simd-everywhere
Several tests fail on FreeBSD with clang-15
Hello @yurivict ; thanks for the report. I've never seen SIMDe tested on FreeBSD before. Did a previous release pass the tests?
I'm seeing some SIGBUS errors, that is because FreeBSD doesn't allow unaligned reads/writes like Linus does, correct? I would accept patches to fix that, of course. And a CI config to prevent regressions.
simde-0.7.2 had fewer failures:
7/874 x86/avx512/abs/emul/c ERROR 0.39s killed by signal 11 SIGSEGV
8/874 x86/avx512/abs/native/c ERROR 0.28s killed by signal 11 SIGSEGV
9/874 x86/avx512/abs/emul/cpp ERROR 0.90s killed by signal 11 SIGSEGV
10/874 x86/avx512/abs/native/cpp ERROR 0.86s killed by signal 11 SIGSEGV
371/874 arm/neon/abs/emul/c ERROR 0.81s killed by signal 11 SIGSEGV
373/874 arm/neon/abs/emul/cpp ERROR 0.59s killed by signal 11 SIGSEGV
So this is a regression.
I'm seeing some SIGBUS errors, that is because FreeBSD doesn't allow unaligned reads/writes like Linus does, correct?
No, FreeBSD allows unaligned reads just like Linux does.
Hmm.. I'm not able to reproduce using a VM: https://github.com/mr-c/simde/actions/runs/4918777831/jobs/8785545316#step:3:7057
Maybe you can correct my configuration to reproduce what you are seeing: https://github.com/mr-c/simde/blob/ci/gha/freebsd/.github/workflows/ci.yml#L36
In a VM please have an updated ports tree:
git clone https://git.FreeBSD.org/ports.git /usr/ports
Then:
- cd /usr/ports/devel/simde
- make test
This should reproduce the problem.
@yurivict Thanks. make test tried to build the ninja package from source, which failed, so I'll try again using the binary package for meson and its dependencies
https://github.com/mr-c/simde/actions/runs/4922970383/jobs/8794277813#step:3:5370
Seems that VM is FreeBSD 13.1 which is using clang-13 which has extra complaints https://github.com/mr-c/simde/actions/runs/4923301187/jobs/8794964385#step:2:1588
Trying again with -Wno-pass-failed: https://github.com/mr-c/simde/actions/runs/4923955421/jobs/8796401951
https://github.com/mr-c/simde/actions/runs/4923955421/jobs/8796401951 has a couple "emul" tests failing, but those can be skipped.
@yurivict can you try changing the test command to be something like meson test -C _build --print-errorlogs $(meson test -C _build --list | grep -v emul) ?
Any other failures appear to be compiler regressions, so I would have to encourage you to file bugs with the clang people, or try -O2 instead.
@yurivict As show in https://github.com/simd-everywhere/simde/actions/runs/5143791700/jobs/9259255771?pr=1034#step:2:5440 the only failing tests are "emul" tests, when we disable the use of native intrinsincs on the target architecture. As that does not represent real-world usage of SIMDe, I recommend that you skip all those tests: meson test -C _build --print-errorlogs $(meson test -C _build --list | grep -v emul)
Please let me know when the port is updated and I'll merge #1034 to add freebsd to our test matrix
@yurivict FYI: We got Travis CI working again, and they support testing on FreeBSD 12.1, but I'm guessing that version of FreeBSD is a too old to be useful for CI?
https://docs.travis-ci.com/user/reference/freebsd/
@yurivict Can you test again using version 0.8.0-rc2 of SIMDe ? https://github.com/simd-everywhere/simde/releases/tag/v0.8.0-rc2
In 0.8.2 tasts fail to link:
FAILED: test/x86/avx512/round-emul-c
cc -o test/x86/avx512/round-emul-c test/x86/avx512/round-emul-c.p/round.c.o -Wl,--as-needed -Wl,--no-undefined -lm
ld: error: undefined symbol: roundevenf
>>> referenced by round.c
>>> test/x86/avx512/round-emul-c.p/round.c.o:(test_simde_x_mm512_round_ps)
>>> referenced by round.c
>>> test/x86/avx512/round-emul-c.p/round.c.o:(test_simde_x_mm512_round_ps)
>>> referenced by round.c
>>> test/x86/avx512/round-emul-c.p/round.c.o:(test_simde_x_mm512_round_ps)
>>> referenced 1 more times
ld: error: undefined symbol: roundeven
>>> referenced by round.c
>>> test/x86/avx512/round-emul-c.p/round.c.o:(test_simde_x_mm512_round_pd)
>>> referenced by round.c
>>> test/x86/avx512/round-emul-c.p/round.c.o:(test_simde_x_mm512_round_pd)
>>> referenced by round.c
>>> test/x86/avx512/round-emul-c.p/round.c.o:(test_simde_x_mm512_round_pd)
>>> referenced 1 more times
FreeBSD 14.0
I am currently experiencing the same problem while building with Clang-Cl on Windows (Clang 17.0.3 with MSVC-like command-line) while linking some code using the simde_mm_floor_ps and simde_mm_floor_pd:
windows-clang-cl-x64-debug\lld-link : error : undefined symbol: roundeven
and
windows-clang-cl-x64-debug\lld-link : error : undefined symbol: roundevenf
Weirdly enough, it links fine in release, so it probably has something to do with O2? Compiler bug perhaps?
It ends up using __builtin_roundeven and __builtin_roundevenf, so I suspect that this guard gives a false-positive for clang:
https://github.com/simd-everywhere/simde/blob/17416b157fcd88bb498660e3e2f32ce822f436dd/simde/simde-math.h#L1268-L1271
If I force it to use simde_math_round and simde_math_roundf fallbacks it links fine.
The problem seems to persist even in the old SIMDe version I'm using (5e7c4d4ec10855ebe8f19af8a80fdfa26caad5e9) https://github.com/simd-everywhere/simde/blob/5e7c4d4ec10855ebe8f19af8a80fdfa26caad5e9/simde/simde-math.h#L1268-L1271
@yurivict and @Epixu Huh, I know that emscripten had some issues with clang related to those functions. Maybe https://github.com/simd-everywhere/simde/issues/1001#issuecomment-1498050395 (use of musl instead of the gnu c lib) is applicable here?