GH-100762: Don't call gen.throw() in gen.close(), unless necessary.

[markshannon]

Change oparg of YIELD_VALUE to hold the exception handler depth.
We can then test that in gen.close() to avoid calling gen.throw() unless necessary.

• Issue: Generator finalization is slower in 3.11 vs 3.10 #100762

[markshannon]

Using the benchmarks in #101011 (comment), but using timeit instead of pyperf and a non-pgo and no-lto build because I'm impatient.
I get no significant change on "comprehension", "generator exhaust", and "all", but runtime is about 60% for "generator" and "all with comparison"

[gvanrossum]

I guess this will not be easy to backport, as we won't want to change the meaning of oparg in 3.11? Then again it appears to be unused except for the assert. Maybe it should just be a bit indicating whether it's safe to optimize out the throw?

[brandtbucher]

I guess this will not be easy to backport, as we won't want to change the meaning of oparg in 3.11?

I think a similar change in 3.11 would actually be simpler: just check if an exception table exists. If not, don't throw.

The only reason we need the YIELD_VALUE oparg in this 3.12 patch is because every generator now has an implicit try/except around the entire body, so it's a bit trickier to figure out if we're in that handler (which we can safely ignore) or something else.

[gvanrossum]

Great, it'll just be different. :-)

[miss-islington]

Thanks @markshannon for the PR 🌮🎉.. I'm working now to backport this PR to: 3.11.
🐍🍒⛏🤖

[miss-islington]

Sorry, @markshannon, I could not cleanly backport this to 3.11 due to a conflict.
Please backport using cherry_picker on command line.
cherry_picker f02fa64bf2d03ef7a28650c164e17a5fb5d8543d 3.11

[brandtbucher]

I don’t think this should be backported as-is, right? (See my comment above.)

[markshannon]

@brandtbucher agreed.
The changes to genobject.c need backporting (with the change you suggest for determining when to skip the call to gen.throw(), but the other changes are 3.12 specific.