Klara
Klara is a static analysis tools to automatic generate test case, based on SMT (z3) solver, with a powerful ast level inference system. Klara will take python file as input and generate corresponding test file in pytest format, that attempt to cover all return values. For example, following function in file test.py
def triangle(x: int, y: int, z: int) -> str:
if x == y == z:
return "Equilateral triangle"
elif x == y or y == z or x == z:
return "Isosceles triangle"
else:
return "Scalene triangle"
will generate
import test
def test_triangle_0():
assert test.triangle(0, 0, 0) == 'Equilateral triangle'
assert test.triangle(0, 0, 1) == 'Isosceles triangle'
assert test.triangle(2, 0, 1) == 'Scalene triangle'
See the Klara's documentation at https://klara-py.readthedocs.io
Installing
Klara can be installed via pip tool by using:
pip install klara
Usage
We can invoke klara on any python source file, and it will generate a corresponding pytest test file.
$ cat source.py
def foo(x: int, y: int, z: str):
if x + y > 2:
return x + y + 12
elif x < y:
return x + y
elif (z + "me") == "some":
return z + "thing"
else:
return x - y
$ klara source.py
$ cat test_source.py
import contract_test
def test_foo_0():
assert contract_test.foo(0, 3, \'\') == 15
assert contract_test.foo(0, 1, \'\') == 1
assert contract_test.foo(0, 0, \'so\') == \'sothing\'
assert contract_test.foo(0, 0, \'\') == 0
Consult the quick start manual for more examples and guidance. To use it as a static analysis library, go to Inference.
Why Klara?
Klara works on ast level and it doesn't execute user code in any way, which is a very important difference compared to similar tool like Crosshair and Pynguin that utilize concolic symbolic execution that required user code execution that might cause unwanted side effects. Klara work on ast level, combine with data flow analysis that utilize Control Flow Graph(CFG), Static Single Assignment(SSA), use-def chain, etc... to build a powerful python inference system that leverages Z3-solver for constraints solving and path feasibility check. Because of this, Klara is able to operate on both python2/3 source code with the help of typed_ast. To specify the source code is in python 2, pass in -py 2 argument. It's python 3 by default.
Klara can also be used as a static analysis tool, allow user to define custom rule to identify programming bugs, error or enforcing coding standard. With SMT solver support, analysis will be more accurate and greatly reduce false-positive case. For example
import klara
tree = klara.parse("""
def foo(v1: int):
if v1 > 4:
if v1 < 3:
z = 1
else:
z = 2
else:
z = 3
s = z
""")
with klara.MANAGER.initialize_z3_var_from_func(tree.body[0]):
print(list(tree.body[0].body[-1].value.infer()))
Will print out:
[2, 3]
Because z = 1 is not possible due to v1 > 4 and v1 < 3 is unsatisfiable
The inference system architecture and api is largely inspired by Astroid, a static inference library used by Pylint.
Klara utilize the inference system to generate test case, in other words, it generate test case for all possible return values of the function, instead of generate test case for all control path of the function.
To illustrate the point, consider the function below, with divide by zero vulnerabilities at line 3
def foo(v1: int, v2: float):
if v1 > 10000:
s = v1 / 0 # unused statement
if v1 > v2:
s = v1
else:
s = v2
return s
Klara will generate test inputs below
import contract_test
def test_foo_0():
assert contract_test.foo(0, -1.0) == 0
assert contract_test.foo(0, 0.0) == 0.0
It doesn't generate input v1 > 10000, so the test case would not be able to find out the exceptions. This is because the s at line 3 is unused in the return value.
If we modify the second if statement to elif, which we'll be able to return the [s]{.title-ref} at line 3, klara will generate test inputs that cover v1 > 10000 case.
This is an important distinction with other automatic test case generation available now, because by only generate test case for return values, we can generate a minimal test case, and it's easier to customize how do Klara cover the function.
For example, say we are composing a complex system
def main(number: int, cm: int, dc: int, wn: int):
mc = 0
if wn > 2:
if number > 2 and number > 2 or number > 2:
if number > 0:
if wn > 2 or wn > 2:
mc = 2
else:
mc = 5
else:
mc = 100
else:
mc = 1
nnn = number * cm
if cm <= 4:
num_incr = 4
else:
num_incr = cm
n_num_incr = nnn / num_incr
nnn_left = dc * num_incr * (n_num_incr / 2 + n_num_incr % 2)
nnn_right = nnn - nnn_left
is_flag = nnn_right
if is_flag:
cell = Component(nnn_right, options=[mc])
else:
cell = Component(nnn_right)
return cell
It isn't immediately clear to us how many possible return values there are. But we can utilize Klara to generate inputs instantly, below is the generated test
import contract_test
def test_main_0():
assert contract_test.main(2, 4, 1, 3) is not None
assert contract_test.main(2, 4, -1, 6) is not None
assert contract_test.main(2, 4, 1, 4) is not None
assert contract_test.main(-2, 4, 3, 4) is not None
assert contract_test.main(-1, -1, -1, 2) is not None
assert contract_test.main(0, 0, 0, 3) is not None
assert contract_test.main(0, 0, 0, 6) is not None
assert contract_test.main(0, 0, 0, 4) is not None
assert contract_test.main(-2, 0, 0, 4) is not None
assert contract_test.main(0, 0, 0, 0) is not None
Above generated 10 total results, which is product of nnn_right which have 2 possibilities and mc which have 5 possibilities.
Suppose that 10 tests input is too much, and we have determine that the options argument to Component is redundant to test, we can use Klara's custom plugin to selectively determine which part to ignore in test generation. Go to customize coverage strategy for more information.
After we have setup the plugin, Klara will generate following test
import contract_test
def test_main_0():
assert contract_test.main(1, 3, 0, 0) == 3.0
assert contract_test.main(0, 0, 0, 0) == 0.0
Which is only 2 combinations of nnn_right
Because Klara can't dynamically execute the code, it will provide extension to specify how to infer specific ast node or user defined type to make Klara 'smarter'. It's described in extending, extending user type and customize coverage strategy.
Contributing
We use Poetry to manage dependencies. After poetry is installed, run:
$ poetry shell
$ poetry install
To run the test case, do:
$ poetry run pytest test
Acknowledgements
- The architecture of the inference system is largely inspired by Astroid.
- Special thanks to Dr. Poh for guiding the early stages of the project.
License
This project is licensed under the terms of the GNU Lesser General Public License.
2 Feb 26, 2022
67 Dec 01, 2022
305 Dec 23, 2022
756 Dec 22, 2022
48 Dec 12, 2022
14 Jan 25, 2022
4 Nov 20, 2022
4 Jan 21, 2022
27 Jan 07, 2023
1 Nov 08, 2021
502 Dec 30, 2022
745 Dec 28, 2022
1.6k Dec 09, 2022
2.2k Jan 04, 2023
55 Nov 25, 2022
2 Nov 21, 2021
5 Sep 19, 2022
23 Dec 05, 2022
30 Oct 26, 2022