gh-127750: Fix singledispatchmethod caching (v2)

[eendebakpt]

Version based on idea from @dg-pb in #127839. This version

• Fixes the issue of collisions between different objects with equal __hash__/__eq__ Regression in Django with singledispatchmethod on models #127750
• Fixes the issue of keeping object instances alive
• Adds two regression tests

There is still a cache (stored on the object instances). Quick benchmark (windows, non-pgo):

bench singledispatchmethod: Mean +- std dev: [main] 798 ns +- 64 ns -> [prx] 495 ns +- 38 ns: 1.61x faster

Benchmark hidden because not significant (1): bench singledispatchmethod slots

Geometric mean: 1.26x faster

(note that the alternative to this PR is not to keep main, but to revert #107148)

• Issue: Regression in Django with singledispatchmethod on models #127750

[dg-pb]

Check cached_property.__set_name__, it has some more stuff in it - might be needed here as well.

[eendebakpt]

Hmm. The additions there prevent something like this:

@dataclass(frozen=True)
class A:
    value: int

    @singledispatchmethod
    def dispatch(self, x):
        return id(self)

    renamed_dispatch = dispatch # allowed? if so, how should it behave

The corresponding test for the cached_property for this is

cpython/Lib/test/test_functools.py

Line 3315 in 34e840f

def test_reuse_different_names(self):

.

But on current main renaming is allowed for the singledispatchmethod.

I am not sure here what the desired behavior is (and why)

[dg-pb]

If this implementation is desirable, maybe later someone who knows more about this can comment.

[vodik]

As far as I know, the only reason cached properties can't be renamed is because the cache is keyed by the attribute's name.

Allowing a rebind would disconnect the cached property from it's cached value.

[vodik]

Actually, I think you might want to either ignore renames or do something along these lines (ignoring error handling):

    if self.attrname:
        cache[name] = cache.pop(self.attrname)
    self.attrname = name

As far as I know, each binding shares the same instance of the descriptor, so as long as the cache key is constant, it should work no matter how many times it's been renamed.

[dg-pb]

Allowing a rebind would disconnect the cached property from it's cached value.

This is kind of the same situation.

If rename is allowed, then it would simply cache to the last attrname. Drawback is that there is a small risk for unused cached methods.

I think it might be most straight forward to copy+paste cached_property.__set_name__. It does seem a sensible restriction. It comes at expense of flexibility, but personally, I have never run into that TypeError.

Also, it will be easier to address changes/improvements when 2 implementations that use the same caching approach are aligned.