path: root/Lib/fontTools/cffLib/specializer.py

# -*- coding: utf-8 -*-

"""T2CharString operator specializer and generalizer.

PostScript glyph drawing operations can be expressed in multiple different
ways. For example, as well as the ``lineto`` operator, there is also a
``hlineto`` operator which draws a horizontal line, removing the need to
specify a ``dx`` coordinate, and a ``vlineto`` operator which draws a
vertical line, removing the need to specify a ``dy`` coordinate. As well
as decompiling :class:`fontTools.misc.psCharStrings.T2CharString` objects
into lists of operations, this module allows for conversion between general
and specific forms of the operation.

"""

from fontTools.cffLib import maxStackLimit


def stringToProgram(string):
    if isinstance(string, str):
        string = string.split()
    program = []
    for token in string:
        try:
            token = int(token)
        except ValueError:
            try:
                token = float(token)
            except ValueError:
                pass
        program.append(token)
    return program


def programToString(program):
    return " ".join(str(x) for x in program)


def programToCommands(program, getNumRegions=None):
    """Takes a T2CharString program list and returns list of commands.
    Each command is a two-tuple of commandname,arg-list.  The commandname might
    be empty string if no commandname shall be emitted (used for glyph width,
    hintmask/cntrmask argument, as well as stray arguments at the end of the
    program (🤷).
    'getNumRegions' may be None, or a callable object. It must return the
    number of regions. 'getNumRegions' takes a single argument, vsindex. If
    the vsindex argument is None, getNumRegions returns the default number
    of regions for the charstring, else it returns the numRegions for
    the vsindex.
    The Charstring may or may not start with a width value. If the first
    non-blend operator has an odd number of arguments, then the first argument is
    a width, and is popped off. This is complicated with blend operators, as
    there may be more than one before the first hint or moveto operator, and each
    one reduces several arguments to just one list argument. We have to sum the
    number of arguments that are not part of the blend arguments, and all the
    'numBlends' values. We could instead have said that by definition, if there
    is a blend operator, there is no width value, since CFF2 Charstrings don't
    have width values. I discussed this with Behdad, and we are allowing for an
    initial width value in this case because developers may assemble a CFF2
    charstring from CFF Charstrings, which could have width values.
    """

    seenWidthOp = False
    vsIndex = None
    lenBlendStack = 0
    lastBlendIndex = 0
    commands = []
    stack = []
    it = iter(program)

    for token in it:
        if not isinstance(token, str):
            stack.append(token)
            continue

        if token == "blend":
            assert getNumRegions is not None
            numSourceFonts = 1 + getNumRegions(vsIndex)
            # replace the blend op args on the stack with a single list
            # containing all the blend op args.
            numBlends = stack[-1]
            numBlendArgs = numBlends * numSourceFonts + 1
            # replace first blend op by a list of the blend ops.
            stack[-numBlendArgs:] = [stack[-numBlendArgs:]]
            lenBlendStack += numBlends + len(stack) - 1
            lastBlendIndex = len(stack)
            # if a blend op exists, this is or will be a CFF2 charstring.
            continue

        elif token == "vsindex":
            vsIndex = stack[-1]
            assert type(vsIndex) is int

        elif (not seenWidthOp) and token in {
            "hstem",
            "hstemhm",
            "vstem",
            "vstemhm",
            "cntrmask",
            "hintmask",
            "hmoveto",
            "vmoveto",
            "rmoveto",
            "endchar",
        }:
            seenWidthOp = True
            parity = token in {"hmoveto", "vmoveto"}
            if lenBlendStack:
                # lenBlendStack has the number of args represented by the last blend
                # arg and all the preceding args. We need to now add the number of
                # args following the last blend arg.
                numArgs = lenBlendStack + len(stack[lastBlendIndex:])
            else:
                numArgs = len(stack)
            if numArgs and (numArgs % 2) ^ parity:
                width = stack.pop(0)
                commands.append(("", [width]))

        if token in {"hintmask", "cntrmask"}:
            if stack:
                commands.append(("", stack))
            commands.append((token, []))
            commands.append(("", [next(it)]))
        else:
            commands.append((token, stack))
        stack = []
    if stack:
        commands.append(("", stack))
    return commands


def _flattenBlendArgs(args):
    token_list = []
    for arg in args:
        if isinstance(arg, list):
            token_list.extend(arg)
            token_list.append("blend")
        else:
            token_list.append(arg)
    return token_list


def commandsToProgram(commands):
    """Takes a commands list as returned by programToCommands() and converts
    it back to a T2CharString program list."""
    program = []
    for op, args in commands:
        if any(isinstance(arg, list) for arg in args):
            args = _flattenBlendArgs(args)
        program.extend(args)
        if op:
            program.append(op)
    return program


def _everyN(el, n):
    """Group the list el into groups of size n"""
    if len(el) % n != 0:
        raise ValueError(el)
    for i in range(0, len(el), n):
        yield el[i : i + n]


class _GeneralizerDecombinerCommandsMap(object):
    @staticmethod
    def rmoveto(args):
        if len(args) != 2:
            raise ValueError(args)
        yield ("rmoveto", args)

    @staticmethod
    def hmoveto(args):
        if len(args) != 1:
            raise ValueError(args)
        yield ("rmoveto", [args[0], 0])

    @staticmethod
    def vmoveto(args):
        if len(args) != 1:
            raise ValueError(args)
        yield ("rmoveto", [0, args[0]])

    @staticmethod
    def rlineto(args):
        if not args:
            raise ValueError(args)
        for args in _everyN(args, 2):
            yield ("rlineto", args)

    @staticmethod
    def hlineto(args):
        if not args:
            raise ValueError(args)
        it = iter(args)
        try:
            while True:
                yield ("rlineto", [next(it), 0])
                yield ("rlineto", [0, next(it)])
        except StopIteration:
            pass

    @staticmethod
    def vlineto(args):
        if not args:
            raise ValueError(args)
        it = iter(args)
        try:
            while True:
                yield ("rlineto", [0, next(it)])
                yield ("rlineto", [next(it), 0])
        except StopIteration:
            pass

    @staticmethod
    def rrcurveto(args):
        if not args:
            raise ValueError(args)
        for args in _everyN(args, 6):
            yield ("rrcurveto", args)

    @staticmethod
    def hhcurveto(args):
        if len(args) < 4 or len(args) % 4 > 1:
            raise ValueError(args)
        if len(args) % 2 == 1:
            yield ("rrcurveto", [args[1], args[0], args[2], args[3], args[4], 0])
            args = args[5:]
        for args in _everyN(args, 4):
            yield ("rrcurveto", [args[0], 0, args[1], args[2], args[3], 0])

    @staticmethod
    def vvcurveto(args):
        if len(args) < 4 or len(args) % 4 > 1:
            raise ValueError(args)
        if len(args) % 2 == 1:
            yield ("rrcurveto", [args[0], args[1], args[2], args[3], 0, args[4]])
            args = args[5:]
        for args in _everyN(args, 4):
            yield ("rrcurveto", [0, args[0], args[1], args[2], 0, args[3]])

    @staticmethod
    def hvcurveto(args):
        if len(args) < 4 or len(args) % 8 not in {0, 1, 4, 5}:
            raise ValueError(args)
        last_args = None
        if len(args) % 2 == 1:
            lastStraight = len(args) % 8 == 5
            args, last_args = args[:-5], args[-5:]
        it = _everyN(args, 4)
        try:
            while True:
                args = next(it)
                yield ("rrcurveto", [args[0], 0, args[1], args[2], 0, args[3]])
                args = next(it)
                yield ("rrcurveto", [0, args[0], args[1], args[2], args[3], 0])
        except StopIteration:
            pass
        if last_args:
            args = last_args
            if lastStraight:
                yield ("rrcurveto", [args[0], 0, args[1], args[2], args[4], args[3]])
            else:
                yield ("rrcurveto", [0, args[0], args[1], args[2], args[3], args[4]])

    @staticmethod
    def vhcurveto(args):
        if len(args) < 4 or len(args) % 8 not in {0, 1, 4, 5}:
            raise ValueError(args)
        last_args = None
        if len(args) % 2 == 1:
            lastStraight = len(args) % 8 == 5
            args, last_args = args[:-5], args[-5:]
        it = _everyN(args, 4)
        try:
            while True:
                args = next(it)
                yield ("rrcurveto", [0, args[0], args[1], args[2], args[3], 0])
                args = next(it)
                yield ("rrcurveto", [args[0], 0, args[1], args[2], 0, args[3]])
        except StopIteration:
            pass
        if last_args:
            args = last_args
            if lastStraight:
                yield ("rrcurveto", [0, args[0], args[1], args[2], args[3], args[4]])
            else:
                yield ("rrcurveto", [args[0], 0, args[1], args[2], args[4], args[3]])

    @staticmethod
    def rcurveline(args):
        if len(args) < 8 or len(args) % 6 != 2:
            raise ValueError(args)
        args, last_args = args[:-2], args[-2:]
        for args in _everyN(args, 6):
            yield ("rrcurveto", args)
        yield ("rlineto", last_args)

    @staticmethod
    def rlinecurve(args):
        if len(args) < 8 or len(args) % 2 != 0:
            raise ValueError(args)
        args, last_args = args[:-6], args[-6:]
        for args in _everyN(args, 2):
            yield ("rlineto", args)
        yield ("rrcurveto", last_args)


def _convertBlendOpToArgs(blendList):
    # args is list of blend op args. Since we are supporting
    # recursive blend op calls, some of these args may also
    # be a list of blend op args, and need to be converted before
    # we convert the current list.
    if any([isinstance(arg, list) for arg in blendList]):
        args = [
            i
            for e in blendList
            for i in (_convertBlendOpToArgs(e) if isinstance(e, list) else [e])
        ]
    else:
        args = blendList

    # We now know that blendList contains a blend op argument list, even if
    # some of the args are lists that each contain a blend op argument list.
    # 	Convert from:
    # 		[default font arg sequence x0,...,xn] + [delta tuple for x0] + ... + [delta tuple for xn]
    # 	to:
    # 		[ [x0] + [delta tuple for x0],
    #                 ...,
    #          [xn] + [delta tuple for xn] ]
    numBlends = args[-1]
    # Can't use args.pop() when the args are being used in a nested list
    # comprehension. See calling context
    args = args[:-1]

    numRegions = len(args) // numBlends - 1
    if not (numBlends * (numRegions + 1) == len(args)):
        raise ValueError(blendList)

    defaultArgs = [[arg] for arg in args[:numBlends]]
    deltaArgs = args[numBlends:]
    numDeltaValues = len(deltaArgs)
    deltaList = [
        deltaArgs[i : i + numRegions] for i in range(0, numDeltaValues, numRegions)
    ]
    blend_args = [a + b + [1] for a, b in zip(defaultArgs, deltaList)]
    return blend_args


def generalizeCommands(commands, ignoreErrors=False):
    result = []
    mapping = _GeneralizerDecombinerCommandsMap
    for op, args in commands:
        # First, generalize any blend args in the arg list.
        if any([isinstance(arg, list) for arg in args]):
            try:
                args = [
                    n
                    for arg in args
                    for n in (
                        _convertBlendOpToArgs(arg) if isinstance(arg, list) else [arg]
                    )
                ]
            except ValueError:
                if ignoreErrors:
                    # Store op as data, such that consumers of commands do not have to
                    # deal with incorrect number of arguments.
                    result.append(("", args))
                    result.append(("", [op]))
                else:
                    raise

        func = getattr(mapping, op, None)
        if not func:
            result.append((op, args))
            continue
        try:
            for command in func(args):
                result.append(command)
        except ValueError:
            if ignoreErrors:
                # Store op as data, such that consumers of commands do not have to
                # deal with incorrect number of arguments.
                result.append(("", args))
                result.append(("", [op]))
            else:
                raise
    return result


def generalizeProgram(program, getNumRegions=None, **kwargs):
    return commandsToProgram(
        generalizeCommands(programToCommands(program, getNumRegions), **kwargs)
    )


def _categorizeVector(v):
    """
    Takes X,Y vector v and returns one of r, h, v, or 0 depending on which
    of X and/or Y are zero, plus tuple of nonzero ones.  If both are zero,
    it returns a single zero still.

    >>> _categorizeVector((0,0))
    ('0', (0,))
    >>> _categorizeVector((1,0))
    ('h', (1,))
    >>> _categorizeVector((0,2))
    ('v', (2,))
    >>> _categorizeVector((1,2))
    ('r', (1, 2))
    """
    if not v[0]:
        if not v[1]:
            return "0", v[:1]
        else:
            return "v", v[1:]
    else:
        if not v[1]:
            return "h", v[:1]
        else:
            return "r", v


def _mergeCategories(a, b):
    if a == "0":
        return b
    if b == "0":
        return a
    if a == b:
        return a
    return None


def _negateCategory(a):
    if a == "h":
        return "v"
    if a == "v":
        return "h"
    assert a in "0r"
    return a


def _convertToBlendCmds(args):
    # return a list of blend commands, and
    # the remaining non-blended args, if any.
    num_args = len(args)
    stack_use = 0
    new_args = []
    i = 0
    while i < num_args:
        arg = args[i]
        if not isinstance(arg, list):
            new_args.append(arg)
            i += 1
            stack_use += 1
        else:
            prev_stack_use = stack_use
            # The arg is a tuple of blend values.
            # These are each (master 0,delta 1..delta n, 1)
            # Combine as many successive tuples as we can,
            # up to the max stack limit.
            num_sources = len(arg) - 1
            blendlist = [arg]
            i += 1
            stack_use += 1 + num_sources  # 1 for the num_blends arg
            while (i < num_args) and isinstance(args[i], list):
                blendlist.append(args[i])
                i += 1
                stack_use += num_sources
                if stack_use + num_sources > maxStackLimit:
                    # if we are here, max stack is the CFF2 max stack.
                    # I use the CFF2 max stack limit here rather than
                    # the 'maxstack' chosen by the client, as the default
                    #  maxstack may have been used unintentionally. For all
                    # the other operators, this just produces a little less
                    # optimization, but here it puts a hard (and low) limit
                    # on the number of source fonts that can be used.
                    break
            # blendList now contains as many single blend tuples as can be
            # combined without exceeding the CFF2 stack limit.
            num_blends = len(blendlist)
            # append the 'num_blends' default font values
            blend_args = []
            for arg in blendlist:
                blend_args.append(arg[0])
            for arg in blendlist:
                assert arg[-1] == 1
                blend_args.extend(arg[1:-1])
            blend_args.append(num_blends)
            new_args.append(blend_args)
            stack_use = prev_stack_use + num_blends

    return new_args


def _addArgs(a, b):
    if isinstance(b, list):
        if isinstance(a, list):
            if len(a) != len(b) or a[-1] != b[-1]:
                raise ValueError()
            return [_addArgs(va, vb) for va, vb in zip(a[:-1], b[:-1])] + [a[-1]]
        else:
            a, b = b, a
    if isinstance(a, list):
        assert a[-1] == 1
        return [_addArgs(a[0], b)] + a[1:]
    return a + b


def specializeCommands(
    commands,
    ignoreErrors=False,
    generalizeFirst=True,
    preserveTopology=False,
    maxstack=48,
):
    # We perform several rounds of optimizations.  They are carefully ordered and are:
    #
    # 0. Generalize commands.
    #    This ensures that they are in our expected simple form, with each line/curve only
    #    having arguments for one segment, and using the generic form (rlineto/rrcurveto).
    #    If caller is sure the input is in this form, they can turn off generalization to
    #    save time.
    #
    # 1. Combine successive rmoveto operations.
    #
    # 2. Specialize rmoveto/rlineto/rrcurveto operators into horizontal/vertical variants.
    #    We specialize into some, made-up, variants as well, which simplifies following
    #    passes.
    #
    # 3. Merge or delete redundant operations, to the extent requested.
    #    OpenType spec declares point numbers in CFF undefined.  As such, we happily
    #    change topology.  If client relies on point numbers (in GPOS anchors, or for
    #    hinting purposes(what?)) they can turn this off.
    #
    # 4. Peephole optimization to revert back some of the h/v variants back into their
    #    original "relative" operator (rline/rrcurveto) if that saves a byte.
    #
    # 5. Combine adjacent operators when possible, minding not to go over max stack size.
    #
    # 6. Resolve any remaining made-up operators into real operators.
    #
    # I have convinced myself that this produces optimal bytecode (except for, possibly
    # one byte each time maxstack size prohibits combining.)  YMMV, but you'd be wrong. :-)
    # A dynamic-programming approach can do the same but would be significantly slower.
    #
    # 7. For any args which are blend lists, convert them to a blend command.

    # 0. Generalize commands.
    if generalizeFirst:
        commands = generalizeCommands(commands, ignoreErrors=ignoreErrors)
    else:
        commands = list(commands)  # Make copy since we modify in-place later.

    # 1. Combine successive rmoveto operations.
    for i in range(len(commands) - 1, 0, -1):
        if "rmoveto" == commands[i][0] == commands[i - 1][0]:
            v1, v2 = commands[i - 1][1], commands[i][1]
            commands[i - 1] = ("rmoveto", [v1[0] + v2[0], v1[1] + v2[1]])
            del commands[i]

    # 2. Specialize rmoveto/rlineto/rrcurveto operators into horizontal/vertical variants.
    #
    # We, in fact, specialize into more, made-up, variants that special-case when both
    # X and Y components are zero.  This simplifies the following optimization passes.
    # This case is rare, but OCD does not let me skip it.
    #
    # After this round, we will have four variants that use the following mnemonics:
    #
    #  - 'r' for relative,   ie. non-zero X and non-zero Y,
    #  - 'h' for horizontal, ie. zero X and non-zero Y,
    #  - 'v' for vertical,   ie. non-zero X and zero Y,
    #  - '0' for zeros,      ie. zero X and zero Y.
    #
    # The '0' pseudo-operators are not part of the spec, but help simplify the following
    # optimization rounds.  We resolve them at the end.  So, after this, we will have four
    # moveto and four lineto variants:
    #
    #  - 0moveto, 0lineto
    #  - hmoveto, hlineto
    #  - vmoveto, vlineto
    #  - rmoveto, rlineto
    #
    # and sixteen curveto variants.  For example, a '0hcurveto' operator means a curve
    # dx0,dy0,dx1,dy1,dx2,dy2,dx3,dy3 where dx0, dx1, and dy3 are zero but not dx3.
    # An 'rvcurveto' means dx3 is zero but not dx0,dy0,dy3.
    #
    # There are nine different variants of curves without the '0'.  Those nine map exactly
    # to the existing curve variants in the spec: rrcurveto, and the four variants hhcurveto,
    # vvcurveto, hvcurveto, and vhcurveto each cover two cases, one with an odd number of
    # arguments and one without.  Eg. an hhcurveto with an extra argument (odd number of
    # arguments) is in fact an rhcurveto.  The operators in the spec are designed such that
    # all four of rhcurveto, rvcurveto, hrcurveto, and vrcurveto are encodable for one curve.
    #
    # Of the curve types with '0', the 00curveto is equivalent to a lineto variant.  The rest
    # of the curve types with a 0 need to be encoded as a h or v variant.  Ie. a '0' can be
    # thought of a "don't care" and can be used as either an 'h' or a 'v'.  As such, we always
    # encode a number 0 as argument when we use a '0' variant.  Later on, we can just substitute
    # the '0' with either 'h' or 'v' and it works.
    #
    # When we get to curve splines however, things become more complicated...  XXX finish this.
    # There's one more complexity with splines.  If one side of the spline is not horizontal or
    # vertical (or zero), ie. if it's 'r', then it limits which spline types we can encode.
    # Only hhcurveto and vvcurveto operators can encode a spline starting with 'r', and
    # only hvcurveto and vhcurveto operators can encode a spline ending with 'r'.
    # This limits our merge opportunities later.
    #
    for i in range(len(commands)):
        op, args = commands[i]

        if op in {"rmoveto", "rlineto"}:
            c, args = _categorizeVector(args)
            commands[i] = c + op[1:], args
            continue

        if op == "rrcurveto":
            c1, args1 = _categorizeVector(args[:2])
            c2, args2 = _categorizeVector(args[-2:])
            commands[i] = c1 + c2 + "curveto", args1 + args[2:4] + args2
            continue

    # 3. Merge or delete redundant operations, to the extent requested.
    #
    # TODO
    # A 0moveto that comes before all other path operations can be removed.
    # though I find conflicting evidence for this.
    #
    # TODO
    # "If hstem and vstem hints are both declared at the beginning of a
    # CharString, and this sequence is followed directly by the hintmask or
    # cntrmask operators, then the vstem hint operator (or, if applicable,
    # the vstemhm operator) need not be included."
    #
    # "The sequence and form of a CFF2 CharString program may be represented as:
    # {hs* vs* cm* hm* mt subpath}? {mt subpath}*"
    #
    # https://www.microsoft.com/typography/otspec/cff2charstr.htm#section3.1
    #
    # For Type2 CharStrings the sequence is:
    # w? {hs* vs* cm* hm* mt subpath}? {mt subpath}* endchar"

    # Some other redundancies change topology (point numbers).
    if not preserveTopology:
        for i in range(len(commands) - 1, -1, -1):
            op, args = commands[i]

            # A 00curveto is demoted to a (specialized) lineto.
            if op == "00curveto":
                assert len(args) == 4
                c, args = _categorizeVector(args[1:3])
                op = c + "lineto"
                commands[i] = op, args
                # and then...

            # A 0lineto can be deleted.
            if op == "0lineto":
                del commands[i]
                continue

            # Merge adjacent hlineto's and vlineto's.
            # In CFF2 charstrings from variable fonts, each
            # arg item may be a list of blendable values, one from
            # each source font.
            if i and op in {"hlineto", "vlineto"} and (op == commands[i - 1][0]):
                _, other_args = commands[i - 1]
                assert len(args) == 1 and len(other_args) == 1
                try:
                    new_args = [_addArgs(args[0], other_args[0])]
                except ValueError:
                    continue
                commands[i - 1] = (op, new_args)
                del commands[i]
                continue

    # 4. Peephole optimization to revert back some of the h/v variants back into their
    #    original "relative" operator (rline/rrcurveto) if that saves a byte.
    for i in range(1, len(commands) - 1):
        op, args = commands[i]
        prv, nxt = commands[i - 1][0], commands[i + 1][0]

        if op in {"0lineto", "hlineto", "vlineto"} and prv == nxt == "rlineto":
            assert len(args) == 1
            args = [0, args[0]] if op[0] == "v" else [args[0], 0]
            commands[i] = ("rlineto", args)
            continue

        if op[2:] == "curveto" and len(args) == 5 and prv == nxt == "rrcurveto":
            assert (op[0] == "r") ^ (op[1] == "r")
            if op[0] == "v":
                pos = 0
            elif op[0] != "r":
                pos = 1
            elif op[1] == "v":
                pos = 4
            else:
                pos = 5
            # Insert, while maintaining the type of args (can be tuple or list).
            args = args[:pos] + type(args)((0,)) + args[pos:]
            commands[i] = ("rrcurveto", args)
            continue

    # 5. Combine adjacent operators when possible, minding not to go over max stack size.
    for i in range(len(commands) - 1, 0, -1):
        op1, args1 = commands[i - 1]
        op2, args2 = commands[i]
        new_op = None

        # Merge logic...
        if {op1, op2} <= {"rlineto", "rrcurveto"}:
            if op1 == op2:
                new_op = op1
            else:
                if op2 == "rrcurveto" and len(args2) == 6:
                    new_op = "rlinecurve"
                elif len(args2) == 2:
                    new_op = "rcurveline"

        elif (op1, op2) in {("rlineto", "rlinecurve"), ("rrcurveto", "rcurveline")}:
            new_op = op2

        elif {op1, op2} == {"vlineto", "hlineto"}:
            new_op = op1

        elif "curveto" == op1[2:] == op2[2:]:
            d0, d1 = op1[:2]
            d2, d3 = op2[:2]

            if d1 == "r" or d2 == "r" or d0 == d3 == "r":
                continue

            d = _mergeCategories(d1, d2)
            if d is None:
                continue
            if d0 == "r":
                d = _mergeCategories(d, d3)
                if d is None:
                    continue
                new_op = "r" + d + "curveto"
            elif d3 == "r":
                d0 = _mergeCategories(d0, _negateCategory(d))
                if d0 is None:
                    continue
                new_op = d0 + "r" + "curveto"
            else:
                d0 = _mergeCategories(d0, d3)
                if d0 is None:
                    continue
                new_op = d0 + d + "curveto"

        # Make sure the stack depth does not exceed (maxstack - 1), so
        # that subroutinizer can insert subroutine calls at any point.
        if new_op and len(args1) + len(args2) < maxstack:
            commands[i - 1] = (new_op, args1 + args2)
            del commands[i]

    # 6. Resolve any remaining made-up operators into real operators.
    for i in range(len(commands)):
        op, args = commands[i]

        if op in {"0moveto", "0lineto"}:
            commands[i] = "h" + op[1:], args
            continue

        if op[2:] == "curveto" and op[:2] not in {"rr", "hh", "vv", "vh", "hv"}:
            op0, op1 = op[:2]
            if (op0 == "r") ^ (op1 == "r"):
                assert len(args) % 2 == 1
            if op0 == "0":
                op0 = "h"
            if op1 == "0":
                op1 = "h"
            if op0 == "r":
                op0 = op1
            if op1 == "r":
                op1 = _negateCategory(op0)
            assert {op0, op1} <= {"h", "v"}, (op0, op1)

            if len(args) % 2:
                if op0 != op1:  # vhcurveto / hvcurveto
                    if (op0 == "h") ^ (len(args) % 8 == 1):
                        # Swap last two args order
                        args = args[:-2] + args[-1:] + args[-2:-1]
                else:  # hhcurveto / vvcurveto
                    if op0 == "h":  # hhcurveto
                        # Swap first two args order
                        args = args[1:2] + args[:1] + args[2:]

            commands[i] = op0 + op1 + "curveto", args
            continue

    # 7. For any series of args which are blend lists, convert the series to a single blend arg.
    for i in range(len(commands)):
        op, args = commands[i]
        if any(isinstance(arg, list) for arg in args):
            commands[i] = op, _convertToBlendCmds(args)

    return commands


def specializeProgram(program, getNumRegions=None, **kwargs):
    return commandsToProgram(
        specializeCommands(programToCommands(program, getNumRegions), **kwargs)
    )


if __name__ == "__main__":
    import sys

    if len(sys.argv) == 1:
        import doctest

        sys.exit(doctest.testmod().failed)

    import argparse

    parser = argparse.ArgumentParser(
        "fonttools cffLib.specialer",
        description="CFF CharString generalizer/specializer",
    )
    parser.add_argument("program", metavar="command", nargs="*", help="Commands.")
    parser.add_argument(
        "--num-regions",
        metavar="NumRegions",
        nargs="*",
        default=None,
        help="Number of variable-font regions for blend opertaions.",
    )

    options = parser.parse_args(sys.argv[1:])

    getNumRegions = (
        None
        if options.num_regions is None
        else lambda vsIndex: int(options.num_regions[0 if vsIndex is None else vsIndex])
    )

    program = stringToProgram(options.program)
    print("Program:")
    print(programToString(program))
    commands = programToCommands(program, getNumRegions)
    print("Commands:")
    print(commands)
    program2 = commandsToProgram(commands)
    print("Program from commands:")
    print(programToString(program2))
    assert program == program2
    print("Generalized program:")
    print(programToString(generalizeProgram(program, getNumRegions)))
    print("Specialized program:")
    print(programToString(specializeProgram(program, getNumRegions)))