enh: reduce binary size across all libraries #5292 by matejk · Pull Request #5297 · pocoproject/poco

matejk · 2026-04-05T20:08:54Z

Summary

Move = default constructors/destructors from headers to .cpp files for 42 non-template classes across Util, Prometheus, ActiveRecord, MongoDB, Net, and Redis
Add extern template declarations for BasicEvent types used by AbstractConfiguration
Add -Wl,-dead_strip on macOS and visibility-aware -ffunction-sections/gc-sections
Document binary size reduction options in README

Binary Size (stripped, default options, RelWithDebInfo)

macOS arm64 (Apple Clang)

Library	main	fix	diff	%
Util	647,104	620,656	-26,448	-4.1%
Foundation	3,372,384	3,347,584	-24,800	-0.7%
DataODBC	1,269,088	1,244,496	-24,592	-1.9%
Data	2,821,152	2,788,992	-32,160	-1.1%
Net	1,431,824	1,414,000	-17,824	-1.2%
JSON	467,248	460,480	-6,768	-1.4%
MongoDB	433,520	426,976	-6,544	-1.5%
TOTAL (17 libs)	15,227,944	15,077,608	-150,336	-1.0%

Linux x86_64 (GCC)

Library	main	fix	diff	%
Util	793,752	728,464	-65,288	-8.2%
TOTAL (17 libs)	17,395,512	17,330,504	-65,008	-0.4%

On Linux with default visibility, GCC already deduplicates weak template symbols effectively. The Util savings (-65 KB) come from the BasicEvent extern template declarations eliminating duplicate instantiations across 11 Configuration subclasses.

Closes #5292

Test plan

matejk · 2026-04-05T20:09:05Z

Binary Size Comparison: 1.14.2 vs main vs fix branch

Stripped library sizes in bytes. "fix" = code changes only (default visibility).
"fix+hidden" = code changes + -DCMAKE_CXX_VISIBILITY_PRESET=hidden -DCMAKE_VISIBILITY_INLINES_HIDDEN=ON.

macOS arm64 (Apple Clang, RelWithDebInfo)

Library	1.14.2	main	fix	fix+hidden	fix-main	hidden-main
ActiveRecord	50,800	50,608	50,752	51,248	+144	+640
Crypto	361,176	361,992	358,504	356,904	-3,488	-5,088
Data	3,102,680	3,151,848	3,119,592	3,183,640	-32,256	+31,792
DataODBC	1,435,384	1,428,872	1,404,168	1,441,368	-24,704	+12,496
DataSQLite	1,823,336	1,838,776	1,837,416	1,847,224	-1,360	+8,448
Encodings	945,208	945,080	945,016	945,240	-64	+160
Foundation	3,233,544	3,850,760	3,825,368	3,741,672	-25,392	-109,088
JSON	525,720	524,568	517,784	524,072	-6,784	-496
JWT	277,112	258,216	256,440	260,296	-1,776	+2,080
MongoDB	298,392	487,848	483,656	463,096	-4,192	-24,752
Net	1,589,144	1,692,696	1,674,968	1,666,696	-17,728	-26,000
NetSSL	410,376	412,120	410,024	409,976	-2,096	-2,144
Prometheus	198,008	196,440	201,432	202,040	+4,992	+5,600
Redis	175,864	202,248	201,944	200,360	-304	-1,888
Util	624,008	722,424	699,096	695,896	-23,328	-26,528
XML	762,856	776,280	774,456	774,952	-1,824	-1,328
Zip	304,152	304,856	302,552	302,136	-2,304	-2,720
TOTAL	16,117,760	17,205,632	17,063,168	17,066,816	-142,464	-138,816

Configuration	vs main	%
fix (default visibility)	-139 KB	-0.8%
fix + visibility hidden	-136 KB	-0.8%

Linux x86_64 (GCC, RelWithDebInfo)

Library	1.14.2	main	fix	fix+hidden	fix-main	hidden-main
ActiveRecord	67,888	67,904	67,912	67,856	+8	-48
Crypto	333,504	333,520	333,520	333,152	+0	-368
Data	3,230,232	3,296,176	3,296,176	2,631,448	+0	-664,728
DataODBC	1,517,216	1,517,920	1,517,920	1,318,064	+0	-199,856
DataSQLite	1,724,328	1,724,560	1,724,560	1,657,680	+0	-66,880
Encodings	921,320	921,320	921,320	921,312	+0	-8
Foundation	3,099,048	3,759,304	3,759,304	3,492,192	+0	-267,112
JSON	529,680	529,600	529,600	463,008	+0	-66,592
JWT	266,608	267,368	267,368	266,664	+0	-704
MongoDB	332,368	531,872	531,952	465,448	+80	-66,424
Net	1,523,176	1,724,312	1,724,312	1,657,448	+0	-66,864
NetSSL	400,224	400,704	400,704	400,408	+0	-296
Prometheus	199,896	200,328	200,504	200,184	+176	-144
Redis	200,048	200,240	200,256	200,024	+16	-216
Util	662,560	793,752	728,464	662,056	-65,288	-131,696
XML	727,824	794,240	794,240	794,040	+0	-200
Zip	332,352	332,392	332,392	332,056	+0	-336
TOTAL	16,068,272	17,395,512	17,330,504	15,863,040	-65,008	-1,532,472

Configuration	vs main	%
fix (default visibility)	-63 KB	-0.4%
fix + visibility hidden	-1,497 KB	-8.8%

Note: on Linux with default visibility, GCC already deduplicates weak template symbols
effectively, so the code-level improvements yield modest savings. The visibility-hidden
option produces dramatic savings because gc-sections can then discard all
unreferenced hidden symbols from shared libraries.

Changes Applied

Moved = default out-of-line: 42 non-template classes across Util (22),
Prometheus (14), ActiveRecord (1), MongoDB (3), Net (1), Redis (2).
BasicEvent explicit instantiations (Util): extern template for the three
BasicEvent types used by AbstractConfiguration.
CMake dead-code stripping: macOS always gets -Wl,-dead_strip. When
CMAKE_CXX_VISIBILITY_PRESET=hidden is set, -ffunction-sections -fdata-sections
and linker gc are enabled automatically.
README: documented binary size reduction options for users.

aleks-f · 2026-04-05T20:16:02Z

I would like to see some benchmarks on intToStr. That function was manually written for a reason (used often in Var) to be faster than standard implementations

Move = default constructors/destructors from headers to .cpp files for 42 non-template classes across Util (22), Prometheus (14), ActiveRecord (1), MongoDB (3), Net (1), and Redis (2) to eliminate template bloat from implicit instantiations in every translation unit. Add extern template declarations for BasicEvent types used by AbstractConfiguration to prevent duplicate instantiations across Util's 11 Configuration subclasses.

aleks-f

intToStr seems to be mostly pessimization. benchmark code here

Aggregate (58 cases across 4 types)

Metric	Value
Total OLD	1933.22 ms
Total NEW	2442.94 ms
Delta	+509.72 ms (+26.4%)
NEW faster	8 cases
OLD faster	49 cases
~Same	1 case
Best NEW gain	-53.2% (uint64/int64_max)
Worst NEW loss	+168.0% (int/zero)

Per-type breakdown

Type	OLD	NEW	Change
int	480.09 ms	687.61 ms	+43.2%
uint	413.37 ms	525.80 ms	+27.2%
int64	575.95 ms	689.48 ms	+19.7%
uint64	463.80 ms	540.05 ms	+16.4%

aleks-f · 2026-04-06T11:19:23Z

Here is an optimized version of the old intToStr, with comparison to old and new (std::to_chars) versions

Type	OLD	NEW	OPT	OPT/OLD	OPT/NEW
int	490.30	692.71	419.53	-14.4%	-39.4%
uint	404.68	521.96	342.54	-15.4%	-34.4%
int64	545.73	699.65	466.30	-14.6%	-33.4%
uint64	470.13	543.89	369.13	-21.5%	-32.1%

matejk · 2026-04-06T17:41:27Z

I will remove numeric parser/formatter changes from this PR. New PR is open for that: #5301.

macOS: add -Wl,-dead_strip for non-Debug builds (Apple linker handles it natively on Mach-O atoms). All platforms: when CMAKE_CXX_VISIBILITY_PRESET=hidden is set, enable -ffunction-sections -fdata-sections and linker gc for non-Debug builds. Use generator expressions for proper multi-config generator support. ICF (--icf=safe) is enabled on non-Apple platforms when gold or lld is detected.

…n of copy constructor and assignment ('Rule of Two') Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com> Potential fix for code scanning alert no. 118: Inconsistent definition of copy constructor and assignment ('Rule of Two') Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>

* enh: move defaulted constructors/destructors out-of-line #5292 Move = default constructors/destructors from headers to .cpp files for 42 non-template classes across Util (22), Prometheus (14), ActiveRecord (1), MongoDB (3), Net (1), and Redis (2) to eliminate template bloat from implicit instantiations in every translation unit. Add extern template declarations for BasicEvent types used by AbstractConfiguration to prevent duplicate instantiations across Util's 11 Configuration subclasses. * enh(cmake): add dead-code stripping and visibility-aware gc-sections macOS: add -Wl,-dead_strip for non-Debug builds (Apple linker handles it natively on Mach-O atoms). All platforms: when CMAKE_CXX_VISIBILITY_PRESET=hidden is set, enable -ffunction-sections -fdata-sections and linker gc for non-Debug builds. Use generator expressions for proper multi-config generator support. ICF (--icf=safe) is enabled on non-Apple platforms when gold or lld is detected. * Potential fix for code scanning alert no. 118: Inconsistent definition of copy constructor and assignment ('Rule of Two') Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com> Potential fix for code scanning alert no. 118: Inconsistent definition of copy constructor and assignment ('Rule of Two') Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com> --------- Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>

…lution The cherry-pick of PRs #5297 and #5276 resolved conflicts by taking the incoming single-brace namespace close (} // namespace Poco::X), but this branch still uses old-style dual namespace opens (namespace Poco { namespace X {), requiring two closing braces.

matejk added this to the Release 1.15.2 milestone Apr 5, 2026

matejk added the enhancement label Apr 5, 2026

github-advanced-security AI found potential problems Apr 5, 2026

View reviewed changes

Comment thread Net/src/HTTPAuthenticationParams.cpp Fixed

aleks-f requested changes Apr 6, 2026

View reviewed changes

matejk mentioned this pull request Apr 6, 2026

enh(Foundation): optimize numeric conversions with digit-pairs algorithm and std::to_chars/from_chars #5301

Open

4 tasks

matejk force-pushed the 5292-reduce-binary-size branch from b2c8f25 to 1bb1c47 Compare April 6, 2026 18:47

matejk force-pushed the 5292-reduce-binary-size branch from ab2087a to 04fb9a7 Compare April 7, 2026 08:19

matejk merged commit 1aa0624 into main Apr 7, 2026
106 checks passed

matejk deleted the 5292-reduce-binary-size branch April 7, 2026 17:07

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

enh: reduce binary size across all libraries #5292#5297

enh: reduce binary size across all libraries #5292#5297
matejk merged 3 commits intomainfrom
5292-reduce-binary-size

matejk commented Apr 5, 2026 •

edited

Loading

Uh oh!

matejk commented Apr 5, 2026 •

edited

Loading

Uh oh!

aleks-f commented Apr 5, 2026

Uh oh!

Uh oh!

aleks-f left a comment

Uh oh!

aleks-f commented Apr 6, 2026

Uh oh!

matejk commented Apr 6, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

matejk commented Apr 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Binary Size (stripped, default options, RelWithDebInfo)

macOS arm64 (Apple Clang)

Linux x86_64 (GCC)

Test plan

Uh oh!

matejk commented Apr 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Binary Size Comparison: 1.14.2 vs main vs fix branch

macOS arm64 (Apple Clang, RelWithDebInfo)

Linux x86_64 (GCC, RelWithDebInfo)

Changes Applied

Uh oh!

aleks-f commented Apr 5, 2026

Uh oh!

Uh oh!

aleks-f left a comment

Choose a reason for hiding this comment

Aggregate (58 cases across 4 types)

Per-type breakdown

Uh oh!

aleks-f commented Apr 6, 2026

Uh oh!

matejk commented Apr 6, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

matejk commented Apr 5, 2026 •

edited

Loading

matejk commented Apr 5, 2026 •

edited

Loading