Add support for building cpython with clang thin lto #88506

holmanb · 2021-06-07T22:38:11Z

BPO	44340
Nosy	@gpshead, @ned-deily, @ambv, @corona10, @FFY00, @holmanb
PRs	bpo-44340: Add support for building with clang thin lto via --with-lto=thin #26585 bpo-44340: Add support for building with clang full/thin lto #27231 bpo-44340: Update whatsnews for ThinLTO #28229

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2021-07-19.10:55:14.358>
created_at = <Date 2021-06-07.22:38:11.007>
labels = ['type-feature', 'build', '3.11']
title = 'Add support for building cpython with clang thin lto'
updated_at = <Date 2021-09-08.17:29:40.230>
user = 'https://github.com/holmanb'

bugs.python.org fields:

activity = <Date 2021-09-08.17:29:40.230>
actor = 'lukasz.langa'
assignee = 'none'
closed = True
closed_date = <Date 2021-07-19.10:55:14.358>
closer = 'corona10'
components = ['Build']
creation = <Date 2021-06-07.22:38:11.007>
creator = 'holmanb'
dependencies = []
files = []
hgrepos = []
issue_num = 44340
keywords = ['patch']
message_count = 9.0
messages = ['395293', '397755', '397765', '397766', '397767', '397785', '397786', '397855', '401415']
nosy_count = 6.0
nosy_names = ['gregory.p.smith', 'ned.deily', 'lukasz.langa', 'corona10', 'FFY00', 'holmanb']
pr_nums = ['26585', '27231', '28229']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue44340'
versions = ['Python 3.11']

holmanb · 2021-06-07T22:38:11Z

The existing --with-lto argument could be extended to pass through a value to select non-default lto compiler options:

CC=clang ./configure --with-lto=thin

This would allow default behavior to remain unchanged, while allowing those that want to use thin lto to opt in.

For what it's worth, the tests (make test) pass using clang 11.1.0 and thinlto.

corona10 · 2021-07-18T16:40:36Z

I am now building the experiment environment to compare between thin-lto and full-lto

corona10 · 2021-07-18T23:18:54Z

FYI, Thin LTO shows enhanced build time.

Full LTO (./configure --with-lto=full CC=clang)
real 2m33.740s
user 8m25.695s
sys 0m13.124s

Thin LTO (./configure --with-lto=thin CC=clang)
real 1m51.867s
user 12m53.694s
sys 0m12.786s

corona10 · 2021-07-18T23:47:11Z

The test is executed under the following environments.
There is no significant performance changed.

MS Azure: D8s v3
CentOS Linux release 8.2.2004 (Core)

[corona10@PythonLinux cpython]$ ./python -m pyperformance compare full.json thin.json
full.json
=========

Performance version: 1.0.2
Report on Linux-4.18.0-193.28.1.el8_2.x86_64-x86_64-with-glibc2.28
Number of logical CPUs: 8
Start date: 2021-07-18 23:18:04.644067
End date: 2021-07-18 23:44:20.951457

thin.json
=========

Performance version: 1.0.2
Report on Linux-4.18.0-193.28.1.el8_2.x86_64-x86_64-with-glibc2.28
Number of logical CPUs: 8
Start date: 2021-07-18 22:46:00.717563
End date: 2021-07-18 23:12:19.376766

### 2to3 ###
Mean +- std dev: 570 ms +- 17 ms -> 568 ms +- 20 ms: 1.00x faster
Not significant

### chameleon ###
Mean +- std dev: 16.9 ms +- 0.6 ms -> 16.9 ms +- 0.9 ms: 1.00x slower
Not significant

### chaos ###
Mean +- std dev: 182 ms +- 7 ms -> 179 ms +- 7 ms: 1.02x faster
Not significant

### crypto_pyaes ###
Mean +- std dev: 198 ms +- 6 ms -> 192 ms +- 6 ms: 1.03x faster
Significant (t=5.26)

### deltablue ###
Mean +- std dev: 13.4 ms +- 0.5 ms -> 13.5 ms +- 0.5 ms: 1.01x slower
Not significant

### django_template ###
Mean +- std dev: 94.0 ms +- 3.2 ms -> 91.8 ms +- 3.7 ms: 1.02x faster
Significant (t=3.53)

### dulwich_log ###
Mean +- std dev: 178 ms +- 6 ms -> 176 ms +- 8 ms: 1.02x faster
Not significant

### fannkuch ###
Mean +- std dev: 764 ms +- 17 ms -> 755 ms +- 15 ms: 1.01x faster
Not significant

### float ###
Mean +- std dev: 194 ms +- 8 ms -> 187 ms +- 6 ms: 1.03x faster
Significant (t=4.95)

### go ###
Mean +- std dev: 388 ms +- 14 ms -> 387 ms +- 14 ms: 1.00x faster
Not significant

### hexiom ###
Mean +- std dev: 17.0 ms +- 0.7 ms -> 17.5 ms +- 0.8 ms: 1.03x slower
Significant (t=-3.40)

### json_dumps ###
Mean +- std dev: 22.5 ms +- 0.9 ms -> 22.3 ms +- 0.7 ms: 1.01x faster
Not significant

### json_loads ###
Mean +- std dev: 45.8 us +- 2.3 us -> 46.5 us +- 1.8 us: 1.02x slower
Not significant

### logging_format ###
Mean +- std dev: 19.1 us +- 0.9 us -> 18.7 us +- 0.7 us: 1.02x faster
Not significant

### logging_silent ###
Mean +- std dev: 336 ns +- 17 ns -> 334 ns +- 18 ns: 1.00x faster
Not significant

### logging_simple ###
Mean +- std dev: 17.1 us +- 0.8 us -> 16.7 us +- 0.8 us: 1.03x faster
Significant (t=3.12)

### mako ###
Mean +- std dev: 27.6 ms +- 1.6 ms -> 26.6 ms +- 0.9 ms: 1.04x faster
Significant (t=4.11)

### meteor_contest ###
Mean +- std dev: 172 ms +- 5 ms -> 169 ms +- 5 ms: 1.01x faster
Not significant

### nbody ###
Mean +- std dev: 232 ms +- 8 ms -> 224 ms +- 8 ms: 1.04x faster
Significant (t=6.03)

### nqueens ###
Mean +- std dev: 167 ms +- 7 ms -> 166 ms +- 7 ms: 1.00x faster
Not significant

### pathlib ###
Mean +- std dev: 38.2 ms +- 1.7 ms -> 37.4 ms +- 1.9 ms: 1.02x faster
Significant (t=2.41)

### pickle ###
Mean +- std dev: 19.4 us +- 0.8 us -> 19.5 us +- 0.8 us: 1.00x slower
Not significant

### pickle_dict ###
Mean +- std dev: 43.5 us +- 1.9 us -> 43.0 us +- 1.9 us: 1.01x faster
Not significant

### pickle_list ###
Mean +- std dev: 6.81 us +- 0.26 us -> 6.81 us +- 0.27 us: 1.00x slower
Not significant

### pickle_pure_python ###
Mean +- std dev: 840 us +- 28 us -> 825 us +- 28 us: 1.02x faster
Not significant

### pidigits ###
Mean +- std dev: 294 ms +- 9 ms -> 294 ms +- 9 ms: 1.00x slower
Not significant

### pyflate ###
Mean +- std dev: 1.17 sec +- 0.02 sec -> 1.16 sec +- 0.03 sec: 1.01x faster
Not significant

### python_startup ###
Mean +- std dev: 15.3 ms +- 0.6 ms -> 15.3 ms +- 0.6 ms: 1.00x faster
Not significant

### python_startup_no_site ###
Mean +- std dev: 10.3 ms +- 0.3 ms -> 10.2 ms +- 0.4 ms: 1.01x faster
Not significant

### raytrace ###
Mean +- std dev: 911 ms +- 19 ms -> 911 ms +- 21 ms: 1.00x faster
Not significant

### regex_compile ###
Mean +- std dev: 314 ms +- 12 ms -> 310 ms +- 10 ms: 1.01x faster
Not significant

### regex_dna ###
Mean +- std dev: 317 ms +- 10 ms -> 299 ms +- 9 ms: 1.06x faster
Significant (t=9.99)

### regex_effbot ###
Mean +- std dev: 6.20 ms +- 0.27 ms -> 5.80 ms +- 0.25 ms: 1.07x faster
Significant (t=8.49)

### regex_v8 ###
Mean +- std dev: 43.0 ms +- 1.3 ms -> 39.9 ms +- 1.9 ms: 1.08x faster
Significant (t=10.22)

### richards ###
Mean +- std dev: 158 ms +- 8 ms -> 157 ms +- 8 ms: 1.01x faster
Not significant

### scimark_fft ###
Mean +- std dev: 727 ms +- 18 ms -> 716 ms +- 18 ms: 1.02x faster
Not significant

### scimark_lu ###
Mean +- std dev: 309 ms +- 11 ms -> 304 ms +- 10 ms: 1.01x faster
Not significant

### scimark_monte_carlo ###
Mean +- std dev: 180 ms +- 6 ms -> 181 ms +- 8 ms: 1.00x slower
Not significant

### scimark_sor ###
Mean +- std dev: 355 ms +- 9 ms -> 352 ms +- 11 ms: 1.01x faster
Not significant

### scimark_sparse_mat_mult ###
Mean +- std dev: 9.51 ms +- 0.32 ms -> 9.19 ms +- 0.34 ms: 1.03x faster
Significant (t=5.27)

### spectral_norm ###
Mean +- std dev: 277 ms +- 10 ms -> 272 ms +- 9 ms: 1.02x faster
Not significant

### sqlalchemy_declarative ###
Mean +- std dev: 273 ms +- 10 ms -> 273 ms +- 10 ms: 1.00x faster
Not significant

### sqlalchemy_imperative ###
Mean +- std dev: 46.2 ms +- 2.3 ms -> 45.6 ms +- 2.1 ms: 1.01x faster
Not significant

### sqlite_synth ###
Mean +- std dev: 4.39 us +- 0.29 us -> 4.37 us +- 0.23 us: 1.00x faster
Not significant

### sympy_expand ###
Mean +- std dev: 996 ms +- 22 ms -> 993 ms +- 31 ms: 1.00x faster
Not significant

### sympy_integrate ###
Mean +- std dev: 42.6 ms +- 1.9 ms -> 43.5 ms +- 2.0 ms: 1.02x slower
Significant (t=-2.47)

### sympy_str ###
Mean +- std dev: 607 ms +- 18 ms -> 599 ms +- 15 ms: 1.01x faster
Not significant

### sympy_sum ###
Mean +- std dev: 350 ms +- 12 ms -> 344 ms +- 11 ms: 1.02x faster
Not significant

### telco ###
Mean +- std dev: 11.3 ms +- 0.6 ms -> 11.2 ms +- 0.6 ms: 1.01x faster
Not significant

### tornado_http ###
Mean +- std dev: 294 ms +- 11 ms -> 296 ms +- 12 ms: 1.01x slower
Not significant

### unpack_sequence ###
Mean +- std dev: 78.4 ns +- 7.1 ns -> 75.7 ns +- 2.5 ns: 1.04x faster
Significant (t=2.86)

### unpickle ###
Mean +- std dev: 26.1 us +- 1.1 us -> 27.3 us +- 1.3 us: 1.05x slower
Significant (t=-5.57)

### unpickle_list ###
Mean +- std dev: 6.65 us +- 0.21 us -> 6.68 us +- 0.31 us: 1.00x slower
Not significant

### unpickle_pure_python ###
Mean +- std dev: 567 us +- 21 us -> 572 us +- 21 us: 1.01x slower
Not significant

### xml_etree_generate ###
Mean +- std dev: 165 ms +- 8 ms -> 166 ms +- 8 ms: 1.00x slower
Not significant

### xml_etree_iterparse ###
Mean +- std dev: 187 ms +- 6 ms -> 187 ms +- 8 ms: 1.00x faster
Not significant

### xml_etree_parse ###
Mean +- std dev: 274 ms +- 10 ms -> 274 ms +- 11 ms: 1.00x slower
Not significant

### xml_etree_process ###
Mean +- std dev: 142 ms +- 6 ms -> 139 ms +- 6 ms: 1.02x faster
Not significant

corona10 · 2021-07-18T23:53:02Z

clang version 11.0.0

corona10 · 2021-07-19T10:53:01Z

New changeset b2cf251 by Dong-hee Na in branch 'main':
bpo-44340: Add support for building with clang full/thin lto (GH-27231)
b2cf251

corona10 · 2021-07-19T10:55:08Z

Now CPython 3.11 supports the Thin LTO,
Thank you for the report and contribution, Brett!

And also thank you Pablo and Gregory for the reviews!

corona10 · 2021-07-20T06:51:58Z

@ned.deily

Can we use the thin-lto option for next macOS Python distribution?
In my local environment, it passes all tests :)

cpython/Mac/BuildScript/build-installer.py

Line 1199 in 366fcba

(' ', "--enable-optimizations --with-lto")[compilerCanOptimize()],

FYI, Gentoo already recommends using the thin LTO instead of the full LTO.
https://wiki.gentoo.org/wiki/Clang#Link-time_optimizations_with_Clang

ambv · 2021-09-08T17:29:40Z

New changeset 84ca5fc by Dong-hee Na in branch 'main':
bpo-44340: Update whatsnews for ThinLTO (GH-28229)
84ca5fc

holmanb mannequin added interpreter-core (Objects, Python, Grammar, and Parser dirs) type-feature A feature request or enhancement labels Jun 7, 2021

ned-deily added build The build process and cross-build and removed interpreter-core (Objects, Python, Grammar, and Parser dirs) labels Jun 8, 2021

corona10 added 3.11 only security fixes labels Jul 18, 2021

corona10 closed this as completed Jul 19, 2021

ezio-melotti transferred this issue from another repository Apr 10, 2022

corona10 mentioned this issue Sep 12, 2022

gh-89536: Use ThinLTO policy if possible #96766

Merged

haampie mentioned this issue Jan 15, 2023

python: 3.11.1, use -flto=thin when clang spack/spack#34951

Merged

sevz17 mentioned this issue Nov 14, 2023

media-video/ffmpeg: add 6.1; wire up vulkan filters gentoo/gentoo#33776

Closed

ned-deily mentioned this issue Sep 3, 2024

Python3.13 performance Issue with python.org macOS installers on ARM Macs #122580

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add support for building cpython with clang thin lto #88506

Add support for building cpython with clang thin lto #88506

holmanb mannequin commented Jun 7, 2021

holmanb mannequin commented Jun 7, 2021

corona10 commented Jul 18, 2021

corona10 commented Jul 18, 2021

corona10 commented Jul 18, 2021

corona10 commented Jul 18, 2021

corona10 commented Jul 19, 2021

corona10 commented Jul 19, 2021

corona10 commented Jul 20, 2021

ambv commented Sep 8, 2021

Add support for building cpython with clang thin lto #88506

Add support for building cpython with clang thin lto #88506

Comments

holmanb mannequin commented Jun 7, 2021

holmanb mannequin commented Jun 7, 2021

corona10 commented Jul 18, 2021

corona10 commented Jul 18, 2021

corona10 commented Jul 18, 2021

corona10 commented Jul 18, 2021

corona10 commented Jul 19, 2021

corona10 commented Jul 19, 2021

corona10 commented Jul 20, 2021

ambv commented Sep 8, 2021