gh-149481: skip FOR_ITER inline specialization for Python __next__

[NekoAsakura]

Thanks for the careful bisect and the empirical fix from @savannahostrowski. ❤️

Benchmark

Benchmark	unpatched	patched	Δ
for_iter_dict_items	185.8 ms ± 1.6 ms	187.0 ms ± 2.8 ms	+0.68%
for_iter_dict_keys	154.1 ms ± 2.1 ms	153.9 ms ± 2.6 ms	-0.11%
for_iter_dict_values	148.2 ms ± 1.0 ms	146.9 ms ± 2.8 ms	-0.85%
for_iter_set	179.1 ms ± 1.9 ms	179.0 ms ± 2.2 ms	-0.07%
for_iter_reversed	136.6 ms ± 1.2 ms	136.0 ms ± 1.9 ms	-0.47%
for_iter_enumerate	201.1 ms ± 2.9 ms	201.1 ms ± 2.4 ms	-0.04%
for_iter_zip	230.6 ms ± 1.2 ms	227.9 ms ± 1.3 ms	-1.15%
for_iter_list	114.2 ms ± 1.8 ms	115.6 ms ± 2.2 ms	+1.26%
for_iter_tuple	104.4 ms ± 1.5 ms	104.7 ms ± 1.7 ms	+0.35%
for_iter_range	137.6 ms ± 2.5 ms	137.8 ms ± 1.9 ms	+0.13%
Geometric mean			-0.03%

Root cause analyse (from opus 4.7)

When __next__ is Python-defined, tp_iternext is the generic slot_tp_iternext slot wrapper that re-enters the eval loop via vectorcall_method. Calling it through the new _ITER_NEXT_INLINE (which captures tp_iternext as a baked-in operand and lives in a tight _JUMP_TO_TOP loop) is correct functionally, but each new outer-iteration class causes _GUARD_TYPE_ITER to fail and warm a side-exit. After ~SIDE_EXIT_INITIAL_VALUE (4000) hits at each side-exit, a side trace forms there; _EXIT_TRACE jumps trace-to-trace via TIER2_TO_TIER2, which doesn't re-check the eval-breaker. With MAX_CHAIN_DEPTH=4 traces possible, the chain accumulates ~4×4000+4002 ≈ 20k inner hits before the system spirals into a tier-2-only loop with no signal checkpoints — matching the observed exact threshold. The pre-PR _FOR_ITER_TIER_TWO path goes through _PyForIter_VirtualIteratorNext, which uses an indirect dispatch on Py_TYPE(iter_o)->tp_iternext and does not produce a per-type guard, so no side-trace storm forms. Falling back to that path for slot iterators sidesteps the issue while keeping _ITER_NEXT_INLINE for C-level iterators (dict, set, enumerate, zip, reversed, etc.), which the existing test_for_iter_direct_* tests exercise.

• Issue: JIT hang in xml.etree.iterparse #149481

[savannahostrowski]

Yep, this was pretty much the exact fix I had locally. Thanks for the quick turnaround on this @NekoAsakura ❤️

I'll wait to see if anyone else wants to have a look.

[markshannon]

Looks good. Thanks for doing this.