After a couple of rounds, this is the benchmark result I got for the attached patch. It seems to be a simple, non-performance degrading fix.

 python -m perf compare_to origin.json patch.json -G --min-speed=2
Slower (1):
- sqlalchemy_declarative: 213 ms +- 4 ms -> 219 ms +- 9 ms: 1.03x slower (+3%)

Faster (9):
- logging_format: 15.1 us +- 0.3 us -> 14.5 us +- 0.3 us: 1.05x faster (-5%)
- hexiom: 15.0 ms +- 0.5 ms -> 14.5 ms +- 0.2 ms: 1.03x faster (-3%)
- json_loads: 39.6 us +- 0.7 us -> 38.5 us +- 0.5 us: 1.03x faster (-3%)
- json_dumps: 17.2 ms +- 0.4 ms -> 16.7 ms +- 0.2 ms: 1.03x faster (-3%)
- fannkuch: 634 ms +- 4 ms -> 619 ms +- 6 ms: 1.02x faster (-2%)
- django_template: 183 ms +- 3 ms -> 178 ms +- 5 ms: 1.02x faster (-2%)
- float: 153 ms +- 2 ms -> 150 ms +- 1 ms: 1.02x faster (-2%)
- raytrace: 761 ms +- 19 ms -> 745 ms +- 8 ms: 1.02x faster (-2%)
- regex_effbot: 3.69 ms +- 0.06 ms -> 3.61 ms +- 0.05 ms: 1.02x faster (-2%)

Benchmark hidden because not significant (47): 2to3, chaos, crypto_pyaes, deltablue, dulwich_log, go, html5lib, logging_silent, logging_simple, mako, meteor_contest, nbody, nqueens, pathlib, pickle, pickle_dict, pickle_list, pickle_pure_python, pidigits, python_startup, python_startup_no_site, regex_compile, regex_dna, regex_v8, richards, scimark_fft, scimark_lu, scimark_monte_carlo, scimark_sor, scimark_sparse_mat_mult, spectral_norm, sqlalchemy_imperative, sqlite_synth, sympy_expand, sympy_integrate, sympy_sum, sympy_str, telco, tornado_http, unpack_sequence, unpickle, unpickle_list, unpickle_pure_python, xml_etree_parse, xml_etree_iterparse, xml_etree_generate, xml_etree_process