Negative number raised to even power gives negative result if exponent is a variable

[bart-nouveau]

Unexpected behavior of the power operator with a variable of even value as exponent. It returns the negative value of the expected result.

mcve

from jinja2 import Environment

env = Environment()
env.globals = {"x": 2}
env.from_string("{{ -1 ** x }}").render()

From which I get '-1'

Expected behavior

I expected the positive value '1' as a result

It instead works as expected in a Sandboxed environment when intercepting the ** operator:

from jinja2.sandbox import SandboxedEnvironment

sandbox_env = SandboxedEnvironment()
sandbox_env.globals = {"x": 2}
sandbox_env.intercepted_binops = frozenset(["**"])
sandbox_env.from_string("{{ -1 ** x }}").render()

From which I get '1'

Environment:

• Python version: 3.8.10
• Jinja version: 3.1.2

[davidism]

Happy to review a PR

[krepysh]

Would you mind if I'll take a look onto this during this weekend?

[davidism]

Anyone is welcome to work on any issue at any time, no need to ask.

[bart-nouveau]

Some more details about this issue.

There are two problems: differences with respect to the results obtained with python and discrepancies between different inputs in jinja.

source code

In the following I will list some examples for different templates obtained with the following code:

env = jinja2.Environment()
env.globals = {"x": 2}
jinja_output = env.from_string(f"{{{{ {expression} }}}}").render()
python_output = eval(f"{expression}",env.globals)

examples for different input expressions

expression: '(-1) ** x'
jinja_output: '-1'
python_output: 1

expression: '-1 ** x'
jinja_output: '-1'
python_output: -1

expression: '(-1) ** 2'
jinja_output: '1'
python_output: 1

expression: '-1 ** 2'
jinja_output: '1'
python_output: -1

conclusion

The cause of the differences with respect to python could be in the parsing stage since, independently from the use of parenthesis, the result of the parser for -1 ** 2 is:
Pow(left=Neg(node=Const(value=1)), right=Name(name='x', ctx='load'))
While I expect the Neg operator to be applied to Pow

Discrepancies between jinja outputs with and without variable exponent could be found in the code generation step, where an optimization step is applied when the exponent is a Const but not when the exponent is a Name

[Troon]

I've raised this as an issue against Home Assistant, and have now seen this related issue.