gh-143667: Specialize FOR_ITER for dict.items(), and scalar replace the pair in the JIT

[Fidget-Spinner]

This PR specializes FOR_ITER for iterating dct.items(). Furthermore, in the JIT, it goes one step further by avoiding allocation/reuse/extra writes to the pair tuple entirely. Choosing to scalar replace the tuple on the stack instead.

According to stats, dict items makes up 33% of specialization failures.

Roughly 33% reduction in execution time for iteration through dictionary items in the JIT on microbenchmark:

def testfunc(dct):
    for k, v in dct.items():
        pass

n = 10000000
dct = dict(zip(range(n), range(n)))

import sys
import time
start = time.time()
for _ in range(20):
    testfunc(dct)
end = time.time()
print(end - start)

JIT on on both
Main: 3.54s
This branch: 2.34s

• Issue: Specialize FOR_ITER for dict.items() #143667

[eendebakpt]

Nice work! Would the same approach work for enumerate?

[Fidget-Spinner]

Nice work! Would the same approach work for enumerate?

Yes. I'm hesitant to do it for that though. For enumerate, I'd rather we write a Python equivalent for the JIT and let it specialize on the generators.

[Fidget-Spinner]

@eendebakpt sorry I can't reproduce the speedup on deepcopy anymore. Deepcopy is too unstable on my machine so I have no trust in the results. From what I can see, it's roughly the same speed.

[markshannon]

I don't think we should do this for 2 reasons. I think the optimization is sound, but I'm opposed to the mechanism.

1. We are getting low on spare opcodes, and while this might be a justifiable use of one of those spares, we can do the same optimization in the JIT provided we have the type information provided by the more generic approach in Broader specialization in the Specializing Adaptive Interpreter for better JIT performance #143732
2. The new code in optimizer_bytecodes.c is fragile, as it depends on the exact layout of uops. Such optimizations belong in the virtualization pass. In a virtualization pass, _ITER_NEXT_DICT_ITEMS would be unconditionally replaced with _ITER_NEXT_DICT_ITEMS_UNPACK, pushing a virtual pair to the (shadow) stack