Compile with -O2 instead of -O3

[brandtbucher]

On a hunch, I tried compiling CPython with -O2… which results in a 1% speedup.

If I had to attribute this to anything, I’d guess that it’s because -O3’s aggressive inlining and loop unrolling hurts i-cache locality (the -O2 executable is ~3.5% smaller than the -O3 one).

Benchmark	main	O2
pickle_dict	29.5 us	27.0 us: 1.09x faster
regex_v8	25.5 ms	23.6 ms: 1.08x faster
regex_dna	223 ms	207 ms: 1.08x faster
regex_effbot	3.31 ms	3.08 ms: 1.07x faster
unpack_sequence	46.0 ns	43.9 ns: 1.05x faster
json_loads	28.0 us	27.1 us: 1.03x faster
chaos	83.8 ms	81.3 ms: 1.03x faster
telco	6.42 ms	6.24 ms: 1.03x faster
float	86.9 ms	85.3 ms: 1.02x faster
fannkuch	425 ms	418 ms: 1.02x faster
json_dumps	12.8 ms	12.6 ms: 1.01x faster
pickle_list	4.64 us	4.57 us: 1.01x faster
crypto_pyaes	90.8 ms	89.6 ms: 1.01x faster
unpickle	13.8 us	13.6 us: 1.01x faster
mako	12.7 ms	12.5 ms: 1.01x faster
spectral_norm	119 ms	117 ms: 1.01x faster
meteor_contest	107 ms	106 ms: 1.01x faster
python_startup	9.47 ms	9.39 ms: 1.01x faster
python_startup_no_site	6.36 ms	6.31 ms: 1.01x faster
scimark_fft	358 ms	356 ms: 1.01x faster
xml_etree_parse	156 ms	155 ms: 1.01x faster
unpickle_pure_python	268 us	266 us: 1.01x faster
hexiom	7.78 ms	7.75 ms: 1.00x faster
pickle	10.2 us	10.1 us: 1.00x faster
pidigits	188 ms	188 ms: 1.00x slower
chameleon	7.56 ms	7.61 ms: 1.01x slower
pathlib	19.2 ms	19.3 ms: 1.01x slower
logging_simple	6.72 us	6.76 us: 1.01x slower
scimark_monte_carlo	80.9 ms	81.5 ms: 1.01x slower
raytrace	350 ms	353 ms: 1.01x slower
sympy_expand	512 ms	517 ms: 1.01x slower
xml_etree_process	61.2 ms	61.9 ms: 1.01x slower
xml_etree_iterparse	107 ms	108 ms: 1.01x slower
logging_format	7.32 us	7.40 us: 1.01x slower
django_template	39.4 ms	39.9 ms: 1.01x slower
richards	59.5 ms	60.4 ms: 1.02x slower
logging_silent	124 ns	126 ns: 1.02x slower
go	175 ms	178 ms: 1.02x slower
pyflate	541 ms	550 ms: 1.02x slower
scimark_sparse_mat_mult	4.95 ms	5.09 ms: 1.03x slower
unpickle_list	4.78 us	4.92 us: 1.03x slower
Geometric mean	(ref)	1.01x faster

[tiran]

Red Hat's and Fedora's Python are compiled with -O2. Florian Weimar's blog post https://developers.redhat.com/blog/2018/03/21/compiler-and-linker-flags-gcc also recommends -O2 for most applications:

For many applications, -O2 is a good choice because the additional inlining and loop unrolling introduced by -O3 increases the instruction cache footprint, which ends up reducing performance

Florian also suggests to try -ftree-loop-vectorize with -O2:

For some applications -O3 or -O2 -free-loop-vectorize will provide a significant speed boost. By default, GCC does not perform loop vectorization.

(The blog post has a typo. I have informed Florian that the correct spelling is -ftree-loop-vectorize.

[brandtbucher]

Awesome, thanks for the link! I'll find time today to try that out.

[brandtbucher]

Looks like -O2 -ftree-loop-vectorize is ~1% slower than vanilla -O3:

Slower (25):
- unpack_sequence: 43.7 ns +- 0.9 ns -> 48.8 ns +- 2.9 ns: 1.12x slower
- pickle_dict: 26.8 us +- 0.1 us -> 28.1 us +- 0.1 us: 1.05x slower
- regex_effbot: 3.15 ms +- 0.03 ms -> 3.25 ms +- 0.04 ms: 1.03x slower
- pickle_list: 4.48 us +- 0.04 us -> 4.63 us +- 0.04 us: 1.03x slower
- spectral_norm: 107 ms +- 1 ms -> 111 ms +- 2 ms: 1.03x slower
- scimark_monte_carlo: 77.5 ms +- 0.8 ms -> 79.9 ms +- 1.3 ms: 1.03x slower
- regex_dna: 212 ms +- 1 ms -> 217 ms +- 1 ms: 1.02x slower
- scimark_sor: 154 ms +- 2 ms -> 158 ms +- 2 ms: 1.02x slower
- richards: 56.9 ms +- 0.9 ms -> 57.9 ms +- 1.5 ms: 1.02x slower
- deltablue: 4.85 ms +- 0.06 ms -> 4.93 ms +- 0.06 ms: 1.02x slower
- go: 165 ms +- 1 ms -> 168 ms +- 2 ms: 1.02x slower
- logging_silent: 116 ns +- 3 ns -> 117 ns +- 1 ns: 1.01x slower
- float: 80.5 ms +- 0.9 ms -> 81.4 ms +- 0.8 ms: 1.01x slower
- regex_v8: 23.2 ms +- 0.1 ms -> 23.4 ms +- 0.1 ms: 1.01x slower
- sqlite_synth: 2.74 us +- 0.04 us -> 2.76 us +- 0.07 us: 1.01x slower
- mako: 11.9 ms +- 0.1 ms -> 12.0 ms +- 0.1 ms: 1.01x slower
- pickle_pure_python: 367 us +- 3 us -> 371 us +- 3 us: 1.01x slower
- scimark_sparse_mat_mult: 4.84 ms +- 0.08 ms -> 4.89 ms +- 0.10 ms: 1.01x slower
- tornado_http: 109 ms +- 3 ms -> 110 ms +- 3 ms: 1.01x slower
- pathlib: 19.0 ms +- 0.3 ms -> 19.1 ms +- 0.2 ms: 1.01x slower
- 2to3: 271 ms +- 1 ms -> 272 ms +- 1 ms: 1.00x slower
- regex_compile: 144 ms +- 2 ms -> 145 ms +- 1 ms: 1.00x slower
- dulwich_log: 67.6 ms +- 0.3 ms -> 67.8 ms +- 0.3 ms: 1.00x slower
- python_startup: 8.51 ms +- 0.01 ms -> 8.52 ms +- 0.01 ms: 1.00x slower
- python_startup_no_site: 5.90 ms +- 0.00 ms -> 5.90 ms +- 0.00 ms: 1.00x slower

Faster (15):
- xml_etree_iterparse: 107 ms +- 3 ms -> 105 ms +- 1 ms: 1.02x faster
- nqueens: 90.1 ms +- 0.7 ms -> 88.2 ms +- 0.9 ms: 1.02x faster
- fannkuch: 427 ms +- 4 ms -> 418 ms +- 4 ms: 1.02x faster
- json_dumps: 12.6 ms +- 0.1 ms -> 12.4 ms +- 0.1 ms: 1.02x faster
- unpickle_list: 5.05 us +- 0.04 us -> 4.95 us +- 0.06 us: 1.02x faster
- pidigits: 190 ms +- 0 ms -> 188 ms +- 0 ms: 1.01x faster
- crypto_pyaes: 87.5 ms +- 0.6 ms -> 86.6 ms +- 0.6 ms: 1.01x faster
- hexiom: 7.53 ms +- 0.12 ms -> 7.46 ms +- 0.07 ms: 1.01x faster
- sqlalchemy_imperative: 18.6 ms +- 0.5 ms -> 18.4 ms +- 0.2 ms: 1.01x faster
- xml_etree_parse: 155 ms +- 3 ms -> 154 ms +- 1 ms: 1.01x faster
- django_template: 37.2 ms +- 0.5 ms -> 37.0 ms +- 0.5 ms: 1.01x faster
- pickle: 9.97 us +- 0.16 us -> 9.89 us +- 0.06 us: 1.01x faster
- xml_etree_generate: 80.4 ms +- 0.7 ms -> 80.0 ms +- 0.9 ms: 1.01x faster
- logging_simple: 6.15 us +- 0.08 us -> 6.12 us +- 0.10 us: 1.01x faster
- raytrace: 331 ms +- 2 ms -> 330 ms +- 2 ms: 1.00x faster

Benchmark hidden because not significant (18): chameleon, chaos, json_loads, logging_format, meteor_contest, nbody, pyflate, scimark_fft, scimark_lu, sqlalchemy_declarative, sympy_expand, sympy_integrate, sympy_sum, sympy_str, telco, unpickle, unpickle_pure_python, xml_etree_process

Geometric mean: 1.01x slower

[tiran]

So -O2 > -O3 > -O2 -ftree-loop-vectorize?