Bigfloat

Python 2.

- Declarations in ``cmpfr.pxd`` fixed to include ``const`` where necessary;
this fixes some compiler warnings.

- Improved tests for hashing.


Miscellaneous
-------------

- Fix documentation link for PyPI upload.

- Docs updated to remove outdated instructions about use under Python 2.5 and
outdated information (dating from the earlier ctypes version) about runtime
location of the MPFR and GMP libraries.

- ``README``, ``setup.py`` updates: include the quick tour in the ``README``,
make content match that of the ``setup.py`` long description.

- ``CHANGELOG``, ``README``, ``INSTALL`` files updated with various
reStructuredText fixes; renamed to add ``.rst`` suffix so that they display
nicely on GitHub.

- Update copyright notices.

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

Library
-------

- Python 3.2 and later are now supported. Python 2 support requires Python 2.6
or later.

- Added support for the new 'divide-by-zero' flag (exposed as ``ZeroDivision``
to bigfloat).

Miscellaneous
-------------

- Development moved from bitbucket (https://bitbucket.org/dickinsm/bigfloat)
to GitHub (https://github.com/mdickinson/bigfloat), and repository
converted from mercurial to Git.

- Travis CI and Read the Docs GitHub hooks added.

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

- Minor documentation fixes.

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


Prerequisites
-------------

- For now, Python 2.7 and MPFR version >= 3.0 are required. Support for
earlier versions of Python is planned. Support for MPFR 2.x will probably
not be added.


Library
-------

- Major rewrite and restructuring of the core code. The core now uses Cython
rather than ctypes.

- mpfr module now separately available as an extension type, with semantics
very close to the Mpfr library.

- Added bigfloat support for new rounding mode MPFR_RNDA (round away from
zero).

- Rounding modes are now represented as named integers (instances of
RoundingMode, an int subclass), rather than strings. This allows them to be
passed directly to Mpfr functions.

- Fixed a bug that meant old context wasn't restored when changing exponents
with one valid exponent and one invalid. (eminmax)

- Many more docstrings than originally; better function signatures.

- Restructured bigfloat package: mpfr wrapper module is now bigfloat.mpfr,
context class and related functions are in bigfloat.context, rounding modes
are defined in bigfloat.rounding_modes, and the main BigFloat type is defined
in bigfloat.core. Everything is importable directory from bigfloat, as
before.


Documentation
-------------

- Most of the documentation is now generated automatically using Sphinx's
autodoc extension.

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

Library
-------

- Simplify weakref callback scheme used to ensure BigFloat instances
are freed properly.

- Make sure all python files in distribution have the LGPL header.

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

Library
-------

- The setup.py file no longer attempts to locate the MPFR library at
install time. The configuration file still exists, and can be edited
to give a hard-coded location. Alternatively, ctypes should find the
library if it's within the usual library search paths. (If not, it
might be necessary to set LD_LIBRARY_PATH or DYLD_LIBRARY_PATH.)

- Use a weakref callback to ensure ``mpfr_clear`` gets called, instead of
relying on ``__del__``.

- Make bigfloat compatible with versions of MPFR as far back as MPFR 2.3.0

- Relicense under LGPL, since this package might be considered a
derived work of MPFR.


Tests
-----

- Skip the hashing consistency tests (those that test whether ``hash(n) ==
hash(BigFloat(n))`` for integers ``n``) on Python 2.5. In rare cases, this
equality fails with Python 2.5, and this is awkward to fix. If this affects
you, upgrade to Python 2.6 or avoid mixing BigFloat instances with ints in
sets or dictionary keys.