CHECK_UNARY test fails on ppc64el with gcc 15.2.0

[drew-parsons]

xsimd tests (CHECK_UNARY) are now failing with release 14.2.0 on ppc64el architecture with gcc 15.2.0.
https://ci.debian.net/data/autopkgtest/unstable/ppc64el/x/xsimd/71330067/log.gz
https://ci.debian.net/data/autopkgtest/testing/ppc64el/x/xsimd/71345568/log.gz

The gcc tests were passing with 14.1.0,
https://ci.debian.net/data/autopkgtest/unstable/ppc64el/x/xsimd/70780399/log.gz
https://ci.debian.net/data/autopkgtest/testing/ppc64el/x/xsimd/71221930/log.gz

Ironically ppc64el previously failed tests with clang, Issue #999 , but that is now fixed in 14.2.0 by PR #1010 .
So 14.2.0 now passes on ppc64el with clang 21.1.8, while failing with gcc 15.2.0.

The failure message from the gcc test is

551s 1: /tmp/autopkgtest-lxc.z0hh1a_o/downtmp/build.21l/src/test/test_load_store.cpp:594:
551s 1: TEST CASE:  [load store]<xsimd::batch<short int>>
551s 1:   store
551s 1: 
551s 1: /tmp/autopkgtest-lxc.z0hh1a_o/downtmp/build.21l/src/test/test_load_store.cpp:422: ERROR: CHECK_UNARY( vectors_are_near ) is NOT correct!
551s 1:   values: CHECK_UNARY( false )
551s 1:   logged: load int16_t unaligned
551s 1:           load int16_t aligned
551s 1:           load int16_t unaligned (store_as)
551s 1:           load int16_t aligned (store_as)
551s 1:           load int16_t stream (batch::store)
551s 1:           load int16_t stream (store_as)
551s 1:           load int16_t masked none aligned store
551s 1:           load int16_t masked none unaligned store
551s 1:           load int16_t masked last half aligned
551s 1:           load int16_t masked last half unaligned
551s 1:           res:{?}
551s 1:           expected_masked:{?}
551s 1: 
551s 1: ===============================================================================
551s 1: [doctest] test cases:  367 |  366 passed | 1 failed | 0 skipped
551s 1: [doctest] assertions: 5660 | 5659 passed | 1 failed |
551s 1: [doctest] Status: FAILURE!
551s 1/1 Test #1: test_xsimd .......................***Failed    0.30 sec

The ppc64le machine particulars are

 33s Linux ci-140-7f005e25 6.12.88+deb13-powerpc64le #1 SMP Debian 6.12.88-1 (2026-05-15) ppc64le GNU/Linux
 33s Architecture:                            ppc64le
 33s Byte Order:                              Little Endian
 33s CPU(s):                                  16
 33s On-line CPU(s) list:                     0-15
 33s Model name:                              POWER10 (architected), altivec supported
 33s Model:                                   2.0 (pvr 0080 0200)
 33s Thread(s) per core:                      1
 33s Core(s) per socket:                      1
 33s Socket(s):                               16
 33s Hypervisor vendor:                       KVM
 33s Virtualization type:                     para
 33s L1d cache:                               512 KiB (16 instances)
 33s L1i cache:                               768 KiB (16 instances)
 33s NUMA node(s):                            1
 33s NUMA node0 CPU(s):                       0-15
 33s Vulnerability Gather data sampling:      Not affected
 33s Vulnerability Indirect target selection: Not affected
 33s Vulnerability Itlb multihit:             Not affected
 33s Vulnerability L1tf:                      Mitigation; RFI Flush, L1D private per thread
 33s Vulnerability Mds:                       Not affected
 33s Vulnerability Meltdown:                  Mitigation; RFI Flush, L1D private per thread
 33s Vulnerability Mmio stale data:           Not affected
 33s Vulnerability Reg file data sampling:    Not affected
 33s Vulnerability Retbleed:                  Not affected
 33s Vulnerability Spec rstack overflow:      Not affected
 33s Vulnerability Spec store bypass:         Mitigation; Kernel entry/exit barrier (eieio)
 33s Vulnerability Spectre v1:                Mitigation; __user pointer sanitization, ori31 speculation barrier enabled
 33s Vulnerability Spectre v2:                Mitigation; Software count cache flush (hardware accelerated), Software link stack flush
 33s Vulnerability Srbds:                     Not affected
 33s Vulnerability Tsa:                       Not affected
 33s Vulnerability Tsx async abort:           Not affected
 33s Vulnerability Vmscape:                   Not affected

[drew-parsons]

I should clarify, the clang tests were fixed more by PR #1337 than by PR 1010.
Could PR 1337 be interfering with the gcc build?

[serge-sans-paille]

@Andreas-Krebbel : could you have a look at this one?

[Andreas-Krebbel]

Seems to work on Power 9 for me. I guess I have to find a Power 10 box then. Also the number of tests does not seem to match.

CXX=${HOME}/install/bin/g++-15.2.0 cmake .. -DBUILD_TESTS=ON -DDOWNLOAD_DOCTEST=ON -DBUILD_BENCHMARK=OFF -DBUILD_EXAMPLES=OFF -DCMAKE_BUILD_TYPE=Release && make -j16 xtest VERBOSE=1

[doctest] test cases:  367 |  367 passed | 0 failed | 0 skipped
[doctest] assertions: 5649 | 5649 passed | 0 failed |
[doctest] Status: SUCCESS!

[Andreas-Krebbel]

I've tested various combinations on Power 10. I don't see any issue with that. As expected XSIMD 14.2.0 fails to build with clang, but this has been fixed with #1337

clang-19.1.7-14.3.0.log:[doctest] test cases:  388 |  388 passed | 0 failed | 0 skipped
clang-19.1.7-14.3.0.log:[doctest] assertions: 5912 | 5912 passed | 0 failed |

clang-21.1.8-14.3.0.log:[doctest] test cases:  388 |  388 passed | 0 failed | 0 skipped
clang-21.1.8-14.3.0.log:[doctest] assertions: 5912 | 5912 passed | 0 failed |

gcc-13.4.0-14.2.0.log:[doctest] test cases:  367 |  367 passed | 0 failed | 0 skipped
gcc-13.4.0-14.2.0.log:[doctest] assertions: 5649 | 5649 passed | 0 failed |

gcc-13.4.0-14.3.0.log:[doctest] test cases:  388 |  388 passed | 0 failed | 0 skipped
gcc-13.4.0-14.3.0.log:[doctest] assertions: 5912 | 5912 passed | 0 failed |

gcc-14.4.0-14.2.0.log:[doctest] test cases:  367 |  367 passed | 0 failed | 0 skipped
gcc-14.4.0-14.2.0.log:[doctest] assertions: 5649 | 5649 passed | 0 failed |

gcc-14.4.0-14.3.0.log:[doctest] test cases:  388 |  388 passed | 0 failed | 0 skipped
gcc-14.4.0-14.3.0.log:[doctest] assertions: 5912 | 5912 passed | 0 failed |

gcc-15.2.0-14.2.0.log:[doctest] test cases:  367 |  367 passed | 0 failed | 0 skipped
gcc-15.2.0-14.2.0.log:[doctest] assertions: 5649 | 5649 passed | 0 failed |

gcc-15.2.0-14.3.0.log:[doctest] test cases:  388 |  388 passed | 0 failed | 0 skipped
gcc-15.2.0-14.3.0.log:[doctest] assertions: 5912 | 5912 passed | 0 failed |

gcc-15.3.0-14.2.0.log:[doctest] test cases:  367 |  367 passed | 0 failed | 0 skipped
gcc-15.3.0-14.2.0.log:[doctest] assertions: 5649 | 5649 passed | 0 failed |

gcc-15.3.0-14.3.0.log:[doctest] test cases:  388 |  388 passed | 0 failed | 0 skipped
gcc-15.3.0-14.3.0.log:[doctest] assertions: 5912 | 5912 passed | 0 failed |

I haven't checked yet. But my guess is that Debian currently uses some in-between version "14.2.0 + the clang fixes" perhaps?

[drew-parsons]

That's right, debian is on 14.2.0 but has applied PR #1351, #1374, #1377 as well your patch #1337.
Our patch list is found at https://salsa.debian.org/science-team/xsimd/-/tree/master/debian/patches?ref_type=heads
i.e. https://salsa.debian.org/science-team/xsimd/-/tree/debian/14.2.0-4/debian/patches?ref_type=tags

You've just released 14.3.0, so I can try again with that.

[drew-parsons]

Looks like 14.3.0 applied more fixes, ppc64el is passing tests now
https://ci.debian.net/packages/x/xsimd/testing/ppc64el/73630413/
https://ci.debian.net/packages/x/xsimd/unstable/ppc64el/73633469/
https://ci.debian.net/packages/x/xsimd/