Toolbox for working with the Python AST

pip install ast_tools

Useful References

• Green Tree Snakes - the missing Python AST docs

Passes

ast_tools provides a number of passes for rewriting function and classes (could also work at the module level however no such pass exists). Passes are applied with the apply_passes decorator:

@apply_passes([pass1(), pass2()])
def foo(...): ...

Each pass takes as arguments an AST, an environment, and metadata and returns (possibly) modified versions of each. apply_passes begins a chain of rewrites by first looking up the ast of the decorated object and gather attempts to gather locals and globals from the call site to build the environment.

After all rewrites have run apply_passes serializes and execute the rewritten ast.

Know Issues

Collecting the AST

apply_passes relies on inspect.getsource to get the source of the decorated definition (which is then parsed to get the initial ast). However, inspect.getsource has many limitations.

Collecting the Environment

apply_passes does its best to infer the environment however there is no way to do this in a fully correct way. Users are encouraged to pass environment explicitly:

@apply_passes(..., env=SymbolTable(locals(), globals()))
def foo(...): ...

Wrapping the apply_passes decorator

The apply_passes decorator must not be wrapped.

As decorators are a part of the AST of the object they are applied to they must be removed from the rewritten AST before it is executed. If they are not removed rewrites will recurse infinitely as

@apply_passes([...])
def foo(...): ...

would become

exec('''\
@apply_passes([...])
def rewritten_foo(...): ...
''')

Note: this would invoke apply_passes([...]) on rewritten_foo

To avoid this the apply_passes decorator filters itself from the decorator list. If, however, the decorator is wrapped inside another decorator, this will fail.

Inner decorators are called multiple times

Decorators that are applied before a rewrite group will be called multiple times. See https://github.com/leonardt/ast_tools/issues/46 for detailed explanation. To avoid this users are encouraged to make rewrites the inner most decorators when possible.

Macros

Loop Unrolling

Unroll loops using the pattern

for <var> in ast_tools.macros.unroll(<iter>):
    ...

<iter> should be an iterable object that produces integers (e.g. range(8)) that can be evaluated at definition time (can refer to variables in the scope of the function definition)

For example,

from ast_tools.passes import apply_passes, loop_unroll

@apply_passes([loop_unroll()])
def foo():
    for i in ast_tools.macros.unroll(range(8)):
        print(i)

is rewritten into

def foo():
    print(0)
    print(1)
    print(2)
    print(3)
    print(4)
    print(5)
    print(6)
    print(7)

You can also use a list of ints, here's an example that also uses a reference to a variable defined in the outer scope:

from ast_tools.passes import apply_passes, loop_unroll

j = [1, 2, 3]
@apply_passes([loop_unroll()])
def foo():
    for i in ast_tools.macros.unroll(j):
        print(i)

becomes

def foo():
    print(1)
    print(2)
    print(3)

Inlining If Statements

This macro allows you to evaluate if statements at function definition time, so the resulting rewritten function will have the if statements marked "inlined" removed from the final code and replaced with the chosen branch based on evaluating the condition in the definition's enclosing scope. if statements are marked by using the form if inline(...): where inline is imported from the ast_tools.macros package. if statements not matching this pattern will be ignored by the rewrite logic.

Here's an example

from ast_tools.macros import inline
from ast_tools.passes import apply_passes, if_inline

y = True

apply_passes([if_inline()])
def foo(x):
    if inline(y):
        return x + 1
    else:
        return x - 1


import inspect
assert inspect.getsource(foo) == f"""\
def foo(x):
    return x + 1
"""