>  I think the patch should not be pushed without such analysis. Perhaps Raymond will found a time to do this.

Ok, I won't push until expert's LGTM.

> One possible optimization is removing freeslot completely. because:
> 
> * freeslot is used only when inserting. finding is more important.
> * insertdict() can find dummy quickly, only looking dk_indices.
>
> But before trying optimization,

I found this is not only optimization, but also refactoring.
There is function for finding insertion point already, and insertdict()
and PyDict_SetDefault() use it in some cases.
So this can be done without adding any code.

+ ../python.default -m perf compare_to default.json patched.json -G --min-speed=2
Slower (5):
- logging_silent: 736 ns +- 7 ns -> 758 ns +- 15 ns: 1.03x slower (+3%)
- unpickle: 32.2 us +- 0.4 us -> 33.0 us +- 0.3 us: 1.03x slower (+3%)
- unpickle_pure_python: 829 us +- 13 us -> 849 us +- 13 us: 1.02x slower (+2%)
- meteor_contest: 200 ms +- 2 ms -> 204 ms +- 2 ms: 1.02x slower (+2%)
- scimark_sor: 482 ms +- 8 ms -> 493 ms +- 9 ms: 1.02x slower (+2%)

Faster (6):
- unpack_sequence: 132 ns +- 2 ns -> 123 ns +- 0 ns: 1.07x faster (-7%)
- call_simple: 14.3 ms +- 0.5 ms -> 13.7 ms +- 0.3 ms: 1.05x faster (-5%)
- sqlite_synth: 10.3 us +- 0.3 us -> 9.91 us +- 0.21 us: 1.03x faster (-3%)
- mako: 40.7 ms +- 0.5 ms -> 39.8 ms +- 0.3 ms: 1.02x faster (-2%)
- xml_etree_parse: 319 ms +- 11 ms -> 311 ms +- 14 ms: 1.02x faster (-2%)
- chameleon: 30.4 ms +- 0.4 ms -> 29.8 ms +- 0.3 ms: 1.02x faster (-2%)

Benchmark hidden because not significant (53): ...