Cache look-up for functions with optional arguments fails ...

[asodeur]

... for non-singleton default arguments (ie booleans and small ints seem ok, floats fail).

from numba import jit

DEFAULT = 123456789

@jit(nopython=True, cache=True)
def x_or_default(x=DEFAULT):
    return x


x_or_default()

# looks like the compile result gets written into a .nbc file
overloads = x_or_default._cache._cache_file._load_index()
print('Cached overloads: ', overloads)

# but you cannot load the compile result from cache
assert x_or_default._cache.load_overload(
        (types.Omitted(DEFAULT),), x_or_default.targetctx
    ) is not None

# the cause seems to be that the unpickled Omitted(DEFAULT) does not equal
# the newly created one, ie he following fails
assert list(overloads)[0][0][0] == types.Omitted(DEFAULT)
assert list(overloads)[0][0][0] == list(
            x_or_default._cache._cache_file._load_index())[0][0][0]

The code above will work for DEFAULT=123 (likely due to Python's small int caching)

[asodeur]

The issue seems to be with the definition of Omitted.key:

@property
def key(self):
   return type(self.value), id(self.value)

To get the above working you had to return type(self.value), self.value. But that will break for any unhashable default value. Maybe need to keep the current behaviour for unhashable default values.

[stuartarchibald]

Thanks for the report. I can reproduce, even the round trip fails:

from numba import jit
from numba import types

DEFAULT = 123456789

@jit(nopython=True, cache=True)
def x_or_default(x=DEFAULT):
    return x


x_or_default()

# looks like the compile result gets written into a .nbc file
overloads = x_or_default._cache._cache_file._load_index()
print('Cached overloads: ', overloads)

# Round trip fails
key = [_ for _ in overloads][0]
assert x_or_default._cache.load_overload(key, x_or_default.targetctx)

[njriasan]

@stuartarchibald is this issue resolved by #7422? Can this issue be closed?

[stuartarchibald]

@stuartarchibald is this issue resolved by #7422? Can this issue be closed?

@njriasan I think so, OP now works as expected. Thanks