This appears to be related to valgrind 3.21.0 rather than Perl 5.38. I did a mock build, keeping the chroot, then "mock ... --remove valgrind" and then "mock ... --install valgrind-3.20.0-3.fc38.x86_64.rpm" and the tests passed with a "mock ... shell" and "make test":

[mockbuild@7e50b388172e4f27b426ac5f2ac2d88a Test-Valgrind-1.19]$ make test
"/usr/bin/perl" -MExtUtils::Command::MM -e 'cp_nonempty' -- Valgrind.bs blib/arch/auto/Test/Valgrind/Valgrind.bs 644
PERL_DL_NONLAZY=1 "/usr/bin/perl" "-MExtUtils::Command::MM" "-MTest::Harness" "-e" "undef *Test::Harness::Switches; test_harness(0, 'blib/lib', 'blib/arch')" t/*.t
t/00-load.t ................... 1/1 # Testing Test::Valgrind 1.19, Perl 5.038000, /usr/bin/perl
t/00-load.t ................... ok   
t/10-good.t ................... # Using valgrind 3.20.0 located at /usr/bin/valgrind
# Generating suppressions...
# Suppressions for this perl stored in /builddir/.perl/Test-Valgrind/suppressions/1.19/memcheck-3.20.0-d58730dcab9961e1c0618f9f9a87aee3.supp
# Using suppressions from /builddir/.perl/Test-Valgrind/suppressions/1.19/memcheck-3.20.0-d58730dcab9961e1c0618f9f9a87aee3.supp
# 1..1
# not ok 1 - dummy test in the child, should not interfere with the actual TAP stream
# 
# #   Failed test 'dummy test in the child, should not interfere with the actual TAP stream'
# #   at t/10-good.t line 30.
# # Looks like you failed 1 test of 1.
t/10-good.t ................... ok    
t/12-good-run-exception.t ..... # Using valgrind 3.20.0 located at /usr/bin/valgrind
# Using suppressions from /builddir/.perl/Test-Valgrind/suppressions/1.19/memcheck-3.20.0-d58730dcab9961e1c0618f9f9a87aee3.supp
# dummy run-time exception, should not cause the test to fail at t/12-good-run-exception.t line 20.
t/12-good-run-exception.t ..... ok     
t/20-bad.t .................... # Using valgrind 3.20.0 located at /usr/bin/valgrind
# Using suppressions from /builddir/.perl/Test-Valgrind/suppressions/1.19/memcheck-3.20.0-d58730dcab9961e1c0618f9f9a87aee3.supp
# # leaking some bytes!
t/20-bad.t .................... 1/18 # The subsequent report was correctly caught:
#   10,000 bytes in 1 blocks are still reachable in loss record 18 of 18
#     malloc (/usr/libexec/valgrind/vgpreload_memcheck-amd64-linux.so) [vg_replace_malloc.c:393]
#     tv_leak (/builddir/build/BUILD/Test-Valgrind-1.19/blib/arch/auto/Test/Valgrind/Valgrind.so) [Valgrind.xs:34]
#     XS_Test__Valgrind_leak (/builddir/build/BUILD/Test-Valgrind-1.19/blib/arch/auto/Test/Valgrind/Valgrind.so) [Valgrind.xs:54]
#     Perl_pp_entersub (/usr/lib64/libperl.so.5.38.0) [pp_hot.c:5555]
#     Perl_runops_standard (/usr/lib64/libperl.so.5.38.0) [run.c:41]
#     UnknownInlinedFun (/usr/lib64/libperl.so.5.38.0) [perl.c:2812]
#     perl_run (/usr/lib64/libperl.so.5.38.0) [perl.c:2727]
#     main (/usr/bin/perl) [perlmain.c:127]
t/20-bad.t .................... ok     
t/30-skip.t ................... # Using IO::Handle 1.52
# Using IO::Select 1.52
# Using IPC::Open3 1.22
t/30-skip.t ................... ok   
t/60-version.t ................ ok       
t/70-session.t ................ ok   
t/71-session-command.t ........ ok   
t/80-suppressions.t ........... 1/4 # The default suppression file contains 12 suppressions, of which 12 are not empty and 9 apply to perl
t/80-suppressions.t ........... ok   
t/81-suppressions-demangle.t .. ok     
All tests successful.
Files=10, Tests=200, 32 wallclock secs ( 0.03 usr  0.00 sys + 17.83 cusr  0.98 csys = 18.84 CPU)
Result: PASS
[mockbuild@7e50b388172e4f27b426ac5f2ac2d88a Test-Valgrind-1.19]$ 

Reassign to valgrind?

Well, Fedora 38 also has valgrind 3.21.0 and that still builds OK, so it looks like it's the combination of Perl 5.38 and valgrind 3.21.0 that's the problem. Or maybe even something else, like the compiler in Rawhide.

OK, the test is failing because the trace report from valgrind does not include the symbol 'Perl_pp_entersub' (see the Fedora 38 trace for comparison). Not sure why that is, or if it would be there if debuginfo was installed, so I don't know if this is an upstream issue or not.

In any case, I'm able to work around it by having the test look for the 'XS_Test__Valgrind_leak' and 'Perl_runops_standard' symbols surrounding where 'Perl_pp_entersub' was expected.

FEDORA-2023-6f45694481 has been submitted as an update to Fedora 39. https://bodhi.fedoraproject.org/updates/FEDORA-2023-6f45694481

FEDORA-2023-6f45694481 has been pushed to the Fedora 39 stable repository.
If problem still persists, please make note of it in this bug report.

It might be related to this <https://github.com/Perl/perl5/commit/214a9432e04871da2b88eba02ea3f1148a99f9c8> "call_sv, amagic_call: call pp_entersub via runops" Perl commit.

(In reply to Petr Pisar from comment #6)
> It might be related to this <https://github.com/Perl/perl5/commit/214a9432e04871da2b88eba02ea3f1148a99f9c8> "call_sv, amagic_call: call pp_entersub via runops" Perl commit.

Perhaps. That would then be an issue for upstream. The CPAN Testers matrix isn't helping since the testers don't generally have valgrind installed so the failing test just gets skipped.

I installed perl-debuginfo and perl-libs-debuginfo then re-ran the tests and the Perl_pp_entersub symbol showed up again:

#   10,000 bytes in 1 blocks are still reachable in loss record 18 of 18
#     malloc (/usr/libexec/valgrind/vgpreload_memcheck-amd64-linux.so) [vg_replace_malloc.c:431]
#     tv_leak (/builddir/build/BUILD/Test-Valgrind-1.19/blib/arch/auto/Test/Valgrind/Valgrind.so) [Valgrind.xs:34]
#     XS_Test__Valgrind_leak (/builddir/build/BUILD/Test-Valgrind-1.19/blib/arch/auto/Test/Valgrind/Valgrind.so) [Valgrind.xs:54]
#     Perl_pp_entersub (/usr/lib64/libperl.so.5.38.0) [pp_hot.c:5555]
#     Perl_runops_standard (/usr/lib64/libperl.so.5.38.0) [run.c:41]
#     UnknownInlinedFun (/usr/lib64/libperl.so.5.38.0) [perl.c:2812]
#     perl_run (/usr/lib64/libperl.so.5.38.0) [perl.c:2727]
#     ? (/usr/bin/perl) [?:?]
#     (below main) (/usr/lib64/libc.so.6) [?:?]

So that looks like a downstream issue rather than an upstream one.

Do you say debuginfo? Then it's because the symbol is not exported anymore:

$ nm -D /usr/lib64/libperl.so.5.36 | grep Perl_pp_entersub
000000000011f0f0 T Perl_pp_entersub

That was an intentional change in the upstream <https://github.com/Perl/perl5/commit/0351a629e71de127cbfd1b142e9eaa6069deabf5> "hide private functions with __attribute__((visibility("hidden"))
)". Valgrind translates call addresses into function names based on a symbol table in an executable, and if not found, then from a symbol table in debuginfo data.