optimization: -O2 vs -Os

[sisyphus]

Following on from some ideas that came up in PDLPorters/pdl#522, I decided to try building the devel release of 5.41.8 (released overnight) with -O2 optimization.
It was built using gcc version 14.2.0 (MinGW-W64 x86_64-ucrt-mcf-seh, built by Brecht Sanders, r1, but a number of tests failed.

So I reverted to optimizing with -Os -falign-functions -falign-jumps -falign-labels -falign-loops -freorder-blocks -freorder-blocks-algorithm=stc -freorder-blocks-and-partition - same as SP has been doing, and same as I normally do each month when a new devel release is made available..
And that worked fine, of course.

I did three 64-bit builds - with nvtypes of double, long double and quadmath, respectively.
In each of those 3 installations I then reset $Config{OPTIMIZE} to -O2 (by hacking lib/Config_heavy.pl) before building any additional modules.
With optimization at -O2, I then installed some perl modules without any issues - apart from some minor issues that I also strike when I build with the usual "-Os ..." optimization:
Among those that I installed were PDL, Image::Magick, Inline::C, Math::LongDouble, Math::Float128, Math::Decimal64, Math::Decimal128, Math::GMP, Math::GMPf, Math::GMPq, "Math::GMPz, Math::MPFR, Math::Random::BlumBlumShub, Math::Random::MicaliSchnorr, Math-FakeDD, Win32::IPC, Win32::GenRandom, Math::MPFI, and a variety of prerequisite modules that are also pulled in.

My hunch is that the "-Os ...." optimization is needed only for the building of perl's guts, and that all of the Core modules could then be built with simply -O2 optimization.
I'd like to test that out, but I don't know how to orchestrate such a build. Any ideas on that ?
(I suppose I could test out the dual-life modules by re-installing them using "cpan -fi ...", but I don't think the ones in "ext" can readily be rebuilt outside of the building of perl itself.)

For my own builds, I intend to keep doing that post-build hack of re-setting $Config{optimize} to "-O2" until such time as it becomes evident that it causes a problem.

It would be a whole lot simpler if perl could just be built flawlessly with "-O2", as was once the case.

[shawnlaffan]

Thanks Rob - all good info.

What happens if you compile perl with -O2 -fno-inline-functions? For gcc-13, inline-functions is the only one of the -O2 optimisations not used for the SP 5.40 build.

Also, SP is using POSIX threads instead of MCF. Does that make a difference?

[sisyphus]

What happens if you compile perl with -O2 -fno-inline-functions?

I tried that without success a couple of nights back - but I think the source I used was corrupted.
Tonight, with perl-5.41.8 source. gmake runs to completion, though I did notice (for the first time) some compilation warnings that are also produced when building with -Os -falign-functions -falign-jumps -falign-labels -falign-loops -freorder-blocks -freorder-blocks-algorithm=stc -freorder-blocks-and-partition:

..\..\lib\CORE/inline.h: In function 'Perl_utf8_hop_forward_overshoot':
..\..\lib\CORE/inline.h:2707:34: warning: unknown conversion type character 'z' in format [-Wformat=]
 2707 |             Perl_croak_nocontext("panic: Start of forward hop (0x%p) is %zd"
      |                                  ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
..\..\lib\CORE/inline.h:2707:34: warning: format '%p' expects argument of type 'void *', but argument 3 has type 'long long int' [-Wformat=]
 2707 |             Perl_croak_nocontext("panic: Start of forward hop (0x%p) is %zd"
      |                                  ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 2708 |                                  " bytes beyond legal end position (0x%p)",
 2709 |                                  s, 1 + s - end, end);
      |                                     ~~~~~~~~~~~
      |                                           |
      |                                           long long int
..\..\lib\CORE/inline.h:2707:34: warning: too many arguments for format [-Wformat-extra-args]

I'll look more at that tomorrow, and file a perl issue if I have trouble unravelling it.
The test suite is the killer, but:

Test Summary Report
-------------------
porting/copyright.t                                                (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
  Parse errors: No plan found in TAP output
run/switches.t                                                     (Wstat: 0 Tests: 140 Failed: 1)
  Failed test:  35
run/todo.t                                                         (Wstat: 0 Tests: 6 Failed: 0)
  TODO passed:   3-4
../cpan/bignum/t/backend-gmp-bignum.t                              (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
../cpan/Encode/t/enc_eucjp.t                                       (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
  Parse errors: No plan found in TAP output
../cpan/Encode/t/enc_utf8.t                                        (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
  Parse errors: No plan found in TAP output
../cpan/Encode/t/encoding.t                                        (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
  Parse errors: No plan found in TAP output
../cpan/Encode/t/piconv.t                                          (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
  Parse errors: No plan found in TAP output
../cpan/ExtUtils-MakeMaker/t/MM_BeOS.t                             (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
../cpan/ExtUtils-MakeMaker/t/MM_VMS.t                              (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
../dist/threads/t/no_threads.t                                     (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
  Parse errors: No plan found in TAP output
../ext/IPC-Open3/t/IPC-Open3.t                                     (Wstat: 0 Tests: 53 Failed: 0)
  TODO passed:   33
../lib/User/pwent.t                                                (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
  Parse errors: No plan found in TAP output
../lib/warnings.t                                                  (Wstat: 0 Tests: 955 Failed: 3)
  Failed tests:  401, 877, 944
Files=2907, Tests=1201800, 1481 wallclock secs (44.61 usr +  5.12 sys = 49.73 CPU)
Result: FAIL

Also, SP is using POSIX threads instead of MCF. Does that make a difference?

I think not - but should be checked.

BTW (and unrelated), Shawn, are you aware that perls 5.38.3 and 5.40.1 were released a couple of days ago ?

[shawnlaffan]

My hunch is that the "-Os ...." optimization is needed only for the building of perl's guts, and that all of the Core modules could then be built with simply -O2 optimization.
I'd like to test that out, but I don't know how to orchestrate such a build. Any ideas on that ?

If it comes to it, it should be possible to add a P::D::S step to modify lib/Config_heavy.pl. Would be good to discover what part of the code is tickling the problem, though.

BTW (and unrelated), Shawn, are you aware that perls 5.38.3 and 5.40.1 were released a couple of days ago ?

Yes, I started on it yesterday. Progress is being tracked under #231

[shawnlaffan]

Likely useful data point: #85 (comment)

[shawnlaffan]

Perl 5.40.1 built with -O2 -fno-inline-functions. GCC is 13.2, same as comes with SP 5.40.

Test Summary Report
-------------------
comp/final_line_num.t                                              (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
  Parse errors: No plan found in TAP output
op/refstack.t                                                      (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
op/setpgrpstack.t                                                  (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
porting/checkcfgvar.t                                              (Wstat: 0 Tests: 14 Failed: 1)
  Failed test:  12
win32/popen.t                                                      (Wstat: 2304 (exited 9) Tests: 0 Failed: 0)
  Non-zero exit status: 9
  Parse errors: No plan found in TAP output
win32/stat.t                                                       (Wstat: 0 Tests: 50 Failed: 0)
  TODO passed:   39
../cpan/autodie/t/import-into.t                                    (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
../cpan/IPC-Cmd/t/01_IPC-Cmd.t                                     (Wstat: 1536 (exited 6) Tests: 459 Failed: 6)
  Failed tests:  129-130, 241-242, 351-352
  Non-zero exit status: 6
../cpan/IPC-Cmd/t/02_Interactive.t                                 (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
../cpan/Test2-Suite/t/regression/247_check_ref_bool.t              (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
../dist/encoding-warnings/t/3-normal.t                             (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
  Parse errors: No plan found in TAP output
../dist/Net-Ping/t/250_ping_hires.t                                (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
  Parse errors: No plan found in TAP output
../ext/IPC-Open3/t/IPC-Open3.t                                     (Wstat: 0 Tests: 45 Failed: 0)
  TODO passed:   25
../lib/Net/servent.t                                               (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
  Parse errors: No plan found in TAP output
../lib/User/grent.t                                                (Wstat: 1280 (exited 5) Tests: 0 Failed: 0)
  Non-zero exit status: 5
../lib/warnings.t                                                  (Wstat: 0 Tests: 952 Failed: 1)
  Failed test:  713
Files=2898, Tests=1195931, 1495 wallclock secs (14.88 usr +  4.73 sys = 19.61 CPU)
Result: FAIL
Finished test run at Wed Jan 22 12:27:00 2025.
gmake: *** [GNUmakefile:1833: test] Error 1

[sisyphus]

Would be good to discover what part of the code is tickling the problem, though.

I've managed to build perl-5.41.8 using gcc-14.2 such that all of the shipped modules are built with -O2 optimization, while still building the rest of perl with -Os -falign-functions -falign-jumps -falign-labels -falign-loops -freorder-blocks -freorder-blocks-algorithm=stc -freorder-blocks-and-partition optimization.
It required a very simple hack to make_ext.pl, located in the top level perl source directory.
It was just a matter of changing line 513 of that file from:

my @args = ("-I$lib_dir", 'Makefile.PL');
to:
my @args = ("-I$lib_dir", 'Makefile.PL "OPTIMIZE=-O2"');

You should find it pretty quickly if you search for my @args.
All tests passed.

I guess that narrows it down a bit .... though only to the point that it proves what the smart punters will reckon they already knew ;-)

I'll give the same a go with SP's gcc-13.2.0, and provide the result as an update.
Why does it have to take so long to run the test suite :-(

UPDATE - The same procedure worked splendidly with SP's gcc-13.2.0 compiler and perl-5.41.8 source.
I then installed this build, reset optimize to -O2 in Config_heavy.pl, and then installed the modules (listed in my opening post) into its perl/site/lib folder.

[shawnlaffan]

I've added some proof-of-concept code in cc5cf66

It only changes Config_heavy.pl as that way there are fewer changes. I could also change make_ext.pl as described but these modules build happily with the -Os ... flags.

As noted in the commit message, it should either be its own step or an optional argument, but for now it works.

Would be good to get more opinions on this process given the degree of hackery involved.

[shawnlaffan]

ExtUtils::Config fails tests if there are comments in the config data, as done by cc5cf66. I'll switch to creating a backup file instead.

[sisyphus]

Yeah, I didn't change Config_heavy.pl until after it had been installed, as I didn't see any point in doing it any earlier than that. I don't think there are any compilations done during gmake test - that is, I think all compilations are done during the gmake step ... but I could be wrong about that.

I do like the idea of patching make_ext.pl prior to the running of gmake. Admittedly, -O2 might not make any improvements over the current -Os ... selection, but a simple -O2 is a whole lot cleaner, a lot easier to grasp, should have no detrimental effect, and just maybe does provide a runtime improvement in some places.

[shawnlaffan]

I do like the idea of patching make_ext.pl prior to the running of gmake.

I'll have a look at patching that also.

Update: But not in time for the first dev release.

[shawnlaffan]

A dev release is at https://github.com/StrawberryPerl/Perl-Dist-Strawberry/releases/tag/dev_5401_20250124_gcc13

Core and dual-life modules are compiled using the same optimisations as perl itself, but anything installed or updated from CPAN uses -O2. The latter are under the vendor dir if they are dual-life.

[sisyphus]

Cool ... and good to know that none of those vendor modules choked on the "-O2".
Better still, perhaps, if someone can locate a module that does choke on the "-O2", as that might help us learn a bit more about the issue.