Hashing of jitclass doesn't work

[louisabraham]

Reporting a bug

This is my code

import numba
from numba import njit, jitclass


@jitclass([("a", numba.int64), ("b", numba.int64)])
class A:
    def __init__(self, a, b):
        self.a = a
        self.b = b

    def __hash__(self):
        return self.a + self.b


print(A.class_type())

x = A(1, 2)
y = A(1, 2)

print(hash(x))
print(hash(y))

It appears that:

• the override of __hash__ doesn't work
• the hash is not the same even for objects that are the same (like in Python)

Is there a way to get a value-dependent hashing?

Or is there a way to subclass namedtuples and still compile?

[stuartarchibald]

Thanks for the report.

It appears that:

• the override of __hash__ doesn't work
• the hash is not the same even for objects that are the same (like in Python)

I think this needs qualifying, it seems like the override of the __hash__ method does not work as expected from the CPython interpreter, it does however work fine from compiled code, e.g.:

import numba
from numba import njit, jitclass


@jitclass([("a", numba.int64), ("b", numba.int64)])
class A:
    def __init__(self, a, b):
        self.a = a
        self.b = b

    def __hash__(self):
        return self.a + self.b


@njit
def printhash(z):
    print(hash(z))

print(A.class_type())

x = A(1, 2)
y = A(1, 2)

print(hash(x))
print(hash(y))

printhash(x)
printhash(y)

gives:

() -> jitclass.A#55d3d1eac8c8<a:int64,b:int64>
8766512762915
8766512086525
3
3

I think that this works because the compiled call to hash() forces resolution of the method on the compiled class whereas calling hash() on the jitclass instance object does not.

@sklam any thoughts on whether we ought to support this?

[louisabraham]

Interesting! Does it mean that code relying on dicts won't produce the same results after compilation?

[sklam]

This should be a quick one. It should just be a matter for the interface not being exported. I will need to check on this.