Pytorch/Tools/scripts/find_recursionlimit.py

#! /usr/bin/env python3
"""Find the maximum recursion limit that prevents interpreter termination.

This script finds the maximum safe recursion limit on a particular
platform.  If you need to change the recursion limit on your system,
this script will tell you a safe upper bound.  To use the new limit,
call sys.setrecursionlimit().

This module implements several ways to create infinite recursion in
Python.  Different implementations end up pushing different numbers of
C stack frames, depending on how many calls through Python's abstract
C API occur.

After each round of tests, it prints a message:
"Limit of NNNN is fine".

The highest printed value of "NNNN" is therefore the highest potentially
safe limit for your system (which depends on the OS, architecture, but also
the compilation flags). Please note that it is practically impossible to
test all possible recursion paths in the interpreter, so the results of
this test should not be trusted blindly -- although they give a good hint
of which values are reasonable.

NOTE: When the C stack space allocated by your system is exceeded due
to excessive recursion, exact behaviour depends on the platform, although
the interpreter will always fail in a likely brutal way: either a
segmentation fault, a MemoryError, or just a silent abort.

NB: A program that does not use __methods__ can set a higher limit.
"""

import sys
import itertools

class RecursiveBlowup1:
    def __init__(self):
        self.__init__()

def test_init():
    return RecursiveBlowup1()

class RecursiveBlowup2:
    def __repr__(self):
        return repr(self)

def test_repr():
    return repr(RecursiveBlowup2())

class RecursiveBlowup4:
    def __add__(self, x):
        return x + self

def test_add():
    return RecursiveBlowup4() + RecursiveBlowup4()

class RecursiveBlowup5:
    def __getattr__(self, attr):
        return getattr(self, attr)

def test_getattr():
    return RecursiveBlowup5().attr

class RecursiveBlowup6:
    def __getitem__(self, item):
        return self[item - 2] + self[item - 1]

def test_getitem():
    return RecursiveBlowup6()[5]

def test_recurse():
    return test_recurse()

def test_cpickle(_cache={}):
    import io
    try:
        import _pickle
    except ImportError:
        print("cannot import _pickle, skipped!")
        return
    k, l = None, None
    for n in itertools.count():
        try:
            l = _cache[n]
            continue  # Already tried and it works, let's save some time
        except KeyError:
            for i in range(100):
                l = [k, l]
                k = {i: l}
        _pickle.Pickler(io.BytesIO(), protocol=-1).dump(l)
        _cache[n] = l

def test_compiler_recursion():
    # The compiler uses a scaling factor to support additional levels
    # of recursion. This is a sanity check of that scaling to ensure
    # it still raises RecursionError even at higher recursion limits
    compile("()" * (10 * sys.getrecursionlimit()), "<single>", "single")

def check_limit(n, test_func_name):
    sys.setrecursionlimit(n)
    if test_func_name.startswith("test_"):
        print(test_func_name[5:])
    else:
        print(test_func_name)
    test_func = globals()[test_func_name]
    try:
        test_func()
    # AttributeError can be raised because of the way e.g. PyDict_GetItem()
    # silences all exceptions and returns NULL, which is usually interpreted
    # as "missing attribute".
    except (RecursionError, AttributeError):
        pass
    else:
        print("Yikes!")

if __name__ == '__main__':

    limit = 1000
    while 1:
        check_limit(limit, "test_recurse")
        check_limit(limit, "test_add")
        check_limit(limit, "test_repr")
        check_limit(limit, "test_init")
        check_limit(limit, "test_getattr")
        check_limit(limit, "test_getitem")
        check_limit(limit, "test_cpickle")
        check_limit(limit, "test_compiler_recursion")
        print("Limit of %d is fine" % limit)
        limit = limit + 100