"""
Atom classes for describing dataset contents.

:Author: Ivan Vilata i Balaguer
:Contact: ivan at selidor dot net
:License: BSD
:Created: December 16, 2004
:Revision: $Id$

See the docstrings of `Atom` classes for more info.

Variables
=========

`__docformat`__
    The format of documentation strings in this module.
`__version__`
    Repository version of this file.
`all_types`
    Set of all PyTables types.
`atom_map`
    Maps atom kinds to item sizes and atom classes.

    If there is a fixed set of possible item sizes for a given kind,
    the kind maps to another mapping from item size in bytes to atom
    class.  Otherwise, the kind maps directly to the atom class.

`deftype_from_kind`
    Maps atom kinds to their default atom type (if any).
"""

# Imports
# =======
import re
import sys
import inspect
import cPickle

import numpy

from tables.utils import SizeType
from tables.misc.enum import Enum


# Public variables
# ================
__docformat__ = 'reStructuredText'
"""The format of documentation strings in this module."""

__version__ = '$Revision$'
"""Repository version of this file."""

all_types = set()  # filled as atom classes are created
"""Set of all PyTables types."""

atom_map = {}  # filled as atom classes are created
"""
Maps atom kinds to item sizes and atom classes.

If there is a fixed set of possible item sizes for a given kind, the
kind maps to another mapping from item size in bytes to atom class.
Otherwise, the kind maps directly to the atom class.
"""

deftype_from_kind = {}  # filled as atom classes are created
"""Maps atom kinds to their default atom type (if any)."""


# Public functions
# ================
_type_re = re.compile(r'^([a-z]+)([0-9]*)$')
def split_type(type):
    """
    Split a PyTables `type` into a PyTables kind and an item size.

    Returns a tuple of ``(kind, itemsize)``.  If no item size is
    present in the `type` (in the form of a precision), the returned
    item size is `None`.

    >>> split_type('int32')
    ('int', 4)
    >>> split_type('string')
    ('string', None)
    >>> split_type('int20')
    Traceback (most recent call last):
      ...
    ValueError: precision must be a multiple of 8: 20
    >>> split_type('foo bar')
    Traceback (most recent call last):
      ...
    ValueError: malformed type: 'foo bar'
    """

    match = _type_re.match(type)
    if not match:
        raise ValueError("malformed type: %r" % type)
    kind, precision = match.groups()
    itemsize = None
    if precision:
        precision = int(precision)
        itemsize, remainder = divmod(precision, 8)
        if remainder:  # 0 could be a valid item size
            raise ValueError( "precision must be a multiple of 8: %d"
                              % precision )
    return (kind, itemsize)


# Private functions
# =================
def _invalid_itemsize_error(kind, itemsize, itemsizes):
    isizes = sorted(itemsizes)
    return ValueError( "invalid item size for kind ``%s``: %r; "
                       "it must be one of ``%r``"
                       % (kind, itemsize, isizes) )

def _abstract_atom_init(deftype, defvalue):
    """Return a constructor for an abstract `Atom` class."""
    defitemsize = split_type(deftype)[1]
    def __init__(self, itemsize=defitemsize, shape=(), dflt=defvalue):
        assert self.kind in atom_map
        try:
            atomclass = atom_map[self.kind][itemsize]
        except KeyError:
            raise _invalid_itemsize_error( self.kind, itemsize,
                                           atom_map[self.kind] )
        self.__class__ = atomclass
        atomclass.__init__(self, shape, dflt)
    return __init__

def _normalize_shape(shape):
    """Check that the `shape` is safe to be used and return it as a tuple."""

    if isinstance(shape, (int, numpy.integer, long)):
        if shape < 1:
            raise ValueError( "shape value must be greater than 0: %d"
                              % shape )
        shape = (shape,)  # N is a shorthand for (N,)
    try:
        shape = tuple(shape)
    except TypeError:
        raise TypeError( "shape must be an integer or sequence: %r"
                         % (shape,) )

    ## XXX Get from HDF5 library if possible.
    # HDF5 does not support ranks greater than 32
    if len(shape) > 32:
        raise ValueError(
            "shapes with rank > 32 are not supported: %r" % (shape,) )

    return tuple(SizeType(s) for s in shape)

def _normalize_default(value, dtype):
    """Return `value` as a valid default of NumPy type `dtype`."""
    # Create NumPy objects as defaults
    # This is better in order to serialize them as attributes
    if value is None:
        value = 0
    basedtype = dtype.base
    try:
        default = numpy.array(value, dtype=basedtype)
    except ValueError:
        array = numpy.array(value)
        if array.shape != basedtype.shape:
            raise
        # Maybe nested dtype with "scalar" value.
        default = numpy.array(value, dtype=basedtype.base)
    # 0-dim arrays will be representented as NumPy scalars
    # (PyTables attribute convention)
    if default.shape == ():
        default = default[()]
    return default

def _cmp_dispatcher(other_method_name):
    """
    Dispatch comparisons to a method of the *other* object.

    Returns a new *rich comparison* method which dispatches calls to
    the method `other_method_name` of the *other* object.  If there is
    no such method in the object, ``False`` is returned.

    This is part of the implementation of a double dispatch pattern.
    """
    def dispatched_cmp(self, other):
        try:
            other_method = getattr(other, other_method_name)
        except AttributeError:
            return False
        return other_method(self)
    return dispatched_cmp


# Helper classes
# ==============
class MetaAtom(type):

    """
    Atom metaclass.

    This metaclass ensures that data about atom classes gets inserted
    into the suitable registries.
    """

    def __init__(class_, name, bases, dict_):
        super(MetaAtom, class_).__init__(name, bases, dict_)

        kind = dict_.get('kind')
        itemsize = dict_.get('itemsize')
        type_ = dict_.get('type')
        deftype = dict_.get('_deftype')

        if kind and deftype:
            deftype_from_kind[kind] = deftype

        if type_:
            all_types.add(type_)

        if kind and itemsize and not hasattr(itemsize, '__int__'):
            # Atom classes with a non-fixed item size do have an
            # ``itemsize``, but it's not a number (e.g. property).
            atom_map[kind] = class_
            return

        if kind:  # first definition of kind, make new entry
            atom_map[kind] = {}

        if itemsize and hasattr(itemsize, '__int__'):  # fixed
            kind = class_.kind  # maybe from superclasses
            atom_map[kind][int(itemsize)] = class_


# Atom classes
# ============
class Atom(object):

    """
    Defines the type of atomic cells stored in a dataset.

    The meaning of *atomic* is that individual elements of a cell can
    not be extracted directly by indexing (i.e. ``__getitem__()``) the
    dataset; e.g. if a dataset has shape (2, 2) and its atoms have
    shape (3,), to get the third element of the cell at (1, 0) one
    should use ``dataset[1,0][2]`` instead of ``dataset[1,0,2]``.

    The `Atom` class is meant to declare the different properties of
    the *base element* (also known as *atom*) of `CArray`, `EArray`
    and `VLArray` datasets, although they are also used to describe
    the base elements of `Array` datasets.  Atoms have the property
    that their length is always the same.  However, you can grow
    datasets along the extensible dimension in the case of `EArray` or
    put a variable number of them on a `VLArray` row.  Moreover, atoms
    are not restricted to scalar values, and they can be *fully
    multidimensional objects*.

    A series of descendant classes are offered in order to make the
    use of these element descriptions easier.  You should use a
    particular `Atom` descendant class whenever you know the exact
    type you will need when writing your code.  Otherwise, you may use
    one of the ``Atom.from_*()`` factory methods.

    Public instance variables
    -------------------------

    dflt
        The default value of the atom.

        If the user does not supply a value for an element while
        filling a dataset, this default value will be written to
        disk. If the user supplies a scalar value for a
        multidimensional atom, this value is automatically *broadcast*
        to all the items in the atom cell.  If ``dflt`` is not
        supplied, an appropriate zero value (or *null* string) will be
        chosen by default.  Please note that default values are kept
        internally as NumPy objects.

    dtype
        The NumPy ``dtype`` that most closely matches this atom.
    itemsize
        Size in bytes of a sigle item in the atom.

        Specially useful for atoms of the ``string`` kind.

    kind
        The PyTables kind of the atom (a string).
    recarrtype
        String type to be used in ``numpy.rec.array()``.
    shape
        The shape of the atom (a tuple, ``()`` for scalar atoms).
    size
        Total size in bytes of the atom.
    type
        The PyTables type of the atom (a string).

    Atoms can be compared with atoms and other objects for strict
    (in)equality without having to compare individual attributes:

    >>> atom1 = StringAtom(itemsize=10)  # same as ``atom2``
    >>> atom2 = Atom.from_kind('string', 10)  # same as ``atom1``
    >>> atom3 = IntAtom()
    >>> atom1 == 'foo'
    False
    >>> atom1 == atom2
    True
    >>> atom2 != atom1
    False
    >>> atom1 == atom3
    False
    >>> atom3 != atom2
    True

    Public methods
    --------------

    copy(**override)
        Get a copy of the atom, possibly overriding some arguments.

    Factory methods
    ---------------

    from_dtype(dtype[, dflt])
        Create an `Atom` from a NumPy ``dtype``.
    from_kind(kind[, itemsize][, shape][, dflt])
        Create a `Atom` from a PyTables ``kind``.
    from_sctype(sctype[, shape][, dflt])
        Create a `Atom` from a NumPy scalar type ``sctype``.
    from_type(type[, shape][, dflt])
        Create a `Atom` from a PyTables ``type``.

    Constructors
    ------------

    There are some common arguments for most `Atom` -derived
    constructors:

    itemsize
        For types with a non-fixed size, this sets the size in bytes
        of individual items in the atom.

    shape
        Sets the shape of the atom.  An integer shape of ``N`` is
        equivalent to the tuple ``(N,)``.

    dflt
        Sets the default value for the atom.

    """

    # Register data for all subclasses.
    __metaclass__ = MetaAtom

    # Class methods
    # ~~~~~~~~~~~~~
    @classmethod
    def prefix(class_):
        """Return the atom class prefix."""
        cname = class_.__name__
        return cname[:cname.rfind('Atom')]

    @classmethod
    def from_sctype(class_, sctype, shape=(), dflt=None):
        """
        Create an `Atom` from a NumPy scalar type `sctype`.

        Optional shape and default value may be specified as the
        `shape` and `dflt` arguments, respectively.  Information in
        the `sctype` not represented in an `Atom` is ignored.

        >>> import numpy
        >>> Atom.from_sctype(numpy.int16, shape=(2, 2))
        Int16Atom(shape=(2, 2), dflt=0)
        >>> Atom.from_sctype('S5', dflt='hello')
        Traceback (most recent call last):
          ...
        ValueError: unknown NumPy scalar type: 'S5'
        >>> Atom.from_sctype('Float64')
        Float64Atom(shape=(), dflt=0.0)
        """
        if ( not isinstance(sctype, type)
             or not issubclass(sctype, numpy.generic) ):
            if sctype not in numpy.sctypeDict:
                raise ValueError("unknown NumPy scalar type: %r" % (sctype,))
            sctype = numpy.sctypeDict[sctype]
        return class_.from_dtype(numpy.dtype((sctype, shape)), dflt)

    @classmethod
    def from_dtype(class_, dtype, dflt=None):
        """
        Create an `Atom` from a NumPy `dtype`.

        An optional default value may be specified as the `dflt`
        argument.  Information in the `dtype` not represented in an
        `Atom` is ignored.

        >>> import numpy
        >>> Atom.from_dtype(numpy.dtype((numpy.int16, (2, 2))))
        Int16Atom(shape=(2, 2), dflt=0)
        >>> Atom.from_dtype(numpy.dtype('S5'), dflt='hello')
        StringAtom(itemsize=5, shape=(), dflt='hello')
        >>> Atom.from_dtype(numpy.dtype('Float64'))
        Float64Atom(shape=(), dflt=0.0)
        """
        basedtype = dtype.base
        if basedtype.names:
            raise ValueError( "compound data types are not supported: %r"
                              % dtype )
        if basedtype.shape != ():
            raise ValueError( "nested data types are not supported: %r"
                              % dtype )
        if basedtype.kind == 'S':  # can not reuse something like 'string80'
            itemsize = basedtype.itemsize
            return class_.from_kind('string', itemsize, dtype.shape, dflt)
        # Most NumPy types have direct correspondence with PyTables types.
        return class_.from_type(basedtype.name, dtype.shape, dflt)

    @classmethod
    def from_type(class_, type, shape=(), dflt=None):
        """
        Create an `Atom` from a PyTables `type`.

        Optional shape and default value may be specified as the
        `shape` and `dflt` arguments, respectively.

        >>> Atom.from_type('bool')
        BoolAtom(shape=(), dflt=False)
        >>> Atom.from_type('int16', shape=(2, 2))
        Int16Atom(shape=(2, 2), dflt=0)
        >>> Atom.from_type('string40', dflt='hello')
        Traceback (most recent call last):
          ...
        ValueError: unknown type: 'string40'
        >>> Atom.from_type('Float64')
        Traceback (most recent call last):
          ...
        ValueError: unknown type: 'Float64'
        """
        if type not in all_types:
            raise ValueError("unknown type: %r" % (type,))
        kind, itemsize = split_type(type)
        return class_.from_kind(kind, itemsize, shape, dflt)

    @classmethod
    def from_kind(class_, kind, itemsize=None, shape=(), dflt=None):
        """
        Create an `Atom` from a PyTables `kind`.

        Optional item size, shape and default value may be specified
        as the `itemsize`, `shape` and `dflt` arguments, respectively.
        Bear in mind that not all atoms support a default item size.

        >>> Atom.from_kind('int', itemsize=2, shape=(2, 2))
        Int16Atom(shape=(2, 2), dflt=0)
        >>> Atom.from_kind('int', shape=(2, 2))
        Int32Atom(shape=(2, 2), dflt=0)
        >>> Atom.from_kind('int', shape=1)
        Int32Atom(shape=(1,), dflt=0)
        >>> Atom.from_kind('string', itemsize=5, dflt='hello')
        StringAtom(itemsize=5, shape=(), dflt='hello')
        >>> Atom.from_kind('string', dflt='hello')
        Traceback (most recent call last):
          ...
        ValueError: no default item size for kind ``string``
        >>> Atom.from_kind('Float')
        Traceback (most recent call last):
          ...
        ValueError: unknown kind: 'Float'

        Moreover, some kinds with atypical constructor signatures are
        not supported; you need to use the proper constructor:

        >>> Atom.from_kind('enum')  #doctest: +ELLIPSIS
        Traceback (most recent call last):
          ...
        ValueError: the ``enum`` kind is not supported...
        """

        kwargs = {'shape': shape}
        if kind not in atom_map:
            raise ValueError("unknown kind: %r" % (kind,))
        # This incompatibility detection may get out-of-date and is
        # too hard-wired, but I couldn't come up with something
        # smarter.  -- Ivan (2007-02-08)
        if kind in ['enum']:
            raise ValueError( "the ``%s`` kind is not supported; "
                              "please use the appropriate constructor"
                              % kind )
        # If no `itemsize` is given, try to get the default type of the
        # kind (which has a fixed item size).
        if itemsize is None:
            if kind not in deftype_from_kind:
                raise ValueError( "no default item size for kind ``%s``"
                                  % kind )
            type_ = deftype_from_kind[kind]
            kind, itemsize = split_type(type_)
        kdata = atom_map[kind]
        # Look up the class and set a possible item size.
        if hasattr(kdata, 'kind'):  # atom class: non-fixed item size
            atomclass = kdata
            kwargs['itemsize'] = itemsize
        else:  # dictionary: fixed item size
            if itemsize not in kdata:
                raise _invalid_itemsize_error(kind, itemsize, kdata)
            atomclass = kdata[itemsize]
        # Only set a `dflt` argument if given (`None` may not be understood).
        if dflt is not None:
            kwargs['dflt'] = dflt

        return atomclass(**kwargs)

    # Properties
    # ~~~~~~~~~~
    size = property(
        lambda self: self.dtype.itemsize,
        None, None, "Total size in bytes of the atom." )
    recarrtype = property(
        lambda self: str(self.dtype.shape) + self.dtype.base.str[1:],
        None, None, "String type to be used in ``numpy.rec.array()``." )

    # Special methods
    # ~~~~~~~~~~~~~~~
    def __init__(self, nptype, shape, dflt):
        if not hasattr(self, 'type'):
            raise NotImplementedError( "``%s`` is an abstract class; "
                                       "please use one of its subclasses"
                                       % self.__class__.__name__ )
        self.shape = shape = _normalize_shape(shape)
        # Curiously enough, NumPy isn't generally able to accept NumPy
        # integers in a shape. ;(
        npshape = tuple(int(s) for s in shape)
        self.dtype = dtype = numpy.dtype((nptype, npshape))
        self.dflt = _normalize_default(dflt, dtype)

    def __repr__(self):
        args = 'shape=%s, dflt=%r' % (self.shape, self.dflt)
        if not hasattr(self.__class__.itemsize, '__int__'):  # non-fixed
            args = 'itemsize=%s, %s' % (self.itemsize, args)
        return '%s(%s)' % (self.__class__.__name__, args)

    __eq__ = _cmp_dispatcher('_is_equal_to_atom')

    def __ne__(self, other):
        return not self.__eq__(other)

    # Public methods
    # ~~~~~~~~~~~~~~
    def copy(self, **override):
        """
        Get a copy of the atom, possibly overriding some arguments.

        Constructor arguments to be overridden must be passed as
        keyword arguments.

        >>> atom1 = StringAtom(itemsize=12)
        >>> atom2 = atom1.copy()
        >>> print atom1
        StringAtom(itemsize=12, shape=(), dflt='')
        >>> print atom2
        StringAtom(itemsize=12, shape=(), dflt='')
        >>> atom1 is atom2
        False
        >>> atom3 = atom1.copy(itemsize=100, shape=(2, 2))
        >>> print atom3
        StringAtom(itemsize=100, shape=(2, 2), dflt='')
        >>> atom1.copy(foobar=42)
        Traceback (most recent call last):
          ...
        TypeError: __init__() got an unexpected keyword argument 'foobar'
        """
        newargs = self._get_init_args()
        newargs.update(override)
        return self.__class__(**newargs)

    # Private methods
    # ~~~~~~~~~~~~~~~
    def _get_init_args(self):
        """
        Get a dictionary of instance constructor arguments.

        This implementation works on classes which use the same names
        for both constructor arguments and instance attributes.
        """
        return dict( (arg, getattr(self, arg))
                     for arg in inspect.getargspec(self.__init__)[0]
                     if arg != 'self' )

    def _is_equal_to_atom(self, atom):
        """Is this object equal to the given `atom`?"""
        return ( self.type == atom.type and self.shape == atom.shape
                 and self.itemsize == atom.itemsize
                 and numpy.all(self.dflt == atom.dflt) )


class StringAtom(Atom):
    """
    Defines an atom of type ``string``.

    The item size is the *maximum* length in characters of strings.
    """
    kind = 'string'
    itemsize = property(
        lambda self: self.dtype.base.itemsize,
        None, None, "Size in bytes of a sigle item in the atom." )
    type = 'string'
    _defvalue = ''

    def __init__(self, itemsize, shape=(), dflt=_defvalue):
        if not hasattr(itemsize, '__int__') or int(itemsize) < 0:
            raise ValueError( "invalid item size for kind ``%s``: %r; "
                              "it must be a positive integer"
                              % ('string', itemsize) )
        Atom.__init__(self, 'S%d' % itemsize, shape, dflt)


class BoolAtom(Atom):
    """Defines an atom of type ``bool``."""
    kind = 'bool'
    itemsize = 1
    type = 'bool'
    _deftype = 'bool8'
    _defvalue = False
    def __init__(self, shape=(), dflt=_defvalue):
        Atom.__init__(self, self.type, shape, dflt)


class IntAtom(Atom):
    """Defines an atom of a signed integral type (``int`` kind)."""
    kind = 'int'
    signed = True
    _deftype = 'int32'
    _defvalue = 0
    __init__ = _abstract_atom_init(_deftype, _defvalue)

class UIntAtom(Atom):
    """Defines an atom of an unsigned integral type (``uint`` kind)."""
    kind = 'uint'
    signed = False
    _deftype = 'uint32'
    _defvalue = 0
    __init__ = _abstract_atom_init(_deftype, _defvalue)

class FloatAtom(Atom):
    """Defines an atom of a floating point type (``float`` kind)."""
    kind = 'float'
    _deftype = 'float64'
    _defvalue = 0.0
    __init__ = _abstract_atom_init(_deftype, _defvalue)


def _create_numeric_class(baseclass, itemsize):
    """
    Create a numeric atom class with the given `baseclass` and an
    `itemsize`.
    """
    prefix = '%s%d' % (baseclass.prefix(), itemsize * 8)
    type_ = prefix.lower()
    classdict = { 'itemsize': itemsize, 'type': type_,
                  '__doc__': "Defines an atom of type ``%s``." % type_ }
    def __init__(self, shape=(), dflt=baseclass._defvalue):
        Atom.__init__(self, self.type, shape, dflt)
    classdict['__init__'] = __init__
    return type('%sAtom' % prefix, (baseclass,), classdict)


def _generate_integral_classes():
    """Generate all integral classes."""
    for baseclass in [IntAtom, UIntAtom]:
        for itemsize in [1, 2, 4, 8]:
            newclass = _create_numeric_class(baseclass, itemsize)
            yield newclass

def _generate_floating_classes():
    """Generate all floating classes."""
    for itemsize in [4, 8]:
        newclass = _create_numeric_class(FloatAtom, itemsize)
        yield newclass

# Create all numeric atom classes.
for _classgen in [_generate_integral_classes, _generate_floating_classes]:
    for _newclass in _classgen():
        exec '%s = _newclass' % _newclass.__name__
del _classgen, _newclass


class ComplexAtom(Atom):
    """
    Defines an atom of a complex type.

    Allowed item sizes are 8 (single precision) and 16 (double
    precision).  This class must be used instead of more concrete ones
    to avoid confusions with ``numarray`` -like precision
    specifications used in PyTables 1.X.
    """

    # This definition is a little more complex (no pun intended)
    # because, although the complex kind is a normal numerical one,
    # the usage of bottom-level classes is artificially forbidden.
    # Everything will be back to normality when people has stopped
    # using the old bottom-level complex classes.

    kind = 'complex'
    itemsize = property(
        lambda self: self.dtype.base.itemsize,
        None, None, "Size in bytes of a sigle item in the atom." )
    _deftype = 'complex128'
    _defvalue = 0j

    # Only instances have a `type` attribute, so complex types must be
    # registered by hand.
    all_types.add('complex64')
    all_types.add('complex128')

    def __init__(self, itemsize, shape=(), dflt=_defvalue):
        isizes = [8, 16]
        if itemsize not in isizes:
            raise _invalid_itemsize_error('complex', itemsize, isizes)
        self.type = '%s%d' % (self.kind, itemsize * 8)
        Atom.__init__(self, self.type, shape, dflt)

class _ComplexErrorAtom(ComplexAtom):
    """Reminds the user to stop using the old complex atom names."""
    __metaclass__ = type  # do not register anything about this class
    def __init__(self, shape=(), dflt=ComplexAtom._defvalue):
        raise TypeError(
            "to avoid confusions with PyTables 1.X complex atom names, "
            "please use ``ComplexAtom(itemsize=N)``, "
            "where N=8 for single precision complex atoms, "
            "and N=16 for double precision complex atoms" )
Complex32Atom = Complex64Atom = Complex128Atom = _ComplexErrorAtom


class TimeAtom(Atom):

    """
    Defines an atom of time type (``time`` kind).

    There are two distinct supported types of time: a 32 bit integer
    value and a 64 bit floating point value.  Both of them reflect the
    number of seconds since the Unix epoch.  This atom has the
    property of being stored using the HDF5 time datatypes.
    """
    kind = 'time'
    _deftype = 'time32'
    _defvalue = 0
    __init__ = _abstract_atom_init(_deftype, _defvalue)

class Time32Atom(TimeAtom):
    """Defines an atom of type ``time32``."""
    itemsize = 4
    type = 'time32'
    _defvalue = 0
    def __init__(self, shape=(), dflt=_defvalue):
        Atom.__init__(self, 'int32', shape, dflt)

class Time64Atom(TimeAtom):
    """Defines an atom of type ``time64``."""
    itemsize = 8
    type = 'time64'
    _defvalue = 0.0
    def __init__(self, shape=(), dflt=_defvalue):
        Atom.__init__(self, 'float64', shape, dflt)


class EnumAtom(Atom):

    """
    Description of an atom of an enumerated type.

    Instances of this class describe the atom type used to store
    enumerated values.  Those values belong to an enumerated type,
    defined by the first argument (``enum``) in the constructor of the
    atom, which accepts the same kinds of arguments as the ``Enum``
    class.  The enumerated type is stored in the ``enum`` attribute of
    the atom.

    A default value must be specified as the second argument
    (``dflt``) in the constructor; it must be the *name* (a string) of
    one of the enumerated values in the enumerated type.  When the
    atom is created, the corresponding concrete value is broadcast and
    stored in the ``dflt`` attribute (setting different default values
    for items in a multidimensional atom is not supported yet).  If
    the name does not match any value in the enumerated type, a
    ``KeyError`` is raised.

    Another atom must be specified as the ``base`` argument in order
    to determine the base type used for storing the values of
    enumerated values in memory and disk.  This *storage atom* is kept
    in the ``base`` attribute of the created atom.  As a shorthand,
    you may specify a PyTables type instead of the storage atom,
    implying that this has a scalar shape.

    The storage atom should be able to represent each and every
    concrete value in the enumeration.  If it is not, a ``TypeError``
    is raised.  The default value of the storage atom is ignored.

    The ``type`` attribute of enumerated atoms is always ``'enum'``.

    Enumerated atoms also support comparisons with other objects:

    >>> enum = ['T0', 'T1', 'T2']
    >>> atom1 = EnumAtom(enum, 'T0', 'int8')  # same as ``atom2``
    >>> atom2 = EnumAtom(enum, 'T0', Int8Atom())  # same as ``atom1``
    >>> atom3 = EnumAtom(enum, 'T0', 'int16')
    >>> atom4 = Int8Atom()
    >>> atom1 == enum
    False
    >>> atom1 == atom2
    True
    >>> atom2 != atom1
    False
    >>> atom1 == atom3
    False
    >>> atom1 == atom4
    False
    >>> atom4 != atom1
    True

    Examples
    --------

    The next C ``enum`` construction::

      enum myEnum {
        T0,
        T1,
        T2
      };

    would correspond to the following PyTables declaration:

    >>> myEnumAtom = EnumAtom(['T0', 'T1', 'T2'], 'T0', 'int32')

    Please note the ``dflt`` argument with a value of ``'T0'``.  Since
    the concrete value matching ``T0`` is unknown right now (we have
    not used explicit concrete values), using the name is the only
    option left for defining a default value for the atom.

    The chosen representation of values for this enumerated atom uses
    unsigned 32-bit integers, which surely wastes quite a lot of
    memory.  Another size could be selected by using the ``base``
    argument (this time with a full-blown storage atom):

    >>> myEnumAtom = EnumAtom(['T0', 'T1', 'T2'], 'T0', UInt8Atom())

    You can also define multidimensional arrays for data elements:

    >>> myEnumAtom = EnumAtom(
    ...    ['T0', 'T1', 'T2'], 'T0', base='uint32', shape=(3,2))

    for 3x2 arrays of ``uint32``.
    """

    # Registering this class in the class map may be a little wrong,
    # since the ``Atom.from_kind()`` method fails miserably with
    # enumerations, as they don't support an ``itemsize`` argument.
    # However, resetting ``__metaclass__`` to ``type`` doesn't seem to
    # work and I don't feel like creating a subclass of ``MetaAtom``.

    kind = 'enum'
    type = 'enum'

    # Properties
    # ~~~~~~~~~~
    itemsize = property(
        lambda self: self.dtype.base.itemsize,
        None, None, "Size in bytes of a sigle item in the atom." )

    # Private methods
    # ~~~~~~~~~~~~~~~
    def _checkBase(self, base):
        """Check the `base` storage atom."""

        if base.kind == 'enum':
            raise TypeError( "can not use an enumerated atom "
                             "as a storage atom: %r" % base )

        # Check whether the storage atom can represent concrete values
        # in the enumeration...
        basedtype = base.dtype
        pyvalues = [value for (name, value) in self.enum]
        try:
            npgenvalues = numpy.array(pyvalues)
        except ValueError:
            raise TypeError("concrete values are not uniformly-shaped")
        try:
            npvalues = numpy.array(npgenvalues, dtype=basedtype.base)
        except ValueError:
            raise TypeError( "storage atom type is incompatible with "
                             "concrete values in the enumeration" )
        if npvalues.shape[1:] != basedtype.shape:
            raise TypeError( "storage atom shape does not match that of "
                             "concrete values in the enumeration" )
        if npvalues.tolist() != npgenvalues.tolist():
            raise TypeError( "storage atom type lacks precision for "
                             "concrete values in the enumeration" )

        # ...with some implementation limitations.
        if not npvalues.dtype.kind in ['i', 'u']:
            raise NotImplementedError( "only integer concrete values "
                                       "are supported for the moment, sorry" )
        if len(npvalues.shape) > 1:
            raise NotImplementedError( "only scalar concrete values "
                                       "are supported for the moment, sorry" )

    def _get_init_args(self):
        """Get a dictionary of instance constructor arguments."""
        return dict( enum=self.enum, dflt=self._defname,
                     base=self.base, shape=self.shape )

    def _is_equal_to_atom(self, atom):
        """Is this object equal to the given `atom`?"""
        return False

    def _is_equal_to_enumatom(self, enumatom):
        """Is this object equal to the given `enumatom`?"""
        return ( self.enum == enumatom.enum and self.shape == enumatom.shape
                 and numpy.all(self.dflt == enumatom.dflt)
                 and self.base == enumatom.base )

    # Special methods
    # ~~~~~~~~~~~~~~~
    def __init__(self, enum, dflt, base, shape=()):
        if not isinstance(enum, Enum):
            enum = Enum(enum)
        self.enum = enum

        if type(base) is str:
            base = Atom.from_type(base)
        self._checkBase(base)
        self.base = base

        default = enum[dflt]  # check default value
        self._defname = dflt  # kept for representation purposes

        # These are kept to ease dumping this particular
        # representation of the enumeration to storage.
        names, values = [], []
        for (name, value) in enum:
            names.append(name)
            values.append(value)
        basedtype = self.base.dtype

        self._names = names
        self._values = numpy.array(values, dtype=basedtype.base)

        Atom.__init__(self, basedtype, shape, default)

    def __repr__(self):
        return ( 'EnumAtom(enum=%r, dflt=%r, base=%r, shape=%r)'
                 % (self.enum, self._defname, self.base, self.shape) )

    __eq__ = _cmp_dispatcher('_is_equal_to_enumatom')


# Pseudo-atom classes
# ===================
#
# Now, there come three special classes, `ObjectAtom`, `VLStringAtom`
# and `VLUnicodeAtom`, that actually do not descend from `Atom`, but
# which goal is so similar that they should be described here.
# Pseudo-atoms can only be used with `VLArray` datasets, and they do
# not support multidimensional values, nor multiple values per row.
#
# They can be recognised because they also have ``kind``, ``type`` and
# ``shape`` attributes, but no ``size``, ``itemsize`` or ``dflt``
# ones.  Instead, they have a ``base`` atom which defines the elements
# used for storage.
#
# See ``examples/vlarray1.py`` and ``examples/vlarray2.py`` for
# further examples on `VLArray` datasets, including object
# serialization and string management.


class PseudoAtom(object):
    """
    Pseudo-atoms can only be used in ``VLArray`` nodes.

    They can be recognised because they also have `kind`, `type` and
    `shape` attributes, but no `size`, `itemsize` or `dflt` ones.
    Instead, they have a `base` atom which defines the elements used
    for storage.
    """
    def __repr__(self):
        return '%s()' % self.__class__.__name__

    def toarray(self, object_):
        """Convert an `object_` into an array of base atoms."""
        raise NotImplementedError

    def fromarray(self, array):
        """Convert an `array` of base atoms into an object."""
        raise NotImplementedError

class _BufferedAtom(PseudoAtom):
    """Pseudo-atom which stores data as a buffer (flat array of uints)."""
    shape = ()

    def toarray(self, object_):
        buffer_ = self._tobuffer(object_)
        array = numpy.ndarray( buffer=buffer_, dtype=self.base.dtype,
                               shape=len(buffer_) )
        return array

    def _tobuffer(self, object_):
        """Convert an `object_` into a buffer."""
        raise NotImplementedError

class VLStringAtom(_BufferedAtom):
    """
    Defines an atom of type ``vlstring``.

    This class describes a *row* of the `VLArray` class, rather than
    an atom.  It differs from the `StringAtom` class in that you can
    only add *one instance of it to one specific row*, i.e. the
    `VLArray.append()` method only accepts one object when the base
    atom is of this type.

    Like `StringAtom`, this class does not make assumptions on the
    encoding of the string, and raw bytes are stored as is.  Unicode
    strings are supported as long as no character is out of the ASCII
    set; otherwise, you will need to *explicitly* convert them to
    strings before you can save them.  For full Unicode support, using
    `VLUnicodeAtom` is recommended.

    Variable-length string atoms do not accept parameters and they
    cause the reads of rows to always return Python strings.  You can
    regard ``vlstring`` atoms as an easy way to save generic variable
    length strings.
    """
    kind = 'vlstring'
    type = 'vlstring'
    base = UInt8Atom()

    def _tobuffer(self, object_):
        if not isinstance(object_, basestring):
            raise TypeError("object is not a string: %r" % (object_,))
        return numpy.string0(object_)

    def fromarray(self, array):
        return array.tostring()

class VLUnicodeAtom(_BufferedAtom):
    """
    Defines an atom of type ``vlunicode``.

    This class describes a *row* of the `VLArray` class, rather than
    an atom.  It is very similar to `VLStringAtom`, but it stores
    Unicode strings (using 32-bit characters a la UCS-4, so all
    strings of the same length also take up the same space).

    This class does not make assumptions on the encoding of plain
    input strings.  Plain strings are supported as long as no
    character is out of the ASCII set; otherwise, you will need to
    *explicitly* convert them to Unicode before you can save them.

    Variable-length Unicode atoms do not accept parameters and they
    cause the reads of rows to always return Python Unicode strings.
    You can regard ``vlunicode`` atoms as an easy way to save variable
    length Unicode strings.
    """
    kind = 'vlunicode'
    type = 'vlunicode'
    base = UInt32Atom()

    if sys.maxunicode <= 0xffff:
        # When the Python build is UCS-2, we need to promote the
        # Unicode string to UCS-4.  We *must* use a 0-d array since
        # NumPy scalars inherit the UCS-2 encoding from Python (see
        # NumPy ticket #525).  Since ``_tobuffer()`` can't return an
        # array, we must override ``toarray()`` itself.
        def toarray(self, object_):
            if not isinstance(object_, basestring):
                raise TypeError("object is not a string: %r" % (object_,))
            ustr = unicode(object_)
            uarr = numpy.array(ustr, dtype='U')
            return numpy.ndarray(
                buffer=uarr, dtype=self.base.dtype, shape=len(ustr) )

    def _tobuffer(self, object_):
        # This works (and is used) only with UCS-4 builds of Python,
        # where the width of the internal representation of a
        # character matches that of the base atoms.
        if not isinstance(object_, basestring):
            raise TypeError("object is not a string: %r" % (object_,))
        return numpy.unicode0(object_)

    def fromarray(self, array):
        length = len(array)
        if length == 0:
            return u''  # ``array.view('U0')`` raises a `TypeError`
        return array.view('U%d' % length).item()

class ObjectAtom(_BufferedAtom):
    """
    Defines an atom of type ``object``.

    This class is meant to fit *any* kind of Python object in a row of
    a `VLArray` dataset by using ``cPickle`` behind the scenes.  Due
    to the fact that you can not foresee how long will be the output
    of the ``cPickle`` serialization (i.e. the atom already has a
    *variable* length), you can only fit *one object per row*.
    However, you can still group several objects in a single tuple or
    list and pass it to the `VLArray.append()` method.

    Object atoms do not accept parameters and they cause the reads of
    rows to always return Python objects.  You can regard ``object``
    atoms as an easy way to save an arbitrary number of generic Python
    objects in a `VLArray` dataset.
    """
    kind = 'object'
    type = 'object'
    base = UInt8Atom()

    def _tobuffer(self, object_):
        return cPickle.dumps(object_, cPickle.HIGHEST_PROTOCOL)

    def fromarray(self, array):
        # We have to check for an empty array because of a possible
        # bug in HDF5 which makes it claim that a dataset has one
        # record when in fact it is empty.
        if array.size == 0:
            return None
        return cPickle.loads(array.tostring())


# Main part
# =========
def _test():
    """Run ``doctest`` on this module."""
    import doctest
    doctest.testmod()

if __name__ == '__main__':
    _test()