Tools/unicode/makeunicodedata.py

#
# (re)generate unicode property and type databases
#
# this script converts a unicode 3.0 database file to
# Modules/unicodedata_db.h, Modules/unicodename_db.h,
# and Objects/unicodetype_db.h
#
# history:
# 2000-09-24 fl   created (based on bits and pieces from unidb)
# 2000-09-25 fl   merged tim's splitbin fixes, separate decomposition table
# 2000-09-25 fl   added character type table
# 2000-09-26 fl   added LINEBREAK, DECIMAL, and DIGIT flags/fields (2.0)
# 2000-11-03 fl   expand first/last ranges
# 2001-01-19 fl   added character name tables (2.1)
#
# written by Fredrik Lundh (fredrik@pythonware.com), September 2000
#

import sys

SCRIPT = sys.argv[0]
VERSION = "2.1"

UNICODE_DATA = "UnicodeData-Latest.txt"

CATEGORY_NAMES = [ "Cn", "Lu", "Ll", "Lt", "Mn", "Mc", "Me", "Nd",
    "Nl", "No", "Zs", "Zl", "Zp", "Cc", "Cf", "Cs", "Co", "Cn", "Lm",
    "Lo", "Pc", "Pd", "Ps", "Pe", "Pi", "Pf", "Po", "Sm", "Sc", "Sk",
    "So" ]

BIDIRECTIONAL_NAMES = [ "", "L", "LRE", "LRO", "R", "AL", "RLE", "RLO",
    "PDF", "EN", "ES", "ET", "AN", "CS", "NSM", "BN", "B", "S", "WS",
    "ON" ]

# note: should match definitions in Objects/unicodectype.c
ALPHA_MASK = 0x01
DECIMAL_MASK = 0x02
DIGIT_MASK = 0x04
LOWER_MASK = 0x08
LINEBREAK_MASK = 0x10
SPACE_MASK = 0x20
TITLE_MASK = 0x40
UPPER_MASK = 0x80

def maketables(trace=0):

    print "--- Reading", UNICODE_DATA, "..."

    unicode = UnicodeData(UNICODE_DATA)

    print len(filter(None, unicode.table)), "characters"

    makeunicodedata(unicode, trace)
    makeunicodetype(unicode, trace)
    makeunicodename(unicode, trace)

# --------------------------------------------------------------------
# unicode character properties

def makeunicodedata(unicode, trace):

    dummy = (0, 0, 0, 0)
    table = [dummy]
    cache = {0: dummy}
    index = [0] * len(unicode.chars)

    FILE = "Modules/unicodedata_db.h"

    print "--- Preparing", FILE, "..."

    # 1) database properties

    for char in unicode.chars:
        record = unicode.table[char]
        if record:
            # extract database properties
            category = CATEGORY_NAMES.index(record[2])
            combining = int(record[3])
            bidirectional = BIDIRECTIONAL_NAMES.index(record[4])
            mirrored = record[9] == "Y"
            item = (
                category, combining, bidirectional, mirrored
                )
            # add entry to index and item tables
            i = cache.get(item)
            if i is None:
                cache[item] = i = len(table)
                table.append(item)
            index[char] = i

    # 2) decomposition data

    # FIXME: <fl> using the encoding stuff from unidb would save
    # another 50k or so, but I'll leave that for 2.1...

    decomp_data = [""]
    decomp_index = [0] * len(unicode.chars)

    for char in unicode.chars:
        record = unicode.table[char]
        if record:
            if record[5]:
                try:
                    i = decomp_data.index(record[5])
                except ValueError:
                    i = len(decomp_data)
                    decomp_data.append(record[5])
            else:
                i = 0
            decomp_index[char] = i

    print len(table), "unique properties"
    print len(decomp_data), "unique decomposition entries"

    print "--- Writing", FILE, "..."

    fp = open(FILE, "w")
    print >>fp, "/* this file was generated by %s %s */" % (SCRIPT, VERSION)
    print >>fp
    print >>fp, "/* a list of unique database records */"
    print >>fp, \
          "const _PyUnicode_DatabaseRecord _PyUnicode_Database_Records[] = {"
    for item in table:
        print >>fp, "    {%d, %d, %d, %d}," % item
    print >>fp, "};"
    print >>fp

    # FIXME: <fl> the following tables could be made static, and
    # the support code moved into unicodedatabase.c

    print >>fp, "/* string literals */"
    print >>fp, "const char *_PyUnicode_CategoryNames[] = {"
    for name in CATEGORY_NAMES:
        print >>fp, "    \"%s\"," % name
    print >>fp, "    NULL"
    print >>fp, "};"

    print >>fp, "const char *_PyUnicode_BidirectionalNames[] = {"
    for name in BIDIRECTIONAL_NAMES:
        print >>fp, "    \"%s\"," % name
    print >>fp, "    NULL"
    print >>fp, "};"

    print >>fp, "static const char *decomp_data[] = {"
    for name in decomp_data:
        print >>fp, "    \"%s\"," % name
    print >>fp, "    NULL"
    print >>fp, "};"

    # split record index table
    index1, index2, shift = splitbins(index, trace)

    print >>fp, "/* index tables for the database records */"
    print >>fp, "#define SHIFT", shift
    Array("index1", index1).dump(fp)
    Array("index2", index2).dump(fp)

    # split decomposition index table
    index1, index2, shift = splitbins(decomp_index, trace)

    print >>fp, "/* index tables for the decomposition data */"
    print >>fp, "#define DECOMP_SHIFT", shift
    Array("decomp_index1", index1).dump(fp)
    Array("decomp_index2", index2).dump(fp)

    fp.close()

# --------------------------------------------------------------------
# unicode character type tables

def makeunicodetype(unicode, trace):

    FILE = "Objects/unicodetype_db.h"

    print "--- Preparing", FILE, "..."

    # extract unicode types
    dummy = (0, 0, 0, 0, 0, 0)
    table = [dummy]
    cache = {0: dummy}
    index = [0] * len(unicode.chars)

    for char in unicode.chars:
        record = unicode.table[char]
        if record:
            # extract database properties
            category = record[2]
            bidirectional = record[4]
            flags = 0
            if category in ["Lm", "Lt", "Lu", "Ll", "Lo"]:
                flags |= ALPHA_MASK
            if category == "Ll":
                flags |= LOWER_MASK
            if category == "Zl" or bidirectional == "B":
                flags |= LINEBREAK_MASK
            if category == "Zs" or bidirectional in ("WS", "B", "S"):
                flags |= SPACE_MASK
            if category == "Lt":
                flags |= TITLE_MASK
            if category == "Lu":
                flags |= UPPER_MASK
            # use delta predictor for upper/lower/title
            if record[12]:
                upper = (int(record[12], 16) - char) & 0xffff
            else:
                upper = 0
            if record[13]:
                lower = (int(record[13], 16) - char) & 0xffff
            else:
                lower = 0
            if record[14]:
                title = (int(record[14], 16) - char) & 0xffff
            else:
                title = 0
            # decimal digit, integer digit
            decimal = 0
            if record[6]:
                flags |= DECIMAL_MASK
                decimal = int(record[6])
            digit = 0
            if record[7]:
                flags |= DIGIT_MASK
                digit = int(record[7])
            item = (
                flags, upper, lower, title, decimal, digit
                )
            # add entry to index and item tables
            i = cache.get(item)
            if i is None:
                cache[item] = i = len(table)
                table.append(item)
            index[char] = i

    print len(table), "unique character type entries"

    print "--- Writing", FILE, "..."

    fp = open(FILE, "w")
    print >>fp, "/* this file was generated by %s %s */" % (SCRIPT, VERSION)
    print >>fp
    print >>fp, "/* a list of unique character type descriptors */"
    print >>fp, "const _PyUnicode_TypeRecord _PyUnicode_TypeRecords[] = {"
    for item in table:
        print >>fp, "    {%d, %d, %d, %d, %d, %d}," % item
    print >>fp, "};"
    print >>fp

    # split decomposition index table
    index1, index2, shift = splitbins(index, trace)

    print >>fp, "/* type indexes */"
    print >>fp, "#define SHIFT", shift
    Array("index1", index1).dump(fp)
    Array("index2", index2).dump(fp)

    fp.close()

# --------------------------------------------------------------------
# unicode name database

def makeunicodename(unicode, trace):

    FILE = "Modules/unicodename_db.h"

    print "--- Preparing", FILE, "..."

    # collect names
    names = [None] * len(unicode.chars)

    for char in unicode.chars:
        record = unicode.table[char]
        if record:
            name = record[1].strip()
            if name and name[0] != "<":
                names[char] = name + chr(0)

    print len(filter(lambda n: n is not None, names)), "distinct names"

    # collect unique words from names (note that we differ between
    # words inside a sentence, and words ending a sentence.  the
    # latter includes the trailing null byte.

    words = {}
    n = b = 0
    for char in unicode.chars:
        name = names[char]
        if name:
            w = name.split()
            b = b + len(name)
            n = n + len(w)
            for w in w:
                l = words.get(w)
                if l:
                    l.append(None)
                else:
                    words[w] = [len(words)]

    print n, "words in text;", b, "bytes"

    wordlist = words.items()

    # sort on falling frequency
    wordlist.sort(lambda a, b: len(b[1])-len(a[1]))

    # statistics
    n = 0
    for i in range(128):
        n = n + len(wordlist[i][1])
    print n, "short words (7-bit indices)"

    # pick the 128 most commonly used words, and sort the rest on
    # falling length (to maximize overlap)

    wordlist, wordtail = wordlist[:128], wordlist[128:]
    wordtail.sort(lambda a, b: len(b[0])-len(a[0]))
    wordlist.extend(wordtail)

    # generate lexicon from words

    lexicon_offset = [0]
    lexicon = ""
    words = {}

    # build a lexicon string
    offset = 0
    for w, x in wordlist:
        # encoding: bit 7 indicates last character in word (chr(128)
        # indicates the last character in an entire string)
        ww = w[:-1] + chr(ord(w[-1])+128)
        # reuse string tails, when possible
        o = string.find(lexicon, ww)
        if o < 0:
            o = offset
            lexicon = lexicon + ww
            offset = offset + len(w)
        words[w] = len(lexicon_offset)
        lexicon_offset.append(offset)

    print len(words), "words in lexicon;", len(lexicon), "bytes"

    assert len(words) < 32768 # 15-bit word indices

    lexicon = map(ord, lexicon)

    # generate phrasebook from names and lexicon
    phrasebook = [0]
    phrasebook_offset = [0] * len(unicode.chars)
    for char in unicode.chars:
        name = names[char]
        if name:
            w = name.split()
            phrasebook_offset[char] = len(phrasebook)
            for w in w:
                i = words[w]
                if i < 128:
                    phrasebook.append(128+i)
                else:
                    phrasebook.append(i>>8)
                    phrasebook.append(i&255)

    #
    # unicode name hash table

    # extract names
    data = []
    for char in unicode.chars:
        record = unicode.table[char]
        if record:
            name = record[1].strip()
            if name and name[0] != "<":
                data.append((name, char))

    # the magic number 47 was chosen to minimize the number of
    # collisions on the current data set.  if you like, change it
    # and see what happens...

    codehash = Hash("code", data, 47)

    print "--- Writing", FILE, "..."

    fp = open(FILE, "w")
    print >>fp, "/* this file was generated by %s %s */" % (SCRIPT, VERSION)
    print >>fp
    print >>fp, "#define NAME_MAXLEN", 256
    print >>fp
    print >>fp, "/* lexicon */"
    Array("lexicon", lexicon).dump(fp)
    Array("lexicon_offset", lexicon_offset).dump(fp)

    # split decomposition index table
    offset1, offset2, shift = splitbins(phrasebook_offset, trace)

    print >>fp, "/* code->name phrasebook */"
    print >>fp, "#define phrasebook_shift", shift

    Array("phrasebook", phrasebook).dump(fp)
    Array("phrasebook_offset1", offset1).dump(fp)
    Array("phrasebook_offset2", offset2).dump(fp)

    print >>fp, "/* name->code dictionary */"
    codehash.dump(fp)

    fp.close()

# --------------------------------------------------------------------
# the following support code is taken from the unidb utilities
# Copyright (c) 1999-2000 by Secret Labs AB

# load a unicode-data file from disk

import string, sys

class UnicodeData:

    def __init__(self, filename, expand=1):
        file = open(filename)
        table = [None] * 65536
        while 1:
            s = file.readline()
            if not s:
                break
            s = string.split(string.strip(s), ";")
            char = string.atoi(s[0], 16)
            table[char] = s

        # expand first-last ranges (ignore surrogates and private use)
        if expand:
            field = None
            for i in range(0, 0xD800):
                s = table[i]
                if s:
                    if s[1][-6:] == "First>":
                        s[1] = ""
                        field = s[:]
                    elif s[1][-5:] == "Last>":
                        s[1] = ""
                        field = None
                elif field:
                    field[0] = hex(i)
                    table[i] = field

        # public attributes
        self.filename = filename
        self.table = table
        self.chars = range(65536) # unicode

    def uselatin1(self):
        # restrict character range to ISO Latin 1
        self.chars = range(256)

# hash table tools

# this is a straight-forward reimplementation of Python's built-in
# dictionary type, using a static data structure, and a custom string
# hash algorithm.

def myhash(s, magic):
    h = 0
    for c in map(ord, string.upper(s)):
        h = (h * magic) + c
        ix = h & 0xff000000
        if ix:
            h = (h ^ ((ix>>24) & 0xff)) & 0x00ffffff
    return h

SIZES = [
    (4,3), (8,3), (16,3), (32,5), (64,3), (128,3), (256,29), (512,17),
    (1024,9), (2048,5), (4096,83), (8192,27), (16384,43), (32768,3),
    (65536,45), (131072,9), (262144,39), (524288,39), (1048576,9),
    (2097152,5), (4194304,3), (8388608,33), (16777216,27)
]

class Hash:
    def __init__(self, name, data, magic):
        # turn a (key, value) list into a static hash table structure

        # determine table size
        for size, poly in SIZES:
            if size > len(data):
                poly = size + poly
                break
        else:
            raise AssertionError, "ran out of polynominals"

        print size, "slots in hash table"

        table = [None] * size

        mask = size-1

        n = 0

        hash = myhash

        # initialize hash table
        for key, value in data:
            h = hash(key, magic)
            i = (~h) & mask
            v = table[i]
            if v is None:
                table[i] = value
                continue
            incr = (h ^ (h >> 3)) & mask;
            if not incr:
                incr = mask
            while 1:
                n = n + 1
                i = (i + incr) & mask
                v = table[i]
                if v is None:
                    table[i] = value
                    break
                incr = incr << 1
                if incr > mask:
                    incr = incr ^ poly

        print n, "collisions"
        self.collisions = n

        for i in range(len(table)):
            if table[i] is None:
                table[i] = 0

        self.data = Array(name + "_hash", table)
        self.magic = magic
        self.name = name
        self.size = size
        self.poly = poly

    def dump(self, file):
        # write data to file, as a C array
        self.data.dump(file)
        file.write("#define %s_magic %d\n" % (self.name, self.magic))
        file.write("#define %s_size %d\n" % (self.name, self.size))
        file.write("#define %s_poly %d\n" % (self.name, self.poly))

# stuff to deal with arrays of unsigned integers

class Array:

    def __init__(self, name, data):
        self.name = name
        self.data = data

    def dump(self, file):
        # write data to file, as a C array
        size = getsize(self.data)
        # print >>sys.stderr, self.name+":", size*len(self.data), "bytes"
        file.write("static ")
        if size == 1:
            file.write("unsigned char")
        elif size == 2:
            file.write("unsigned short")
        else:
            file.write("unsigned int")
        file.write(" " + self.name + "[] = {\n")
        if self.data:
            s = "    "
            for item in self.data:
                i = str(item) + ", "
                if len(s) + len(i) > 78:
                    file.write(s + "\n")
                    s = "    " + i
                else:
                    s = s + i
            if string.strip(s):
                file.write(s + "\n")
        file.write("};\n\n")

def getsize(data):
    # return smallest possible integer size for the given array
    maxdata = max(data)
    if maxdata < 256:
        return 1
    elif maxdata < 65536:
        return 2
    else:
        return 4

def splitbins(t, trace=0):
    """t, trace=0 -> (t1, t2, shift).  Split a table to save space.

    t is a sequence of ints.  This function can be useful to save space if
    many of the ints are the same.  t1 and t2 are lists of ints, and shift
    is an int, chosen to minimize the combined size of t1 and t2 (in C
    code), and where for each i in range(len(t)),
        t[i] == t2[(t1[i >> shift] << shift) + (i & mask)]
    where mask is a bitmask isolating the last "shift" bits.

    If optional arg trace is non-zero (default zero), progress info
    is printed to sys.stderr.  The higher the value, the more info
    you'll get.
    """

    import sys
    if trace:
        def dump(t1, t2, shift, bytes):
            print >>sys.stderr, "%d+%d bins at shift %d; %d bytes" % (
                len(t1), len(t2), shift, bytes)
        print >>sys.stderr, "Size of original table:", len(t)*getsize(t), \
                            "bytes"
    n = len(t)-1    # last valid index
    maxshift = 0    # the most we can shift n and still have something left
    if n > 0:
        while n >> 1:
            n >>= 1
            maxshift += 1
    del n
    bytes = sys.maxint  # smallest total size so far
    t = tuple(t)    # so slices can be dict keys
    for shift in range(maxshift + 1):
        t1 = []
        t2 = []
        size = 2**shift
        bincache = {}
        for i in range(0, len(t), size):
            bin = t[i:i+size]
            index = bincache.get(bin)
            if index is None:
                index = len(t2)
                bincache[bin] = index
                t2.extend(bin)
            t1.append(index >> shift)
        # determine memory size
        b = len(t1)*getsize(t1) + len(t2)*getsize(t2)
        if trace > 1:
            dump(t1, t2, shift, b)
        if b < bytes:
            best = t1, t2, shift
            bytes = b
    t1, t2, shift = best
    if trace:
        print >>sys.stderr, "Best:",
        dump(t1, t2, shift, bytes)
    if __debug__:
        # exhaustively verify that the decomposition is correct
        mask = ~((~0) << shift) # i.e., low-bit mask of shift bits
        for i in xrange(len(t)):
            assert t[i] == t2[(t1[i >> shift] << shift) + (i & mask)]
    return best

if __name__ == "__main__":
    maketables(1)