

  iv) Add file.c to nodist_libdummy_la_SOURCES in mpn/Makefile.am (in
      order to get an ansi2knr rule).  If the file is only in
      assembler then this step is unnecessary, but do it anyway so as
      not to forget if later a .c version is added.

  v) If the function can be provided by a multi-function file, then
     add to the "case" statement in configure.in which lists each
     multi-function filename and what function files it can provide.


** Adding a test program

  i) Tests to be run early in the testing can be added to the main
     "tests" sub-directory.

  ii) Tests for mpn, mpz, mpq and mpf can be added under the
      corresponding tests subdirectory.

  iii) Generic tests for late in the testing can be added to
       "tests/misc".  printf and scanf tests currently live there too.

  iv) Random number function tests can be added to "tests/rand".  That
      directory has some development-time programs too.

  v) C++ test programs can be added to "tests/cxx".  A line like the
     following must be added for each, since by default automake looks
     for a .c file.

             t_foo_SOURCES = t-foo.cc

In all cases the name of the program should be added to check_PROGRAMS
in the Makefile.am.  TESTS is equal to check_PROGRAMS, so all those
programs get run.

"tests/devel" has a number of programs which are only for development
purposes and are not for use in "make check".  These should be listed
in EXTRA_PROGRAMS to get Makefile rules created, but they're never
built or run unless an explicit "make someprog" is used.

* Adding a new CPU

In general it's policy to use proper names for each CPU type
supported.  If two CPUs are quite similar and perhaps don't have any
actual differences in MPIR then they're still given separate names, for
example alphaev67 and alphaev68.

Canonical names:

  i) Decide the canonical CPU names MPIR will accept.

  ii) Add these to the config.sub wrapper if configfsf.sub doesn't
      already accept them.

  iii) Document the names in gmp.texi.

Aliases (optional):

  i) Any aliases can be added to the config.sub wrapper, unless
     configfsf.sub already does the right thing with them.

  ii) Leave configure.in and everywhere else using only the canonical
      names.  Aliases shouldn't appear anywhere except config.sub.

  iii) Document in gmp.texi, if desired.  Usually this isn't a good
       idea, better encourage users to know just the canonical
       names.

Configure:

  i) Add patterns to configure.in for the new CPU names.  Include the
     following (see configure.in for the variables to set up),

     a) ABI choices (if any).
     b) Compiler choices.
     c) mpn path for CPU specific code.
     d) Good default CFLAGS for each likely compiler.
     d) Any special tests necessary on the compiler or assembler
        capabilities.

  ii) M4 macros to be shared by asm files in a CPU family are by
      convention in a foo-defs.m4 like mpn/x86/x86-defs.m4.  They're
      likely to use settings from config.m4 generated by configure.

Fat binaries:

  i) In configure.in, add CPU specific directory(s) to fat_path.

  ii) In mpn/<cpu>/fat.c, identify the CPU at runtime and use suitable
      CPUVEC_SETUP_subdir macros to select the function pointers for it.

  iii) For the x86s, add to the "$tmp_prefix" setups in configure.in
       which abbreviates subdirectory names to fit an 8.3 filesystem.
       (No need to restrict to 8.3, just ensure uniqueness when
       truncated.)


* The configure system

** Installing tools

The current versions of automake, autoconf and libtool in use can be
checked in the ChangeLog.  Look for "Update to ...".  Patches may have
been applied, look for "Regenerate ...".

The MPIR build system is in places somewhat dependent on the internals
of the build tools.  Obviously that's avoided as much as possible, but
where it can't it creates a problem when upgrading or attempting to
use different tools versions.

** Updating mpir

The following files need to be updated when going to a new version of
the build tools.  Unfortunately the tools generally don't identify
when an out-of-date version is present.

aclocal.m4 is updated by running "aclocal".  (Only needed for a new
automake or libtool.)

INSTALL.autoconf can be copied from INSTALL in autoconf.

ltmain.sh comes from libtool.  Remove it and run "libtoolize --copy",
or just copy the file by hand.

ansi2knr.c, ansi2knr.1, install-sh and doc/mdate-sh come from automake
and can be updated by copying or by removing and running "automake
--add-missing --copy".

texinfo.tex can be updated from ftp.gnu.org.  Check it still works
with "make gmp.dvi", "make gmp.ps" and "make gmp.pdf".

configfsf.guess and configfsf.sub can be updated from ftp.gnu.org (or
from the "config" cvs module at subversions.gnu.org).  The gmp
config.guess and config.sub wrappers are supposed to make such an
update fairly painless.

depcomp from automake is not needed because configure.in specifies
automake with "no-dependencies".

** How it works

During development:

    Input files                       Tool       Output files
    ---------------------------------------------------------
                                 
                                     aclocal
    $prefix/share/aclocal*/*.m4 ----------------> aclocal.m4

                   
    configure.in \                   autoconf
    aclocal.m4   / -----------------------------> configure

                  
    */Makefile.am \                  automake
    configure.in  | ----------------------------> Makefile.in
    aclocal.m4    /
                    
    configure.in \                  autoheader
    aclocal.m4   / -----------------------------> config.in

At build time:

    Input files          Tool       Output files
    --------------------------------------------

    */Makefile.in  \   configure    / */Makefile
    config.in      | -------------> | config.h
    gmp-h.in       |                | config.m4
    mp-h.in        /                | gmp.h
                                    | mp.h
                                    \ fat.h  (fat binary build only)

When configured with --enable-maintainer-mode the Makefiles include
rules to re-run the necessary tools if the input files are changed.
This can end up running a lot more things than are really necessary.

If a build tree is in too much of a mess for those rules to work
properly then a bootstrap can be done from the source directory with

	aclocal
	autoconf
	automake
	autoheader

The autom4te.cache directory is created by autoconf to save some work
in subsequent automake or autoheader runs.  It's recreated
automatically if removed, it doesn't get distributed.

** C++ configuration

It's intended that the contents of libgmp.la won't vary according to
whether --enable-cxx is selected.  This means that if C++ shared
libraries don't work properly then a shared+static with --disable-cxx
can be done for the C parts, then a static-only with --enable-cxx to
get libgmpxx.

libgmpxx.la uses some internals from libgmp.la, in order to share code
between C and C++.  It's intended that libgmpxx can only be expected
to work with libgmp from the same version of MPIR.  If some of the
shared internals change their interface, then it's proposed to rename
them, for instance __gmp_doprint2 or the like, so as to provoke link
errors rather than mysterious failures from a mismatch.

* Development setups

** General

--disable-shared will make builds go much faster, though of course
shared or shared+static should be tested too.

--enable-mpbsd grabs various bits of mpz, which might need to be
adjusted if things in those routines are changed.  Building mpbsd all
the time doesn't cost much.

--prefix to a dummy directory followed by "make install" will show
what's installed.

"make check" acts on the libgmp just built, and will ignore any other
/usr/lib/libgmp, or at least it should do.  Libtool does various hairy
things to ensure it hits the just-built library.

** Long long limb testing

On systems where gcc supports long long, but a limb is normally just a
long, the following can be used to force long long for testing
purposes.  It will probably run quite slowly.

	./configure --host=none ABI=longlong

** Function argument conversions

When using gcc, configuring with something like

	./configure CFLAGS="-g -Wall -Wconversion -Wno-sign-compare"

can show where function parameters are being converted due to having
function prototypes available, which won't happen in a K&R compiler.
Doing this in combination with the long long limb setups above is
good.

Conversions between int and long aren't warned about by gcc when
they're the same size, which is unfortunate because casts should be
used in such cases, for the benefit of K&R compilers with int!=long
and where the difference matters in function calls.

** K&R support

Function definitions must be in the GNU stylized form to work.  See
the ansi2knr.1 man page (included in the MPIR sources).

__GMP_PROTO is used for function prototypes, other ANSI / K&R
differences are conditionalized in various places.

Proper testing of the K&R support requires a compiler which gives an
error for ANSI-isms.  Configuring with --host=none is a good idea, to
test all the generic C code.

When using an ANSI compiler, the ansi2knr setups can be partially
tested with

	./configure am_cv_prog_cc_stdc=no ac_cv_prog_cc_stdc=no

This will test the use of $U and the like in the makefiles, but not
much else.

Forcing the cache variables can be used with a compiler like HP C
which is K&R by default but to which configure normally adds ANSI mode
flags.  This then should be a good full K&R test.

* Other Notes

** Compatibility

compat.c is the home of functions retained for binary compatibility,
    but now done by other means (like a macro).

struct __mpz_struct etc - this must be retained for C++ compatibility.
    C++ applications defining functions taking mpz_t etc parameters
    will get this in the mangled name because C++ "sees though" the
    typedef mpz_t to the underlying struct.

    Incidentally, this probably means for C++ that our mp.h is not
    compatible with an original BSD mp.h, since we use struct
    __mpz_struct for MINT in ours.  Maybe we could change to whatever
    the original did, but it seems unlikely anyone would be using C++
    with mp.h.

__gmpn - note that glibc defines some __mpn symbols, old versions of
    some mpn routines, which it uses for floating point printfs.




Local variables:
mode: outline
fill-column: 70
End:
/* eof doc/configuration */