# Licensed under a 3-clause BSD style license - see LICENSE.rst
"""
This module handles the conversion of various VOTABLE datatypes
to/from TABLEDATA_ and BINARY_ formats.
"""

from __future__ import absolute_import, division, print_function, unicode_literals
from ...extern import six
from ...extern.six.moves import xrange

# STDLIB
import re
import sys
from struct import unpack as _struct_unpack
from struct import pack as _struct_pack

# THIRD-PARTY
import numpy as np
from numpy import ma

# ASTROPY
from ...utils.xml.writer import xml_escape_cdata

# LOCAL
from .exceptions import (vo_raise, vo_warn, warn_or_raise, W01,
    W30, W31, W39, W46, W47, W49, W51, E01, E02, E03, E04, E05, E06)


__all__ = ['get_converter', 'Converter', 'table_column_to_votable_datatype']


pedantic_array_splitter = re.compile(r" +")
array_splitter = re.compile(r"\s+|(?:\s*,\s*)")
"""
A regex to handle splitting values on either whitespace or commas.

SPEC: Usage of commas is not actually allowed by the spec, but many
files in the wild use them.
"""

_zero_int = b'\0\0\0\0'
_empty_bytes = b''
_zero_byte = b'\0'


if six.PY3:
    struct_unpack = _struct_unpack
    struct_pack = _struct_pack
else:
    def struct_unpack(format, s):
        return _struct_unpack(format.encode('ascii'), s)

    def struct_pack(format, *args):
        return _struct_pack(format.encode('ascii'), *args)


if sys.byteorder == 'little':
    def _ensure_bigendian(x):
        if x.dtype.byteorder != '>':
            return x.byteswap()
        return x
else:
    def _ensure_bigendian(x):
        if x.dtype.byteorder == '<':
            return x.byteswap()
        return x


def _make_masked_array(data, mask):
    """
    Masked arrays of zero length that also have a mask of zero length
    cause problems in Numpy (at least in 1.6.2).  This function
    creates a masked array from data and a mask, unless it is zero
    length.
    """
    # np.ma doesn't like setting mask to []
    if len(data):
        return ma.array(
            np.array(data),
            mask=np.array(mask, dtype='bool'))
    else:
        return ma.array(np.array(data))


def bitarray_to_bool(data, length):
    """
    Converts a bit array (a string of bits in a bytes object) to a
    boolean Numpy array.

    Parameters
    ----------
    data : bytes
        The bit array.  The most significant byte is read first.

    length : int
        The number of bits to read.  The least significant bits in the
        data bytes beyond length will be ignored.

    Returns
    -------
    array : numpy bool array
    """
    results = []
    for byte in data:
        if not six.PY3:
            byte = ord(byte)
        for bit_no in range(7, -1, -1):
            bit = byte & (1 << bit_no)
            bit = (bit != 0)
            results.append(bit)
            if len(results) == length:
                break
        if len(results) == length:
            break

    return np.array(results, dtype='b1')


def bool_to_bitarray(value):
    """
    Converts a numpy boolean array to a bit array (a string of bits in
    a bytes object).

    Parameters
    ----------
    value : numpy bool array

    Returns
    -------
    bit_array : bytes
        The first value in the input array will be the most
        significant bit in the result.  The length will be `floor((N +
        7) / 8)` where `N` is the length of `value`.
    """
    value = value.flat
    bit_no = 7
    byte = 0
    bytes = []
    for v in value:
        if v:
            byte |= 1 << bit_no
        if bit_no == 0:
            bytes.append(byte)
            bit_no = 7
            byte = 0
        else:
            bit_no -= 1
    if bit_no != 7:
        bytes.append(byte)

    return struct_pack("%sB" % len(bytes), *bytes)


class Converter(object):
    """
    The base class for all converters.  Each subclass handles
    converting a specific VOTABLE data type to/from the TABLEDATA_ and
    BINARY_ on-disk representations.

    Parameters
    ----------
    field : `~astropy.io.votable.tree.Field`
        object describing the datatype

    config : dict
        The parser configuration dictionary

    pos : tuple
        The position in the XML file where the FIELD object was
        found.  Used for error messages.

    """
    def __init__(self, field, config=None, pos=None):
        pass

    @staticmethod
    def _parse_length(read):
        return struct_unpack(">I", read(4))[0]

    @staticmethod
    def _write_length(length):
        return struct_pack(">I", int(length))

    def supports_empty_values(self, config):
        """
        Returns True when the field can be completely empty.
        """
        return config.get('version_1_3_or_later')

    def parse(self, value, config=None, pos=None):
        """
        Convert the string *value* from the TABLEDATA_ format into an
        object with the correct native in-memory datatype and mask flag.

        Parameters
        ----------
        value : str
            value in TABLEDATA format

        Returns
        -------
        native : tuple (value, mask)
            The value as a Numpy array or scalar, and *mask* is True
            if the value is missing.
        """
        raise NotImplementedError(
            "This datatype must implement a 'parse' method.")

    def parse_scalar(self, value, config=None, pos=None):
        """
        Parse a single scalar of the underlying type of the converter.
        For non-array converters, this is equivalent to parse.  For
        array converters, this is used to parse a single
        element of the array.

        Parameters
        ----------
        value : str
            value in TABLEDATA format

        Returns
        -------
        native : tuple (value, mask)
            The value as a Numpy array or scalar, and *mask* is True
            if the value is missing.
        """
        return self.parse(value, config, pos)

    def output(self, value, mask):
        """
        Convert the object *value* (in the native in-memory datatype)
        to a unicode string suitable for serializing in the TABLEDATA_
        format.

        Parameters
        ----------
        value : native type corresponding to this converter
            The value

        mask : bool
            If `True`, will return the string representation of a
            masked value.

        Returns
        -------
        tabledata_repr : unicode
        """
        raise NotImplementedError(
            "This datatype must implement a 'output' method.")

    def binparse(self, read):
        """
        Reads some number of bytes from the BINARY_ format
        representation by calling the function *read*, and returns the
        native in-memory object representation for the datatype
        handled by *self*.

        Parameters
        ----------
        read : function
            A function that given a number of bytes, returns a byte
            string.

        Returns
        -------
        native : tuple (value, mask)
            The value as a Numpy array or scalar, and *mask* is True
            if the value is missing.
        """
        raise NotImplementedError(
            "This datatype must implement a 'binparse' method.")

    def binoutput(self, value, mask):
        """
        Convert the object *value* in the native in-memory datatype to
        a string of bytes suitable for serialization in the BINARY_
        format.

        Parameters
        ----------
        value : native type corresponding to this converter
            The value

        mask : bool
            If `True`, will return the string representation of a
            masked value.

        Returns
        -------
        bytes : byte string
            The binary representation of the value, suitable for
            serialization in the BINARY_ format.
        """
        raise NotImplementedError(
            "This datatype must implement a 'binoutput' method.")


class Char(Converter):
    """
    Handles the char datatype. (7-bit unsigned characters)

    Missing values are not handled for string or unicode types.
    """
    default = _empty_bytes

    def __init__(self, field, config=None, pos=None):
        if config is None:
            config = {}

        Converter.__init__(self, field, config, pos)

        if field.arraysize is None:
            vo_warn(W47, (), config, pos)
            field.arraysize = '1'

        if field.arraysize == '*':
            self.format = 'O'
            self.binparse = self._binparse_var
            self.binoutput = self._binoutput_var
            self.arraysize = '*'
        else:
            if field.arraysize.endswith('*'):
                field.arraysize = field.arraysize[:-1]
            try:
                self.arraysize = int(field.arraysize)
            except ValueError:
                vo_raise(E01, (field.arraysize, 'char', field.ID), config)
            self.format = 'S%d' % self.arraysize
            self.binparse = self._binparse_fixed
            self.binoutput = self._binoutput_fixed
            self._struct_format = ">%ds" % self.arraysize

        if config.get('pedantic'):
            self.parse = self._ascii_parse
        else:
            self.parse = self._str_parse

    def supports_empty_values(self, config):
        return True

    def _ascii_parse(self, value, config=None, pos=None):
        if self.arraysize != '*' and len(value) > self.arraysize:
            vo_warn(W46, ('char', self.arraysize), config, pos)
        return value.encode('ascii'), False

    def _str_parse(self, value, config=None, pos=None):
        if self.arraysize != '*' and len(value) > self.arraysize:
            vo_warn(W46, ('char', self.arraysize), config, pos)
        return value.encode('utf-8'), False

    def output(self, value, mask):
        if mask:
            return ''
        if not isinstance(value, str):
            value = value.decode('ascii')
        return xml_escape_cdata(value)

    def _binparse_var(self, read):
        length = self._parse_length(read)
        return read(length), False

    def _binparse_fixed(self, read):
        s = struct_unpack(self._struct_format, read(self.arraysize))[0]
        end = s.find(_zero_byte)
        if end != -1:
            return s[:end], False
        return s, False

    def _binoutput_var(self, value, mask):
        if mask or value is None or value == '':
            return _zero_int
        return self._write_length(len(value)) + value

    def _binoutput_fixed(self, value, mask):
        if mask:
            value = _empty_bytes
        return struct_pack(self._struct_format, value)


class UnicodeChar(Converter):
    """
    Handles the unicodeChar data type. UTF-16-BE.

    Missing values are not handled for string or unicode types.
    """
    default = ''

    def __init__(self, field, config=None, pos=None):
        Converter.__init__(self, field, config, pos)

        if field.arraysize is None:
            vo_warn(W47, (), config, pos)
            field.arraysize = '1'

        if field.arraysize == '*':
            self.format = 'O'
            self.binparse = self._binparse_var
            self.binoutput = self._binoutput_var
            self.arraysize = '*'
        else:
            try:
                self.arraysize = int(field.arraysize)
            except ValueError:
                vo_raise(E01, (field.arraysize, 'unicode', field.ID), config)
            self.format = 'U%d' % self.arraysize
            self.binparse = self._binparse_fixed
            self.binoutput = self._binoutput_fixed
            self._struct_format = ">%ds" % (self.arraysize * 2)

    def parse(self, value, config=None, pos=None):
        if self.arraysize != '*' and len(value) > self.arraysize:
            vo_warn(W46, ('unicodeChar', self.arraysize), config, pos)
        return value, False

    def output(self, value, mask):
        if mask:
            return ''
        return xml_escape_cdata(six.text_type(value))

    def _binparse_var(self, read):
        length = self._parse_length(read)
        return read(length * 2).decode('utf_16_be'), False

    def _binparse_fixed(self, read):
        s = struct_unpack(self._struct_format, read(self.arraysize * 2))[0]
        s = s.decode('utf_16_be')
        end = s.find('\0')
        if end != -1:
            return s[:end], False
        return s, False

    def _binoutput_var(self, value, mask):
        if mask or value is None or value == '':
            return _zero_int
        encoded = value.encode('utf_16_be')
        return self._write_length(len(encoded) / 2) + encoded

    def _binoutput_fixed(self, value, mask):
        if mask:
            value = ''
        return struct_pack(self._struct_format, value.encode('utf_16_be'))


class Array(Converter):
    """
    Handles both fixed and variable-lengths arrays.
    """
    def __init__(self, field, config=None, pos=None):
        if config is None:
            config = {}
        Converter.__init__(self, field, config, pos)
        if config.get('pedantic'):
            self._splitter = self._splitter_pedantic
        else:
            self._splitter = self._splitter_lax

    def parse_scalar(self, value, config=None, pos=0):
        return self._base.parse_scalar(value, config, pos)

    @staticmethod
    def _splitter_pedantic(value, config=None, pos=None):
        return pedantic_array_splitter.split(value)

    @staticmethod
    def _splitter_lax(value, config=None, pos=None):
        if ',' in value:
            vo_warn(W01, (), config, pos)
        return array_splitter.split(value)


class VarArray(Array):
    """
    Handles variable lengths arrays (i.e. where *arraysize* is '*').
    """
    format = 'O'

    def __init__(self, field, base, arraysize, config=None, pos=None):
        Array.__init__(self, field, config)

        self._base = base
        self.default = np.array([], dtype=self._base.format)

    def output(self, value, mask):
        output = self._base.output
        result = [output(x, m) for x, m in np.broadcast(value, mask)]
        return ' '.join(result)

    def binparse(self, read):
        length = self._parse_length(read)

        result = []
        result_mask = []
        binparse = self._base.binparse
        for i in xrange(length):
            val, mask = binparse(read)
            result.append(val)
            result_mask.append(mask)

        return _make_masked_array(result, result_mask), False

    def binoutput(self, value, mask):
        if value is None or len(value) == 0:
            return _zero_int

        length = len(value)
        result = [self._write_length(length)]
        binoutput = self._base.binoutput
        for x, m in zip(value, value.mask):
            result.append(binoutput(x, m))
        return _empty_bytes.join(result)


class ArrayVarArray(VarArray):
    """
    Handles an array of variable-length arrays, i.e. where *arraysize*
    ends in '*'.
    """
    def parse(self, value, config=None, pos=None):
        if value.strip() == '':
            return ma.array([]), False

        parts = self._splitter(value, config, pos)
        items = self._base._items
        parse_parts = self._base.parse_parts
        if len(parts) % items != 0:
            vo_raise(E02, (items, len(parts)), config, pos)
        result = []
        result_mask = []
        for i in xrange(0, len(parts), items):
            value, mask = parse_parts(parts[i:i+items], config, pos)
            result.append(value)
            result_mask.append(mask)

        return _make_masked_array(result, result_mask), False


class ScalarVarArray(VarArray):
    """
    Handles a variable-length array of numeric scalars.
    """
    def parse(self, value, config=None, pos=None):
        if value.strip() == '':
            return ma.array([]), False

        parts = self._splitter(value, config, pos)

        parse = self._base.parse
        result = []
        result_mask = []
        for x in parts:
            value, mask = parse(x, config, pos)
            result.append(value)
            result_mask.append(mask)

        return _make_masked_array(result, result_mask), False


class NumericArray(Array):
    """
    Handles a fixed-length array of numeric scalars.
    """
    vararray_type = ArrayVarArray

    def __init__(self, field, base, arraysize, config=None, pos=None):
        Array.__init__(self, field, config, pos)

        self._base = base
        self._arraysize = arraysize
        self.format = "%s%s" % (tuple(arraysize), base.format)

        self._items = 1
        for dim in arraysize:
            self._items *= dim

        self._memsize = np.dtype(self.format).itemsize
        self._bigendian_format = '>' + self.format

        self.default = np.empty(arraysize, dtype=self._base.format)
        self.default[...] = self._base.default

    def parse(self, value, config=None, pos=None):
        if config is None:
            config = {}
        if config['version_1_3_or_later'] and value == '':
            return np.zeros(self._arraysize, dtype=self._base.format), True
        parts = self._splitter(value, config, pos)
        if len(parts) != self._items:
            warn_or_raise(E02, E02, (self._items, len(parts)), config, pos)
        if config.get('pedantic'):
            return self.parse_parts(parts, config, pos)
        else:
            if len(parts) == self._items:
                pass
            elif len(parts) > self._items:
                parts = parts[:self._items]
            else:
                parts = (parts +
                         ([self._base.default] * (self._items - len(parts))))
            return self.parse_parts(parts, config, pos)

    def parse_parts(self, parts, config=None, pos=None):
        base_parse = self._base.parse
        result = []
        result_mask = []
        for x in parts:
            value, mask = base_parse(x, config, pos)
            result.append(value)
            result_mask.append(mask)
        result = np.array(result, dtype=self._base.format).reshape(
            self._arraysize)
        result_mask = np.array(result_mask, dtype='bool').reshape(
            self._arraysize)
        return result, result_mask

    def output(self, value, mask):
        base_output = self._base.output
        value = np.asarray(value)
        mask = np.asarray(mask)
        return ' '.join(base_output(x, m) for x, m in
                         zip(value.flat, mask.flat))

    def binparse(self, read):
        result = np.fromstring(read(self._memsize),
                               dtype=self._bigendian_format)[0]
        result_mask = self._base.is_null(result)
        return result, result_mask

    def binoutput(self, value, mask):
        filtered = self._base.filter_array(value, mask)
        filtered = _ensure_bigendian(filtered)
        return filtered.tostring()


class Numeric(Converter):
    """
    The base class for all numeric data types.
    """
    array_type = NumericArray
    vararray_type = ScalarVarArray
    null = None

    def __init__(self, field, config=None, pos=None):
        Converter.__init__(self, field, config, pos)

        self._memsize = np.dtype(self.format).itemsize
        self._bigendian_format = '>' + self.format
        if field.values.null is not None:
            self.null = np.asarray(field.values.null, dtype=self.format)
            self.default = self.null
            self.is_null = self._is_null
        else:
            self.is_null = np.isnan

    def binparse(self, read):
        result = np.fromstring(read(self._memsize),
                               dtype=self._bigendian_format)
        return result[0], self.is_null(result[0])

    def _is_null(self, value):
        return value == self.null


class FloatingPoint(Numeric):
    """
    The base class for floating-point datatypes.
    """
    default = np.nan

    def __init__(self, field, config=None, pos=None):
        if config is None:
            config = {}

        Numeric.__init__(self, field, config, pos)

        precision = field.precision
        width = field.width
        format_parts = ['%']

        if width is not None:
            format_parts.append(six.text_type(width))

        if precision is None:
            format_parts.append('r')
        elif precision.startswith("E"):
            format_parts.append('.%dg' % int(precision[1:]))
        elif precision.startswith("F"):
            format_parts.append('.%df' % int(precision[1:]))
        else:
            format_parts.append('.%df' % int(precision))

        self._output_format = ''.join(format_parts)

        self.nan = np.array(np.nan, self.format)

        if self.null is None:
            self._null_output = 'NaN'
            self._null_binoutput = self.binoutput(self.nan, False)
            self.filter_array = self._filter_nan
        else:
            self._null_output = self.output(np.asarray(self.null), False)
            self._null_binoutput = self.binoutput(np.asarray(self.null), False)
            self.filter_array = self._filter_null

        if config.get('pedantic'):
            self.parse = self._parse_pedantic
        else:
            self.parse = self._parse_permissive

    def supports_empty_values(self, config):
        return True

    def _parse_pedantic(self, value, config=None, pos=None):
        if value.strip() == '':
            return self.null, True
        f = float(value)
        return f, self.is_null(f)

    def _parse_permissive(self, value, config=None, pos=None):
        try:
            f = float(value)
            return f, self.is_null(f)
        except ValueError:
            # IRSA VOTables use the word 'null' to specify empty values,
            # but this is not defined in the VOTable spec.
            if value.strip() != '':
                vo_warn(W30, value, config, pos)
            return self.null, True

    @property
    def output_format(self):
        return self._output_format

    def output(self, value, mask):
        if mask:
            return self._null_output
        if np.isfinite(value):
            if not np.isscalar(value):
                value = value.dtype.type(value)
            result = self._output_format % value
            if result.startswith('array'):
                raise RuntimeError()
            if (self._output_format[-1] == 'r' and
                result.endswith('.0')):
                result = result[:-2]
            return result
        elif np.isnan(value):
            return 'NaN'
        elif np.isposinf(value):
            return '+InF'
        elif np.isneginf(value):
            return '-InF'
        # Should never raise
        vo_raise("Invalid floating point value '%s'" % value)

    def binoutput(self, value, mask):
        if mask:
            return self._null_binoutput

        value = _ensure_bigendian(value)
        return value.tostring()

    def _filter_nan(self, value, mask):
        return np.where(mask, np.nan, value)

    def _filter_null(self, value, mask):
        return np.where(mask, self.null, value)


class Double(FloatingPoint):
    """
    Handles the double datatype.  Double-precision IEEE
    floating-point.
    """
    format = 'f8'


class Float(FloatingPoint):
    """
    Handles the float datatype.  Single-precision IEEE floating-point.
    """
    format = 'f4'


class Integer(Numeric):
    """
    The base class for all the integral datatypes.
    """
    default = 0

    def __init__(self, field, config=None, pos=None):
        Numeric.__init__(self, field, config, pos)

    def parse(self, value, config=None, pos=None):
        if config is None:
            config = {}
        mask = False
        if isinstance(value, six.string_types):
            value = value.lower()
            if value == '':
                if config['version_1_3_or_later']:
                    mask = True
                else:
                    warn_or_raise(W49, W49, (), config, pos)
                if self.null is not None:
                    value = self.null
                else:
                    value = self.default
            elif value == 'nan':
                mask = True
                if self.null is None:
                    warn_or_raise(W31, W31, (), config, pos)
                    value = self.default
                else:
                    value = self.null
            elif value.startswith('0x'):
                value = int(value[2:], 16)
            else:
                value = int(value, 10)
        else:
            value = int(value)
        if self.null is not None and value == self.null:
            mask = True

        if value < self.val_range[0]:
            warn_or_raise(W51, W51, (value, self.bit_size), config, pos)
            value = self.val_range[0]
        elif value > self.val_range[1]:
            warn_or_raise(W51, W51, (value, self.bit_size), config, pos)
            value = self.val_range[1]

        return value, mask

    def output(self, value, mask):
        if mask:
            if self.null is None:
                warn_or_raise(W31, W31)
                return 'NaN'
            return six.text_type(self.null)
        return six.text_type(value)

    def binoutput(self, value, mask):
        if mask:
            if self.null is None:
                vo_raise(W31)
            else:
                value = self.null

        value = _ensure_bigendian(value)
        return value.tostring()

    def filter_array(self, value, mask):
        if np.any(mask):
            if self.null is not None:
                return np.where(mask, self.null, value)
            else:
                vo_raise(W31)
        return value


class UnsignedByte(Integer):
    """
    Handles the unsignedByte datatype.  Unsigned 8-bit integer.
    """
    format = 'u1'
    val_range = (0, 255)
    bit_size = '8-bit unsigned'


class Short(Integer):
    """
    Handles the short datatype.  Signed 16-bit integer.
    """
    format = 'i2'
    val_range = (-32768, 32767)
    bit_size = '16-bit'


class Int(Integer):
    """
    Handles the int datatype.  Signed 32-bit integer.
    """
    format = 'i4'
    val_range = (-2147483648, 2147483647)
    bit_size = '32-bit'


class Long(Integer):
    """
    Handles the long datatype.  Signed 64-bit integer.
    """
    format = 'i8'
    val_range = (-9223372036854775808, 9223372036854775807)
    bit_size = '64-bit'


class ComplexArrayVarArray(VarArray):
    """
    Handles an array of variable-length arrays of complex numbers.
    """
    def parse(self, value, config=None, pos=None):
        if value.strip() == '':
            return ma.array([]), True

        parts = self._splitter(value, config, pos)
        items = self._base._items
        parse_parts = self._base.parse_parts
        if len(parts) % items != 0:
            vo_raise(E02, (items, len(parts)), config, pos)
        result = []
        result_mask = []
        for i in xrange(0, len(parts), items):
            value, mask = parse_parts(parts[i:i + items], config, pos)
            result.append(value)
            result_mask.append(mask)

        return _make_masked_array(result, result_mask), False


class ComplexVarArray(VarArray):
    """
    Handles a variable-length array of complex numbers.
    """
    def parse(self, value, config=None, pos=None):
        if value.strip() == '':
            return ma.array([]), True

        parts = self._splitter(value, config, pos)
        parse_parts = self._base.parse_parts
        result = []
        result_mask = []
        for i in xrange(0, len(parts), 2):
            value = [float(x) for x in parts[i:i + 2]]
            value, mask = parse_parts(value, config, pos)
            result.append(value)
            result_mask.append(mask)

        return _make_masked_array(
            np.array(result, dtype=self._base.format), result_mask), False


class ComplexArray(NumericArray):
    """
    Handles a fixed-size array of complex numbers.
    """
    vararray_type = ComplexArrayVarArray

    def __init__(self, field, base, arraysize, config=None, pos=None):
        NumericArray.__init__(self, field, base, arraysize, config, pos)
        self._items *= 2

    def parse(self, value, config=None, pos=None):
        parts = self._splitter(value, config, pos)
        if parts == ['']:
            parts = []
        return self.parse_parts(parts, config, pos)

    def parse_parts(self, parts, config=None, pos=None):
        if len(parts) != self._items:
            vo_raise(E02, (self._items, len(parts)), config, pos)
        base_parse = self._base.parse_parts
        result = []
        result_mask = []
        for i in xrange(0, self._items, 2):
            value = [float(x) for x in parts[i:i + 2]]
            value, mask = base_parse(value, config, pos)
            result.append(value)
            result_mask.append(mask)
        result = np.array(
            result, dtype=self._base.format).reshape(self._arraysize)
        result_mask = np.array(
            result_mask, dtype='bool').reshape(self._arraysize)
        return result, result_mask


class Complex(FloatingPoint, Array):
    """
    The base class for complex numbers.
    """
    array_type = ComplexArray
    vararray_type = ComplexVarArray
    default = np.nan

    def __init__(self, field, config=None, pos=None):
        FloatingPoint.__init__(self, field, config, pos)
        Array.__init__(self, field, config, pos)

    def parse(self, value, config=None, pos=None):
        stripped = value.strip()
        if stripped == '' or stripped.lower() == 'nan':
            return np.nan, True
        splitter = self._splitter
        parts = [float(x) for x in splitter(value, config, pos)]
        if len(parts) != 2:
            vo_raise(E03, (value,), config, pos)
        return self.parse_parts(parts, config, pos)
    _parse_permissive = parse
    _parse_pedantic = parse

    def parse_parts(self, parts, config=None, pos=None):
        value = complex(*parts)
        return value, self.is_null(value)

    def output(self, value, mask):
        if mask:
            if self.null is None:
                return 'NaN'
            else:
                value = self.null
        real = self._output_format % float(value.real)
        imag = self._output_format % float(value.imag)
        if self._output_format[-1] == 'r':
            if real.endswith('.0'):
                real = real[:-2]
            if imag.endswith('.0'):
                imag = imag[:-2]
        return real + ' ' + imag


class FloatComplex(Complex):
    """
    Handle floatComplex datatype.  Pair of single-precision IEEE
    floating-point numbers.
    """
    format = 'c8'


class DoubleComplex(Complex):
    """
    Handle doubleComplex datatype.  Pair of double-precision IEEE
    floating-point numbers.
    """
    format = 'c16'


class BitArray(NumericArray):
    """
    Handles an array of bits.
    """
    vararray_type = ArrayVarArray

    def __init__(self, field, base, arraysize, config=None, pos=None):
        NumericArray.__init__(self, field, base, arraysize, config, pos)

        self._bytes = ((self._items - 1) // 8) + 1

    @staticmethod
    def _splitter_pedantic(value, config=None, pos=None):
        return list(re.sub('\s', '', value))

    @staticmethod
    def _splitter_lax(value, config=None, pos=None):
        if ',' in value:
            vo_warn(W01, (), config, pos)
        return list(re.sub('\s|,', '', value))

    def output(self, value, mask):
        if np.any(mask):
            vo_warn(W39)
        value = np.asarray(value)
        mapping = {False: '0', True: '1'}
        return ''.join(mapping[x] for x in value.flat)

    def binparse(self, read):
        data = read(self._bytes)
        result = bitarray_to_bool(data, self._items)
        result = result.reshape(self._arraysize)
        result_mask = np.zeros(self._arraysize, dtype='b1')
        return result, result_mask

    def binoutput(self, value, mask):
        if np.any(mask):
            vo_warn(W39)

        return bool_to_bitarray(value)


class Bit(Converter):
    """
    Handles the bit datatype.
    """
    format = 'b1'
    array_type = BitArray
    vararray_type = ScalarVarArray
    default = False
    binary_one = b'\x08'
    binary_zero = b'\0'

    def parse(self, value, config=None, pos=None):
        if config is None:
            config = {}
        mapping = {'1': True, '0': False}
        if value is False or value.strip() == '':
            if not config['version_1_3_or_later']:
                warn_or_raise(W49, W49, (), config, pos)
            return False, True
        else:
            try:
                return mapping[value], False
            except KeyError:
                vo_raise(E04, (value,), config, pos)

    def output(self, value, mask):
        if mask:
            vo_warn(W39)

        if value:
            return '1'
        else:
            return '0'

    def binparse(self, read):
        data = read(1)
        return (ord(data) & 0x8) != 0, False

    def binoutput(self, value, mask):
        if mask:
            vo_warn(W39)

        if value:
            return self.binary_one
        return self.binary_zero


class BooleanArray(NumericArray):
    """
    Handles an array of boolean values.
    """
    vararray_type = ArrayVarArray

    def binparse(self, read):
        data = read(self._items)
        binparse = self._base.binparse_value
        result = []
        result_mask = []
        for char in data:
            if not six.PY3:
                char = ord(char)
            value, mask = binparse(char)
            result.append(value)
            result_mask.append(mask)
        result = np.array(result, dtype='b1').reshape(
            self._arraysize)
        result_mask = np.array(result_mask, dtype='b1').reshape(
            self._arraysize)
        return result, result_mask

    def binoutput(self, value, mask):
        binoutput = self._base.binoutput
        value = np.asarray(value)
        mask = np.asarray(mask)
        result = [binoutput(x, m)
                  for x, m in np.broadcast(value.flat, mask.flat)]
        return _empty_bytes.join(result)


class Boolean(Converter):
    """
    Handles the boolean datatype.
    """
    format = 'b1'
    array_type = BooleanArray
    vararray_type = ScalarVarArray
    default = False
    binary_question_mark = b'?'
    binary_true = b'T'
    binary_false = b'F'

    def parse(self, value, config=None, pos=None):
        if value == '':
            return False, True
        if value is False:
            return False, True
        mapping = {'TRUE'  : (True, False),
                   'FALSE' : (False, False),
                   '1'     : (True, False),
                   '0'     : (False, False),
                   'T'     : (True, False),
                   'F'     : (False, False),
                   '\0'    : (False, True),
                   ' '     : (False, True),
                   '?'     : (False, True),
                   ''      : (False, True)}
        try:
            return mapping[value.upper()]
        except KeyError:
            vo_raise(E05, (value,), config, pos)

    def output(self, value, mask):
        if mask:
            return '?'
        if value:
            return 'T'
        return 'F'

    def binparse(self, read):
        value = ord(read(1))
        return self.binparse_value(value)

    _binparse_mapping = {
        ord('T')  : (True, False),
        ord('t')  : (True, False),
        ord('1')  : (True, False),
        ord('F')  : (False, False),
        ord('f')  : (False, False),
        ord('0')  : (False, False),
        ord('\0') : (False, True),
        ord(' ')  : (False, True),
        ord('?')  : (False, True)}

    def binparse_value(self, value):
        try:
            return self._binparse_mapping[value]
        except KeyError:
            vo_raise(E05, (value,))

    def binoutput(self, value, mask):
        if mask:
            return self.binary_question_mark
        if value:
            return self.binary_true
        return self.binary_false


converter_mapping = {
    'double'        : Double,
    'float'         : Float,
    'bit'           : Bit,
    'boolean'       : Boolean,
    'unsignedByte'  : UnsignedByte,
    'short'         : Short,
    'int'           : Int,
    'long'          : Long,
    'floatComplex'  : FloatComplex,
    'doubleComplex' : DoubleComplex,
    'char'          : Char,
    'unicodeChar'   : UnicodeChar }


def get_converter(field, config=None, pos=None):
    """
    Get an appropriate converter instance for a given field.

    Parameters
    ----------
    field : astropy.io.votable.tree.Field

    config : dict, optional
        Parser configuration dictionary

    pos : tuple
        Position in the input XML file.  Used for error messages.

    Returns
    -------
    converter : astropy.io.votable.converters.Converter
    """
    if config is None:
        config = {}

    if field.datatype not in converter_mapping:
        vo_raise(E06, (field.datatype, field.ID), config)

    cls = converter_mapping[field.datatype]
    converter = cls(field, config, pos)

    arraysize = field.arraysize

    # With numeric datatypes, special things need to happen for
    # arrays.
    if (field.datatype not in ('char', 'unicodeChar') and
        arraysize is not None):
        if arraysize[-1] == '*':
            arraysize = arraysize[:-1]
            last_x = arraysize.rfind('x')
            if last_x == -1:
                arraysize = ''
            else:
                arraysize = arraysize[:last_x]
            fixed = False
        else:
            fixed = True

        if arraysize != '':
            arraysize = [int(x) for x in arraysize.split("x")]
            arraysize.reverse()
        else:
            arraysize = []

        if arraysize != []:
            converter = converter.array_type(
                field, converter, arraysize, config)

        if not fixed:
            converter = converter.vararray_type(
                field, converter, arraysize, config)

    return converter


numpy_dtype_to_field_mapping = {
    np.float64().dtype.num    : 'double',
    np.float32().dtype.num    : 'float',
    np.bool_().dtype.num      : 'bit',
    np.uint8().dtype.num      : 'unsignedByte',
    np.int16().dtype.num      : 'short',
    np.int32().dtype.num      : 'int',
    np.int64().dtype.num      : 'long',
    np.complex64().dtype.num  : 'floatComplex',
    np.complex128().dtype.num : 'doubleComplex',
    np.unicode_().dtype.num   : 'unicodeChar'
}


numpy_dtype_to_field_mapping[np.bytes_().dtype.num] = 'char'


def _all_bytes(column):
    for x in column:
        if not isinstance(x, bytes):
            return False
    return True


def _all_unicode(column):
    for x in column:
        if not isinstance(x, six.text_type):
            return False
    return True


def _all_matching_dtype(column):
    first_dtype = False
    first_shape = ()
    for x in column:
        if not isinstance(x, np.ndarray) or len(x) == 0:
            continue

        if first_dtype is False:
            first_dtype = x.dtype
            first_shape = x.shape[1:]
        elif first_dtype != x.dtype:
            return False, ()
        elif first_shape != x.shape[1:]:
            first_shape = ()
    return first_dtype, first_shape


def numpy_to_votable_dtype(dtype, shape):
    """
    Converts a numpy dtype and shape to a dictionary of attributes for
    a VOTable FIELD element and correspond to that type.

    Parameters
    ----------
    dtype : Numpy dtype instance

    shape : tuple

    Returns
    -------
    attributes : dict
       A dict containing 'datatype' and 'arraysize' keys that can be
       set on a VOTable FIELD element.
    """
    if dtype.num not in numpy_dtype_to_field_mapping:
        raise TypeError(
            "{0!r} can not be represented in VOTable".format(dtype))

    if dtype.char == 'S':
        return {'datatype': 'char',
                'arraysize': str(dtype.itemsize)}
    elif dtype.char == 'U':
        return {'datatype': 'unicodeChar',
                'arraysize': str(dtype.itemsize // 4)}
    else:
        result = {
            'datatype': numpy_dtype_to_field_mapping[dtype.num]}
        if len(shape):
            result['arraysize'] = 'x'.join(str(x) for x in shape)

        return result


def table_column_to_votable_datatype(column):
    """
    Given a `astropy.table.Column` instance, returns the attributes
    necessary to create a VOTable FIELD element that corresponds to
    the type of the column.

    This necessarily must perform some heuristics to determine the
    type of variable length arrays fields, since they are not directly
    supported by Numpy.

    If the column has dtype of "object", it performs the following
    tests:

       - If all elements are byte or unicode strings, it creates a
         variable-length byte or unicode field, respectively.

       - If all elements are numpy arrays of the same dtype and with a
         consistent shape in all but the first dimension, it creates a
         variable length array of fixed sized arrays.  If the dtypes
         match, but the shapes do not, a variable length array is
         created.

    If the dtype of the input is not understood, it sets the data type
    to the most inclusive: a variable length unicodeChar array.

    Parameters
    ----------
    column : `astropy.table.Column` instance

    Returns
    -------
    attributes : dict
       A dict containing 'datatype' and 'arraysize' keys that can be
       set on a VOTable FIELD element.
    """
    if column.dtype.char == 'O':
        if isinstance(column[0], bytes):
            if _all_bytes(column[1:]):
                return {'datatype': 'char', 'arraysize': '*'}
        elif isinstance(column[0], six.text_type):
            if _all_unicode(column[1:]):
                return {'datatype': 'unicodeChar', 'arraysize': '*'}
        elif isinstance(column[0], np.ndarray):
            dtype, shape = _all_matching_dtype(column)
            if dtype is not False:
                result = numpy_to_votable_dtype(dtype, shape)
                if 'arraysize' not in result:
                    result['arraysize'] = '*'
                else:
                    result['arraysize'] += '*'
                return result

        # All bets are off, do the most generic thing
        return {'datatype': 'unicodeChar', 'arraysize': '*'}

    return numpy_to_votable_dtype(column.dtype, column.shape[1:])
