use -flto=thin for clang-cl on Windows

[chris-eibl]

This started off as a build time analysis (#130090 (comment)), but since I now have the infrastructure, I tried -flto=thin, too:

• faster in building 520.6 vs 651.2 seconds
• is neutral on the pyperformance benchmarks
• would bring us in sync with Linux, because there CONFIGURE_CFLAGS_NODIST and CONFIGURE_LDFLAGS_NOLTO both use -flto=thin when I configure for clang in WSL Ubuntu-24.04. See also the discussion why not to use full -flto in Revert to default fullLTO on Clang #130048

Benchmark	clang.pgo.20.1.0-rc2	clang.pgo.thin.20.1.0-rc2
Geometric mean	(ref)	1.00x faster

Detailed pybenchmark results

Benchmark	clang.pgo.20.1.0-rc2	clang.pgo.thin.20.1.0-rc2
float	95.0 ms	89.7 ms: 1.06x faster
json_loads	29.8 us	28.6 us: 1.04x faster
mdp	2.86 sec	2.77 sec: 1.03x faster
html5lib	68.3 ms	66.2 ms: 1.03x faster
async_tree_none_tg	330 ms	320 ms: 1.03x faster
pyflate	518 ms	505 ms: 1.03x faster
sqlite_synth	3.21 us	3.13 us: 1.03x faster
pidigits	228 ms	223 ms: 1.02x faster
bench_mp_pool	168 ms	165 ms: 1.02x faster
async_tree_eager_io	742 ms	727 ms: 1.02x faster
generators	34.5 ms	33.8 ms: 1.02x faster
comprehensions	18.3 us	17.9 us: 1.02x faster
async_tree_cpu_io_mixed	641 ms	629 ms: 1.02x faster
scimark_sparse_mat_mult	4.51 ms	4.43 ms: 1.02x faster
async_tree_memoization	425 ms	417 ms: 1.02x faster
sympy_expand	538 ms	529 ms: 1.02x faster
unpack_sequence	57.0 ns	56.0 ns: 1.02x faster
regex_dna	209 ms	205 ms: 1.02x faster
async_generators	465 ms	458 ms: 1.02x faster
scimark_sor	140 ms	137 ms: 1.02x faster
sympy_str	319 ms	314 ms: 1.02x faster
async_tree_io_tg	751 ms	740 ms: 1.01x faster
regex_effbot	3.14 ms	3.10 ms: 1.01x faster
async_tree_eager_tg	272 ms	268 ms: 1.01x faster
pickle_dict	27.3 us	27.0 us: 1.01x faster
async_tree_eager_memoization_tg	363 ms	359 ms: 1.01x faster
sympy_integrate	22.5 ms	22.2 ms: 1.01x faster
sympy_sum	181 ms	179 ms: 1.01x faster
2to3	390 ms	386 ms: 1.01x faster
hexiom	6.68 ms	6.61 ms: 1.01x faster
docutils	3.03 sec	3.00 sec: 1.01x faster
sqlglot_normalize	121 ms	120 ms: 1.01x faster
async_tree_memoization_tg	392 ms	389 ms: 1.01x faster
async_tree_cpu_io_mixed_tg	614 ms	609 ms: 1.01x faster
tomli_loads	2.20 sec	2.18 sec: 1.01x faster
spectral_norm	102 ms	101 ms: 1.01x faster
python_startup_no_site	34.4 ms	34.2 ms: 1.01x faster
genshi_text	24.6 ms	24.5 ms: 1.01x faster
dulwich_log	119 ms	118 ms: 1.00x faster
go	128 ms	128 ms: 1.00x faster
deltablue	3.62 ms	3.63 ms: 1.00x slower
unpickle_pure_python	247 us	248 us: 1.00x slower
xml_etree_generate	107 ms	107 ms: 1.01x slower
django_template	39.2 ms	39.4 ms: 1.01x slower
coroutines	24.8 ms	25.0 ms: 1.01x slower
mako	13.3 ms	13.5 ms: 1.01x slower
unpickle	15.9 us	16.1 us: 1.01x slower
nbody	119 ms	121 ms: 1.01x slower
fannkuch	465 ms	472 ms: 1.01x slower
crypto_pyaes	81.3 ms	82.6 ms: 1.02x slower
json_dumps	11.5 ms	11.7 ms: 1.02x slower
deepcopy	285 us	291 us: 1.02x slower
pprint_safe_repr	858 ms	876 ms: 1.02x slower
xml_etree_iterparse	136 ms	139 ms: 1.02x slower
gc_traversal	5.03 ms	5.14 ms: 1.02x slower
meteor_contest	115 ms	117 ms: 1.02x slower
deepcopy_memo	33.8 us	34.7 us: 1.03x slower
richards_super	51.1 ms	52.6 ms: 1.03x slower
scimark_fft	327 ms	337 ms: 1.03x slower
richards	44.9 ms	46.3 ms: 1.03x slower
pickle_list	4.83 us	4.99 us: 1.03x slower
deepcopy_reduce	2.93 us	3.03 us: 1.03x slower
pprint_pformat	1.74 sec	1.80 sec: 1.03x slower
logging_simple	10.9 us	11.4 us: 1.05x slower
logging_format	12.1 us	12.6 us: 1.05x slower
xml_etree_parse	197 ms	208 ms: 1.05x slower
Geometric mean	(ref)	1.00x faster

	pgo_clang_20.1.0-rc2	pgo_clang_thin_20.1.0-rc2
pginstr	297.2	219.3
pgo	70.0	69.0
kill	1.2	0.5
pgupd	282.8	231.7
total time	651.2	520.6

Details pginstrument

	pgo_clang_20.1.0-rc2	pgo_clang_thin_20.1.0-rc2
_freeze_module	38.5	40.0
python314	141.5	81.3
pyexpat	52.7	3.9
_elementtree	51.8	5.3
sqlite3	46.0	42.4
liblzma	18.2	16.5
_decimal	12.4	7.7
_testcapi	8.3	7.1
_bz2	7.0	4.9
_ctypes	6.9	7.5
_testlimitedcapi	4.9	4.3
_wmi	4.5	3.0
_overlapped	4.5	3.2
_asyncio	4.0	5.2
_lzma	3.8	1.8
_ssl	3.7	5.5
_ctypes_test	3.7	3.4
_multiprocessing	3.5	2.7
_sqlite3	3.4	2.8
venvwlauncher	3.3	2.7
_zoneinfo	3.1	3.4
unicodedata	2.7	3.0
pyshellext	2.7	2.6
pyw	2.7	2.7
py	2.6	2.5
_socket	2.4	3.7
_testinternalcapi	2.4	2.2
_tkinter	2.2	4.1
_testclinic	2.0	1.9
_hashlib	1.8	3.1
select	1.8	2.2
venvlauncher	1.8	1.7
winsound	1.7	3.3
_uuid	1.6	3.2
_queue	1.6	2.3
_testembed	1.5	1.5
_testbuffer	1.4	1.3
pythonw	1.1	1.1
_testconsole	1.1	1.1
_testmultiphase	1.0	1.0
_testsinglephase	1.0	1.0
python	1.0	0.9
_testclinic_limited	0.9	0.9
_testimportmultiple	0.9	0.9
python3	0.5	0.5
total	465.8	303.3

Details pgupdate

	pgo_clang_20.1.0-rc2	pgo_clang_thin_20.1.0-rc2
_freeze_module	38.0	39.5
python314	141.9	95.4
sqlite3	44.4	42.9
liblzma	17.3	16.5
_decimal	11.2	8.7
_testcapi	8.6	7.3
_ctypes	8.0	7.2
_bz2	7.8	5.5
_ssl	5.2	5.6
_testlimitedcapi	5.0	4.2
pyexpat	4.6	3.6
_asyncio	4.5	4.6
_socket	4.3	3.5
_tkinter	4.0	4.2
_ctypes_test	3.7	3.4
_overlapped	3.5	3.7
_elementtree	3.5	4.5
_wmi	3.5	3.1
_zoneinfo	3.2	3.2
_lzma	3.2	1.9
unicodedata	3.2	3.0
_sqlite3	3.1	2.7
_hashlib	3.1	3.3
venvwlauncher	3.1	3.0
_multiprocessing	2.8	2.6
pyshellext	2.7	2.6
pyw	2.6	2.6
_uuid	2.6	2.8
py	2.6	2.7
_testinternalcapi	2.4	2.2
_testclinic	2.0	1.9
_queue	1.9	2.2
winsound	1.8	3.0
venvlauncher	1.7	1.5
select	1.6	2.0
_testembed	1.5	1.4
_testbuffer	1.4	1.3
_testconsole	1.1	1.0
pythonw	1.1	1.1
_testmultiphase	1.0	1.1
_testsinglephase	1.0	1.0
python	1.0	0.9
_testclinic_limited	0.9	0.9
_testimportmultiple	0.9	0.9
python3	0.5	0.5
total	372.9	316.8

Linked PRs

• GH-131035: use -flto=thin for clang-cl on Windows #131036

[chris-eibl]

@zooba: yippie, that was a fast one 🚀

PR is merged. Shall I close the issue (or is this a core-dev task)? I see, I get the option to "close as completed".

[Fidget-Spinner]

You can close your own issues whenever you think it's done.