gh-141510: Update specializer to support frozendict

[corona10]

• Issue: PEP 814: Add built-in frozendict type #141510

[corona10]

@Fidget-Spinner,
Since @markshannon prefers to use the current specializer pipeline (and I am also fine with this option), I opened a new PR.

[vstinner]

LGTM. Your new PR optimizes two bytecodes instead of just one, and the change is smaller, nice!

In the future, we might use different opcodes for specialized frozendict operations to avoid locking, but we should first measure the performance benefits before considering going that far in terms in optimization. This change is a nice tradeoff between code complexity and performance.

[markshannon]

@vstinner@corona10

This breaks STORE_SUBSCR_DICT as it allows stores to frozen dicts.
Please revert or fix.

To fix this you'll need to add a new guard for any dict. Use that for loads, and keep the old dict-only guard for stores.

And please add tests that specialized stores to frozen dicts fail.

[corona10]

I will fix this today

[markshannon]

BTW, it's quite tricky to write the test, you'll need a non-branching way to get a dict for specializing and a frozendict once specialized. This is a pattern we've used before:

dicts = [ {...}, frozendict(...) ]
...
for i in range(...):
d = dicts[i == 100]
d[...] = ...

Actually, that's only necessary for JIT tests, a simple d = frozendict() if i == 100 else {} should work as well.

[corona10]

#145039 is PR for fix.