Optimizing `super().meth()` via adaptive superinstructions

[Fidget-Spinner]

Please see #239 for the discussion and https://bugs.python.org/issue46564 for the BPO.

[Fidget-Spinner]

CPython PR is at python/cpython#30992.

[Fidget-Spinner]

@markshannon your 3.11 TODO reminded me that this is dead in the water. Do you think this will ever be worth it?

With the new inline cache infrastructure. It is easier to share caches across multiple instructions. But IMO the complexity involved makes it unpalatable.

For the record, the PR above measured no speedups in pyperformance, but in super() microbenchmarks, it saw super().meth() speeding up by nearly 2.2x. With super().meth() now only having a ~10% slowdown over self.meth() (basically, non-existent in the real world), versus a 2x slowdown previously. In fact, the new super.meth() is as fast as or even faster than pre-3.11 plain self.meth() calls thanks to caching of method lookups.

[gvanrossum]

Hi @Fidget-Spinner I'm a little confused about what you're saying at the end. I understand the PR (that won't ever be merged) saw a 2.2 speedup. After that, are you saying that in the main branch, super().meth() is now 10% slower than self.meth(), which is an improvement over that ratio for 3.10?

[Fidget-Spinner]

@gvanrossum Sorry, I meant specifically in that PR, super().meth() is only 10% slower than self.meth(). In 3.10, super().meth() was around 2x slower than self.meth().

On main, it should still be around 2x slower. Although we optimized super() with vectorcall support, we also optimized self.meth() via specialization. Most likely the latter outstrips the former. I have yet to benchmark main recently.

[gvanrossum]

@Fidget-Spinner Do we still need this issue?

[Fidget-Spinner]

@gvanrossum Nope. Carl has has merged his version, so it's done.