gh-154115: Fix typing._tp_cache calling the wrapped function twice on TypeError

[fedonman]

_tp_cache wraps typing's __getitem__ methods with an lru_cache and falls back to calling the original function when it sees a TypeError, on the assumption that the TypeError came from hashing an unhashable argument. A TypeError raised by the wrapped function itself was caught the same way, so with hashable arguments the function ran once inside the cache and then a second time through the fallback before the error propagated.

You can see it directly:

import typing

calls = []

@typing._tp_cache
def f(x):
    calls.append(x)
    raise TypeError("unrelated to hashability")

try:
    f("value")
except TypeError:
    pass

print(calls)  # ['value', 'value']

and through the public API, for example typing.List[int, str] runs its whole __getitem__ twice before raising.

The fix only takes the fallback when the arguments really are unhashable. It rebuilds the same cache key that lru_cache builds (functools._make_key, which also folds in the argument types when typed=True) and re-raises when that key is hashable, so a TypeError coming from the wrapped function is no longer turned into a second call. The successful path is unchanged.

• Issue: typing._tp_cache runs the wrapped function twice when it raises TypeError #154115

[fedonman]

@zware part of EuroPython sprint.

[JelleZijlstra]

I'm not convinced we should change anything here. The issue description doesn't show any evidence that this causes practical problems, and the proposed code change makes things more complicated.

[fedonman]

@JelleZijlstra Should I close the PR then?

[JelleZijlstra]

We can wait for some other opinions.