33. Python compiler package

버전 2.6부터 폐지: The compiler package has been removed in Python 3.

The Python compiler package is a tool for analyzing Python source code and generating Python bytecode. The compiler contains libraries to generate an abstract syntax tree from Python source code and to generate Python bytecode from the tree.

The compiler package is a Python source to bytecode translator written in Python. It uses the built-in parser and standard parser module to generate a concrete syntax tree. This tree is used to generate an abstract syntax tree (AST) and then Python bytecode.

The full functionality of the package duplicates the built-in compiler provided with the Python interpreter. It is intended to match its behavior almost exactly. Why implement another compiler that does the same thing? The package is useful for a variety of purposes. It can be modified more easily than the built-in compiler. The AST it generates is useful for analyzing Python source code.

This chapter explains how the various components of the compiler package work. It blends reference material with a tutorial.

33.1. The basic interface

The top-level of the package defines four functions. If you import compiler, you will get these functions and a collection of modules contained in the package.

compiler.parse(buf)

Returns an abstract syntax tree for the Python source code in buf. The function raises SyntaxError if there is an error in the source code. The return value is a compiler.ast.Module instance that contains the tree.

compiler.parseFile(path)

Return an abstract syntax tree for the Python source code in the file specified by path. It is equivalent to parse(open(path).read()).

compiler.walk(ast, visitor[, verbose])

Do a pre-order walk over the abstract syntax tree ast. Call the appropriate method on the visitor instance for each node encountered.

compiler.compile(source, filename, mode, flags=None, dont_inherit=None)

Compile the string source, a Python module, statement or expression, into a code object that can be executed by the exec statement or eval(). This function is a replacement for the built-in compile() function.

The filename will be used for run-time error messages.

The mode must be 〈exec〉 to compile a module, 〈single〉 to compile a single (interactive) statement, or 〈eval〉 to compile an expression.

The flags and dont_inherit arguments affect future-related statements, but are not supported yet.

compiler.compileFile(source)

Compiles the file source and generates a .pyc file.

The compiler package contains the following modules: ast, consts, future, misc, pyassem, pycodegen, symbols, transformer, and visitor.

33.2. Limitations

There are some problems with the error checking of the compiler package. The interpreter detects syntax errors in two distinct phases. One set of errors is detected by the interpreter’s parser, the other set by the compiler. The compiler package relies on the interpreter’s parser, so it get the first phases of error checking for free. It implements the second phase itself, and that implementation is incomplete. For example, the compiler package does not raise an error if a name appears more than once in an argument list: def f(x, x): ...

A future version of the compiler should fix these problems.

33.3. Python Abstract Syntax

The compiler.ast module defines an abstract syntax for Python. In the abstract syntax tree, each node represents a syntactic construct. The root of the tree is Module object.

The abstract syntax offers a higher level interface to parsed Python source code. The parser module and the compiler written in C for the Python interpreter use a concrete syntax tree. The concrete syntax is tied closely to the grammar description used for the Python parser. Instead of a single node for a construct, there are often several levels of nested nodes that are introduced by Python’s precedence rules.

The abstract syntax tree is created by the compiler.transformer module. The transformer relies on the built-in Python parser to generate a concrete syntax tree. It generates an abstract syntax tree from the concrete tree.

The transformer module was created by Greg Stein and Bill Tutt for an experimental Python-to-C compiler. The current version contains a number of modifications and improvements, but the basic form of the abstract syntax and of the transformer are due to Stein and Tutt.

33.3.1. AST Nodes

The compiler.ast module is generated from a text file that describes each node type and its elements. Each node type is represented as a class that inherits from the abstract base class compiler.ast.Node and defines a set of named attributes for child nodes.

class compiler.ast.Node

The Node instances are created automatically by the parser generator. The recommended interface for specific Node instances is to use the public attributes to access child nodes. A public attribute may be bound to a single node or to a sequence of nodes, depending on the Node type. For example, the bases attribute of the Class node, is bound to a list of base class nodes, and the doc attribute is bound to a single node.

Each Node instance has a lineno attribute which may be None. XXX Not sure what the rules are for which nodes will have a useful lineno.

All Node objects offer the following methods:

getChildren()

Returns a flattened list of the child nodes and objects in the order they occur. Specifically, the order of the nodes is the order in which they appear in the Python grammar. Not all of the children are Node instances. The names of functions and classes, for example, are plain strings.

getChildNodes()

Returns a flattened list of the child nodes in the order they occur. This method is like getChildren(), except that it only returns those children that are Node instances.

Two examples illustrate the general structure of Node classes. The while statement is defined by the following grammar production:

while_stmt:     "while" expression ":" suite
               ["else" ":" suite]

The While node has three attributes: test, body, and else_. (If the natural name for an attribute is also a Python reserved word, it can’t be used as an attribute name. An underscore is appended to the word to make it a legal identifier, hence else_ instead of else.)

The if statement is more complicated because it can include several tests.

if_stmt: 'if' test ':' suite ('elif' test ':' suite)* ['else' ':' suite]

The If node only defines two attributes: tests and else_. The tests attribute is a sequence of test expression, consequent body pairs. There is one pair for each if/elif clause. The first element of the pair is the test expression. The second elements is a Stmt node that contains the code to execute if the test is true.

The getChildren() method of If returns a flat list of child nodes. If there are three if/elif clauses and no else clause, then getChildren() will return a list of six elements: the first test expression, the first Stmt, the second text expression, etc.

The following table lists each of the Node subclasses defined in compiler.ast and each of the public attributes available on their instances. The values of most of the attributes are themselves Node instances or sequences of instances. When the value is something other than an instance, the type is noted in the comment. The attributes are listed in the order in which they are returned by getChildren() and getChildNodes().

Node type

Attribute

Value

Add

left

left operand

right

right operand

And

nodes

list of operands

AssAttr

attribute as target of assignment

expr

expression on the left-hand side of the dot

attrname

the attribute name, a string

flags

XXX

AssList

nodes

list of list elements being assigned to

AssName

name

name being assigned to

flags

XXX

AssTuple

nodes

list of tuple elements being assigned to

Assert

test

the expression to be tested

fail

the value of the AssertionError

Assign

nodes

a list of assignment targets, one per equal sign

expr

the value being assigned

AugAssign

node

op

expr

Backquote

expr

Bitand

nodes

Bitor

nodes

Bitxor

nodes

Break

CallFunc

node

expression for the callee

args

a list of arguments

star_args

the extended *-arg value

dstar_args

the extended **-arg value

Class

name

the name of the class, a string

bases

a list of base classes

doc

doc string, a string or None

code

the body of the class statement

Compare

expr

ops

Const

value

Continue

Decorators

nodes

List of function decorator expressions

Dict

items

Discard

expr

Div

left

right

Ellipsis

Expression

node

Exec

expr

locals

globals

FloorDiv

left

right

For

assign

list

body

else_

From

modname

names

Function

decorators

Decorators or None

name

name used in def, a string

argnames

list of argument names, as strings

defaults

list of default values

flags

xxx

doc

doc string, a string or None

code

the body of the function

GenExpr

code

GenExprFor

assign

iter

ifs

GenExprIf

test

GenExprInner

expr

quals

Getattr

expr

attrname

Global

names

If

tests

else_

Import

names

Invert

expr

Keyword

name

expr

Lambda

argnames

defaults

flags

code

LeftShift

left

right

List

nodes

ListComp

expr

quals

ListCompFor

assign

list

ifs

ListCompIf

test

Mod

left

right

Module

doc

doc string, a string or None

node

body of the module, a Stmt

Mul

left

right

Name

name

Not

expr

Or

nodes

Pass

Power

left

right

Print

nodes

dest

Printnl

nodes

dest

Raise

expr1

expr2

expr3

Return

value

RightShift

left

right

Slice

expr

flags

lower

upper

Sliceobj

nodes

list of statements

Stmt

nodes

Sub

left

right

Subscript

expr

flags

subs

TryExcept

body

handlers

else_

TryFinally

body

final

Tuple

nodes

UnaryAdd

expr

UnarySub

expr

While

test

body

else_

With

expr

vars

body

Yield

value

33.3.2. Assignment nodes

There is a collection of nodes used to represent assignments. Each assignment statement in the source code becomes a single Assign node in the AST. The nodes attribute is a list that contains a node for each assignment target. This is necessary because assignment can be chained, e.g. a = b = 2. Each Node in the list will be one of the following classes: AssAttr, AssList, AssName, or AssTuple.

Each target assignment node will describe the kind of object being assigned to: AssName for a simple name, e.g. a = 1. AssAttr for an attribute assigned, e.g. a.x = 1. AssList and AssTuple for list and tuple expansion respectively, e.g. a, b, c = a_tuple.

The target assignment nodes also have a flags attribute that indicates whether the node is being used for assignment or in a delete statement. The AssName is also used to represent a delete statement, e.g. del x.

When an expression contains several attribute references, an assignment or delete statement will contain only one AssAttr node – for the final attribute reference. The other attribute references will be represented as Getattr nodes in the expr attribute of the AssAttr instance.

33.3.3. Examples

This section shows several simple examples of ASTs for Python source code. The examples demonstrate how to use the parse() function, what the repr of an AST looks like, and how to access attributes of an AST node.

The first module defines a single function. Assume it is stored in doublelib.py.

"""This is an example module.

This is the docstring.
"""

def double(x):
    "Return twice the argument"
    return x * 2

In the interactive interpreter session below, I have reformatted the long AST reprs for readability. The AST reprs use unqualified class names. If you want to create an instance from a repr, you must import the class names from the compiler.ast module.

>>> import compiler
>>> mod = compiler.parseFile("doublelib.py")
>>> mod
Module('This is an example module.\n\nThis is the docstring.\n',
       Stmt([Function(None, 'double', ['x'], [], 0,
                      'Return twice the argument',
                      Stmt([Return(Mul((Name('x'), Const(2))))]))]))
>>> from compiler.ast import *
>>> Module('This is an example module.\n\nThis is the docstring.\n',
...    Stmt([Function(None, 'double', ['x'], [], 0,
...                   'Return twice the argument',
...                   Stmt([Return(Mul((Name('x'), Const(2))))]))]))
Module('This is an example module.\n\nThis is the docstring.\n',
       Stmt([Function(None, 'double', ['x'], [], 0,
                      'Return twice the argument',
                      Stmt([Return(Mul((Name('x'), Const(2))))]))]))
>>> mod.doc
'This is an example module.\n\nThis is the docstring.\n'
>>> for node in mod.node.nodes:
...     print node
...
Function(None, 'double', ['x'], [], 0, 'Return twice the argument',
         Stmt([Return(Mul((Name('x'), Const(2))))]))
>>> func = mod.node.nodes[0]
>>> func.code
Stmt([Return(Mul((Name('x'), Const(2))))])

33.4. Using Visitors to Walk ASTs

The visitor pattern is … The compiler package uses a variant on the visitor pattern that takes advantage of Python’s introspection features to eliminate the need for much of the visitor’s infrastructure.

The classes being visited do not need to be programmed to accept visitors. The visitor need only define visit methods for classes it is specifically interested in; a default visit method can handle the rest.

XXX The magic visit() method for visitors.

compiler.visitor.walk(tree, visitor[, verbose])
class compiler.visitor.ASTVisitor

The ASTVisitor is responsible for walking over the tree in the correct order. A walk begins with a call to preorder(). For each node, it checks the visitor argument to preorder() for a method named 〈visitNodeType,〉 where NodeType is the name of the node’s class, e.g. for a While node a visitWhile() would be called. If the method exists, it is called with the node as its first argument.

The visitor method for a particular node type can control how child nodes are visited during the walk. The ASTVisitor modifies the visitor argument by adding a visit method to the visitor; this method can be used to visit a particular child node. If no visitor is found for a particular node type, the default() method is called.

ASTVisitor objects have the following methods:

XXX describe extra arguments

default(node[, ...])
dispatch(node[, ...])
preorder(tree, visitor)

33.5. Bytecode Generation

The code generator is a visitor that emits bytecodes. Each visit method can call the emit() method to emit a new bytecode. The basic code generator is specialized for modules, classes, and functions. An assembler converts that emitted instructions to the low-level bytecode format. It handles things like generation of constant lists of code objects and calculation of jump offsets.