I did some more playing around locally and found we could improve performance even more with a few tweaks. Proposed a PR to your branch ateendebakpt#5, which also adds some more test coverage.

@AlexWaygood I merged your PR and added one more optimization: we can skip theobj is not None check since forobj=None the weakref indexing will fail with aTypeError.

One more optimization is inhttps://github.com/eendebakpt/cpython/pull/6/files. It eliminates the caching variable. The assumption there is that ifself._method_cache[obj] returns aKeyError, the expressionself._method_cache[obj] = ... is safe.

I still have to do benchmarking on this.

Latest benchmarks:

bench singledispatchmethod: Mean +- std dev: [main] 2.48 us +- 0.02 us -> [pr] 1.04 us +- 0.02 us: 2.38x fasterbench classmethod: Mean +- std dev: [main] 3.45 us +- 0.06 us -> [pr] 3.50 us +- 0.07 us: 1.01x slowerbench classmethod on instance: Mean +- std dev: [main] 3.49 us +- 0.06 us -> [pr] 1.12 us +- 0.01 us: 3.11x fasterbench slotted class: Mean +- std dev: [main] 2.48 us +- 0.03 us -> [pr] 2.53 us +- 0.03 us: 1.02x slowerGeometric mean: 1.64x faster

So the performance improves for the common case (normal methods) and is more or less equal for class methods and slotted classes.

One more variation is to eliminate attribute lookups in the dispatch wrapper:eendebakpt/cpython@singledispatchmethod3...eendebakpt:cpython:singledispatchmethod3b
This results in (compared to this PR):

bench singledispatchmethod: Mean +- std dev: [pr] 1.04 us +- 0.02 us -> [pr_lookup] 979 ns +- 15 ns: 1.06x fasterbench classmethod on instance: Mean +- std dev: [pr] 1.12 us +- 0.01 us -> [pr_lookup] 1.06 us +- 0.01 us: 1.06x fasterbench slotted class: Mean +- std dev: [pr] 2.53 us +- 0.03 us -> [pr_lookup] 2.52 us +- 0.03 us: 1.01x fasterBenchmark hidden because not significant (1): bench classmethodGeometric mean: 1.03x faster

One more variation is to eliminate attribute lookups in the dispatch wrapper:eendebakpt/cpython@singledispatchmethod3...eendebakpt:cpython:singledispatchmethod3b

I can reproduce the speedup locally (great spot!), and it seems like a reasonable thing to do. Please name the variabledispatch rather thandp, though :)

One other optimisation that shaves around 5 microseconds off the benchmark for slotted classes and other objects that can't be weakref'd (but doesn't do much for those that can):

--- a/Lib/functools.py+++ b/Lib/functools.py@@ -959,7 +959,7 @@ def _method(*args, **kwargs):             method = self.dispatcher.dispatch(args[0].__class__)             return method.__get__(obj, cls)(*args, **kwargs)-        _method.__isabstractmethod__ = self.__isabstractmethod__+        _method.__isabstractmethod__ = getattr(self.func, "__isabstractmethod__", False)

But if you want to do ^that, you should probably add a comment about why we're duplicating the logic in the__isabstractmethod__ property.

One other optimisation that shaves around 5 microseconds off the benchmark for slotted classes and other objects that can't be weakref'd (but doesn't do much for those that can):

--- a/Lib/functools.py+++ b/Lib/functools.py@@ -959,7 +959,7 @@ def _method(*args, **kwargs):             method = self.dispatcher.dispatch(args[0].__class__)             return method.__get__(obj, cls)(*args, **kwargs)-        _method.__isabstractmethod__ = self.__isabstractmethod__+        _method.__isabstractmethod__ = getattr(self.func, "__isabstractmethod__", False)

Hmm, I thought I could see a speedup from this earlier, but now I can no longer reproduce it. I assume it was only noise -- best to leave it.

@mental32 and@eendebakpt, thanks so much for all your work on this!