If this addresses the "class attr simple" case, then according to the stats it should improve the hit rate by 4-5% (13.3% * 0.36).
So, something isn't right there. Are the stats right?

So, something isn't right there. Are the stats right?

I assume so. Here are the stats for the base commit and this branch.

For LOAD_ATTR, deferreds fell by 1.3%, hits rose by 0.7%, and misses rose by 0.6%.

I'm guessing that there are two reasons why the number of hits didn't increase by the full 5% you're expecting:

• About 1/3 of the new instructions that do specialize end up missing (so we should really be expecting a ~3% increase in hits).
• More importantly, it looks like nearly all of the cases that used to fail with "class attr simple" are now failing with "has managed dict" (line 860 in this PR). You can see that, before, we had ~600k each of "class attr simple" and "has managed dict" failures. Now, we have 0 "class attr simple" failures, and 1.1M "has managed dict" failures.

So, it appears that "class attr simple" is not so "simple" after all... 5 out of 6 instances that reach it have a __dict__!

When (if?) you fix the merge conflict, could you add those new opcodes to the LOAD_ATTR family and properly declare their cache/stack effects, letting the generator take care of those? I recently modernized all the existing family members (gh-101488).