persistent_dict can't store NaNs

[matthiasdiener]

from pytools.persistent_dict import PersistentDict
import numpy as np

pdict = PersistentDict("pytools-test", safe_sync=True)

pdict[np.nan] = 42
assert pdict[np.nan] == 42

$ python t.py
/Users/mdiener/Work/e12test/pytools/pytools/persistent_dict.py:602: CollisionWarning: pytools-test: key collision in cache at '/Users/mdiener/Work/e12test/loopy/xdg-cache/pytools' -- these are sufficiently unlikely that they're often indicative of a broken hash key implementation (that is not considering some elements relevant for equality comparison)
  return self.fetch(key)
Traceback (most recent call last):
  File "/Users/mdiener/Work/e12test/pytools/t.py", line 8, in <module>
    assert pdict[np.nan] == 42
           ~~~~~^^^^^^^^
  File "/Users/mdiener/Work/e12test/pytools/pytools/persistent_dict.py", line 602, in __getitem__
    return self.fetch(key)
           ^^^^^^^^^^^^^^^
  File "/Users/mdiener/Work/e12test/pytools/pytools/persistent_dict.py", line 840, in fetch
    self._collision_check(key, stored_key)
  File "/Users/mdiener/Work/e12test/pytools/pytools/persistent_dict.py", line 552, in _collision_check
    raise NoSuchEntryCollisionError(key)
pytools.persistent_dict.NoSuchEntryCollisionError: nan

🤦

Perhaps we should also check if the value is a Nan in the collision check? Are there other cases that could be at issue here (inf seems to be fine)?

cc inducer/loopy#828 [major plot twist!]

[inducer]

We can't really check for NaN in persistent dict, as we would have to traverse the entire key (expensive!). We'll need to rely on the user to not have NaN values. It might be worth a word of warning in the persistent dict documentation though.

Beyond that, I think this is more a loopy bug. It should replace NaN values with pymbolic NaNs.

[matthiasdiener]

Could we emit a warning while generating the key when we encounter a NaN? (done in #288)

Edit: Or perhaps, set a flag somewhere for the key that stores whether key contained a NaN, and perform the key traversal when that flag is set in the collision check.

[inducer]

"Walking a value" is not really well-defined in Python, so I'd rather not try.