Significant slow down on njit'd functions when optional arguments not provided

Hello I ran into a problem where an njit’d function that has optional arguments is almost 200 times slower then when the arguments are provided.

I saw a similar issue from a few years ago which was resolved. Not sure if this is related or not. My numba and python versions are below.

import sys
import numba
import numpy as np

@numba.njit
def calc(x_squared, order_l: int = 2, init_l: int = 10):

    z = x_squared / (2. * init_l + 3.)
    for l_fake in range(order_l - 1, init_l - 1):
        l = init_l - l_fake
        l += 1
        z = x_squared / ((2. * l + 1.) - z)

    return z

%timeit calc(0.1)
%timeit calc(0.1, order_l=2, init_l=10)

# Perhaps an initial compile issue (even though timeit should catch that.), try running again
print('2nd Run')
%timeit calc(0.1)
%timeit calc(0.1, order_l=2, init_l=10)

print('\n')
print('python version:', sys.version)
print('numba version:', numba.__version__)
print('numpy version:', np.__version__)

Running this leads to the following output

54.3 µs ± 554 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
332 ns ± 3.34 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)
2nd Run
54 µs ± 258 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)
331 ns ± 2.76 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)


python version: 3.9.7 (default, Sep 16 2021, 16:59:28) [MSC v.1916 64 bit (AMD64)]
numba version: 0.54.1
numpy version: 1.20.3

As you can see the non argument provided version runs in around 54 us versus the provided version which is only 0.3 us!

[seibert]

Correct, this is definitely a performance bug, although a long standing one, apparently.

[sklam]

The problem seems to have existed for as far back as numba 0.35.0. That’s the earliest accessible version in the current version of conda.