Mod on uint64 near max values gives wrong result

[miloslen]

Reproducible test case:

import numpy as np
import numba

def test_max_mod():
    return np.mod(np.uint64(np.iinfo(np.uint64).max), np.uint64(16)) 

@numba.jit
def test_max_mod_jit():
    return np.mod(np.uint64(np.iinfo(np.uint64).max), np.uint64(16)) 

test_max_mod(), test_max_mod_jit()

(15, 18446744073709551615)

numba.__version__

'0.45.1'

Took me quite some time to strip down the issue to this, for smaller values of uint64 mod works fine so only a fraction of results were broken.

Any suggestion for a workaround? (any bitwise right rotation or something)?
It was my first usage of numba and the speedup was great (>10x on GBs of data), but this prevents me using it...

[ghost]

Changing the mod from the numpy mod to the build in % mod produces the correct result as a heads up.

import numpy as np
import numba

def test_max_mod():
return np.uint64(np.iinfo(np.uint64).max)%np.uint64(16)

@numba.jit
def test_max_mod_jit():
return np.uint64(np.iinfo(np.uint64).max)%np.uint64(16)

print(test_max_mod(), test_max_mod_jit())

[miloslen]

Changing the mod from the numpy mod to the build in % mod produces the correct result as a heads up.

import numpy as np
import numba

def test_max_mod():
return np.uint64(np.iinfo(np.uint64).max)%np.uint64(16)

@numba.jit
def test_max_mod_jit():
return np.uint64(np.iinfo(np.uint64).max)%np.uint64(16)

print(test_max_mod(), test_max_mod_jit())

Thanks for the (partial) workaround...

[stuartarchibald]

Thanks for the report, I can reproduce. Looks like @guilhermeleobas has a fix in progress in #4667.