parser -- Access Python parse trees

Emulation of compile() #

While many useful operations may take place between parsing and bytecode generation, the simplest operation is to do nothing. For this purpose, using the parser module to produce an intermediate data structure is equivalent to the code

>>> code = compile('a + 5', 'file.py', 'eval')
>>> a = 5
>>> eval(code)
10

The equivalent operation using the parser module is somewhat longer, and allows the intermediate internal parse tree to be retained as an AST object:

>>> import parser
>>> ast = parser.expr('a + 5')
>>> code = ast.compile('file.py')
>>> a = 5
>>> eval(code)
10

An application which needs both AST and code objects can package this code into readily available functions:

import parser

def load_suite(source_string):
    ast = parser.suite(source_string)
    return ast, ast.compile()

def load_expression(source_string):
    ast = parser.expr(source_string)
    return ast, ast.compile()

Information Discovery #

Some applications benefit from direct access to the parse tree. The remainder of this section demonstrates how the parse tree provides access to module documentation defined in docstrings without requiring that the code being examined be loaded into a running interpreter via import. This can be very useful for performing analyses of untrusted code.

Generally, the example will demonstrate how the parse tree may be traversed to distill interesting information. Two functions and a set of classes are developed which provide programmatic access to high level function and class definitions provided by a module. The classes extract information from the parse tree and provide access to the information at a useful semantic level, one function provides a simple low-level pattern matching capability, and the other function defines a high-level interface to the classes by handling file operations on behalf of the caller. All source files mentioned here which are not part of the Python installation are located in the Demo/parser/ directory of the distribution.

The dynamic nature of Python allows the programmer a great deal of flexibility, but most modules need only a limited measure of this when defining classes, functions, and methods. In this example, the only definitions that will be considered are those which are defined in the top level of their context, e.g., a function defined by a def statement at column zero of a module, but not a function defined within a branch of an if … else construct, though there are some good reasons for doing so in some situations. Nesting of definitions will be handled by the code developed in the example.

To construct the upper-level extraction methods, we need to know what the parse tree structure looks like and how much of it we actually need to be concerned about. Python uses a moderately deep parse tree so there are a large number of intermediate nodes. It is important to read and understand the formal grammar used by Python. This is specified in the file Grammar/Grammar in the distribution. Consider the simplest case of interest when searching for docstrings: a module consisting of a docstring and nothing else. (See file docstring.py.)

"""Some documentation.
"""

Using the interpreter to take a look at the parse tree, we find a bewildering mass of numbers and parentheses, with the documentation buried deep in nested tuples.

>>> import parser
>>> import pprint
>>> ast = parser.suite(open('docstring.py').read())
>>> tup = ast.totuple()
>>> pprint.pprint(tup)
(257,
 (264,
  (265,
   (266,
    (267,
     (307,
      (287,
       (288,
        (289,
         (290,
          (292,
           (293,
            (294,
             (295,
              (296,
               (297,
                (298,
                 (299,
                  (300, (3, '"""Some documentation.\n"""'))))))))))))))))),
   (4, ''))),
 (4, ''),
 (0, ''))

The numbers at the first element of each node in the tree are the node types; they map directly to terminal and non-terminal symbols in the grammar. Unfortunately, they are represented as integers in the internal representation, and the Python structures generated do not change that. However, the symbol and token modules provide symbolic names for the node types and dictionaries which map from the integers to the symbolic names for the node types.

In the output presented above, the outermost tuple contains four elements: the integer 257 and three additional tuples. Node type 257 has the symbolic name file_input. Each of these inner tuples contains an integer as the first element; these integers, 264, 4, and 0, represent the node types stmt, NEWLINE, and ENDMARKER, respectively. Note that these values may change depending on the version of Python you are using; consult symbol.py and token.py for details of the mapping. It should be fairly clear that the outermost node is related primarily to the input source rather than the contents of the file, and may be disregarded for the moment. The stmt node is much more interesting. In particular, all docstrings are found in subtrees which are formed exactly as this node is formed, with the only difference being the string itself. The association between the docstring in a similar tree and the defined entity (class, function, or module) which it describes is given by the position of the docstring subtree within the tree defining the described structure.

By replacing the actual docstring with something to signify a variable component of the tree, we allow a simple pattern matching approach to check any given subtree for equivalence to the general pattern for docstrings. Since the example demonstrates information extraction, we can safely require that the tree be in tuple form rather than list form, allowing a simple variable representation to be ['variable_name']. A simple recursive function can implement the pattern matching, returning a Boolean and a dictionary of variable name to value mappings. (See file example.py.)

from types import ListType, TupleType

def match(pattern, data, vars=None):
    if vars is None:
        vars = {}
    if type(pattern) is ListType:
        vars[pattern[0]] = data
        return 1, vars
    if type(pattern) is not TupleType:
        return (pattern == data), vars
    if len(data) != len(pattern):
        return 0, vars
    for pattern, data in map(None, pattern, data):
        same, vars = match(pattern, data, vars)
        if not same:
            break
    return same, vars

Using this simple representation for syntactic variables and the symbolic node types, the pattern for the candidate docstring subtrees becomes fairly readable. (See file example.py.)

import symbol
import token

DOCSTRING_STMT_PATTERN = (
    symbol.stmt,
    (symbol.simple_stmt,
     (symbol.small_stmt,
      (symbol.expr_stmt,
       (symbol.testlist,
        (symbol.test,
         (symbol.and_test,
          (symbol.not_test,
           (symbol.comparison,
            (symbol.expr,
             (symbol.xor_expr,
              (symbol.and_expr,
               (symbol.shift_expr,
                (symbol.arith_expr,
                 (symbol.term,
                  (symbol.factor,
                   (symbol.power,
                    (symbol.atom,
                     (token.STRING, ['docstring'])
                     )))))))))))))))),
     (token.NEWLINE, '')
     ))

Using the match function with this pattern, extracting the module docstring from the parse tree created previously is easy:

>>> found, vars = match(DOCSTRING_STMT_PATTERN, tup[1])
>>> found
1
>>> vars
{'docstring': '"""Some documentation.\n"""'}

Once specific data can be extracted from a location where it is expected, the question of where information can be expected needs to be answered. When dealing with docstrings, the answer is fairly simple: the docstring is the first stmt node in a code block (file_input or suite node types). A module consists of a single file_input node, and class and function definitions each contain exactly one suite node. Classes and functions are readily identified as subtrees of code block nodes which start with (stmt, (compound_stmt, (classdef, ... or (stmt, (compound_stmt, (funcdef, .... Note that these subtrees cannot be matched by match since it does not support multiple sibling nodes to match without regard to number. A more elaborate matching function could be used to overcome this limitation, but this is sufficient for the example.

Given the ability to determine whether a statement might be a docstring and extract the actual string from the statement, some work needs to be performed to walk the parse tree for an entire module and extract information about the names defined in each context of the module and associate any docstrings with the names. The code to perform this work is not complicated, but bears some explanation.

The public interface to the classes is straightforward and should probably be somewhat more flexible. Each “major” block of the module is described by an object providing several methods for inquiry and a constructor which accepts at least the subtree of the complete parse tree which it represents. The ModuleInfo constructor accepts an optional name parameter since it cannot otherwise determine the name of the module.

The public classes include ClassInfo, FunctionInfo, and ModuleInfo. All objects provide the methods get_name, get_docstring, get_class_names, and get_class_info. The ClassInfo objects support get_method_names and get_method_info while the other classes provide get_function_names and get_function_info.

Within each of the forms of code block that the public classes represent, most of the required information is in the same form and is accessed in the same way, with classes having the distinction that functions defined at the top level are referred to as “methods.” Since the difference in nomenclature reflects a real semantic distinction from functions defined outside of a class, the implementation needs to maintain the distinction. Hence, most of the functionality of the public classes can be implemented in a common base class, SuiteInfoBase, with the accessors for function and method information provided elsewhere. Note that there is only one class which represents function and method information; this parallels the use of the def statement to define both types of elements.

Most of the accessor functions are declared in SuiteInfoBase and do not need to be overridden by subclasses. More importantly, the extraction of most information from a parse tree is handled through a method called by the SuiteInfoBase constructor. The example code for most of the classes is clear when read alongside the formal grammar, but the method which recursively creates new information objects requires further examination. Here is the relevant part of the SuiteInfoBase definition from example.py:

class SuiteInfoBase:
    _docstring = ''
    _name = ''

    def __init__(self, tree = None):
        self._class_info = {}
        self._function_info = {}
        if tree:
            self._extract_info(tree)

    def _extract_info(self, tree):
        # extract docstring
        if len(tree) == 2:
            found, vars = match(DOCSTRING_STMT_PATTERN[1], tree[1])
        else:
            found, vars = match(DOCSTRING_STMT_PATTERN, tree[3])
        if found:
            self._docstring = eval(vars['docstring'])
        # discover inner definitions
        for node in tree[1:]:
            found, vars = match(COMPOUND_STMT_PATTERN, node)
            if found:
                cstmt = vars['compound']
                if cstmt[0] == symbol.funcdef:
                    name = cstmt[2][1]
                    self._function_info[name] = FunctionInfo(cstmt)
                elif cstmt[0] == symbol.classdef:
                    name = cstmt[2][1]
                    self._class_info[name] = ClassInfo(cstmt)

After initializing some internal state, the constructor calls the _extract_info method. This method performs the bulk of the information extraction which takes place in the entire example. The extraction has two distinct phases: the location of the docstring for the parse tree passed in, and the discovery of additional definitions within the code block represented by the parse tree.

The initial if test determines whether the nested suite is of the “short form” or the “long form.” The short form is used when the code block is on the same line as the definition of the code block, as in

def square(x): "Square an argument."; return x ** 2

while the long form uses an indented block and allows nested definitions:

def make_power(exp):
    "Make a function that raises an argument to the exponent `exp'."
    def raiser(x, y=exp):
        return x ** y
    return raiser

When the short form is used, the code block may contain a docstring as the first, and possibly only, small_stmt element. The extraction of such a docstring is slightly different and requires only a portion of the complete pattern used in the more common case. As implemented, the docstring will only be found if there is only one small_stmt node in the simple_stmt node. Since most functions and methods which use the short form do not provide a docstring, this may be considered sufficient. The extraction of the docstring proceeds using the match function as described above, and the value of the docstring is stored as an attribute of the SuiteInfoBase object.

After docstring extraction, a simple definition discovery algorithm operates on the stmt nodes of the suite node. The special case of the short form is not tested; since there are no stmt nodes in the short form, the algorithm will silently skip the single simple_stmt node and correctly not discover any nested definitions.

Each statement in the code block is categorized as a class definition, function or method definition, or something else. For the definition statements, the name of the element defined is extracted and a representation object appropriate to the definition is created with the defining subtree passed as an argument to the constructor. The representation objects are stored in instance variables and may be retrieved by name using the appropriate accessor methods.

The public classes provide any accessors required which are more specific than those provided by the SuiteInfoBase class, but the real extraction algorithm remains common to all forms of code blocks. A high-level function can be used to extract the complete set of information from a source file. (See file example.py.)

def get_docs(fileName):
    import os
    import parser

    source = open(fileName).read()
    basename = os.path.basename(os.path.splitext(fileName)[0])
    ast = parser.suite(source)
    return ModuleInfo(ast.totuple(), basename)

This provides an easy-to-use interface to the documentation of a module. If information is required which is not extracted by the code of this example, the code may be extended at clearly defined points to provide additional capabilities.